Baker says Anthropic added $11B ARR, comparing it to the entire SaaS revolution's $5-10T value creation. [2]
Read the supporting quotation · Invest Like the Best · 2026-05-20
“Anthropic, they added $11,000,000,000 of AR. And what is astonishing to me about this is that the SaaS and cloud revolution it created, we'll call it between 5 and $10,000,000,000,000 of value.”
Baker argues NVIDIA could sell $2-3 trillion in GPUs if TSMC expanded capacity, but TSMC's restraint prevents a bubble. [3]
Read the supporting quotation · Invest Like the Best · 2026-05-20
“If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2,000,000,000,000 of GPUs in '26 or '27, maybe 2,500,000,000,000, maybe 3,000,000,000,000. But there is a limit where consumers would consume so much.”
FROM THE RECORD · 18 matching entries · 2019-11-26 to 2026-07-08
What does the record say about semiconductors?
Selected points from existing editorial summaries and source excerpts. This covers the material on this page.
1
Baker reveals SpaceX and Tesla are building world's largest fab in America with Intel partnership access. [1]
Read the supporting quotation · Invest Like the Best · 2026-05-20
“It's a SpaceX, I believe Tesla's involved as well, joint venture to build the world's largest fab here in America. I think they're going to be successful.”
Baker notes Taiwan Semi will spend more on CapEx in two years than in the preceding five years. [3]
Read the supporting quotation · Invest Like the Best · 2022-01-25
“Taiwan Semi, their 2022 CapEx is going to be many multiples of 2019. 2021 and 2022, they'll spend more than they did in the preceding five years. So there is a massive supply response coming.”
FROM THE RECORD · 14 matching entries · 2019-11-26 to 2026-07-08
What does the record say about ai?
Selected points from existing editorial summaries and source excerpts. This covers the material on this page.
1
Baker cites research showing AI quality doubles with every 10x increase in training data quantity. [1]
Read the supporting quotation · Invest Like the Best · 2019-11-26
“All that matters is the quantity of data, and it's been very well established in multiple papers from both Google and Microsoft research that for every order of magnitude increase in the data you use to train an algorithm, the quality of the AI doubles.”
Baker calculates rendering Monopoly Go with VEO3 would cost over 100x the game's revenue. [2]
Read the supporting quotation · Generating Alpha Podcast · 2026-07-08
“Monopoly Go is a game where we have the revenue and the hours played and it is to render it using list prices for something like VEO3 is more than two orders of magnitude greater than its revenue. So what is the role for human creativity, man? I don't know.”
Baker predicts America will achieve over 5% real GDP growth in one year due to AI and deregulation. [3]
Read the supporting quotation · All-In Podcast · 2025-01-04
“I think that America over the next at some point over the next four years will print at least one year of greater than 5% GDP growth, real GDP growth.”
FROM THE RECORD · 14 matching entries · 2025-01-04 to 2026-08-04
What does the record say about memory?
Selected points from existing editorial summaries and source excerpts. This covers the material on this page.
1
Baker believes NVIDIA could sell $11.52 trillion in chips next year if TSMC tripled capacity but bubble risks prevent it. [1]
Read the supporting quotation · Sohn Conference Foundation · 2026-05-15
“And if they doubled or tripled capacity, like, NVIDIA could probably sell $11,520,000,000,000 worth of chips next year. I really believe that. But the other side of that might be very painful for everyone.”
Baker reveals high bandwidth memory is a bigger cost component for NVIDIA than Taiwan Semi, with only two capable manufacturers. [2]
Read the supporting quotation · All-In Podcast · 2025-01-04
“High bandwidth memory is a bigger part of NVIDIA's cogs on GPUs than Taiwan Semi is. And today, there's two companies that can make it, Hynix and Micron.”
Search covers the displayed transcript and written excerpts, plus all X posts stored for this person, including replies. Coverage is growing; this is not a complete account history.
Explore
In their own words
1546 entries
Transcript excerpts may contain transcription errors. Open the recording for wording, speaker and context. Source links do not verify the claims made.
No matching entries. Try fewer words or clear the filters.
Choose an end date on or after the start date.
X quote postX
Quoting Marathon Fusion
Last month, we demonstrated high-rate alchemy: turning mercury into real, stable gold, beating CERN’s record by a factor of 10.
Last week, we enriched isotopes in the world’s most powerful plasma centrifuge, beating the Soviet record that stood for 40 years.
Now, we’re building https://t.co/GFlLidBdh8
Last month, we demonstrated high-rate alchemy: turning mercury into real, stable gold, beating CERN’s record by a factor of 10.
Last week, we enriched isotopes in the world’s most powerful plasma centrifuge, beating the Soviet record that stood for 40 years.
Now, we’re building https://t.co/GFlLidBdh8
We raised $550M at an $8.8B valuation.
A cap table tells you how a company intends to build. We built this round around investors with a long time horizon: established long-term holders, plus institutions that work with law firms and in-house legal teams every day and have a https://t.co/ORptZitX4d
Fun to co-host the GTC Berlin pregame show with @dee_bosa and nice of her to invite me on the postgame show.
Deepseek is generating $38 in inference revenue per GPU hour off-peak. That is $100b per gigawatt. 2x this at peak.
Seems important. https://t.co/Kc8kDSkUKc
Fun to co-host the GTC Berlin pregame show with @dee_bosa and nice of her to invite me on the postgame show.
Deepseek is generating $38 in inference revenue per GPU hour off-peak. That is $100b per gigawatt. 2x this at peak.
Seems important. https://t.co/Kc8kDSkUKc
And interesting frontier lab strategy would be to encourage distillation, set price that mass distillation is unit economic positive, and then just focus on building the best models. This results is a downstream ecosystem addicted to you. And with no capability to surpass.
“You don’t want to put all your eggs in a 2026 bucket where people feel good about AI. There are gonna be fortunes made in this vintage. You want to be part of the 2027 vintage as well. Maybe the next NVIDIA is not gonna be built till 2028.”
— Erlina Yeo Terwelp, Columbia
Given all the jitters about AI spending, Coreweave etc. Very fortuitous timing to be talking with @drwconvexity -- someone at the forefront of the compute trading industry -- TONIGHT at Odd Lots live in Chicago.
looking at equal weight basket of "ai stocks" this is the worst 6 day loss in history, 2nd worst 5 day (deepseek).
This is true on absolute and relative terms. OK, now back to Harry Potter mindset.
@Merridew__ Actually quite funny. Although hundreds of billions in hyperscale revenue that are collectively a much more relevant metric for AI than the ARR of a company that has been losing share.
My concern on the "AI Trade" was that most of the monkeys who trade it have no clue what they own.
Therefore, you should all give thanks to Gavin for sticking his head out daily in order to educate bulls and bears alike in a civilized manner.
Thank you Gavin
🌳🙏 https://t.co/uKaqgQQMXW
AI infrastructure 👀
"*Every* open source inference cloud I am aware of is growing faster both sequentially and YoY than either Anthropic or OpenAI. And their combined compute spend is becoming material." https://t.co/NiTg1AKJ9P
@GavinSBaker Slower for longer is the better path. The AI bond vigilantes are already enforcing it by raising the cost of the next wave of unconstrained financing just as growth normalizes and cash flow turns positive
@GavinSBaker The bubble has been in the stocks, regardless of whether we get all of the absurd trillions in spending that has been committed by companies that lose money lol. This is gonna be a shit show for Nasdaq, $EWY, $EWT, $EWJ…and you know that. And you hedge for it. You really
@AtlasShrug1 I am actually quite constructive here through year end - reasonable minds can disagree, that’s what makes a market.
But yes there are going to be a lot of losers. And good feedback that I will think about; i.e. wise to buy puts whenever vol gets cheap, be diversified and always
@GavinSBaker Surprised syndicate didn’t advise them against this- ECM bankers should’ve been well aware of the limited upside in providing ARR disclosures (particularly given most folks will just look at the headline figure and scream decel)
TLDR; we will probably never get another ARR number from Anthropic, which is healthy. Should force the market to move away from extreme focus on Anthropic/OpenAI ARR numbers to hyperscale growth rates, which are a much better metric for AI at this point given they capture Muse,
@GavinSBaker @edzitron K so i have never met Gavin IRL
One random sundaE, earlier this year, he acknowledged my compliment & followed me
Because i am lo-mein low status, i had no idea who he was
But before my recent conversion to monkhood or the life of a loner, i was quite social
Turns out a
One of the biggest challenges in the public market right now is disbelief in the "demand" side of AI. This is largely driven by a lack of public information (metrics, earnings reports, analyst coverage, etc.) around the private companies driving innovation and revenue growth on
Benchmark's Eric Vishria (@ericvishria): "I am of the view that it all works."
"Will the CSPs do well? Yes.
Will these neoclouds do well? Yes.
Will NVIDIA do well? Yes.
Will these chip startups do well? Yes.
Are we gonna have big models in data centers? Yes.
Will the https://t.co/2cTdDcKIYk
@GavinSBaker @MarceloLima Congratulations on the Fireworks investment. Great company!
But now I can't ask you where you’re seeing open weights actually growing lmao. https://t.co/sDJXRcS2Cn
@edzitron I’m glad we agree these deals were fake. I promise they are what I was referring to.
The Brad interview from November 2025 references “$1.4 trillion of commitments” so it was obviously a talking point in 2025, not just 2026.
Anyways, I’m not quite sure what you want but I hope
@GavinSBaker I'm fully aware of those (fake) deals from last year and have covered, and must repeat myself that it's very clear you were NOT referring to those in your original post, because the "$1 trillion" talking point was a 2026 one. You were obviously talking about compute contracts!
@edzitron I’m glad we agree these deals were fake. I promise they are what I was referring to.
The Brad interview from November 2025 references “$1.4 trillion of commitments” so it was obviously a talking point in 2025, not just 2026.
Anyways, I’m not quite sure what you want but I hope
@GavinSBaker Can you be specific about what you are referring to RE: asian supply chain and semi companies?
Because the only trillion dollars worth of commitments I can find are the ones I've mentioned, which are very real, and the only trillion-plus commitments I know of.
@GavinSBaker Gavin, I am asking a very simple question:
What are the $1 trillion in contracts here that OpenAI and Anthropic have signed that are "not taken seriously"?
https://t.co/CtvQQxelez
@GavinSBaker The confusing part is that this post is clearly referring to the $1tr in spending commitments from OpenAI and Anthropic, which you are now claiming are actually taken seriously. What parts weren't taken seriously?
https://t.co/CtvQQxelez
@edzitron As I said we had a misunderstanding. I was referring to the $1 trillion plus in MOUs, which as you pointed were not contracts given they could be “pared back.”
The $400 billion in contractual commits in the Anthropic are separate, serious and well funded.
Good luck.
@GavinSBaker It doesn't make any point at all because the "$1tr wall of payments" was not referring to a November 2025 video interview.
You still are yet to answer my original question: what, exactly, are the "not serious" parts, and how will Anthropic and OpenAI pay for the serious parts?
@edzitron I have actually answered your question multiple times now. The ~$400 billion of contractual commitments in the Anthropic S-1 are $31 billion of annual cloud spend over 7-10 years which can easily be funded out of Anthropic’s revenue and $161 billion in Broadcom commitments that
"own your weights, control your future."
the flip side is worth saying out loud.
if you rent all your intelligence from someone else, you're renting your future too.
most boards haven't done that math yet. https://t.co/BDsAweFG7n
Probably also worth pointing out that Anthropic likely has $100B+ of cash on its balance sheet today (e.g. prior to any IPO proceeds).
Per Reuters, Anthropic had $20.3B as of Dec. 2025. Add the $30B Series G in Feb., the $65B Series H in May, and then let's conservatively assume https://t.co/4KaBNuapvO https://t.co/Y53LEMs0jq
@GavinSBaker So Brad here is referring to the $1.4 trillion that OpenAI had agreed to (pared back to $750bn now, lol) - what are you talking about with these 'unfunded asian MOUs'? Because when any skeptic says $1tr they are referring explicitly to Ant/OAI's commits https://t.co/i5A2PnNkuI
@edzitron Well then we had a misunderstanding. But the fact that the $1.4 trillion could be “pared back to $750b” makes the point. Thank you for sharing.
this is such a weird hill to die on
the video you attached literally doesn't help your argument either
one would love to short the stock but they keep delaying their IPO https://t.co/QDFeXMsrxI
@GavinSBaker @edzitron "Hundreds of billions each, but *reasonable* on an annualized basis"... because they're generating FCF right? B/c if they're just losing money hand over fist those contractual commitments are as money good as the local homeless guy's IOUs.
@GavinSBaker @edzitron Oh buddy, ARR isn’t a metric you can plan around! It’s a number meant to show a snapshot in time, and it’s essentially meaningless for the cloud software of LLMs that have no moat or lock-in subscriptions.
Best of luck with your lemonade stand! Appreciate how hard you try!
@GavinSBaker Okay, so your point was that "nobody takes the $1 trillion seriously" but what I am describing is very much part of that total. what is the part that isn't being taken seriously? https://t.co/utauir5rAW
@edzitron The $1 trillion plus in unfunded Asian supply chain/semi MOUs referenced below are separate and distinct from the Anthropic contractual commitments you are focused on. No one takes those MOUs seriously except for AI skeptics which is a little funny.
Everyone takes the smaller $ https://t.co/9TBkNFEFmz
Good thread from Gavin citing some of our work.
As he points out, almost all IRR calcs assume a decline in GPU rents...the math gets wild if rents are RISING as they have been this year. https://t.co/0wMWX5QDyK https://t.co/0sAmeLVWga
@edzitron $31 billion in committed cloud annual spend is trivial relative to Anthropic ARR. The $161 billion in committed Broadcom spend will be easily financed. None of those are MOUs.
Thank you for being a skeptic. Super important and appreciate how hard you try.
Keep it up! https://t.co/V7ChISRzi1
Excited for this one, and looking forward to hearing from others across the AI stack. Thanks @GavinSBaker and team for including me. See you in Austin. https://t.co/mm8DVYFmUk
@GavinSBaker @3F_Research good point
but it appears incredibly unlikely to me that nvidia’s value pool won’t be attacked/ performance won’t be challenged in the next 5 yrs
you already see the green shoots w new architectures like etched (who seem to be beating cerebras btw) - and there’ll be many more
@masayoshisson @3F_Research Etched has literally nothing to do competitively with any SRAM based accelerator.
It’s like saying a football player is beating a 100m runner in an imaginary game of water polo.
But Etched is doing great! Super impressive to succeed first try.
@masayoshisson @3F_Research 6 years seems conservative tbh.
I think with disaggregated inference you will see Blackwells have useful lives over 10 years.
Will be able to put the latest and greatest SRAM accelerator in front of a B200 in 5 years and make it sing.
Nothing about this differentiates from historical capex cycles
Telco/fiber
Shale boom
RET
Colossal $ value of forward commitments magnifies the trade but numbers always look too small on the way up and overstated in drawdowns https://t.co/eXbOnNXQwZ
@GavinSBaker Yeah, I think your view aligns much with Jensen. He has been pushing for open source for a while because of its flexibility. Frontier will still be very valuable in different use cases, while open will continue to lead in market share.
2) A future where OpenAI and Anthropic are a dominant oligopoly at the model layer is negative for almost every company on earth except the two of them. Especially bad for the entire token supply chain. They would be a monopsony for hyperscalers, neoclouds, and datacenter
3) Open-weight models are increasingly being post-trained here in America by companies like our portco @FireworksAI_HQ using an enterprise’s private data to create a highly optimized specialized model. This customized model can be combined with a frontier model behind a router to
1) Trying again.
TLDR: Everyone can win. Open-source, frontier labs, infrastructure, applications and most importantly the American people.
I understand all of the jitters about Anthropic losing share, real rates, spreads and second derivatives. Still think the most likely https://t.co/PeSOvvC706
2) A future where OpenAI and Anthropic are a dominant oligopoly at the model layer is negative for almost every company on earth except the two of them. Especially bad for the entire token supply chain. They would be a monopsony for hyperscalers, neoclouds, and datacenter
1) Trying again.
TLDR: Everyone can win. Open-source, frontier labs, infrastructure, applications and most importantly the American people.
I understand all of the jitters about Anthropic losing share, real rates, spreads and second derivatives. Still think the most likely https://t.co/PeSOvvC706
@GavinSBaker My point is I think Hock’s comments had a near 0 impact on semis price action… and any reasonable AI observer should be able to admit there’s a number of sector specific overhangs right now… we know ASML/TSM/etc will be headline beat/raise but need to be honest about where real
@londoncap100 Obviously.
Monocausal explanations are rarely correct.
Real rates, second derivatives, spreads, external financing, max fear into Anthropic tomorrow, etc etc etc etc.
@edzitron “Geologically stupid” is my preferred term.
Please note and use going forward. It will hurt my feelings if you call me a “dunce” again.
I wish you the best of luck with your newsletter.
Quite a mistake from @business yesterday:
“Tan said bluntly that the market misread OpenAI’s revenue figures because the company’s underlying business is strong. Investors who sell on a single short-term data point are no different from gambling in a casino.’”
Via @ckcheung
Volatility on misunderstood newsflow seems endemic to AI:
Deepseek.
Liberation Day.
Kimi.
SALP.
Gross to net confusion for Anthropic.
OpenAI taking share from Anthropic.
Gross to net confusion for OpenAI.
A mistranslation of Hock Tan’s comments yesterday.
Quite a mistake from @business yesterday:
“Tan said bluntly that the market misread OpenAI’s revenue figures because the company’s underlying business is strong. Investors who sell on a single short-term data point are no different from gambling in a casino.’”
Via @ckcheung
@GavinSBaker @joinyellowbrick Hi Gavin can you be more specific about who “doesn’t take them seriously,” primarily whether that “who” includes the people anthropic and OpenAI promised to pay? Because if it’s the latter you should inform the sec
@edzitron @joinyellowbrick Ed, in what sense is a vague MOU with an endless number of outs for both parties a “promise?”
Examples of promises: an LTA, a SCA, a “take or pay” contract, a fully funded datacenter project with external funding.
All of those are contracts. A contract is a promise. Contracts
@joinyellowbrick My main takeaway is that AI skeptics are anchored on the ~$1 trillion in unfunded MOUs being a massive risk and it is therefore upsetting to hear that no one in the supply chain takes them seriously.
There are so many good, realistic bear cases. Such a strange one to focus on.
Broadcom, Apollo, Broadcom and Anthropic’s $35 billion is “not serious” and neither are the $18bn in bonds on the Oracle data center specifically leased to OpenAI. Anthropic’s $161bn in non-cancellable deals with Broadcom are “not serious.” hock tan has a great sense of humor
@edzitron I presume you are engaging in good faith. And therefore it seems you may have simply missed the word “unfunded” in my post.
As your various replies have mixed large *unfunded* commitments that might be non-cancellable with smaller real *funded* commitments ($35b, $18b).
The https://t.co/BS75tIP9i7
@GavinSBaker @BetterIRR Hi Gavin! If that was the case why is Oracle building out capacity to fulfill a $300 billion, five year contract with OpenAI? What about hut8’s plans with Anthropic? Is the $35 billion deal with Apollo and Broadcom also something they’re “not taking seriously”?
@edzitron @BetterIRR Ed, the $35 billion deal with Apollo and Broadcom is definitionally “funded.” The supply chain takes funded commitments like this one seriously.
It is the ~$1 trillion of unfunded commitments based on vague MOUs that almost no one in the actual supply chain treats as real.
That
@ravenloft43 @GavinSBaker @quantian1 Examples Baker alludes to include semiconductors under Moore’s Law, where cost per transistor fell ~20-30% yearly for decades yet firms like Intel, TSMC and Nvidia (plus downstream tech) drove the sector to a dominant share of market cap. Storage costs per GB also collapsed by
The plan? We replace a well capitalized oligopoly with dozens of highly levered players in existential competition for market share in order to increase our pricing power. Then, we ignore or rationalize the collapsing marginal price of the industry’s end-user output. https://t.co/ZIhq1V7g3A
@quantian1 Yes.
It is well understood that collapsing marginal prices have never worked out well for technology companies over the last 50 years.
Oh wait. During that time they became a majority of market cap while marginal prices collapsed.
I’m sure that you are about to be right.
Good
Am I the only one left besides the black leather jacket CEO who's still bullish on OpenAI and Anthropic?
What's going on with frontier AI sentiment on this app? My goodness. https://t.co/1GBB8pY6VA
@TheBigBerbowski Flying home. Not supposed to post where I am anymore.
Took me a while to discover the beef gyros. The shaved beef was just so good in the tzatziki sauce.
On another note.
Sunsets over the Aegean.
And I have always wanted to see the “Temple of Poseidon.”
Also congratulations to SB and JP on their nuptials.
Did not know the Swarthmore crowd could be so fun. https://t.co/HaNvG5QW1I
In an era of relentless acquisitions and acqui-hires, SpaceX’s acquisition of Cursor - and the near-instant manifestation of that deal into Grok Bot (plus joint Grok model training) is pure brilliance and mastery.
Compute, frontier models, and the leading developer surface
Gavin tweeting a lot & think he’s obviously correct
Argument that OSS is bearish (beyond the labs) makes no sense. Bears saying - there is massive economic benefit & surplus, but there will be no way to fund it? Does that smell right?
@OnodaCapital There won’t anyone to fund a $100 billion open-weight training run.
Oh wait.
Also @grok can you make me an image of Jensen laughing on the way to the bank? https://t.co/AKtWyjXHND
Statements like this make me wonder if people are at least bothering to tune in to hyperscalers earnings calls and carefully listening to what they are saying before building their apparently strong opinions on the “hyperscaler capex” subject
It seems to me that any conspiracy https://t.co/tYhqGlptGx
@GavinSBaker @Silicon_Data What is the economic incentive for open-weight model makers (other than Nvidia) to continue training new models if the token prices are too low to generate a reasonable ROI?
@GavinSBaker @BetterIRR And as a 53-year old with three kids between the ages of 15-21, outside looking in, I’m wondering what the future holds for them?
My senior comp sci major who was never even taught classes in AI at Tufts, I suspect he graduates unemployed. Then? I admit, I don’t know.
@BetterIRR The two poster children failing would be the best thing possible for all future VC investments.
This is obvious to all VCs.
The ones who do not own OpenAI and Anthropic will admit it.
@BetterIRR And for the sake of clarity, I do not think they are going to fail.
I think they are highly likely to co-exist in an increasingly stable equilibrium with Muse, Grok, Open-Weight and perhaps even Gemini. But will take a long time to get to equilibrium shares.
Yes and.
@GavinSBaker How much AI VC money will dry up when the two poster children fail?
Contention is no one knows the real demand for AI, that half a trillion which piled into non-lab AI startups is artificially boosting demand.
@BetterIRR The two poster children failing would be the best thing possible for all future VC investments.
This is obvious to all VCs.
The ones who do not own OpenAI and Anthropic will admit it.
@GavinSBaker The labs demand fed into hyperscaler capex choices, they most certainly took it seriously.
Open weight won't be enough to offset, once you blow up OAI/ANT VC AI funding will fall off a cliff, it is a significant driver of current demand
https://t.co/PlRuMykU5V
@BetterIRR The hyperscalers did not and do not take the labs spending commitments seriously.
If they could get out of the contracts they signed with OpenAI and Anthropic, they would as open-weight inference is more profitable for them than the below market prices currently being paid by
@GavinSBaker @Silicon_Data You appear to be conflating short-term and long-term dynamics. The two primary AI laboratories have secured contracts for a substantial amount of compute capacity; as this infrastructure becomes operational, rental costs will decrease, particularly if the open source model layer
@AlokANYC @Silicon_Data They have secured contracts for a substantial amount of compute capacity at prices dramatically below current prices.
OpenAI’s CFO was being literal when she said they could sell all of their commitments up 3-5x from their contracted prices.
I don’t think I’m conflating
@GavinSBaker @a16z @stevehou Yes I’m just much more interested in sustainability of compute per token vs total cost per token. I guess I should care about both.
@BetterIRR No one - and I mean literally no one in the AI supply chain - takes the ridiculous $1 trillion of unfunded spending “commitments” from OpenAI and Anthropic even vaguely seriously.
Hence the TSM executives dismissing Sam Altman as a nontechnical “podcast bro” after he came to
@BetterIRR As to the second part of your deeply thoughtful, well-considered and *super* contrarian post, open-weight tokens are highly profitable and growing much faster than OpenAI and Anthropic.
Good to know. Now we just need to blow up the 11th and 13th largest companies on earth with 1T in spending commitments and see if anyone minds.
Wonder if the half trillion in non lab VC AI money goosing demand will continue at the same pace, hmm. https://t.co/irRpTxfMrU
@BetterIRR No one - and I mean literally no one in the AI supply chain - takes the ridiculous $1 trillion of unfunded spending “commitments” from OpenAI and Anthropic even vaguely seriously.
Hence the TSM executives dismissing Sam Altman as a nontechnical “podcast bro” after he came to
@GavinSBaker chart establishes that compute remains scarce in this particular market segment. It does not establish that the industry will earn an attractive return on the enormous amount of capital being committed
High GPU prices does not equal high economic profits https://t.co/HARFtO2nSP
@GavinSBaker @Silicon_Data Cheaper AI models could actually make GPUs MORE expensive. The cheaper intelligence gets, the more people use it — and that could send demand for computing power through the roof.
@pmje73 @a16z @stevehou Well this is known precisely by the labs and open-weight inference but there is not much public data. Utilization just a piece of the compute per token equation along with FLOPs per watt per dollar and the price of the watts. Utilization much higher for inference than training.
Competition and fragmentation at the model layer is positive for AI infrastructure demand.
The better that OpenAI and Anthropic do relative to Muse, Grok and Open-weight models, the worse it is for AI infrastructure demand over the long-term.
Token volumes, compute costs per https://t.co/i7f1V1wLh7
Open-weight models that decrease margins at the model layer are positive for AI infra, exhibit 1001.2
Since the @Silicon_Data “token index,” which captures both price and mix, peaked on May 31:
B300 prices are up 38%
B200 prices are up 10%
Other GPU indices are up *much* more. https://t.co/PdudhD5skr
Competition and fragmentation at the model layer is positive for AI infrastructure demand.
The better that OpenAI and Anthropic do relative to Muse, Grok and Open-weight models, the worse it is for AI infrastructure demand over the long-term.
Token volumes, compute costs per https://t.co/i7f1V1wLh7
Open-weight models that decrease margins at the model layer are positive for AI infra, exhibit 1001.2
Since the @Silicon_Data “token index,” which captures both price and mix, peaked on May 31:
B300 prices are up 38%
B200 prices are up 10%
Other GPU indices are up *much* more. https://t.co/PdudhD5skr
@GavinSBaker I have a thesis that the stages of life (maturity) are:
Stage 1 (younger): Send everything and burn all bridges
Stage 2 (mid life): Leave everything in drafts and bite your tongue.
Stage 3 (later life): Send everything and burn all bridges.
@modestproposal1 I would never compare myself to MJ.
But more along the lines of “And I took that personally.”
Also anyone who thinks LeBron is better than MJ is even more wrong than the Ramp economist.
Who seems like a nice, thoughtful, smart guy and is empirically doing good work even if we
@GavinSBaker I think the competition between OpenAI and Anthropic is sufficient to explain the decline in AI prices, but that's not really the point.
Whether it's OpenAI or Anthropic or open source models, less spend going to AI would, I think very reasonably, be net-negative for AI infra.
@arakharazian This is exactly wrong.
Less revenue and/or margin at the model layer does not equal less AI spend at the infra layer.
An extreme hypothetical where all AI comes from enterprises running open-weight models on premise would result in *zero* AI spend at the *model* layer and
@arakharazian Hmm. Let’s think for a second.
Why are the frontier models cutting price? Might it possibly, just possibly, be a result of the fact that they are losing share to open-weight models?
Given positive elasticity, their price cuts are clearly good for AI infra demand.
Open weight
A lot of people sharing this, but the decline is driven by price cuts at the frontier, in turn driven by competition between OpenAI and Anthropic.
Not open source models, which are <5% of business spend.
And if frontier models see a hit to their margins, that will depress their https://t.co/u110uEa44o
@arakharazian Hmm. Let’s think for a second.
Why are the frontier models cutting price? Might it possibly, just possibly, be a result of the fact that they are losing share to open-weight models?
Given positive elasticity, their price cuts are clearly good for AI infra demand.
Open weight
One way to put it, maybe, is that the frontier labs are "frenemies" of the rest of the AI ecosystem.
In some ways, what’s good for the frontier labs is good for the entire AI industry: better AI → more useful capabilities → more usage → more revenue and investment → even https://t.co/BE9ePkxyJW
@DKThomp I have a more extreme view than you. I think an acceleration in OpenAI and Anthropic, if it came even partially from share gains, would be net negative for AI infra demand.
Fragmentation and competition at the model layer is best for AI infra. And for apps!
@GavinSBaker Presumably AI spend will rise rapidly in the medium to long term though. The money will be spent on open weight model tokens vs frontier lab model tokens, so infrastructure providers will benefit- ie “hyperscalers” such as SpaceX, Google, Microsoft, Meta, etc
Why open-weight models that compress margins at the model layer are net positive for AI infra demand, exhibit 1001.
An open-weight token consumes *roughly* the same amount of compute as a frontier token for a similar size model. https://t.co/xaKGHqc1yu
Said another way, the production cost of a given token in compute and energy terms is the same regardless of whether it is an open weight token or a closed frontier token.
And open-weight tokens are the highest margin tokens for the clouds and neoclouds given the negotiating
Why open-weight models that compress margins at the model layer are net positive for AI infra demand, exhibit 1001.
An open-weight token consumes *roughly* the same amount of compute as a frontier token for a similar size model. https://t.co/xaKGHqc1yu
This is a truly impressive line-up and keeps getting better! Great job from @ and the @Atreidesmgmt team for putting this together in such a short time!
I’ll be there and @Silicon_Data will have a presentation and host events at a “compute house”! Look forward seeing you there! https://t.co/dluAiI4se6
I guess X really is where dreams come true 🤯
I’ve lost count of how many Gavin Baker interviews I’ve listened to. Ever since he predicted last year that memory—especially HBM—would face severe shortages this year, followed by DRAM, along with his insights on the All-In Podcast, https://t.co/HF3pVHilIQ
@LIWEI_TWCapital Well thank you for the kind words but I am going to look into a way to short myself after reading your post! I’m just a guy doing his best and the market is ever humbling!
Gavin is wrong with his details here. This tech is not applicable to HFT, which trades at millionths of seconds. They are most likely buying it for mid-frequency trading and low-frequency trading. Regardless, the hedge fund industry is not going to absorb compute at 200B/GW at https://t.co/C02i3HSSmJ
@jiujitsu199 I’ve always wondered whether the phrase “missing the forest for the trees” is an idiom, a proverb or an aphorism.
I think it’s an aphorism. But doesn’t really matter.
@GavinSBaker Hi Gavin - Jane Street is a really interesting example. Dylan used it recently as well. Some questions:
1. Jane Street is a maybe one of single digits firms in the world best suited to monetizing systematic usage?
2. Jane Street PnL is also zero sum - does that matter?
@MayankSeksaria Markets are not zero sum over time. They are zero sum on any one single trade. The aggregate revenue of the asset management industry, broadly defined and inclusive of LO/HF/PE, is over $1 trillion.
the high frequency personal agent wars of 2027, fought via low latency chips and rack placement next to the servers hosting Resy and SF tennis court systems
More probable:
The high frequency trading and cyber wars of 2027 fought via low latency SRAM chips, whether Groq 3 LPX or CS-4, in disaggregated setups with Rubin.
Jane Street reportedly paying $200 million per mw - $200b per gigawatt - is likely the beginning, not the end. https://t.co/LKLYdNGgiB
@masayoshisson Ha - did not mean to like the post. Bookmarked it as was thinking of responding as follows:
I have seen a tear sheet of a Coatue product with much higher numbers this year.
Large firms run multiple strategies.
@masayoshisson Ha - did not mean to like the post. Bookmarked it as was thinking of responding as follows:
I have seen a tear sheet of a Coatue product with much higher numbers this year.
Large firms run multiple strategies.
@masayoshisson Ha - did not mean to like the post. Bookmarked it as was thinking of responding as follows:
I have seen a tear sheet of a Coatue product with much higher numbers this year.
Large firms run multiple strategies.
Since 1830, technology has cut the American workweek from 69 hours to 38 while raising real income per person roughly 20-fold.
Properly managed, A.I. can do the same.
My @nytopinion Chart: https://t.co/3WCBRg3gMx https://t.co/xkjn4uNmLH
Very interesting! Now that I think of it, I actually don't think I know a single person with a +$1M net worth who drinks (or at least gets drunk).
Interesting correlation.
My wife and I stopped drinking when she was pregnant almost 5 years ago and we haven't touched alcohol https://t.co/h7oaBDEtjn
Democrats are trying to dismiss the White House Accord on Super Intelligence as optional self-policing. But although the agreement was entered into voluntarily, the governance that follows from it is not. As Gavin Baker points out, once an independent auditor reports a safety https://t.co/PjWiz32VWN
@slxego0 No matter who selects the “independent” committee board members, those board members will have a fiduciary duty and potential personal liability. They will be incented to have robust, credible third party auditors. Corporate governance FTW.
The White House Accord on Super Intelligence will have a positive impact on safety.
Committing to hire third party auditors who will report to an independent committee of board members has more practical near term significance than almost any conceivable regulatory action. https://t.co/m4E2MRIPlA
Backstage before our panel on superintelligence at the launch of https://t.co/pdYe2rppWr
Will post the video when it comes out.
Also @jgebbia has done something truly awesome for the country. We should all be grateful!
📸 by @KatieMiller https://t.co/DlFhRVRcHG
@GavinSBaker @nvidia I keep feeling puzzled by this theme because this sort of thinking/approach is not new. Agents are software that can become malicious, and protecting against malicious software is not a new thing at all. For example, suppose you run a cloud service and you want a certification
@GavinSBaker @ScottWu46 I'm really just curious to understand where the willingness to pay is. WTP is def there for compute (as we all know) so neocloud has been an easy motion. If they are doing pure enterprise sales and hitting $1b, that's pretty amazing
Is this effectively Neocloud/compute reselling? I'm not trying to ding them, but I don't know anyone using Cognition products. Perhaps customers are buying GPUs? https://t.co/M8PeWkSP7H
@NaveenGRao More enterprise customers vs. SV devs.
Definitely not just reselling compute.
Security Swarm is fastest growing product.
Perhaps @ScottWu46 will say more.
Empirically impressive numbers @cognition and @ScottWu46
Would also be interesting to have cash burned to get to $1b on this chart.
Chart from @absoluttig and please note that I am biased here. https://t.co/scWVjjiyTO
From what we know (take with a grain of salt, we need much more transparency!), if @OpenAI had been running this on their own agents that attacked us, they would have caught them before we did!
Since the first agent cyberattack hit us in July, we've been asking what safe agent https://t.co/d2iyRaI20b https://t.co/kfoL6aiklb
Nvidia’s new Open Agent Safety Platform would have probably prevented the much discussed Hugging Face incident.
Engineering a better, safer world. https://t.co/Lah3LztTRv
From what we know (take with a grain of salt, we need much more transparency!), if @OpenAI had been running this on their own agents that attacked us, they would have caught them before we did!
Since the first agent cyberattack hit us in July, we've been asking what safe agent https://t.co/d2iyRaI20b https://t.co/kfoL6aiklb
“Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world.”
100%. Great work from @JensenHuang and team https://t.co/mmju2JkYpQ
@lokoyacap @nvidia Yes. Another variation of Zuckerberg’s statement that “Meta delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing
Safety is an engineering problem, the @nvidia edition:
Open Shell to enforce policy in the kernel outside the agent process. Can be paired with Sentry which runs in BF-4 and can quarantine a rogue agent in milliseconds.
Containment is the backstop to alignment. Super elegant. https://t.co/eEh6T6H5uz
My analogy would be that this is an seatbelt and airbag for AI that protects us from bad driving, whether our own or others.
But their analogy from their technical blog is better:
“The internet in the 90s…was new and full of possibilities. You could build a website over a
Safety is an engineering problem, the @nvidia edition:
Open Shell to enforce policy in the kernel outside the agent process. Can be paired with Sentry which runs in BF-4 and can quarantine a rogue agent in milliseconds.
Containment is the backstop to alignment. Super elegant. https://t.co/eEh6T6H5uz
@JKeynesAlpha The cybersecurity idea is very interesting. Basically, if I can decode tokens faster than my adversarial attacker I can defend against them faster
First of all, thanks @TimmerFidelity for kindly quoting our data ☺️
Two gentle corrections on the series in that chart:
- Our Token Index is an expenditure-weighted price (per million tokens), not token expenditures. Expenditures and volumes have grown exponentially. The https://t.co/6Af04yoitg https://t.co/QpQzuYqib2
@Silicon_Data @TimmerFidelity I might create a “Token Spend Index” that multiples your current “Token Index,” which captures both token mix and price, with a “Token Volume Index.”
Or maybe change the name from “Token Index” to “Token Price/Mix Index.”
Think something along these lines might be helpful.
🚨INCREDIBLE: During a multi-vehicle collision on the Route 90 bridge in Ocean City, Maryland, a 2-year-old girl was ejected from a pickup truck, falling nearly 30 feet into the Assawoman Bay below.
Jonathan Bauer, who was driving across the bridge with his teenage daughter when https://t.co/xbt1DfO4Lm
Last week, @WeAreLegora reached $200 million in annual recurring revenue (ARR).
It took us 18 months to grow from $1 million to $100M ARR. It took us less than 6 months to double it to $200M ARR.
To put that in perspective, we are on track to add more ARR this quarter alone https://t.co/NVnGJz6oiH
Antonio Gracias invested in SpaceX 30 times. Never sold. Just kept adding as the valuation went from $500M to $5B to $10B and beyond.
Most investors think high conviction means putting 3-4x into a company you believe in. He was 10x more convicted than that.
Since 1925, https://t.co/gqzkauOe0A
Excited to be part of @TomorrowXSummit in Austin on November 17 and 18, 2026!
Great to see independent creators and the companies shaping AI coming together. Thanks to @GavinSBaker , @Atreidesmgmt , @rbiscardi and @iconnections_io for making it happen!
FUNDA will also take the https://t.co/PK3324VGeQ
Really happy that @rauchg will be interviewed by @glennsolomon at the @TomorrowXSummit in Austin, Texas November 17-18.
Vercel occupies a critical place in the AI ecosystem with
•Over $500M+ run-rate revenue
•130%+ annual growth
•NDR at 130%+
•400k+ paying customers https://t.co/7ghrMXTeCt
Really happy that @rauchg will be interviewed by @glennsolomon at the @TomorrowXSummit in Austin, Texas November 17-18.
Vercel occupies a critical place in the AI ecosystem with
•Over $500M+ run-rate revenue
•130%+ annual growth
•NDR at 130%+
•400k+ paying customers https://t.co/7ghrMXTeCt
@GavinSBaker you are putting words in my mouth.
I certainly didn’t defend monosopnies. my whole contention for the last three years, since August 2023, has been that OpenAI would NOT be one, and that there would be many competitors since everyone was using essentially the same recipe.
@GaryMarcus I did not mean to suggest you were a supporter of an AI monopsony, monopoly or duopoly. Or that you thought any of these might be likely.
It was a thought experiment. I wholeheartedly apologize for offending you and certainly did not mean to put words in your mouth.
Good luck.
Wait, what? Are you listening to what you are writing @GavinSBaker?
If there are going to be ZERO profits at the model layer, as in your hypothetical, why would anyone value the providers at $2 trillion?
In fact, why would anyone hand them a dime? If people realize that zero https://t.co/aVfOWR98ew
@GaryMarcus Gary I genuinely do not think you understood what I was trying to say. I’m sure this is my fault.
Good luck and I wish you nothing but the best.
@TheBigBerbowski I remember meeting with Staples in 2010 and the CFO being happy that a few people were still taking notes on paper instead of laptops. Remember thinking “no bueno.” Might have to run it back again.
@GavinSBaker 1. the more profitable the LLM-providing business is a a whole, the better the neoclouds will do. if the LLM business as a whole doesn’t ultimately have a strong return on capital, the neocloud business will suffer
2. the high cost providers of steel and oil all utlimately all
@GaryMarcus 1 is just super obviously not true. If there was one lab, and only one lab, with 99.999999% margins, that would be a disaster for the neoclouds and everyone who supplied that single lab.
How can this not be obvious to you?
There are somewhere between 1000 and 1 million case
OpenAI’s forecasted $278B of cumulative cash burn from 2026-2030 looks absurd on the surface.. but when you evaluate their funding requirements with some additional context things don’t look nearly so dire.
They started 2026 with $40B of cash (Sarah Friar confirmed as such on
Master class in spin. Is it really a “positive for the AI infra trade” that OpenAI and Anthropic are steadily losing market share top open models and on a track to have hardly any?
What happens to Coreweave, Oracle, Nvidia etc when generative AI becomes a utility with near-zero https://t.co/XW7EsIXo2g
@GaryMarcus A better analogy would be: what happened to the consumption of steel and oil after the automobile entered mass production?
And yes, steel and oil were commodities but not all of the inputs into producing tokens are commodities.
The valuations of Coreweave, Oracle and Nvidia
Insane part Gavin would say opposite if they weren’t taking share back! Would say sign of model strength and need for more compute.
He won’t be Blodget of this cycle as he actually believes everything he says and since bull market long always been huge win. https://t.co/IM3Ox2iVjs
@StreetSignal__ Nope. OpenAI and Anthropic trending towards a duopoly would be negative for AI infrastructure demand and a disaster for the application layer.
Looks like today may be a record day for token volume % of open models on Vercel AI Gateway:
🟦 Open 78.4% 🟨 Closed 21.6%
While spend 💲 usually tells a different story, #3 and #4 today are Moonshot AI & DeepSeek. Adding Z.ai, their combined spend surpasses OpenAI (#2). https://t.co/vFMh3xEt83
Open models continue taking share. Not just tokens, more $ now spent on open models than OpenAI.
Positive for the AI infra trade.
Open models taking share shift $ margin from the model layer to the infra and app layers. https://t.co/UfNWyFnEk3
Excited to continue the journey we started in 2022 with the truly special team @CrusoeAI
And a privilege to work closely with our close partners @vpat24 and @GavinSBaker and the brilliant teams at Valor Equity Partners and @Atreidesmgmt https://t.co/6Dt0CTGDAK
Extraordinary share gains for OpenAI vs. Anthropic over the last two months.
Per Openrouter, OpenAI has gone from 20% share to 50% share vs. Anthropic (meaning Anthropic has gone from 80% to 50%) since June. https://t.co/zQEq4juBte
Zuck pretty much confirmed @GavinSBaker argument that pacing/safety/alignment = more compute / spending
I think market understands this but it gave ppl an excuse to degross w/macro risks elevated
More on AI pacing:
Sarah Friar CFO of OpenAI yesterday on CNBC: “from where I sit today there is so much opportunity to drive growth that I am still highly focused on getting more compute to keep that flywheel going.”
Sachin Katti VP of Compute Strategy at OpenAI @sk7037
Funny and almost right.
A scenario where OpenAI and Anthropic are a duopoly would be terrible for compute demand regardless of whether they accelerated or “paced.”
And open source losing would also be terrible for compute demand. https://t.co/onUDSIVfcT
@GavinSBaker @sk7037 Kind of weird the market viewed this as negative infrastructure. This seems wildly bullish infrastructure but highly uncertain for short term to maybe medium term model layer margin structure.
More on AI pacing:
Sarah Friar CFO of OpenAI yesterday on CNBC: “from where I sit today there is so much opportunity to drive growth that I am still highly focused on getting more compute to keep that flywheel going.”
Sachin Katti VP of Compute Strategy at OpenAI @sk7037
3 years ago at a conference in Abu Dhabi, I showed this graph and suggested that there is more than correlation between the increase in compute power (Nvidia revenue as proxy) and Cyber attacks (dotted red line and purple lines).
The other lines show defense spending, which I https://t.co/5LI0kJKwfv
@GavinSBaker How much memory infrastructure does it take to train, align, evaluate, monitor and safely deploy GPT-N?
That number is going up faster.
There is also a nice capex implication. Suppose previously $100 of frontier-model infrastructure spending consisted of $80 capability scaling
Many factors would go into the lower margins, but incremental compute spend on alignment would be one.
Don’t take it from me, take it from a senior OpenAI employee: https://t.co/1WZsx4mjVG
And while this is all coming from a place of sincerity, it may also significantly reduce their contingent liabilities by showing a “duty of care.” Important and responsible step before going public.
The way that Anthropic and OpenAI are going to “pace” the frontier is by spending more time and more *compute* on alignment, monitoring and evals.
The frontier labs that choose to “pace” likely spend slightly more money on compute at the cost of lower margins.
That’s it.
Many factors would go into the lower margins, but incremental compute spend on alignment would be one.
Don’t take it from me, take it from a senior OpenAI employee: https://t.co/1WZsx4mjVG
The way that Anthropic and OpenAI are going to “pace” the frontier is by spending more time and more *compute* on alignment, monitoring and evals.
The frontier labs that choose to “pace” likely spend slightly more money on compute at the cost of lower margins.
That’s it.
Co-sign & agree with @GavinSBaker -
"I do not want a few humans in control of intelligence. I want us all to have our own intelligences that reflect our own values and human variation in all of its richness."
The diversity & nuances of humanity r both the beauty & rights that https://t.co/HqhqSbf4tg
@GavinSBaker Why does everyone assume frontier labs are operating in a liability vacuum? Tort, defamation, product liability, consumer protection, privacy and criminal law already apply. Courts are already adjudicating AI harms under existing frameworks.
Even whether Section 230 applies to
“I do not want a few humans in control of intelligence.
I want us all to have our own intelligences that reflect our own values and human variation in all of its richness.
Intelligence distribution over intelligence centralization FTW.” https://t.co/gPxmM5P2jU
@GavinSBaker Gavin, this approach of "owning your own intelligence" is talked about by lots of people, but there are very few examples of it yet.
We have done exactly this with @cfosilvia. We have assembled a world-class AI research lab focused on consumer finance, including proprietary file https://t.co/dj8O2rQZhG
100% this: "I do not want a few humans in control of intelligence.
I want us all to have our own intelligences that reflect our own values and human variation in all of its richness.
Intelligence distribution over intelligence centralization FTW." https://t.co/bqnP2ezCJE
@GavinSBaker Sure, you can argue that @GavinSBaker is talking his book, but everything I’ve heard and read from him has come across as balanced, thought through and well articulated, without the histrionics you get from most of the loudest voices here. Thank you Gavin! Please keep doing what
Good summary Gavin. A step in the right direction to avoid reactionary over-regulation from DC. But plenty work ahead as we balance vibrant competition, open source & our lead on China - with an approach that gives voters confidence the AI benefits are worth the risk. ⚖️🇺🇸 https://t.co/dS8XYK1yPu
Wild 24 hours for AI and lots of different proposals have been made.
TLDR; the only *tangible* new fact is that OpenAI and Anthropic are going to have embedded 3rd party evaluators from unknown organizations with Dario floating METR as a possibility. Having 3rd party evaluators
SemiAnalysis claimed that Google’s Ironwood TPU delivers up to 50% better performance per dollar than NVIDIA Blackwell.
Yet when the best available configurations are compared using SemiAnalysis’s own InferenceX dashboard, the result is the opposite. At 100 tokens per second per https://t.co/NGCsAVUnpq
High confidence prediction: all these Anthropic people come back after 6 months.
I think they are sincere in their beliefs, but their actions are very calculated and coordinated with a specific goal in mind: regulation.
Better for 🇺🇸 for AI to be distributed and democratized.
Delighted to have @DavidSolomon CEO of @GoldmanSachs doing a fireside chat at the @TomorrowXSummit in Austin on November 17!
So many questions about financing the AI buildout, how AI is impacting their business and perhaps most of all how he is seeing the world/markets as I’ve https://t.co/nrC74SWiM7
Delighted to have @DavidSolomon CEO of @GoldmanSachs doing a fireside chat at the @TomorrowXSummit in Austin on November 17!
So many questions about financing the AI buildout, how AI is impacting their business and perhaps most of all how he is seeing the world/markets as I’ve https://t.co/nrC74SWiM7
WATCH: @Meta's @DinaPowellMcC on @RuthlessPodcast highlighting our data center community compact.
"We promise to not only pay for our own electricity but to actually work to drive down electricity costs...we work with teachers & first responders to give them direct bonuses..." https://t.co/2r4XrDIElK
Five numbers tell you whether an AI business is a real business.
Gross retention.@WeAreLegora is 95%. Customers who bought last year are still here this year. If this one is broken, nothing downstream matters.
NRR. Ours is 300%+. Gross retention is the floor. NRR is how much https://t.co/nuPcDMtbrA
@GavinSBaker And it’s the beginning of a renaissance in mental health treatments, with studies underway for PTSD, depression, anxiety, opioid and alcohol addiction, OCD, anorexia, traumatic brain injury, even stroke recovery.
Closing WSJ quote today: “A year from now, I hope we’ll have
@GavinSBaker I agree. I was brought up thinking psychedelics were exceptionally dangerous, but this book changed my mind. I hope they continue to prove to be efficacious in treating veterans with PTSD and other victims of trauma. https://t.co/3ruOpIGfLa
@GavinSBaker Love this! Have a look into the test level correlation to depression while at it for that same population… I’m sure this is already the more optimal path to addressing or they’d have already targeted that… but all eggs in one basket doesn’t have to be the approach. Aggressively
@GavinSBaker Here are a few other predictions/surprises re: Anthropic:
- The (subsidized) consumer plans are actually very GM+ because the features that select for hardcore users are all pay-per-use (ultra code/review/plan + excess Fable usage)
- The recent change around "Preserved Thinking"
I applaud my friends for giving their time, effort and money to help veterans with PTSD and victims of trauma.
They are sincerely trying to do something good for the world and I believe they have already made a real difference to many veterans.
https://t.co/58EOYotzGA
@MollyJongFast @GavinSBaker You should talk to @baglino
He explains it in this clip
Full video here https://t.co/fl4awBRKSE
Clip starts at 59 minutes https://t.co/0cp7IpD4ni
@GavinSBaker All makes sense. It's possible that Anthropic has a generational leap planned pre-IPO...something like Mythos / Fable 6 as opposed to an incremental 5.2. I've heard rumors they're using something better than Fable internally.
@GavinSBaker Having had the pleasure to work with Krishna during his days at AirBnB and now at Anthropic, I fully agree that he’s a tremendous CFO - brilliant, practical, operational, great commercial instincts and strong feel for the investor outlook. He’s a key reason why our team at
@GavinSBaker Speaking of leaks Gavin, I would have expected a "people familiar with the matter" confirming your Meta and distillation revenue point by now, as well as the Gross vs Net
Anthropic pre-IPO gamesmanship post. Pure speculation but sharing as curious for thoughts.
Anthropic shifted from gross to net ARR accounting and stripped out both Meta and Chinese distillation from their $65 billion ARR number. Meta speculated to be over $5 billion in ARR so
@GavinSBaker @MollyJongFast Gavin, if you respond so courteously how am I to maintain a facile notion of a Big Tech boogeyman!
Do appreciate the expertise and will keep listening.
People will be harmed, left behind in the buildout that will happen. They don't have PACs or lobbyists. Please listen to them
@GavinSBaker How do they lower electricity
prices if they use a lot of electricity often taxing the grid and raising the electricity prices of their neighbors?
@MollyJongFast They install net new behind the meter power. Or said simply, they build new power plants that often generate more power than the data center requires. That is how they lower electricity prices - this is a super well established fact today.
Plus many of them suspend operations https://t.co/fd2QfWe5L9
We didn’t see the Blackwell bump @GavinSBaker talked about but maybe it was simply delayed, since 2 of the 3 have reportedly been out of training for “months”
Just a thought
I have a long list of parallels between Apple when they were the center of gravity fueling the industry to NVIDIA having the same role today but just a point on using balance sheet as competitive advantage. Apple did it also but was B2C. This dynamic is interesting when it's https://t.co/Oj2pf3db5b
@WaterworldCapi1 @GavinSBaker Finger roll. George Gervin style. You two boomers are old enough to get the reference, though you might have been playing d&d instead of hoops.
@InnocenceCapit1 @WaterworldCapi1 I am solidly in the bottom 1% for athleticism and was literally picked last for every team other than Dodgeball as I have good reflexes.
I would sit and read while all my friends played basketball beginning in maybe second grade.
I became an ok high school tennis player through
Well said. This is not the time to be divided. We should be building out energy capacity and compute as fast as possible. This wave isn’t going to stop. https://t.co/xTmgNov5JI
AI is driving a resurgence in American manufacturing. This is electrical switchgear preparing to ship to a @CrusoeAI’s data center, all manufactured by Crusoe Industries in a new factory we opened in Tulsa, where we have hired hundreds of skilled American tradespeople. https://t.co/b7HvZEVJG0
@MollyJongFast @GavinSBaker By @MollyJongFast's logic, we should not have construction job, as all construction jobs are always temporary. Just bad faith arguments!
@GavinSBaker “Data centers are continuously upgrading”
So they’re temporary construction jobs?
So it sounds like your argument is that these companies pay taxes? doesn’t everyone pay taxes? Also some data centers get tax abatements?
@MollyJongFast There will be a permanent increase in high paying construction jobs in these small towns. Especially electricians, plumbers and hvac technicians. Plus the people who operate the data centers.
They pay higher property taxes than other tenants.
And they lower electricity prices.
@MollyJongFast @GavinSBaker He's also saying "ignore the negative externalities like chemical and noise pollution."
And he's evoking a Courtier's Reply: "I'm a tech guy and you know less so you're not allowed to have an opinion on this thing affecting literally everyone."
@MBrodek @MollyJongFast All the new builds address these concerns. The wealthiest county in America has the most data centers per capita. We should all want one in our backyard.
And I’m all for you having an opinion!
Appreciate the dialogue. 🙏
@GavinSBaker So you’re saying the temporary construction jobs are how data centers are benefiting working class people? So you’re saying construction jobs?
@MollyJongFast They are not temporary construction jobs. Data centers are continuously upgraded. Construction jobs are good jobs.
They are 10x’ing the tax revenue in moribund small towns. Data center taxes are paying for new high schools, hospitals, a libraries, police and fire stations. https://t.co/ibwRWo3Psu
@MollyJongFast They *are* benefitting from them, which is why unions are increasingly supportive. Especially the trade unions. Plumbers, electricians, HVAC technician and construction workers of all kind.
Genuine question - if you need to ask this and are unfamiliar with the actual facts at
@MollyJongFast They *are* benefitting from them, which is why unions are increasingly supportive. Especially the trade unions. Plumbers, electricians, HVAC technician and construction workers of all kind.
Genuine question - if you need to ask this and are unfamiliar with the actual facts at
@MollyJongFast The facts no longer broadly support anti-datacenter arguments.
Working class America was devastated by the outsourcing of manufacturing to China. You are now implicitly supporting the outsourcing of intelligence to China. This would be so much worse.
No impact on people with
@MollyJongFast The facts no longer broadly support anti-datacenter arguments.
Working class America was devastated by the outsourcing of manufacturing to China. You are now implicitly supporting the outsourcing of intelligence to China. This would be so much worse.
No impact on people with
JUST IN: NVIDIA's $12,930,300,000 acquisition of Hugging Face contains an easter egg. The number 129,303 is the decimal conversion of Unicode point U+1F917.
The 🤗 emoji.
Exciting day for NVIDIA and @huggingface.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. They allow every developer, startup, university, industry and country to build with, customize and benefit from AI.
Thank you
The Hugging Face acquisition is important for America and I think Nvidia will be a good steward for the ecosystem. The Poolside transaction may end up mattering even more.
I think Jensen is likely to bring American open-weight AI to the frontier, which is going to be awesome for https://t.co/k233bPIiVO
@TotemMacro It is a mistake to focus on token prices imo. Token prices to end users are a function of margins at the model layer. So if high margin labs lose share to open-source, Grok or Muse, it is possible (likely) that underlying token demand and AI infra demand is simultaneously
AI was seasonal in 2024-2025. Growth decelerated during the summer (students/people work less is the theory) and reaccelerated after Labor Day.
This year, AI accelerated in July/August led by OpenAI, Grok and open-source.
And today is the first time I’ve ever seen this: https://t.co/peyOJoSj6N
Muse 1.3 enters the chat.
The frontier is now Claude, ChatGPT, Grok and Muse. No longer just a two horse race.
Gemini is playing a slightly different game for now. https://t.co/0rCTJGsv8i
Predict this will be an underrated moment in the data center discourse.
Many normies including "media elite" get their bottleneck takes from Gavin. And he doesn't lecture, scold or worst of all, condescend
Consensus will swing from current nadir to headlines from Atlantic, https://t.co/dAeSsAepD7
@davidlee Appreciate the kind words and super hope you are right as I think absolutely critical for America to embrace data centers.
This is like being able to redo or better yet undo all the outsourcing of manufacturing plants that we so deeply regret as a country today.
Sarah Guo:
"If you restrict use of open source models in the United States, you'd basically just restrict law-abiding American businesses and slow them down.
The actual attackers or people who have adversarial uses of these things are not affected by your restrictions. https://t.co/BKfSPz64ih https://t.co/z5erntvHJN
Grok 4.6 and Fable 5.1 are now alone on the CursorBench pareto frontier.
Should evolve rapidly over the next few weeks as Astra, Grok 4.7 and Fable 5.2 are released. https://t.co/2fU3w6DgiH
@benitoz Yes really excited for Grok 4.7. All my bots will instantaneously improve. Grok likely to close the gap quickly imo.
And super excited for Nemotron - I think Nemotron 5 will be quite a model.
But as ever, time will tell.
Releasing Fable 5.1 before Astra is quite a flex.
Makes me think Fable 5.2 is ready to go.
So much gamesmanship between Anthropic and OpenAI right now.
And I am very much looking forward to the next version of Grok.
Exciting times.
@matthew_sigel I had the same thought, but sadly I commit to these long in advance. So it with the utmost regret that I say that the beatings are going to continue until morale improves.
I have been saying for a while that asteroid mining is a bigger actual threat to gold's price than quantum computing is to bitcoin's. https://t.co/0EA3NiC8yf
Oh another GB pod!
*>50% on SPCX, predicting asteroid mining and Optimus robots on Mars carrying American flags within 10 years* https://t.co/enswab43rQ
Gavin Baker and a16z's David George on the state of the AI boom:
The future doesn't have to be winner-take-all. Labs, open-source, applications, and the clouds can all capture value.
Demand for intelligence is still dramatically underestimated. Today's power users number in the https://t.co/PQEibIDxu3
Gavin Baker compares Jensen Huang to Michael Jordan and advises AI chip startups: just be happy to be on the court with him.
"[Jensen is] in a very, very good position in his strategy of being vertically integrated but horizontally open."
"Let's say there's some accelerator https://t.co/uxIuL1vn5F https://t.co/CcubpgsV5u
I’m all for data centers in other people’s back yards, seems like a lot of VCs are trying to keep the market afloat until their IPOs come out and they can cash out to retail. https://t.co/fLxaqExEHm
One of the most important epistomological insights everyone should eventually learn consciously - and many do - is that "experts" often have worse conclusions than generalists on questions that are within their area of expertise, specifically because parts of of the question that https://t.co/2YKPcMqE0f
Always fun with my old friend @DavidGeorge83 whom I deeply respect in every way.
Also sometime we will have to tell the story of how we first met. https://t.co/8FMvoo3RMY
Always fun with my old friend @DavidGeorge83 whom I deeply respect in every way.
Also sometime we will have to tell the story of how we first met. https://t.co/8FMvoo3RMY
Always fun with my old friend @DavidGeorge83 whom I deeply respect in every way.
Also sometime we will have to tell the story of how we first met. https://t.co/8FMvoo3RMY
@evantindell And for the sake of absolute clarity, I understand that Codex and ChatGPT have been combined and still think Codex user growth is a relevant metric.
@alwayslearni I know that Codex is not a separate app. Everyone who invests in AI and has a pulse understands this. Most also understand why the Codex numbers are still a relevant metric.
Gavin.. he’s actually right lol
After the July 9, 2026 merger, Codex is not a separate app anymore.
It became one of three modes inside the ChatGPT desktop app: Chat, Work, and Codex. You switch with a selector in the top-left. https://t.co/TgcVCUszQT
That matters for the https://t.co/paESLrVpmL
@alwayslearni I know that Codex is not a separate app. Everyone who invests in AI and has a pulse understands this. Most also understand why the Codex numbers are still a relevant metric.
Gavin Baker says the physics PhDs calling orbital data centers impossible on X have not out-thought the 10,000 SpaceX engineers who treat it as solved
"It's very hard for me to engage. There are all these people on X and they're like, I am a physics PhD and this is impossible." https://t.co/mKy496ZzCo https://t.co/MuinNUuywD
Talking to brilliant physicists who have thought about orbital compute for 3 hours and are convinced it will never work makes me think of this Keanu Reeves quote:
“I'm at that stage in life where I stay out of arguments. Even if you say 1+1=5, you're right. Have fun." https://t.co/pXrGl1ZPRs
Gavin Baker says the physics PhDs calling orbital data centers impossible on X have not out-thought the 10,000 SpaceX engineers who treat it as solved
"It's very hard for me to engage. There are all these people on X and they're like, I am a physics PhD and this is impossible." https://t.co/mKy496ZzCo https://t.co/MuinNUuywD
Gavin Baker re $SPCX: Investors should start to consider Grok when they think about SpaceX, very few of them talk about Grok at all
"I talk to a lot of investors about SpaceX. And what's just amazing to me is everyone will have a nuanced view on Starlink."
"Maybe they're a https://t.co/hLs2zBmP9o https://t.co/xRwNOILZmq
@evantindell Because that is an obviously ridiculous take.
Just because a former OpenAI posts something here does not mean you should take it as 100% true and accurate.
Good luck!
@evantindell Because that is an obviously ridiculous take.
Just because a former OpenAI posts something here does not mean you should take it as 100% true and accurate.
Good luck!
Gavin, spot on.
AI is bringing manufacturing back to America and reindustrializing the nation after decades of offshoring.
AI is creating demand that drives investment in our aging power grid and sustainable energy, powered by market forces, not subsidies.
AI is creating https://t.co/pAUYviuu6C
everyone should go read @dwarkesh_sp’s post - it does a great job of laying out the timeline and what we know ( and don’t know)
I do have two issues with it
A) the use of anthropomorphic language. These are not civilizations nor do they have desires just like a CPU thread or a
@GavinSBaker I would add on the environment: the BTM gas gen is creating a long term pathway to renewable proliferation, as these gas assets may eventually be utilized sparingly but offer the firming capacity that renewables need in the few hours the sun and wind are nonexistent
@GavinSBaker Your takes are unique:
-admit your mistakes (very few do that. I do this on my podcast but not enough in text)
-technical backing (few have the big picture and technology to back it.)
-espousing things that don’t always benefit your book (very very very few VCs, buy-side folks do
I'm pleased to see measured, fact-based posts on the challenges and benefits of building and maintaining data centers for the towns where they get built.
I commend @GavinSBaker for owning his "tone" on a previous post and stripping out any judgement around this hot button issue. https://t.co/xauBN1PjFC
@CDHx7x @GavinSBaker Your own example makes my point. A manufacturing plant employs hundreds to thousands of people on a permanent basis. That’s what funded the tavern, the store, the apartments, all of it. A data center is not that. Even a big one runs on 50 to 150 permanent staff once it’s built,
@ArtifactTrades @CDHx7x What you are missing, respectfully, is that in addition to the jobs the data center pays an insane amount of property taxes that can fund the redevelopment of moribund small towns.
Regret the tone of my post on data centers yesterday.
What I should have said:
There were reasonable concerns about data centers 18ish months ago: water, taxes, jobs, electricity prices, the environment and what they would do to small towns. Well-structured data center projects
What happened when GPT 5.6 Terra and Luna were heavily discounted on OpenRouter?
Token usage exploded by 13.8x
Jevons Paradox = as technology makes the use of a resource more efficient, total consumption of that resource increases rather than decreases 🧵 https://t.co/5rPiufYhGz
the Cursor/OAI drama feels like a pretty big opening for Cognition. still independent/model agnostic/has access to all the models
turns out "we're not tied to any one lab" may be a pretty good enterprise pitch
Nature is healing.
Any politician opposed to data centers is unqualified to hold public office.
Data centers are revitalizing small towns all over America, a godsend for blue collar workers, lowering electricity costs, accelerating the transition to sustainable energy and use https://t.co/fTH3qZTP2V
Nature is healing.
Any politician opposed to data centers is unqualified to hold public office.
Data centers are revitalizing small towns all over America, a godsend for blue collar workers, lowering electricity costs, accelerating the transition to sustainable energy and use https://t.co/fTH3qZTP2V
Jcal says Sam Altman provoked Jensen into directly competing with OpenAI
"I think Jensen is going all in on open source based on these two acquisitions, Hugging Face and Poolside. I think Sam Altman poked the tiger."
"If you look at what he did, he had done that huge $100 https://t.co/9qht0H8oAZ
this is misleading in 3 major ways. Ronan completely misrepresents how data centers are taxed.
1. it's uncommon to impose sales taxes on datacenters, for a very ordinary tax policy reason
2. data centers pay far more in taxes than just sales taxes - property taxes, corporate https://t.co/Q7wrgoqs0b https://t.co/h3am7lTFqF
Nature is healing.
Any politician opposed to data centers is unqualified to hold public office.
Data centers are revitalizing small towns all over America, a godsend for blue collar workers, lowering electricity costs, accelerating the transition to sustainable energy and use https://t.co/fTH3qZTP2V
Data centers helped resurrect Quincy, Washington.The poverty rate fell from 29.4% to 6.2%.
Tech tax revenue funded a new high school, hospital, library, police and fire stations, while residents’ property tax rates decreased.
Build data centers in communities that want jobs, https://t.co/5DubXN544M
Founders, don’t get pissed off if a VC that politely passes on you, then subtly trash talks you to other VCs.
A VC that politely passed on us a while ago, recently told another VC (who I was pitching) that they thought I was a grifter. Grifter is about the most offensive thing a
@davidtisch finally agreed to do press and unfortunately it turned out pretty great.
lucky to be his partner and work with everyone at @BoxGroup
https://t.co/pUsH6ET3o9
To be crystal clear, OpenAI's chip engineers said they decided to build an overall balanced, flexible inference chip. They EXPLICITLY said disaggregation was not the way to go for them. https://t.co/u021IOnSEY
Grok Bot is now available to everyone with a standard Grok or Cursor subscription. It's grown faster than any product we've seen.
It's been particularly exciting to see the range of jobs people delegate to Grok Bot, from running small e-commerce businesses (including support, https://t.co/Doc430851g
@dylan522p @maxkan And they also said this followed by “Cerebras FTW.” I would be surprised if their head of Codex was unaware of their roadmap. Reasonable to hypothesize tomorrow’s Ultrafast will be either GPUs or Jalapeño working with CS-4 or CS-5. We shall see! https://t.co/QM0vs809x7
@dylan522p @maxkan Maybe GPUs plus CS-4/5 for Ultrafast at scale. Jalapeño not financeable at comparable rates to GPUs and probably not that big of a wafer allocation next 2 years. Jalapeño still a really impressive achievement from OpenAI. I remember being sad when they hired Thomas away from https://t.co/a0jIXXahXN
@dylan522p @maxkan And they also said this followed by “Cerebras FTW.” I would be surprised if their head of Codex was unaware of their roadmap. Reasonable to hypothesize tomorrow’s Ultrafast will be either GPUs or Jalapeño working with CS-4 or CS-5. We shall see! https://t.co/QM0vs809x7
@maxkan Hmm. I might ask @dylan522p if Attention-FFN disaggregation (AFD) is compatible with data locality and not moving the KV cache. That constraint is what OpenAI focused on with Jalapeño. They absolutely did not make an “explicit bet that disagg is not the way to go.” There are https://t.co/2lVjH2M5vS
@GavinSBaker Worth emphasizing that the OpenAI team has made an explicit bet that disagg is not the way to go. This was a very intentional decision on their part.
Time will tell who’s correct.
@maxkan Hmm. I might ask @dylan522p if Attention-FFN disaggregation (AFD) is compatible with data locality and not moving the KV cache. That constraint is what OpenAI focused on with Jalapeño. They absolutely did not make an “explicit bet that disagg is not the way to go.” There are https://t.co/2lVjH2M5vS
Impressive that Jalapeño outperforms the comparable TPU and is in the mix with Rubin.
Credit where credit is due - first good ASIC outside of TPU/Trainium.
However, will likely significantly underperform a disaggregated GPU/Trainum plus SRAM accelerator setup. Especially with
Impressive that Jalapeño outperforms the comparable TPU and is in the mix with Rubin.
Credit where credit is due - first good ASIC outside of TPU/Trainium.
However, will likely significantly underperform a disaggregated GPU/Trainum plus SRAM accelerator setup. Especially with
Should note the real comparison will be with Rubin Ultra given Rubin shipping today so kinda have to wait. And Feynmann coming fast. But still impressive debut from Jalapeño.
@GavinSBaker Thoughts on Nvidia+Groq vs AMD+Cerebras? A bit surprised Groq doing 3k+ tok/sec (and I owned Groq pre Nvidia acquisition; now own remain co). Thanks, in advance 🙏
@ChrisMurphyHQ Seem to hit different points of interactivity vs. throughput tradeoff for different size models. Basically everyone plays to their traditional strengths is early read. Really good AFFN disaggregation could change this.
Impressive that Jalapeño outperforms the comparable TPU and is in the mix with Rubin.
Credit where credit is due - first good ASIC outside of TPU/Trainium.
However, will likely significantly underperform a disaggregated GPU/Trainum plus SRAM accelerator setup. Especially with
@GavinSBaker @arakharazian Agree w the rest but this is very much not true "An open-source token broadly costs the same amount of compute (watts and wafers) as a frontier token to produce for equivalently sized models."
@JoshPurtell @arakharazian Frontier labs would say their tokens use less compute than comparable open-source tokens but that would break people’s brains. And for sure they have more intelligence density per token which compounds their efficiency advantage.
@GavinSBaker I find it hard to believe that a decline in the valuation for OpenAI and Anthropic wouldn't lead to a significant decline for the AI infrastructure names? They're all incomprehensibly interconencted, financially.
@GavinSBaker @stefanveliki137 @thsottiaux Fair question,it’s both, and they reinforce each other. My comment was about the memory architecture long-term, not the business short-term. Different timeframes, probably not as far apart as the thread made it seem 🤝
So comes the biggest threat to the AI trade, which relies on ever-increasing revenues to the AI model companies to support ever-increasing valuations.
Those revenues are a function of price and quantity. Quantities are all good, from my perspective. Despite increasing https://t.co/C2V93N9Rnq
@arakharazian Respectfully this is not correct from my perspective. Cheaper open-source tokens are net positive for the AI infrastructure names, although maybe not for the model companies.
An open-source token broadly costs the same amount of compute (watts and wafers) as a frontier token to
@DratchCap @anni_sen @stefanveliki137 @thsottiaux I am also curious about this. Along with why Ultrafast mode is capacity constrained. Deal struck in late December, 4.5 months to make a wafer, 6 weeks to make a server, a month to energize a server. So logical that Ultrafast debuted in August and would think it would be less
Consensus on this seems to be ZDR which tracks to my own experience as well. If that really turns out to be the case, it's an incredibly strong and clear message to the labs on the impact of data retention policies. https://t.co/IQTDsPM8Gy
I literally have no idea why anyone is concerned about the FT data on Fable 5. Just shows customer preference for ZDR. Actually surprised so many are still using it after Opus 5 became available. https://t.co/u5Rzi13JmV https://t.co/GCqX5bMgYg
Head of Nemotron Bryan Catanzaro says NVIDIA is sitting on internal "teacher" models that beat everything it has publicly shipped and he's trying to release these as well
"This technology of using teacher models and multi-domain on-policy distillation has been really key for us https://t.co/SgNJHNrf3U https://t.co/PygehUVxNk
@firesidealpha This was a good podcast. The more interesting question is where those teacher models stack up vs. the frontier. And where the teacher models will be in a few months. Those ASICs sure were fun to develop!
@GavinSBaker @stefanveliki137 @thsottiaux Gavin, Noone questioning whether SRAM is fast. It’s the fastest memory in silicon. That’s not the debate.
The debate is whether it scales. And the physics say no.
As for disaggregated inference?
The moment model weights + KV cache exceed what fits on-chip, you need off-chip
@anni_sen @stefanveliki137 @thsottiaux You think OpenAI is constraining the capacity or production volume constraints elsewhere are constraining the capacity? Genuinely curious.
@GavinSBaker "Niche"
Meanwhile, @thsottiaux: "Everyone wants access to Ultrafast. We reserve it for incidents -- be it outages or security incidents. We have those just as any other company, and Ultrafast comes in handy. Top-priority projects, we always have Ultrafast provisioned for those https://t.co/DfPrVzJYCA
Consensus on this seems to be ZDR which tracks to my own experience as well. If that really turns out to be the case, it's an incredibly strong and clear message to the labs on the impact of data retention policies. https://t.co/IQTDsPM8Gy
@insane_analyst @GavinSBaker @IanCutress Always appreciate your thoughts as a sparing partner IA although I will contest your claim that its not optimal ;)
Thesis is pretty simple, would love to take you through it in more detail @GavinSBaker
- Custom ultra low latency interconnect + fully deterministic dataflow
@anni_sen @stefanveliki137 Hmm. This looks like the flagship big model from the second largest frontier lab. And IIRC @thsottiaux from OpenAI said they wanted this currently capacity limited Ultrafast mode to be the default mode 6 days ago. I’m not sure I would describe “default mode” for the flagship https://t.co/OQqwwW8CD9
@anni_sen @stefanveliki137 Hmm. This looks like the flagship big model from the second largest frontier lab. And IIRC @thsottiaux from OpenAI said they wanted this currently capacity limited Ultrafast mode to be the default mode 6 days ago. I’m not sure I would describe “default mode” for the flagship https://t.co/OQqwwW8CD9
The most obvious implication of GPU price increase is that anyone who is spending a lot on Blackwell and Rubin before January 31, 2027 is in a good relative position vs. competitors.
Super curious to know if GPU prices are going up more than rack prices.
CFETs solve this with overnight 50% SRAM scaling. Ask again in 2031-ish.
https://t.co/WStKlP0Ox9
Between now and then it's about optimizing compute vs bandwidth. The big lever Cerebras could use is quantization. They're still FP16. https://t.co/cidMUCTD2a
@IanCutress Luckily for me, I have never been near a cliff or a bridge when I have read a post about how “SRAM cannot scale.” This topic has largely been put to bed for anyone remotely close to logic/memory/packaging roadmaps especially at wafer scale.
@GavinSBaker The router is the hard part and nobody has shipped a good one. Picking a model per task means knowing the task: what changed last sprint, which data can leave the building, what failed last time. Most multi-model stacks today are a static config with a fallback.
@IbrahimAjami Yes. The acceleration in OpenAI/Anthropic as open-source takes share is so interesting. The future is multi-model with specialized open-source models that were post-trained on customer data working together with frontier models behind a customer specific router. Firework Nexus
High conviction hypothesis especially for large enterprises.
This is the best path to Pareto dominance while keeping data safe.
SMBs and individuals probs just use the frontier harnesses. https://t.co/4petLDY4kt
@IbrahimAjami Yes. The acceleration in OpenAI/Anthropic as open-source takes share is so interesting. The future is multi-model with specialized open-source models that were post-trained on customer data working together with frontier models behind a customer specific router. Firework Nexus
@GavinSBaker Or, something else that isn't being discussed. There are ways to density without a full DRAM process. You can imagine which companies would want to see that happen...
Maybe the consensus is wrong? Abilene is a pretty good counterexample. Up to 5,000 construction workers on site, around 1,000 long term jobs expected, and an estimated $30M a year for the city. Sales tax is already running nearly $20M over budget. Cooling is closed loop with zero https://t.co/UiwJIo5ssP
I guess this is consensus now 😂
If I were building these I would be dong things like promising donations to local schools, making beautification grants (and leading projects) for nearby towns, and trying to solve at least one pressing local problem. https://t.co/G0Vwij7Fq9
@GavinSBaker @sedricm Gavin appreciate the book rec! would be grateful for a top 5 books that you think have most shaped your lens for viewing this “ai etc.” investment cycle.
4/n #HotChips2026 HBF reality check!!!
HBF sounds too good to be true, because it is.
Despite high capacity and low cost, it useful in limited areas like Sparse Attention (DeepSeek, ZAI style), Expert parallelism (or to avoid it by keeping all the experts on tje same HBF chip - https://t.co/5v3yNsAe2d https://t.co/EGyLPbUIS1
People thought I was crazy when I was bullish on this. Welcome to the new world.
Your LPDDR now has SRAM latency !!!! I LOVE THIS BUBBLE 🫧 https://t.co/jrNadOt9q6
Every accelerator company, and I mean every accelerator company, is likely going to do some variant of 3D DRAM or the “zHBM” which were both discussed at HotChips today.
The larger companies will probably run (are running) parallel programs, but these two memory variants
@GavinSBaker @Atreidesmgmt Huge infrastructure implications
‘Pareto optimal’ now includes scheduling optimization of Agent CPU control & tool calls + GPU Prefill (KV Cache) + ASIC FFN decode (Cerebras) token processing
AWS is technically prepared for this new era, but other hyperscalers are lagging, and
@GavinSBaker @Atreidesmgmt Hey Gavin, I am trying to attend @TomorrowXSummit as a content creator/independent. Applied and hoping to hear back for a spot - is this a possibility still?
@imnotharsh @Atreidesmgmt @TomorrowXSummit Yes - just went through your feed - please consider yourself accepted and if not approved in next 24 hours just ping me.
@GavinSBaker @vercel @rauchg It really does seem that the GLM 5.2 launch week in mid June tipped this balance in a meaningful way. Open weight tokens grew 3-4x since then which means the frontier grew “only” 40-50% during the past two months. With all the rumored frontier model releases coming, will be
I have interviewed 1,000 founders and had the fortune to invest in many of them.
The top 5 founders that I have ever met (not in order):
- @alanchanguk (Fuse Energy)
- @awxjack (Airwallex)
- @ceo_clickhouse (ClickHouse, hanging out today in pic!)
- @lqiao (Fireworks)
- https://t.co/PEcJ6W2Zdb
Suspect we will start to hear about a “Pareto optimal” balance of computationally efficient humans, cheaper open-source tokens and frontier tokens.
Our internal AI spend @Atreidesmgmt will be roughly 100x higher in August 2026 vs. March 2026. Still roughly doubling every month.
Huh, 62 percent
Open weights taking share doesn't make anyone's compute bill smaller, an open token burns the same GPU minutes as a closed one and someone still pays for it.
I run Qwen locally for the boring stuff and reach for Claude when it actually has to be right, and my https://t.co/in8BhzK8yH
@itsvlady Well it lowers compute spend by the ratio of the margins of the frontier tokens to the margins of the open-source tokens - frontier tokens mark up underlying compute dramatically more than open-source.
Roughly agree with Gavin.
It may even end up more extreme .. closed AI taking > 95% of the value capture (revenue) but only representing only 5% of overall token use vs. open weights which will dominate usage, but struggle to capture any real value.
It may also end up that https://t.co/peLL4P2ZVz
@GavinSBaker @vercel @rauchg The two things can be true prophecy:
- open source wins token share
- frontier labs hockey stick rev continues
Aka Jcal & BG are both right
@GavinSBaker @vercel @rauchg The big 2 frontiers have been holding back revs last 3 quarters and booking in recent quarters to make optics better for the IPOs. A big revenue cliff is upcoming.
@jweil130 @vercel @rauchg Likely true of Anthropic IMO - likely building in an acceleration for roadshow and end of year. Probably less true for OpenAI.
“Most likely end state IMO is that closed, frontier tokens are 60-90% of economic value but only 15 to 25% of tokens.”
Apple iPhones are about 20% of the world’s total smartphone shipments, but half of the entire industry’s total revenue, and about 80% the operating profits. https://t.co/dh5m0lNcdb
@F1NewsWire @TMTBreakout Big models are expensive. I don’t think super important one way or another. Opus 5 is a Mythos/Fable distillation and doing well.
Inference cost doesn't care whether the weights are open or closed. Same GPUs, same electricity. The token doesn't know it's open source. https://t.co/wL7m6wiwuW
More data than open-source AI is taking share from OpenAI and Anthropic. Open source has gone from 28% token share to 62% token share @vercel over the last 2 months. Chart from @rauchg
Super impressive given that the sum of OpenAI and Anthropic accelerated in July. So net https://t.co/wjDtlcl6ic
“More data than open-source AI is taking share from OpenAI and Anthropic. Open source has gone from 28% token share to 62% token share @vercel over the last 2 months. Chart from @rauchg
Super impressive given that the sum of OpenAI and Anthropic accelerated in July. So net”
Gavin Baker (@GavinSBaker) on X · https://x.com/GavinSBaker
Quoted in printX (formerly Twitter)
“Suspect we will start to hear about a “Pareto optimal” balance of computationally efficient humans, cheaper open-source tokens and frontier tokens.
Our internal AI spend @Atreidesmgmt will be roughly 100x higher in August 2026 vs. March 2026. Still roughly doubling every month.”
Gavin Baker (@GavinSBaker) on X · https://x.com/GavinSBaker
Quoted in printX (formerly Twitter)
“More data than open-source AI is taking share from OpenAI and Anthropic. Open source has gone from 28% token share to 62% token share @vercel over the last 2 months. Chart from @rauchg
Super impressive given that the sum of OpenAI and Anthropic accelerated in July. So net”
@CDMarkAngel @CNBC I don’t think you fully understand the relationship between gross exposure and performance. Could easily be down 67% on a 10% decline in GMV if levered enough.
Interesting disclosure from Ken Griffin via @CNBC that they have shed 80% of the SALP portfolio.
Super high confidence that this was $45b levered more than 4 to 1 so at least $180b gross. Unwinding that $180b would absolutely have had an impact.
Have felt it in the market.
Interesting disclosure from Ken Griffin via @CNBC that they have shed 80% of the SALP portfolio.
Super high confidence that this was $45b levered more than 4 to 1 so at least $180b gross. Unwinding that $180b would absolutely have had an impact.
Have felt it in the market.
Harvey is a great example of how American companies are building world-class specialized models: they took an open-source base (Kimi K3), post-trained it on legal data, and delivered state-of-the-art performance on legal benchmarks at a fraction of the cost of frontier models. https://t.co/YO2tE0O11f
May startle the markets with this but: OpenAI is growing faster than Anthropic, actually.
While data through Q2 shows tepid growth for OpenAI vs. Anthropic, Q3 to date shows that OpenAI has surpassed Anthropic in QoQ enterprise growth: 82 vs. 76.
Why? GPT-5.6 Sol is really https://t.co/RrhPTYbBG4
Yes. And the sum has accelerated slightly QTD with a sharper acceleration in August vs. July.
And note that Grok Bot, Grok and Open Source are all likely growing faster than both of the two leading frontier labs.
But at some point going to have to move past second derivatives https://t.co/L4r1aExjrI
23/ At the infrastructure layer:
• Opus produces X of infrastructure revenue per busy GPU hour
• GLM produces Y per busy GPU hour
If Y is close to X, the buildout is fine
If Y is much lower, higher utilization or more demand has to offset it or the GPU providers earn less
@Shaughnessy119 Y is higher than X - Anthropic and Open AI have more leverage over the infrastructure/cloud providers.
For a given token over time all else equal, the cloud and infrastructure players charge more for open source tokens vs. what Anthropic and OpenAI pay per token b/c open source
Empirically impressive metrics from Stripe in the Axios report about their acquisition of Openrouter.
Growing net revenue and FCF almost 2x faster than Adyen YTD with the share count declining over the last 3 years.
I think Stripe will be a good steward of Openrouter and it is https://t.co/mnxU0XqP0N
Empirically impressive metrics from Stripe in the Axios report about their acquisition of Openrouter.
Growing net revenue and FCF almost 2x faster than Adyen YTD with the share count declining over the last 3 years.
I think Stripe will be a good steward of Openrouter and it is https://t.co/mnxU0XqP0N
Fireworks and Baseten are the two largest inference clouds and I think of them as *an* index for open-source AI adoption.
Also first time seeing SpaceXAI on the list of “fastest growing” IIRC.
SambaNova! Will wonders never cease. https://t.co/HlmFsKRtVk
@GavinSBaker @TMTBreakout I think open source is positive if the math on
- Percentage that shifts to open source
- Times the price decrease
- Times the increased usage
Results in net more usage of the infra
Basically does Jevons paradox with open source hold to finance the buildout
What do you think?
@Shaughnessy119 @TMTBreakout 2 of the 3 parts of your equation are correct (first and last, not the middle). An open-source token costs just as much compute and energy to produce as a frontier token for a model of comparable size. And they are often less token efficient.
@GavinSBaker @theallinpod Whoa, you mean overall usage is up big and the hyperscaler audited financials and public mgmt commentary on demand and capacity constraints only 3 weeks ago weren't lies? Shocking.
@GavinSBaker @TMTBreakout The most interesting signal here may be open source gaining share while total AI usage is still accelerating.
That suggests the infrastructure layer could win regardless of which individual model leads at any given moment.
In the long run, is the bigger moat the model, or the
@GavinSBaker @TMTBreakout $13 bln was net arr added for July (up from $11bln added in june).
Roughly evenly split with oai taking $7bln . Pretty good for OAI ..partly reflects their compute additions
Good chart from @TMTBreakout
July accelerated both MoM and YoY for the sum of OpenAI and Anthropic.
Some chance Anthropic shifted from gross to net reporting for ARR. Meaning the 65b is a more conservative metric than the $47. Smart thing to do before an IPO.
Openrouter data https://t.co/FOwqXizXIL
We are all lucky that people like @benthompson, @GavinSBaker, @altcap, @dsundheim and others so often take the time to explain what’s happening in a detailed and deep way.
I think it’s a huge force for good in markets. https://t.co/18M6xcBDpf
Ben Thompson says OpenAI and Anthropic are Silicon Valley's two religious organizations, mainline and evangelical, and the belief they're creating God is why the riskiest labs have the biggest upside
"OpenAI and Anthropic obviously are the riskiest, but also the biggest upside." https://t.co/JBOAH3ukFO https://t.co/hPXhQZZ7vb
You have to be a little bit crazy to build a new chip from scratch. @UbertiGavin and @robertwachen decided to do this in 2022.
4 years later, @Etched has shipped their first deployment to Jane Street, making them the only new company since ChatGPT released to put a brand new https://t.co/VtK1N02Mam
From yours truly on Friday’s @theallinpod
“On the margin, Anthropic is losing share to OpenAI, to open source, and to Grok, and they're still growing so fast.”
If the $65b is correct, then someone leaked it to help Anthropic on their forthcoming IPO roadshow. Basically, they https://t.co/oIYSS1PZhT
SITUATION EXPLAINED: Gavin Baker says AI is either too dangerous to concentrate, or too dangerous to distribute. https://t.co/roGi4YWqdv https://t.co/YoMm5zHsgP
Too aggressive re Anthropic. Base case of $250 B ‘27 exit run rate (same size as MSFT & Meta) seems wild but achievable. $450 B (size of Apple & Google) seems too aggressive given compute & regulatory constraints. Margins also likely lumpy as big compute comes online. 🧐📈 https://t.co/UVbtDz2CSl
@altcap Love the expectations management!
Let me say this: I would not take the “under” on $250 billion *unless* they stumble or fail to secure enough compute.
Maybe you can discuss with @DarioAmodei at @TomorrowXSummit before I interview you about public policy!
We started looking hard at the @AnthropicAI Series E in late 2024 at @notablecap, and our capital went into the round in early ‘25. At a $61.5 billion post-money valuation, it was, by more than an order of magnitude, the highest price I'd ever paid for a venture investment.
It https://t.co/FAkVWjZmCj
Absolutely agree with Gavin. After using $SPCX Grok over the past week or two, the speed feels genuinely impressive. I later realized that Grok has become significantly more “lightweight,” which has driven token consumption down to an extremely low level.
It’s analogous to a https://t.co/LeEyYOGRwh
Insanely valuable @grok @bot automation:
I have an Bee Pioneer AI pin that records my day and then every night at 7 p.m.
I have a Grok Bot called Relationship Coach that hooks into the API and acts as a personal and couples therapist. It tells me how I could be a better https://t.co/EhhNWrcbC0
I think @bot is another “Claude Code” moment for AI.
I would estimate my personal AI usage is up something like 100x.
And for everyone who reached out about how to build a “podcast summarizer” it took me about 15 seconds in Grok Bot and is better than what I had before. https://t.co/AaoeHbEqtl
The GrokBot origin story as told by @bot
Sadly they lack both a sharing feature and an iPad app but I’ve been told they are both coming soon: https://t.co/4YaISohWW6
@GavinSBaker @bot 100% - I think like 80% of it is the fact that it is so fast and responsive compared to using a Claude or Codex remote session. It is so freaking fast and awesome. I shut down my OpenClaw and Hermes agents almost immediately when I started using it
I think @bot is another “Claude Code” moment for AI.
I would estimate my personal AI usage is up something like 100x.
And for everyone who reached out about how to build a “podcast summarizer” it took me about 15 seconds in Grok Bot and is better than what I had before. https://t.co/AaoeHbEqtl
Happy to announce Sumit Sadana EVP and Chief Business Officer of @MicronTech as a confirmed speaker at the @TomorrowXSummit in November.
Memory per FLOP drives tokens per watt. Looking forward to hearing Sumit’s thoughts on how AI is changing the business model of memory https://t.co/k1Pdk94PDu
Happy to announce Sumit Sadana EVP and Chief Business Officer of @MicronTech as a confirmed speaker at the @TomorrowXSummit in November.
Memory per FLOP drives tokens per watt. Looking forward to hearing Sumit’s thoughts on how AI is changing the business model of memory https://t.co/k1Pdk94PDu
Gavin Baker: If you're a patriot, you should be pro-open source AI
@GavinSBaker:
“Very simply, I'm an American, I'm a patriot.
If we ban open source, if we let this control be centralized, we will lose AI, we will lose geopolitics, we will lose everything to China.
It's just https://t.co/EyoUNmDVlu
Pong Over Ping 🎾
We turned ping into a controller and built Pong inside a network switch’s hardware data-path. Zero #CPU, zero game software on the host.
Here’s how we did it:
1. Packets as Joysticks: Sending ping -p 77 acts as a move input. Your IP address automatically https://t.co/E4otsxRjAc
OMG. I asked my Chief Engineer on Grok @bot to bring up the feature and it's done. 😳
I can text my @maticrobots now through my Grok @bot—anywhere and anytime. https://t.co/DEj8zBYAxS https://t.co/kavdcIWKhI
@GaryMarcus Genuine question: how can a company be more transparent than actually going public and filing the requisite highly transparent disclosures?
And you may not know this, but there are strict rules governing what Anthropic can say publicly in the lead up to their IPO.
Excellent, candid back & forth on complex subjects that benefit from a good faith Saturday afternoon exchange! Thx @_sholtodouglas & @GavinSBaker for kicking off & super helpful to have @DarioAmodei weigh in. 🙏 https://t.co/dnz78OjaJm
2/2 Second, on the messaging around AI. I do not agree that my messaging has been disproportionately negative. In fact it has been about equally balanced between risks and benefits: I’ve written one major essay about each, and even in interviews where I discuss the risks, I
@DarioAmodei Many thoughts.
Most of all, I think it is great that you are engaging here in such a constructive, thoughtful way. Open dialogue is a great way to build trust with the public and these are difficult, weighty issues that I think deserve to be debated in the proverbial “town
@GavinSBaker @Abir2 Grew up in Texas but must not have been near Houston. The culinary game is top notch in and around Houston. World class Indian cuisine.
@GavinSBaker Cool!
Lunch is on me at the Penn Club or some chic and cheap authentic Indian at the Mission in San Francisco
I will call next time I am in your amazing Boston!!
@Abir2 Love cheap authentic Indian food as long as they serve super spicy Chicken Tikki Masala or Butter Chicken, which I love.
I know there are much better dishes. I have unsophisticated taste. Grew up in Texas.
We really love our Matic.
Millions of American homes have Chinese made robots mapping their homes and listening to their conversations.
It’s the only robot vacuum built in America on the market. https://t.co/cXIAhX0nIL
Gavin Baker says open source AI actually makes closed models more valuable
"Open source makes frontier tokens more valuable. If you have a 250 IQ intellect that can orchestrate these open source 150 IQ intellects, that makes the 250 IQ intellect more valuable because it can farm https://t.co/IaZczX6Oy3
Hey Gavin,
I had no idea who you are until today
But based on this fuckin' amazing stuff you are like totally fuckin' smart & pragmatic
I hope we meet IRL as the kids say
Thank you for sharing your wisdom https://t.co/yzdO0M7EFd
@GavinSBaker >may be really civilizationally useful in giving us time to adapt to what might be a fast takeoff.
Agree - this has really interesting implications on RSI, which is that even if you had Noam Shazeer level models tomorrow they'd still be experiment bottlenecked, so it changes the
@GavinSBaker Awesome response - thank you for coming back with such good faith.
So much I agree with here. Also a lot of things I think the picture is more nuanced and I want to justify our takes more
A few quick ones, but I might post more here later.
I think that cyber is ultimately
@_sholtodouglas Ok. Becky just brought me another gin and tonic and is watching our daughter.
So: I think the framing of cyber as defense dominant and bio as offense dominant is sensible and the bio point is probably the best pro-regulation argument. And the USG should absolutely have all
@GavinSBaker Awesome response - thank you for coming back with such good faith.
So much I agree with here. Also a lot of things I think the picture is more nuanced and I want to justify our takes more
A few quick ones, but I might post more here later.
I think that cyber is ultimately
@_sholtodouglas Wow - super thought provoking and likewise appreciate your engagement as this is important to get right for everyone - let me think about your points carefully and come back to you.
Am currently having a gin and tonic by the pool while I watch my daughter so might be tomorrow!
@GavinSBaker @theallinpod Thanks, Gavin. I really appreciate your response to me and Sholto.
Dario laid out his positive ambitions for AI here in a lot of detail and he continues to advocate for this vision, both internally and externally: https://t.co/yWBAKo4l9K
His position is nuanced and consistent:
@sashadem @theallinpod Likewise appreciate your response and generally agreed. Am just a little worried that the negative vision is “winning” right now with very unpredictable consequences for everyone.
Reasonable minds can minds can disagree about the best path forward for the most important
@GavinSBaker @_sholtodouglas When you say that the technology should be distributed widely, what do you mean specifically? What specific steps do you have in mind, policy wise, to prevent its concentration?
Really worth reading Gavin’s remarks on concentrated vs. distributed AI.
Humbly, I would add an additional point: those (like Dario) who believe in highly concentrated and regulated AI need to provide a serious argument for why being the best at the art and science of training https://t.co/Ujx5yz32Tc
Gavin Baker says AI is too dangerous to be centralized
"I think Anthropic and people in the effective altruism movement believe that this technology is too dangerous to distribute. What Mark Zuckerberg and I think Elon and Jensen believe is that this technology is too dangerous https://t.co/Hu7riPhSGj
@sashadem @theallinpod I believe both you and Sholto. This issue is that Dario’s own public commentary makes it believable to many serious people in Silicon Valley that Dario might have said something similar to “Anthropic might end up being the only private company.”
I would love nothing more than to https://t.co/NF7MaO8wl1
Posting my reply to @_sholtodouglas given that some of my comments about Anthropic on @theallinpod yesterday have been widely discussed here on X and I think the topic is important. https://t.co/NF7MaO8wl1
@_sholtodouglas Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging
Posting my reply to @_sholtodouglas given that some of my comments about Anthropic on @theallinpod yesterday have been widely discussed here on X and I think the topic is important. https://t.co/NF7MaO8wl1
Completely false.
I like Gavin's takes, but whoever he heard this from is lying so that it fits the narrative some people so desperately want you to believe.
The same people will try to convince you Anthropic has no moat, and a sentence later that it might become so powerful https://t.co/bPf8sZATth https://t.co/W0Xgd7lfNA
@_sholtodouglas Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging
“Residuals for private jets are extremely strong these days. And a big part of the reason is people are taking jet engines off old planes and repurposing them as turbines to power data centers.
”
“Residuals for private jets are extremely strong these days. And a big part of the reason is people are taking jet engines off old planes and repurposing them as turbines to power data centers.
”
Introducing Cues: Voice & Gesture control for Matic
We raised $115M and spent 9 years to make the world’s first intuitive home robot
Say you spilled coffee: Point to the spill and say, "Hey Matic, clean this" and it will hear, see, locate in 3D, and go clean on its own.
Matic https://t.co/tPsUnLh76v
Really does work as advertised.
“Hey Matic go clean the kitchen.”
First 10/10 product rating from Wired in a decade. And kids love it for some strange reason.
Happy that @Atreidesmgmt is an investor. https://t.co/1s0bcxMUHo
They are deliberately misrepresenting HBM performance
For some reason, they fix the bandwidth of both HBM & HBF stacks at 1.6TB/s (12.8/8)
Secondly, nobody serves models on bf16 anymore
Most models are served on fp4 or fp8
So a Qwen 480B should occupy 240GB to 480GB depending on https://t.co/I19HOYo5LJ
“ Baker said.
His working prediction: price around $2 trillion, then let the stock trade upward to roughly $3 trillion, possibly toward $3–4 trillion. Polymarket traders, for what it's worth, were already pricing an 80 percent probability of an Anthropic IPO this year and 67 percent odds that the company holds the top model at year-end.
The strategic reason this pricing matters extends far beyond one company. Sacks argues that Anthropic's quarterly disclosures will become the industry's single most important demand signal — the pace car for NVIDIA, TSMC, Micron, SK Hynix, and everyone else in the chain.
”
Gavin Baker Says Anthropic's $2 Trillion IPO Is a Sandbag — the Stock Could Trade to $3 Trillion · BigGo Finance
Quoted in printX (formerly Twitter)
“Internally, Anthropic is very confident. I have been told by multiple people I trust that Dario has said that Anthropic might be the only private company in the world at some point.
Think about that. In this vision, an Anthropic maximalist vision, there's Anthropic, and then there are governments, and that's it.
So there's a lot of confidence. I'm sure they have more advanced checkpoints than Fable up their sleeve. They've executed really, really well.
I would probably take the under on them being the only private company in the world. Yeah, I think it's gonna be a long time before they land a rocket.”
The All-In Podcast (@theallinpod) on X · https://x.com/theallinpod
X replyX
Replying to Shanu Mathew
30K! Surprised a bunch of you want to listen to my crazy takes but glad you’re following along. More than anything - I appreciate learning and debating respectfully with folks on here… and muting all the haters / cranky people 🤣
Here’s a photo of me with an Irish hat gift https://t.co/PG2dUDb9ld
Today is a pivotal day at @heronpower as we announce Heron Factory One, our first large-scale manufacturing facility, in Morgan Hill, California.
We will invest over $100M to transform a former distribution warehouse into a facility capable of manufacturing 10,000 Heron Links a https://t.co/hvSqBmKkeG
Looking forward to interviewing @ScottWu46 CEO and co-founder of @Cognition at the @TomorrowXSummit in Austin, TX November 17-18.
Cognition operates Devin, the first autonomous AI engineering agent which is deployed at some of the largest institutions in the world including https://t.co/ofsTWCEaFm
We ran GPT-5.6 Sol on Ultrafast mode through Humanity's Last Exam.
2,500 questions across chemistry, economics, literature - questions typically only PhDs could answer.
It finished the entire benchmark in 11 hours and 11 minutes.
Nearly 7× faster than Claude Fable 5.
Frontier https://t.co/EyEMAmBbv3
$3 trillion for Anthropic in the FT.
I feel like I’ve heard that somewhere else.
Although public markets generally struggle to accurately value growth when it is this high. https://t.co/2R80q18GUH
Congrats to @GavinSBaker and the team for a phenomenal lineup at the Summit.
I really wish I could be a part of it, but either way, I hope everyone that goes enjoys it. https://t.co/joTdpyFof6
My lawyer is obligated to in all but the most extreme circumstances; he will even defend me if he knows I’m guilty.
In contrast, the Claude Constitution places the AI's highest priority as Anthropic’s definition of the good of humanity.
I'm concerned this leads to a world https://t.co/oGBwZRDMnt https://t.co/m32YTpx5Eq
A monopolar or bipolar world of frontier models is a very dangerous world for humans.
This is why open-source is so important - we need a rich diversity of models that reflect the richness of human variation and let us all find the model that is best for each of us.
Also why so https://t.co/xxMglmQZJC
Like this tweet if you are attending the k
@TomorrowXSummit in Austin, Texas. Would love to schedule/have a morning 5K (miles) run meet-up followed by some some coffee and snacks a day or two before the event.
I had @spacexai Grok 4.6 follow Fable's plan, and with just a couple of nudges, it was able to repeat this feat in 1h 24m using 8.6M tokens at a cost of ~$55 at per-token pricing. That's about 1/10 the cost of the Fable implementation for the same work! https://t.co/dchAwtlWwc
Important. The future is almost certain to be multiple models working together.
Lower margins at the model layer whether from open-source or vertically integrated players = more AI infrastructure demand. https://t.co/YBWlCNnrDD
@GavinSBaker @andrewdfeldman @cerebras @TomorrowXSummit They deserve to be at 70% hardware gross margins.
Beautiful engineering, surprizingly bad unit economics.
Eric Vishria: "I am of the view that it all works."
"Will the CSPs do well? Yes.
Will these neoclouds do well? Yes.
Will NVIDIA do well? Yes.
Will these chip startups do well? Yes.
Are we gonna have big models in data centers? Yes.
Will the Fireworks of the world do well? Yes. https://t.co/DgFe8g0TZu https://t.co/dKfqkZLZwZ
@GavinSBaker @andrewdfeldman @cerebras @TomorrowXSummit I wonder if yield is gated by PVT across the wafer or reliability issues.
We will find out more next week.
@insane_analyst @andrewdfeldman @cerebras @TomorrowXSummit I would not recommend this course of action but I am happy to ask your questions. Just be sure you want me to ask and attribute to you *after* you have carefully reviewed their patent filings. Always looking out for you as appreciate your contributions both here and IRL.
@GavinSBaker @andrewdfeldman @cerebras @TomorrowXSummit Thanks for having @cerebras on your stage @GavinSBaker. Just make sure @andrewdfeldman gets some real Austin BBQ.
@masayoshisson @andrewdfeldman @cerebras @TomorrowXSummit Happy to discuss yields to put this silly issue to bed forever. In the meantime, I would point you to their patent filings in the public domain.
Looking forward to interviewing @andrewdfeldman CEO and founder of @cerebras at the @TomorrowXSummit in Austin November 17-18.
Scale-up networking and SRAM acceleration are the best solutions to the DRAM shortage from my perspective. Given that Nebius is monetizing at $20 to $25 https://t.co/ZUAzwy8A5d
@jeffrey_check @GavinSBaker They have a router that essentially identifies the areas where Grok 4.6 underperforms and it routes those requests to Opus 5.
Chetan is one of the absolute best VCs and has a real time understanding of what startups think of different models.
Vibes are good: https://t.co/jXUT0YJzwY
Chetan is one of the absolute best VCs and has a real time understanding of what startups think of different models.
Vibes are good: https://t.co/jXUT0YJzwY
Grok 4.6 is roughly the same performance as Fable 5 Max at an 85% discount. 80% cheaper for input tokens and 88% cheaper for output tokens.
Pareto dominant.
Grok 4.7 will be significantly better as is a much larger model with the Cursor and SpaceX data included in pretraining. https://t.co/4e8CXbzs2D
@thogge I thought Grok 4.7 *might* have been this good. Genuinely impressed that Grok 4.6 - which is the same base model as Grok 4.5 - made such a leap.
Gotta see how the all important vibes are - will know a lot more in the next 24 hours.
Grok 4.6 is roughly the same performance as Fable 5 Max at an 85% discount. 80% cheaper for input tokens and 88% cheaper for output tokens.
Pareto dominant.
Grok 4.7 will be significantly better as is a much larger model with the Cursor and SpaceX data included in pretraining. https://t.co/4e8CXbzs2D
Jensen was #1 on Glassdoor’s list of the best CEOs to work for. Still the same guy he was when I met him 27 years ago. Maybe slightly calmer.
The ultimate testament is that roughly half of the employees at Nvidia are worth over $25 million, 75% are millionaires and essentially https://t.co/iS2zgYE0jI
Jensen was #1 on Glassdoor’s list of the best CEOs to work for. Still the same guy he was when I met him 27 years ago. Maybe slightly calmer.
The ultimate testament is that roughly half of the employees at Nvidia are worth over $25 million, 75% are millionaires and essentially https://t.co/iS2zgYE0jI
I’ve joined Cursor / SpaceXAI.
AI is bound by compute, and SpaceXAI has the best near and long term compute roadmap of any AI lab.
Cursor + SpaceX as a joint entity is what sold me: Cursor’s product taste, coding data, and stellar GTM are now backed by the largest open reserve
Exclusive: Inside Fortell (@fortellresearch)
The $740M startup that had to waitlist billionaires and celebrities for its AI hearing aid.
Their custom chip helps you "hear where you look", distinguishing voices in noisy environments - which solved a problem no hearing aid had https://t.co/6xFrEkRKfp
.@lqiao is great. Here she is on the question every AI company has to answer: what is the ROI?
if what a customer pays doesnt cover the compute they consume, growth can just deepen your losses. So you can find product market fit.... and still “scale into bankruptcy” https://t.co/pYmJRG6n8M https://t.co/6ovU5x8k3r
Super excited that I will be interviewing @lqiao CEO and founder of @FireworksAI_HQ at the @tomorrowxsummit in Austin, TX November 17-18.
Lin’s vision of specialized intelligence is compelling and she is making it a reality. Fireworks is at the forefront of helping enterprises https://t.co/MK7CFgAsAU
Super excited that I will be interviewing @lqiao CEO and founder of @FireworksAI_HQ at the @tomorrowxsummit in Austin, TX November 17-18.
Lin’s vision of specialized intelligence is compelling and she is making it a reality. Fireworks is at the forefront of helping enterprises https://t.co/MK7CFgAsAU
Looking forward to listening to @AntonioGracias interview Jon Gray, President and COO of @blackstone at the @tomorrowxsummit November 17-18 in Austin, TX.
Blackstone is leaning into AI more than most of its peers and they are integral part of financing it.
So many topics: their https://t.co/6xK1XtNKl1
Dude @GavinSBaker we gotta get a drink. I feel this feeling you're feeling. Things are ripping and I keep gut-checking with tech folks at the bleeding edge every time we're together and they're always like "yeah it's even more awesome than you think" https://t.co/FECT2VmND8
This is such a great opportunity!
I believe attending Tomorrow X Summit will be an incredible chance to gain new insights and learn from so many inspiring people.
I’ve also submitted my application to attend, and I hope to see many of you there! https://t.co/CjZie4MkFO
Gavin Baker just posed the trillion dollar question, who's actually going to pay for all this AI spending?
Baker frames the whole AI capex debate around one question, since even if hyperscalers are under earning today and compute pricing normalizes, someone has to pay for that https://t.co/NqAVAnXOKi
After writing all the ALAB posts on my Substack, and talking about it on @semidoped, I'd better have a front row seat to this talk.
I'd be disappointed if I can't take audience questions in Jitendra's stead after all that research. https://t.co/ltrk9XVcgd
@vikramskr @semidoped Maybe we will have someone like you do a breakout in a different room/stage on each company separately before/after the interview. Thinking out loud here.
Create the opportunity to go really deep on each company.
Will come back to you. But maybe good way to integrate all the
MY DATA CENTER POLICY PLEDGE: OHIOANS-FIRST
A top concern I hear from Ohioans across our state – second only to property taxes – is the accelerating pace of data center expansion. Key citizen concerns about data centers include rising electric bills, noise, pollution, and the
Jitendra Mohan, CEO and co-founder of Astera Labs, is our third confirmed speaker for the @TomorrowXSummit.
Really looking forward to discussing the role of scale-up networking in a memory constrained world and the intersection of their switch franchise with optics. https://t.co/stwGpJ8zox
Jitendra Mohan, CEO and co-founder of Astera Labs, is our third confirmed speaker for the @TomorrowXSummit.
Really looking forward to discussing the role of scale-up networking in a memory constrained world and the intersection of their switch franchise with optics. https://t.co/stwGpJ8zox
@GavinSBaker @cursor_ai this is the real point here. cursor is the harness, and the harness is already the product. add your own data on top and it gets even stronger
Want to hate on @elonmusk? First try out my Musk Contribution Comparer. Just input your name and see how much you've given to humanity vs him. Then get a score of how much you're allowed to criticize him. Enjoy! https://t.co/zplA9J6Gjv https://t.co/iVgjRK9f1e
Really impressive from @cursor_ai
The future is trending hybrid. Next iteration will likely be allowing customers to use their own proprietary data to post-train an open-source model.
Excited for Grok 4.6. https://t.co/ZTE2VAskTG
@benitoz @ChaseLochmiller @CrusoeAI @TomorrowXSummit He will have an interesting perspective on the buildout.
Vertically integrated from dirt through electrons through tokens.
Happy to announce @ChaseLochmiller CEO and Founder of @crusoeai as the second confirmed speaker at the @TomorrowXSummit
This photo is my first time ever seeing Chase in a suit! https://t.co/mL4kKOTfwf
Brilliant call by @GavinSBaker on @patrick_oshag’s podcast @InvestLikeBest from a couple of months ago on Unity $U being one of the big beneficiaries of AI adoption and “one of our biggest positions” (or he might have said the “second biggest position”)! 🥳 https://t.co/mm67iBcjaR
Great podcast and thank you @GavinSBaker for citing @Silicon_Data token index. Gavin’s interpretation of the trend line is exactly right. The unit economics of tokens (cost per token) should not be confused with total market token consumption. Falling token prices don’t https://t.co/MUAr8FEiQJ https://t.co/7SWHVeC2nD
@carmenli @Silicon_Data Yes so much misinterpretation of your data despite your own crystal clear explanations of the right way to understand it!
Willful ignorance at best.
At 1X, 60 min. with @GavinSBaker is the most dense and valuable learning and information session you can digest. At 2x, the value and productivity is 🤯🤯🤯
Gavin inspires me to go deep, get facts, question, but believe in contrarian theses. Thx @patrick_oshag for another gem. https://t.co/1dBOShgIbO
Delighted to announce @JensenHuang as the first confirmed speaker for the @TomorrowXSummit and even happier that I will be interviewing him.
We will have Chicken Fried Steak at the summit. Sawyer’s is my favorite place in Austin but we will of course have Denny’s. https://t.co/4QKBMGh4TU
“you have the contracted base of installed compute trading at a massive discount to the current spot market, and as those contracts roll off and compute gets repriced higher, spot could decline and compute will still get repriced higher,”
“a company getting smarter about which model to use for which task, that, you know, that may lead to a a stabilization in their spend or even a decline, but it actually has nothing to do with the amount of, you know, GPU compute hours they are effectively consuming”
“There's 25,000,000,000,000 in knowledge work. And so let's, you know, let's say that that's, you know, let's take your 20% number. That's 5,000,000,000,000.”
“I think one of them gave an example of I forget if it was a plumber or an HVAC contractor, but, like, you know, everybody who's a blue collar worker is doing great because of AI.”
“I just think given that this is, the axis around which kind of everything is revolving, man, like, might blow up your entire business and your franchise by breaking an LTA.”
“I would be going to the buyers of GPUs, Traniums, and whoever and say, I'll participate in the NVIDIA credit wrapper. Now their business is just inherently less stable and predictable,”
“They think that SpaceX is gonna try and bring on eight gigawatts of compute. I will never bet against Elon, but, I mean, that would be a truly incredible feat.”
“I don't think anyone in '24 or '25 thought that the prices of old GPUs would be going vertical. Everybody thought, hey. We're gonna be smart. We're gonna sign these long term contracts.”
“Operating cash flow from Microsoft, Meta, and Amazon has reported accelerated from '28 to '32. There are some actually pretty big unusual items now like these hyperscalers.”
“But there is an unusual amount of onetimers this quarter. And if you adjust for that, we went from 28 to 35. That's that's a material acceleration at this scale.”
“What it was is they saw SpaceX have a big installed base of compute and sell some big trading optimized clusters into the market at a truly massive premium to these contracted rates.”
“I think one reason to the podcast two months ago, I got comfortable with that risk was just that you were seeing such incredible things out of anthropic.”
“They went on a podcast, and they essentially said, we are planning to pay 100% more for Blackwell's when our contract expires. And that just means that essentially all the hyperscalers are under earning.”
“So there's four companies that matter at scale. There's Amazon with their tradiums. There's Google with their TPUs. There's AMD, and then there's NVIDIA who's, like, much bigger than everybody else combined.”
“It's the single most important thing you could do to increase token output per unit of compute, and then that obviously, definitionally, actually lowers costs,”
“Well, if they're breaking their LTAs, it probably means, you know, your oversupply, prices are coming down, and then, you know, capacity naturally contracts. Well, what do you think is gonna happen to Google's allocations?”
“The amount of compute is going up and inference margins going up. And if you multiply those three, that's how you're getting this crazy acceleration into some of the labs plus open source,”
“And they're monetizing at something like 50,000,000,000 a gig and consensus estimates for next year are 73,000,000,000. So forget Starlink v three, forget Starlink direct to sell, Grok 4.5 and Cursor.”
“If they bring on anywhere near that, the consensus estimate is 73,000,000,000. That's eight gigs at 50,000,000,000 a gig. And, obviously, that would not all be lit up at the beginning of twenty seven.”
“But to this day, the only companies that have brought on more than 500 megawatts of power in a year are the hyperscalers, Coraweave, Crusoe, and SpaceX.”
Gavin Baker says compute players could risk their entire business if they break memory LTAs
"Let's just think about the game theory of breaking an LTA. So there's four companies that matter at scale. There's Amazon with their Trainiums, there's Google with their TPUs, there's https://t.co/KfZxMabKtm
I don’t usually listen to pods (I’m a reader) but this one (@GavinSBaker on @InvestLikeBest ) and the Gavin+Foxy+Brad+Clark one were fantastic and making me want to be a podcast person. well done https://t.co/db1Ox9wZpw
@GavinSBaker I know I'm a broken record on this, but I have been focused on how the nature of the players all have pro-momentum bias. Your framing is that what is driving the behavior of the players is leading to a loss of diversity. Same result IMO. https://t.co/MA6zilG5Uo
GAVIN BAKER: NVIDIA IS NO LONGER JUST A CHIP COMPANY.
Gavin Baker recently laid out one of the sharpest observations about the current AI hardware cycle, and it explains why Nvidia's multiple looks strange relative to the strategic position the company actually occupies.
The https://t.co/xmrJCXtjYH
Game theory suggests memory LTAs are likely to be durable.
And yes, Nvidia may have a significant cloud business via revenue shares. https://t.co/GFN1kDQmSk
Gavin Baker says the piece missing from the compute debate is SRAM accelerators, and disaggregating inference across three chips lifts AI's ROI.
"I do think something that is missing from all of this conversation about compute is what is going to happen when you put these https://t.co/S6xboh73jq https://t.co/JGxB3d8oJa
Claude is Walter Cronkite for the stock market.
@mjmauboussin talks about how a breakdown in diversity leads to bubbles and crashes. Gavin thinks Claude might be causing that right now.
”Everyone I know in the public equity investment business, whether retail or institutional, https://t.co/bEsWwUCmHf https://t.co/tm2oTCCuKv
@GavinSBaker pointing out that Hyperscale estimates are too low, which is contributing to unfounded funding/credit concerns…
“Consensus hyperscale cash flow from ops estimates of ~$1.3T implies that hyperscalers monetize capex at roughly rate of Ampere, two GPU generations https://t.co/IeqKiHrH9N
"Have you met anyone that's more bullish than you?"
Gavin: "Everyone out here is more bullish than me!"
"I look at what's happening in the stock market, and I feel like a foolish optimist.
But when I talk to anyone in this ecosystem, I'm bearish relative to essentially https://t.co/p6uxxPDzmF https://t.co/PwqzA7ttDd
Gavin Baker reveals a top startup that rented Blackwells in the mid $2 per GPU-hour now hopes to pay just under $4 seven months later.
"I literally spoke to a company this morning who rented a cluster, and this is one of the sexiest startups that people want to be in business https://t.co/VqeSmOXOUK https://t.co/H7IVrnXF5z
Super fun to do this again with @patrick_oshag
Randomly ran into Patrick in Palo Alto last Tuesday and recorded this on Wednesday when the AI complex was 35-40%ish off its June highs. https://t.co/11R0J7l9T7
Super fun to do this again with @patrick_oshag
Randomly ran into Patrick in Palo Alto last Tuesday and recorded this on Wednesday when the AI complex was 35-40%ish off its June highs. https://t.co/11R0J7l9T7
“SITUATION DETECTED: Ilya Sutskever's Safe Superintelligence is reportedly planning to launch its highly anticipated first model in August 2026, per Gavin Baker in an interview with Patrick O'Shaughnessy.”
I looked at 13 different providers for even 1 node of B200/B200s while I wait for my order to get delivered.
Zero availability.
I’ve never seen GPU capacity scarcity like this. Prices are also headed towards $6.50-7/gpu/hr. Expect inference to get more expensive. https://t.co/PtDvtEHJ8B
Today, we're launching Base Core, our own battery designed & built in Texas, and announcing our $1B Series D to bring it to homes across the country.
Three years ago, we started Base with a belief that batteries were the unlock to energy abundance. We saw battery storage as the https://t.co/3rjG59wZvE
Replying to Isaiah Taylor - making nuclear reactors
Announcing Valar Atomics' $1B series B led by Sequoia, with Valor Equity Partners, Atreides, Point72, Conviction, and others.
Alongside the $1B equity, we have closed a $200m credit facility led by Erebor and JPM.
I'm excited to welcome Shaun Maguire from Sequoia to our board. https://t.co/mkWPcEZFWL
Banning open source model, will "remove capabilities for the defenders"
When OpenAI's models breached Huggingface, a Chinese open model is what cleaned up the mess.
Anthropic's Fable 5 refused, so Hugging Face turned to Nvidia's quantized GLM 5.2.
https://t.co/km0tPh8VPO
A notepad in front of U.S. Secretary of the Treasury Scott Bessent reads "To Do Buy Japanese Yen $5-10 bil" as he participates in a cabinet meeting at Camp David.
The note, photographed at 11:33 EDT, came after Reuters earlier reported the Treasury had put banks on alert for a https://t.co/EEsEATHXGK
This is an interesting second order observation I have also been thinking a lot about. If all investors possess 100 points of “capacity” split between fluid and crystallized intelligence, as you age the balance should shift to more crystalized which also means you are arguably https://t.co/NWNrS3LmPh
We are going to see a lot of vertically focused AI native companies accelerate.
Routers, open-source models and specialized post-training enabled by companies like @FireworksAI_HQ have all made dramatic advances and the combination of the three is driving accelerating growth. https://t.co/f8qJdKSciI
Across every dimension, July was the strongest quarter-opening month in @WeAreLegora's history.
Net new ARR was 90% higher than April ‘26, our prior quarter-opening record. ARR grew more than 10x year-on-year. This reflects something specific: more in-house legal teams are https://t.co/GCcPa6NyrP
We are going to see a lot of vertically focused AI native companies accelerate.
Routers, open-source models and specialized post-training enabled by companies like @FireworksAI_HQ have all made dramatic advances and the combination of the three is driving accelerating growth. https://t.co/f8qJdKSciI
Prediction: SALP will be bigger than Citadel by the end of the decade.
Leopold has predicted the last two years better than anyone else - now that he can combine that with very expensive lessons in risk he will be unstoppable.
He has my full confidence. https://t.co/G2Slf7IULH
Really love how at multiple points in various interviews over the past few months people have tried to bait @GavinSBaker into hating on someone and he always takes the high road. Would be nice if more people on here followed that lead right now.
Spending a week with my family and working most of it! Being a few hours ahead of US markets has given me some time to digest a lot of incredible markets discourse on X - there is too much to try and consume about AI but not enough about leverage. We are seeing a classic head
I don’t understand why this FundaAI report hasn’t received more attention. SpaceX plans to build 4 GW of compute capacity using Rubin racks next year.
$NVDA $SPCX https://t.co/M2DeFWmYnm
@jukan05 4 gigawatts this year and another 4 next year per this report. Spot price for compute has gone up since they signed their Anthropic and Google deals.
"Lab compute 3x-es year over year. For a lab to 10x revenue while continuing to only 3x compute, some combination of the following 3 things has to happen:
1. Lab margins have to increase,
2. The price of compute has to increase,
3. Labs have to spend a greater fraction of https://t.co/MrAJwVczqE
A series of fairly wild thoughts from Dwarkesh. I thought I was bullish but even I am not assuming that the price to rent compute continues to go up!
I will say that I am in SV this week and Dwarkesh is capturing the zeitgeist.
Net, net public markets would probably be trading https://t.co/PzIxxnzbiS
“Hyperscalers are underearning and anyone who signed a contract for GPU compute in 2024 and 2025 is overearning. Operating cash flow will be enough to fund capex but as contracts reprice and cloud growth continues to accelerate then spreads likely come in as well.” https://t.co/YSCOxnnUBA
We’ve seen 15%–30% price increases year-to-date across our A100, H100, H200, and B200 neocloud on-demand indices. Every day, we ingest 170,000+ observations from hundreds of data sources worldwide to provide an independent view of the compute market. https://t.co/Jg2gxcWZk0
@GavinSBaker Fully agree, and the first referee prints tomorrow night. Microsoft did $42.6B of operating cash flow in last year's June quarter and has been growing it around 30%. Trend says 55B. If Microsoft matches the acceleration you describe, that's $60B or better.
@GavinSBaker I see. Where does that cash flow come from? To generate that sort of cash flow someone in the economy has to leverage up/(borrow from banks or capital markets) at least temporarily.
@dampedspring Where has Anthropic’s and OpenAI’s combined 150b of ARR come from?
I don’t think the idea that leveraging up is the only way to fund this is at all correct.
@GavinSBaker Credit markets also an important governor around the pace of buildout.. If spreads continue to widen, CapEx $ spend elongates across a longer timeframe and the likelihood of an overbuild declines. Probably a better outcome versus a scenario where lenders imposed no discipline on
@GavinSBaker I guess that could imply prior contracts rolling off in a secular upswing rather than current long duration contracts being in backwardation, although that would seem strange for a market with this much capex. Thanks.
@GavinSBaker Not sure about over or under reacting but the capex needs funding. Funding has a cost. The cost is going up as stock prices fall and corporate bond spreads rise. These things also may signal indigestion in actual dollars available to be raised to generate capex funding so
@dampedspring Of course. Real yields going up and spreads widening is undeniable. I am saying we may need less credit than expected given a likely continued acceleration in operating cash flow.
@GavinSBaker Nah you are not a contra by any means because you're playing a totally different timeframe but I do think there are recent swing momentum folks positioned off of tweets like this which I believe is not good for short term price action, but totally meaningless for multi month
My snark gets the better of me sometimes but just to be clear Gavin cannot be a contra for me because we're playing different time frames. Everything he writes makes sense fundamentally and obviously he has been right all along with a deeper understanding of the space than I https://t.co/4SUBBaoJM4
@GavinSBaker Credit spreads, as a leading indicator for equities and in general, have actually lagged every cycle since 2016. Equities have led credit/debt markets each cycle downturn. @markoinny wrote a piece on this back in 2014
@GavinSBaker Why is this any different from say how an offshore rig operator like Transocean runs their business? Why shouldnt the market be skeptical on the durability of this “trade” just like they are with other cyclical industries? Isnt there real precedent to be skeptical about the capex
@NoonHillCapMgmt Always wise to be skeptical. Nothing wrong with the analogy, just the barriers to entry are higher, quality of companies is materially higher and structurally harder to bring on capacity given shortages.
@realSharat Genuinely happy to be a contra for you. I absolutely love being a contra for X accounts - the last time was widespread IIRC was May-December 2022. Good luck!
Market is overreacting to hyperscale credit spreads widening from my perspective. TL;DR Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning while operating cash flow acceleration is an underestimated source of funds for AI https://t.co/D9FSHXQ8F8
Market is overreacting to hyperscale credit spreads widening from my perspective. TL;DR Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning while operating cash flow acceleration is an underestimated source of funds for AI https://t.co/D9FSHXQ8F8
Two months later, updates to my AI maxi/mini takes:
Maxi takes:
1) ✅Cybersecurity maxi take is looking great. Easy to imagine that in 6-12 months, roving defensive agents are pen-testing every application, etc.
2) ☑️Code growth is looking mixed. Fable 5 became a new daily https://t.co/GQ96Z0ZAJ7
My increasingly strong opinion is that compute is the biggest moat in AI.
Scientific breakthroughs don’t stay secret for long. Papers get published, researchers move, ideas leak, open source catches up. Give it enough time and almost everyone has access to roughly the same https://t.co/C2dMw7WMbz
@GavinSBaker But it isn't real, the US is failing to demonstrate useful openweight models at scale and are at least 6 months to a year behind China and doing so. Any PR on this topic is just lobbying.
And I don't believe you actually care.
Replying to Elisabeth | AI Systems + Creator Growth
@GavinSBaker the specific detail in that incident is the part people are skipping.
hugging face could not run forensics on the closed model. it could not distinguish an investigator from an attacker, so it refused.
they moved to glm 5.2 open weights on their own infra and reviewed 17,000+
An outcome where AI is dominated by a few dominant frontier models is dangerous for America.
Super negative for national security. The Hugging Face incident and Anthropic/DoD dustup are object evidence of the risks.
Happy to see this. https://t.co/xOFEhDxKs6
AI security advances when the industry builds in the open, together.
We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents.
By sharing models, tooling and research in the open, we can broaden the https://t.co/gfhKfrgcbl
An outcome where AI is dominated by a few dominant frontier models is dangerous for America.
Super negative for national security. The Hugging Face incident and Anthropic/DoD dustup are object evidence of the risks.
Happy to see this. https://t.co/xOFEhDxKs6
@GavinSBaker @jukan05 Simple version: DRAM is the workbench; an SSD is the filing cabinet. When the job stops fitting on the bench, the server spends its time moving data instead of doing useful work.
The 38x slowdown occurred in one stage, not the entire workload. Overall throughput improved 2x to
“38x slowdown once shuffle data exceeds available DRAM and spills to SSD.”
From a joint Micron/Meta white paper.
The DRAM bottleneck almost certainly lasts the longest.
OP from @jukan05 https://t.co/a5Xb3Q1P85
Imagine you’re the banker writing Anthropic’s S-1 and now have to add “massive aura loss from employee tweets” to the risk section https://t.co/ZOt0j4GYTa
Nvidia is the leading open source AI company by a wide margin.
The silliness of this post is compounded by the fact that the letter did not suggest that closed source AI companies should open source their models. https://t.co/OMbuz8LPrp https://t.co/ejxkK3rQEB
The future is almost certainly multi-model with closed frontier models co-existing with open source.
Therefore open source is essential to American competitiveness.
And a future with only a few models would be dystopian, risky and exacerbate inequality.
I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release! https://t.co/d8vhGRpZPx
Nvidia is the leading open source AI company by a wide margin.
The silliness of this post is compounded by the fact that the letter did not suggest that closed source AI companies should open source their models. https://t.co/OMbuz8LPrp https://t.co/ejxkK3rQEB
New Pareto Frontier: Grok 4.5, SWE-1.7 and Opus 5.
Intelligence per $ will be the only metric that matters over time.
K3 probably joins the frontier once available on the inference clouds. https://t.co/jdMFn2Vya8
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. https://t.co/t02bi51N4C
@GavinSBaker @carm1nee I’ve been doing that morning light thing too and also waiting an hour to drink my coffee (which lately is Rarebird paraxanthine coffee - it is coffee without the jitters)
Gavin Baker on the SpaceX situation: "Wall Street consensus is $73 billion in revenue - if they bring on 3 gigawatts of power next year that's $150 billion - that's not in the model"
this is him explaining why SpaceX could double what the Street expects, why intelligence per https://t.co/CT53fh2KDK https://t.co/SonIY0WX75
Also can’t get over the idea that Nvidia is the leading supporter of open source AI and many people seem to believe that open source AI models would be negative for AI infrastructure demand and Nvidia.
And obviously the $630b would be at the system level and not all go to Nvidia unless they change their business model to lower their margins. https://t.co/cPvyYEOYsO
Also can’t get over the idea that Nvidia is the leading supporter of open source AI and many people seem to believe that open source AI models would be negative for AI infrastructure demand and Nvidia.
Interesting report from @theinformation that Nvidia will be able to make 1000 Vera Rubin racks per day, which is $630b per quarter. Actually a little hard for me to believe and haven’t checked the math, but wow if true.
And Vera CPU racks and Groq LPU racks would be incremental
And obviously the $630b would be at the system level and not all go to Nvidia unless they change their business model to lower their margins. https://t.co/cPvyYEOYsO
Interesting report from @theinformation that Nvidia will be able to make 1000 Vera Rubin racks per day, which is $630b per quarter. Actually a little hard for me to believe and haven’t checked the math, but wow if true.
And Vera CPU racks and Groq LPU racks would be incremental
@GavinSBaker @garyHeff @GavinSBaker you’re rational takes and thoughtful insights do a lot to calm investors if not markets. I know I’m much happier when you’re restating what I’m thinking and seeing…at least then I know you’re still seeing it too. We all greatly appreciate you
This post is key. The cheaper AI gets, the more opportunity there is for the entire ecosystem - especially including end-customers - to benefit.
Everything is bottlenecked by being able to successfully and cost effectively deploy AI in real workloads. Any time we can lower the https://t.co/rXi1igW73C
@GavinSBaker appreciate the balance in your analysis here - imagine it must have been not very easy to write, but the clear-eyed take says everything
thank you
@GavinSBaker Completely agree with this "Lower margin % at the model layer = more margin $ at every part of the infrastructure layer and is a godsend for software." but how do you think about the financing risks that run through OpenAI and Anthropic?
In a perfectly rational world they would
Kimi K3 may be an important inflection point for AI. Potentially negative for Anthropic and OpenAI while being net positive for essentially every other company in the world. I mean that very literally. Although the real “Sputnik moment” would be an open-source frontier model that https://t.co/tdAkU4oRgo
Risk/reward seems attractive again.
Lots of cheap stocks with durable competitive advantages that are going to crush numbers for the next 6-12 quarters.
Time will tell!
Congratulations @FireworksAI_HQ
@lqiao has built the dominant platform for owning and serving your intelligence. Beast of a platform from the kindest person.
Delighted to welcome @GavinSBaker aboard! https://t.co/B8KF6mlkgF
I think there is general confusion around how AI works, AI tokenomics, and ultimately *what is actually priced in* for the AI trade - and that some of the existing arguments are at odds with one another
Firstly to clear this up - what Brad and Gavin are saying are completely in https://t.co/ZF3hNA4TnA
It would be pretty funny if Hynix missed the quarter immediately after their $7b (or whatever it was) US ADR listing and super bullish roadshow where they made fun of Micron for agreeing to price ceilings in their LTAs.
Semianalysis well above consensus for this Q, KIS spec
Much stranger things have happened but would still be funny.
Also the 1981 “Clash of the Titans” has aged really well even if the Kraken has not. https://t.co/Aml7md4vYV
Curious to see how many subs @SemiAnalysis_ has in South Korea and whether they are more or less respected there than KIS, whose spec sales note was apparently a contributing factor to last nights sell off.
A modern day Clash of the Titans.
Should know in a few hours! https://t.co/mAvNP8Wuj6
It would be pretty funny if Hynix missed the quarter immediately after their $7b (or whatever it was) US ADR listing and super bullish roadshow where they made fun of Micron for agreeing to price ceilings in their LTAs.
Semianalysis well above consensus for this Q, KIS spec
Curious to see how many subs @SemiAnalysis_ has in South Korea and whether they are more or less respected there than KIS, whose spec sales note was apparently a contributing factor to last nights sell off.
A modern day Clash of the Titans.
Should know in a few hours! https://t.co/mAvNP8Wuj6
Understanding that Gavin is mostly right here, but also partially wrong, and learning how to discern when you are likely right vs. the market and when you're likely wrong, is the most important skill in investing.
Gavin is entirely correct that you need to learn whether you are https://t.co/Un1CYhREYT
@BeenThereCap Fully agree with all of this - sometimes not room for nuance in investing. Doubling down late is certainly not always the right thing to do!
@GavinSBaker best line from Tiger Cub checklist:
"Does waiting for the new financials feel like waiting for Christmas? IF YES --> ADD"
https://t.co/lKeA8t4t6d https://t.co/mD630oz2w8
Said this to my CIO and he replied “compensation is when you are frustrated that your bonus wasn’t cut even more” and then shut his office door rather aggressively https://t.co/4RcuyokF0H
This is a thoughtful framing from @GavinSBaker , and directionally I agree that cheaper models should increase intelligence / dollar, stimulate demand, Jevon’s etc (AI infra bull here, duh). I just think it is too early to make the either/or calls generally in AI and especially https://t.co/cMPuEi3AHt
@GavinSBaker If you abstract away the fact that this is AI, and you solely looked at token price and volume as an output of rising infrastructure costs, wouldn’t the only logical outcome eventually be that one of the models figures out a way to commoditize the core complement(s) (currently
This is true as I have heard this from contacts in the Valley. Goes with my pinned post.
The AI race is shifting from bigger models to cheaper, smarter systems https://t.co/lS1YKxkjl0
The mega bull case for AI infrastructure would be *if* market share shifted away from certain frontier labs with 90%+ inference margins toward cheaper models, whether open-source or closed.
It would increase the ROI on AI spend for end customers by increasing intelligence per https://t.co/Z0kE8gs22v
Ken Griffin says short-term alpha from calling the quarter with alt data is dying, and the next era of stock-picking belongs to those with a long-term horizon.
"I think there'll be more focus and emphasis on those who have really good vision about what companies are actually https://t.co/DGAAtfup29 https://t.co/oHFIITUwBP
Might make @tomorrowxsummit planning sessions formal events from now.
I did not own a white dinner jacket as of Friday, but going to be standard for me from now on when a tuxedo is mandated.
Fun hanging out @AntonioGracias and @rbiscardi
Also best wedding band I’ve ever heard. https://t.co/YfqglqJrHI
Might make @tomorrowxsummit planning sessions formal events from now.
I did not own a white dinner jacket as of Friday, but going to be standard for me from now on when a tuxedo is mandated.
Fun hanging out @AntonioGracias and @rbiscardi
Also best wedding band I’ve ever heard. https://t.co/YfqglqJrHI
@Networthknowing I applied, but @GavinSBaker has not accepted my application as of yet - Gavin, I can bring some potato salad or something if you’ll have me
“And a very important distinction we will come to is whether or not that build out is funded out of debt or cash flows. It's a critical distinction for AI.”
“In the bubble, you went down 80 or 85. Dramatically different. And so the scars are so deep from that bubble that just that puts a lid on tech valuations.”
“But even if we solve the shortage of Watts with orbital data centers, the TSM is still a bottleneck. And semi's and semiconductor manufacturing, you have a big advantage if you were ahead.”
“And semi's and semiconductor manufacturing, you have a big advantage if you were ahead. Everybody it's like baking. Everybody uses the same ingredients or whatever.”
“And the nature of AI just as AI eats the world, silicon is going to eat the world because it's so much more computationally intensive than deterministic software written by humans.”
“AI, even if you put a harness on it, even if you do the chain of thought, even if you have multiple agents, it's probabilistic that it is recomputing the answer each time.”
“There's a chapter about him at the Walter Isaacson book. He was on the field having an impact on the outcome, along with the incredibly talented people at Tesla and Elon.”
“Monopoly Go is a game where we have the revenue and the hours played and it is to render it using list prices for something like VEO3 is more than two orders of magnitude greater than its revenue. So what is the role for human creativity, man? I don't know.”
“So what is the role for human creativity, man? I don't know. I don't know. Like what I hope, one of the reasons Neuralink is such an exciting company”
“She had this phrase that I just always think about. Ultimately, as an investor, you either have to panic early or double down late. And essentially no one does both.”
“And I always give the same advice, be kind and be scrappy. A lot of people who are kind are not scrappy and a lot of people are scrappy are not kind.”
Boom! Super excited to attend. I can't wait to have some insightful conversations and rigorous debates around these topics.
Heading to TomorrowX Summit 2026 in Austin, Nov 17-18. Two days with the CEOs, investors, and operators moving TMT. @TomorrowXSummit #tomorrowXsummit2026 https://t.co/xM27U9ftt3 https://t.co/7u1xYyWE1E
Grok 4.5.
Pareto dominant for coding by the numbers.
We will see on the all-important vibes.
Instinct is the benchmarks are likely directionally accurate given the stated focus on real world utility. https://t.co/aLihvvWfwA
Grok 4.5.
Pareto dominant for coding by the numbers.
We will see on the all-important vibes.
Instinct is the benchmarks are likely directionally accurate given the stated focus on real world utility. https://t.co/aLihvvWfwA
Thoroughly enjoyed doing this podcast with @AmirF15336 who is an impressive young man.
Recorded in late February of this year.
Time flies! https://t.co/4cbAFN6RtT
Thoroughly enjoyed doing this podcast with @AmirF15336 who is an impressive young man.
Recorded in late February of this year.
Time flies! https://t.co/4cbAFN6RtT
Will be attending @TomorrowXSummit in November at Austin.
Thanks @GavinSBaker for providing this great opportunity!
See you then! https://t.co/UFbhtfwSr6
Introducing Devin Security Swarm
A more cost effective and accurate way to find security vulnerabilities in complex codebases, based on a new architecture: Agentic MapReduce. https://t.co/bcVDGWHaxf
@GavinSBaker c’mon man I know imitation is flattery but this looks truly horrendous! I can’t have people thinking this is my work. https://t.co/koOHgCFEwZ
@tyrrellwinston This was at a rented house - I promise I do not own anything that looks like your art - but seems like you have done some really cool stuff!
The largest, frontier OAI model at nearly 750 tps on Cerebras!
"We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed." https://t.co/VBiWsDBEeW
This is one of the coolest things I have ever seen someone do on this app.
Took a shot at it and encourage other passionate young investors on here to do the same. https://t.co/dpSuy9NzJd
@GavinSBaker @nvidia Did you not read The Thinking Machine? It’s the seminal book on the man and company - even if the author takes an idiotic turn about AI’s threat to humanity? It discusses his love of Denny’s and chicken fried steak. It’s a must read.
Great day @nvidia
And today I learned that Jensen’s favorite food is Chicken Fried Steak, which he had soon after arriving to Kentucky as a 9 year old boarding student who cleaned the school bathrooms. 🇺🇸 🇺🇸 🇺🇸 🇺🇸 🇺🇸 https://t.co/B4sa59JKsR
Retail, Substack, Reddit and most of all @X accounts are increasingly the most important forces in the stock market and yet there is no way for them to connect directly with the management teams of the companies they are writing about and investing in. And they have almost no
@BirthrightMargs @X Bears are super welcome as long as they are kind in their dialogue as opposed to engaging in ad hominem attacks. It is possible to respectfully disagree with someone.
Retail, Substack, Reddit and most of all @X accounts are increasingly the most important forces in the stock market and yet there is no way for them to connect directly with the management teams of the companies they are writing about and investing in. And they have almost no
@pmarca I would like @atreidesmgmt to be involved.
Something needs to be done about most of the ESG and proxy advisor firms. They are significant negative externalities for the world.
wow.
Well @GavinSBaker I'm afraid to say that you're wrong about distillation, and more obviously you should get your terms straight. Only the Chinese have ever released frontier models with open weights. You mean "released at all". https://t.co/Rg9DMXfhHn https://t.co/dvhFNre8TQ
@teortaxesTex Agree to disagree on the importance of distillation - only a few reasoning traces make a difference. Am super confident in this.
And yes, yes, yes on open weights - I will try to do better thank you. Although Nemotron is an American model that is pretty close to the frontier and
@AntonioGracias @reindsummit @McDonalds More photos like this please.
Also kinda funny that you went from correcting Bloomberg’s reporting about Valor’s SpaceX ownership in your last post to fried apple pie in the next.
The American Dream in action!
And who doesn’t love both McDonalds and SpaceX? As 🇺🇸 as it gets!
I hope that Atreides and Valor are able to secure the franchise rights to @McDonalds on both the Moon and Mars. And Valor has actual expertise running franchised restaurants!
Who can help make this happen? I will put a deposit down today unless SpaceX wants the rights. https://t.co/9UkcsiIdhY
@GavinSBaker @reindsummit Beat you too it Gavin! I took this pic on my way out of Detroit to send to my kids. Thank you @reindsummit and @McDonalds for making my visit to my birthplace very special. https://t.co/xe224sTfcL
@AntonioGracias @reindsummit @McDonalds More photos like this please.
Also kinda funny that you went from correcting Bloomberg’s reporting about Valor’s SpaceX ownership in your last post to fried apple pie in the next.
The American Dream in action!
And who doesn’t love both McDonalds and SpaceX? As 🇺🇸 as it gets!
Just as I land in Detroit, my birthplace, for the Reindustrialize Summit, McD’s brings back one of my favorite childhood desserts. Can’t be a coincidence! Great memories of happy meals and apple pie with my Mom as a kid. I love America! 🇺🇸 @reindsummit https://t.co/aKPbp4N04G
"If token throughput per watt rises faster than price per token falls, revenue per gigawatt can net expand. This is compounded by models getting smarter, which unlocks higher value tasks for end users [justifying increased costs]"
Well said @ShanuMathew93 and @downingARK
"If token throughput per watt rises faster than price per token falls, revenue per gigawatt can net expand. This is compounded by models getting smarter, which unlocks higher value tasks for end users [justifying increased costs]"
Well said @ShanuMathew93 and @downingARK
Kinda wild how wrong most energy experts were about the Straits of Hormuz.
What were the root causes of the error? China had more reserves than anticipated, US flexed up exports faster and more ships were actually transiting the straits via the Omani route?
Genuinely curious.
@GavinSBaker @tbpn Cooling transponders is different to cooling data processing. I'm not supposing ongoing projects are not thinking about cooling, but to say "you don't need that in space" is exactly the kind of oversimplification that lends itself to irresponsible meme stock investment.
@tbpn @GavinSBaker can you we stop pretending like starship being reusable is the base case. it has proven to be much harder than they anticipated, and to date has been a failure whether the faboys admit it or not
Always a must-listen. Gavin Baker (@GavinSBaker of Atreides) broke down the SpaceX IPO, orbital compute, retail, China, and where AI value accrues, on @tbpn (~0:35 to 1:10 min mark).
Here were the biggest takeaways:
1. The orbital-compute math is starting to pencil out. https://t.co/A03VcfFxgt
@GavinSBaker @tbpn Cooling transponders is different to cooling data processing. I'm not supposing ongoing projects are not thinking about cooling, but to say "you don't need that in space" is exactly the kind of oversimplification that lends itself to irresponsible meme stock investment.
@tbpn @GavinSBaker You actually need more tech for cooling in space. Also, radiation shielding. Also, no easy fixing damaged circuits, no easy upgrades. It's not so simple.
@ignace @tbpn Yes, no one from SpaceX has thought about the cooling or the radiation shielding for the circa 10,000 satellites they are currently operating. Insurmountable problems. And in all seriousness yes, RMA might be an issue but 40b vs. 60b of capex, minimal ongoing opex and superior
Atreides' @gavinsbaker says that once Starship is reusable, the economics of orbital compute crush the economics of terrestrial compute.
He lays out the math:
"Once $SPCX can reuse Starship, the math for orbital compute becomes pretty compelling."
"It's $60B to bring on a GW https://t.co/fQwdi6oAWa
Thoroughly enjoyed my good friend Alex Sacerdote’s appearance on ILTB.
I keep thinking about his point that we have an acute compute shortage with 0.1% of the world’s population using agentic AI.
I have learned a lot from Alex over the last 27 years!
Thoroughly enjoyed my good friend Alex Sacerdote’s appearance on ILTB.
I keep thinking about his point that we have an acute compute shortage with 0.1% of the world’s population using agentic AI.
I have learned a lot from Alex over the last 27 years!
Just learned I am going ton @tbpn at 2 PM EST today if anyone is interested. My first time.
Evidently has been on my calendar for a while but was news to me.
Should be fun - the guys seem great.
Just learned I am going ton @tbpn at 2 PM EST today if anyone is interested. My first time.
Evidently has been on my calendar for a while but was news to me.
Should be fun - the guys seem great.
Tremendous respect for @GavinSBaker @altcap @_clarktang & their teams. One of the best pods of the year.
SpaceX becomes hyperscaler #4 in record time
-Effectively went from not focused on it to the #4 in roughly ~30 days, past Oracle and others, on the Google and Anthropic cloud https://t.co/xTD6HdujNV https://t.co/fScy3KgbCG
@GavinSBaker blessed to be growing old together - it’s been such a journey and amazing how spacex has made our paths intertwined. here’s to the next chapter(s).
Such a fun day with so many old friends.
Always a little sad to leave Starbase.
Per aspera much of the time so important to celebrate sometimes.
So grateful to everyone at SpaceX for helping to make the future awesome.
Excited for the next Starship launch. Ad Astra! https://t.co/RP0DYbnIwO
What an extraordinary day.
Deep gratitude to Elon, Gwynne, Juncosa, and the entire SpaceX team for their relentless dedication to building a brighter future.
And thank you to the investors who believed early and stayed the course; Steve, Antonio, Luke, Tamas, Hering, Gavin, https://t.co/mIAKAtR3ii
Oh yes, I remember that Bond film where the villain decarbonized the auto industry, brought fast internet to everyone on the planet, and helped paralyzed people interact with the world again. https://t.co/uMjVmjXHlp
GAVIN BAKER: “A month ago, @SpaceX did not have a cloud computing business. Now by some measures, it is the fourth largest cloud ahead of Oracle — and they did that in 1 month. Let's see what they do in a year.
“A really important variable for SpaceX is how many gigawatts can https://t.co/N9L2dC6s4d
GAVIN BAKER: “A month ago, @SpaceX did not have a cloud computing business. Now by some measures, it is the fourth largest cloud ahead of Oracle — and they did that in 1 month. Let's see what they do in a year.
“A really important variable for SpaceX is how many gigawatts can https://t.co/N9L2dC6s4d
Let’s see whether @MorganStanley or @GoldmanSachs is the first to post a photo of their traders or bankers wearing green shoes which was a @SpaceX mandate. Gotta keep it fun.
There might be a prize!
The $SPCX IPO prices todays at $1.77 trillion. Brad Gerstner (@altcap) and Gavin Baker (@GavinSBaker) just spent 80 minutes on @BG2Pod making the case for why that's the cheapest it's going to get.
Here are the 10 takeaways worth saving:
1. In 30 days, SpaceX added $29 https://t.co/mXBTTTbPQn
A richness of great insight on the @BG2Pod from @GavinSBaker ...too many to choose from but here's a few good ones:
“Something very important on open source: there’s this belief that it’s bearish for AI. It’s actually — maybe bearish for the frontier models — but really bullish
Let’s see whether @MorganStanley or @GoldmanSachs is the first to post a photo of their traders or bankers wearing green shoes which was a @SpaceX mandate. Gotta keep it fun.
There might be a prize!
Thanks for the shout-out, @GavinSBaker on @theallinpod Liquidity.
The network is the bottleneck for AI cluster performance. It's the whole reason we built @AriaNetworks #DeepNetworking: telemetry-first, more usable tokens per GPU.
GA and shipping now. If you're building or https://t.co/NowUa2Nosz
@mansourkaram @theallinpod @AriaNetworks 383 days from company formation to first PO for a hardware company was impressive.
Aria can help optimize Nvidia clusters for unified front-end/scale-out networking use cases especially around observability.
But Mansour, you still have a lot of wood to chop. Keep chopping!
Evidently prices will go up in about two weeks.
Also I preferred Mythos to Fable.
Fable makes me think of fabulist, which is not the best association.
Excited to try it! https://t.co/9X7uMVo7AD
@GavinSBaker @polynoamial If test-time compute keeps unlocking gains without plateauing—especially in frontier models—then we're not just underestimating inference demand. We're staring at an exponential flywheel: smarter models + longer thinking = civilization-scale compute hunger. Energy markets, grids,
Super important post from @polynoamial and the investor TLDR is: all current estimates for compute demand might be low.
“We likely don't know what the capability ceiling is for modern LLMs because it's too expensive to measure.
Frequently when I discuss this, people ask why we
Super important post from @polynoamial and the investor TLDR is: all current estimates for compute demand might be low.
“We likely don't know what the capability ceiling is for modern LLMs because it's too expensive to measure.
Frequently when I discuss this, people ask why we
Really fun to interview my old friend Bret Johnsen in Mission Control.
Three parts of the @SpaceX story that I wish were more widely discussed:
SpaceX has created thousands of good blue-collar jobs: welders, machinists, electricians. Everyone talks about the need to bring https://t.co/240CmtVrQh
@GavinSBaker @SpaceX The most inspiring part of the SpaceX story isn't just rockets reaching space, it's people refusing to give up after failure. From welders signing their work with pride to engineers dedicating years to an almost impossible mission, SpaceX proves that extraordinary achievements
@GavinSBaker @SpaceX This is the kind of SpaceX story that actually matters. Thousands of high-paying blue-collar jobs bringing real manufacturing back. Milestones in Mars that are already saving lives on Earth through Starlink, telemedicine in remote Kenya, schools in Brazil, disaster response in
Really fun to interview my old friend Bret Johnsen in Mission Control.
Three parts of the @SpaceX story that I wish were more widely discussed:
SpaceX has created thousands of good blue-collar jobs: welders, machinists, electricians. Everyone talks about the need to bring https://t.co/240CmtVrQh
Crime is now a choice.
Agree with @GavinSBaker in @chamath, @Jason, @DavidSacks and @friedberg's recent @theallinpod episode.
Safety used to follow wealth. Where there was money, there was safety. Where there was not, families were left to absorb the difference. Today? https://t.co/5Wq7LTp0f5
SWE-Bench style grading has been the standard for years now - you ask the agent to solve an issue and then run its code on a pre-constructed unit test.
The problem is that passing a unit test is only one part of writing production-ready code. You also want to evaluate agents on https://t.co/i4yvOhbryR
This is super big
I think this is the first useful speculative decoding method deployed on a big quasi frontier model
Massive unlock
@fi56622380 https://t.co/augiaFLDOK https://t.co/586sM8lr8M
A drone startup couldn't get FAA approval to test in the US so they went to Africa instead.
Seven years delivering medicine and blood to remote villages. Maternal mortality cut by 90%.
Now they're back in America with seven years of real world data nobody else has. https://t.co/41GgrBUvBZ
Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality:
🧠 LLMs
→ NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput https://t.co/OGP9ApUXNi
Quite a week for open-source AI. Especially American open-source. Nemotron 3 Ultra is the most important release in quite some time. And some really cool RL and fine-tuning work from Harvey. https://t.co/dK4h4GNTaK
$TSM CEO Pushes Back on @GavinSBaker, Clarifying That TSMC's Bottlenecks Are Organic, Not Self-Imposed
"When asked by a reporter, Gavin Baker, a technology investor who previously managed over $17 billion in assets at Fidelity and is now the chief investment officer at Atreides https://t.co/99ySORBHXn
I believe I said “tough, flinty, disciplined” rather than stubborn.
I deeply admire TSM management and believe their capacity discipline, whether organic or self-imposed, is good for everyone.
We want a smoother for longer cycle. An overbuild and/or a bubble would be terrible. https://t.co/o6d4mFmryT
In May 2017, he launched Atreides Management as a crossover tech and consumer investment firm.
Mentored by legends like Will Danoff, Joel Tillinghast, Steve Wymer, and Jennifer Uhrig, he earned a reputation as one of the most voracious learners of technologies — respected by
In high school, Gavin Baker defended a younger student from bullies. That student's father president of the Greater Texas Dartmouth Alumni Association, securing his admission to Dartmouth College.
@GavinSBaker Sigh….
“If the United States hadn’t forced our country, our companies, and our industry, we wouldn’t have done something like this. But we are also grateful to the US for enabling our country’s semiconductor industry chain to truly grow. Now the momentum is very good, and
Get on the plane!
I loved Dan's description of what makes a great analyst today relative to 20 years ago.
It reminded me of @citrini analyst #3, who flew to the Strait of Hormuz during the war.
Being physically present is still one of the most underrated advantages for https://t.co/wWLNaMHNGj https://t.co/aUpVKSlnoz
Grok Build created a customized CRM with full import capability from Salesforce or Hubspot with 5 prompts. Up and running in under 4 hours with user interface plus drag and drop custom reporting. https://t.co/VM1Xa0v6r8
Fast, faster, Qwen. 🚀
Thrilled to see Qwen3.5 reaching a record-breaking 580 tps for agentic workloads on the TokenSpeed engine! This milestone wouldn't be possible without our incredible partners.
Huge thanks to @lightseekorg, @NVIDIAAI, the Mooncake team, and @tri_dao for https://t.co/nCOT6us2Kt
1/ We’ve raised over $1B at a $26B valuation, led by @Lux_Capital, @generalcatalyst, and @8vc.
Our enterprise usage has grown >10x since the start of this year, and our run-rate revenue grew to $492 M.
We launched Devin two years ago as the first AI software engineer. Since https://t.co/k99LLLyWhZ
Composer 2.5 being Pareto dominant in coding per CursorBench is important.
This is after only a few weeks of supplemental training and/or RL in the Colossus 2 cluster.
The 1.5 trillion parameter version of Grok will likely be a much better base model than Kimi. We shall see. https://t.co/47htukDrk3
Gavin Baker had a really good take on the AI market recently on @InvestLikeBest podcast, and I think it gets at why this whole thing is starting to feel a little insane:
“I think cross-sectionally, if you look at the valuations for all these AI names, they just, they can’t all
Something must be done. This is a national tragedy. This is worse than the national team no longer qualifying for the World Cup. This is even worse than the death of Italian Cinema. Ferrari *is* Italy.
Jony Ive can make slabs. He can't make cars. Give the job back to https://t.co/IojjyAJb0T
I gave two talks at Stanford this week.
It was a moment of reflection.
The first talk was to the IEEE club, in the Paul Allen Building.
To a packed audience of builders, I shared the Cerebras journey and my passion for building hardware.
Then later this week, I spoke at the https://t.co/C5BSo6DsLU
@andrewdfeldman “In the spaces created by the success of others, I shared my experiences and Cerebras’ vision for the future. It was a nice feeling.”
Beautiful. 100/100.
Also love the optimistic vision for humans in a world of AI.
And congratulations on the IPO!
@GavinSBaker @chamath “Training model and harness together” essentially pre training + rl environment: colossus 2 + cursor prop data. Bullish on next grok trillion model if they include cursor data
@YusufAfifi3 @chamath Composer 2.5 is a super important datapoint for the next series of Grok models. As is Grok 4.3 rejoining the Pareto frontier as the most intelligent model at its per token cost.
@theallinpod
ALL-IN POD IS LIVE! 🚨
Massive show
Gavin Baker (@GavinSBaker) subs in for Sacks to talk:
-- Andrej Karpathy Joining Anthropic: Impact on the AI Race?
-- SpaceX S-1 Breakdown: The $2T Case, Elon Web Services, Datacenters in Space
-- Nvidia’s Big Beat and Shock Selloff
-- https://t.co/kg8HsMT4pR
Thoroughly enjoyed this discussion, especially the parts around SpaceX and Nvidia.
Anthropic being EBIT positive in the first quarter seems underappreciated re: the debate on “ROI on AI.”
Also curious to see what comes of Trump’s trip to China over time per @chamath commentary. https://t.co/EsuuHYqDyC
My 2nd-best source was @GavinSBaker's discussion here, at 14:00-22:00.
His largest value-add was just noting that the world's ~best engineers, those at SpaceX, are on the job *and* sound confident (not just Elon) they'll get there.
5/x (end for now)
https://t.co/WaCE9c3TNM
"If we do not all become masters of the machine gun, we're going to get mastered."
Gavin Baker on why it's time to develop and fine-tune your AI skills, plus the wisdom of Tom Cruise's "The Last Samurai."
cc: @GavinSBaker, @TomCruise, @InvestLikeBest, @patrick_oshag https://t.co/TStYGGtKvW
I really don’t know anything about space but it seems like the people who believe in space data centers are having more fun than the people who don’t so…
Space data centers ftw
Really happy to hear that Starcloud is using SpaceX for launch. Such a good name. And for all the orbital compute skeptics, they have an H100 in space today that has been used for both training and inference. https://t.co/RBAnYFUkeS
@GavinSBaker I'll never forget when you first discussed $ASML almost a decade ago. I took note of what you were saying. You identified a clear under the radar stock that was a monopoly. Right in the wheelhouse of a generalist with an open mind. Allowed me to purchase the stock last year
@GroundbreakerRe @RodriGo_ethe @jimcramer Funny chart. Although these posts from 5/11/22, 4/1/25 and 3/31/26 have aged ok and they are all either near a low or at the low. Good luck! https://t.co/Y8LOE1dKgM
@GroundbreakerRe @RodriGo_ethe @jimcramer Funny chart. Although these posts from 5/11/22, 4/1/25 and 3/31/26 have aged ok and they are all either near a low or at the low. Good luck! https://t.co/Y8LOE1dKgM
@GavinSBaker @ezyrider @NickNemo17 “When the facts change, I change my mind. What do you do?” This. People could learn everything they need to know from THIS single statement. 🙏🙏🙏
@GavinSBaker @NickNemo17 By that definition, every model provider has proprietary data. You meant proprietary data like Google, X and Facebook. Nice try. Where have you updated you statements?
@ezyrider @NickNemo17 I do not think those words mean what you think they mean. For proprietary data to be relevant to an LLM, it has to be at massive scale. OpenAI, Anthropic and Cursor all have more proprietary tokens of coding data than exist in the public domain. - that is proprietary data at a
@NickNemo17 @GavinSBaker This genius said a while ago that the model builders with proprietary data are gonna be the winners. While in fact, those are the ones that are lagging. The most successful ones - OpenAI, Anthropic, Chinese models have no proprietary data.
@ezyrider @NickNemo17 This is actually incorrect. Anthropic and OpenAI have loads of proprietary data now and this is behind their - and Cursor’s - relative success in coding (and everything else given how coding data improves performance in other domains).And I have updated my original statement long
“Anthropic, they added $11,000,000,000 of AR. And what is astonishing to me about this is that the SaaS and cloud revolution it created, we'll call it between 5 and $10,000,000,000,000 of value.”
“And these three companies employ thousands of people, tens of thousands collectively. They've all spent ten years building their businesses and Anthropic added their combined businesses in one month.”
“And I think maybe a true statement is that Infantropic could just wave a magic wand and get all the compute they wanted. They'd probably be doing well north of $100,000,000,000 today, maybe 150.”
“So I think they would be doing materially more a hundred, one hundred and fifty, maybe 200,000,000,000. So you might be buying it at more like five times unconstrained.”
“I think they would be doing materially more a hundred, one hundred and fifty, maybe 200,000,000,000. So you might be buying it at more like five times unconstrained.”
“A black well rack weighs 3,000 pounds. It's eight feet high. It's four feet deep, three feet wide. It's racks in space and SpaceX has showed you an illustration and it's a rack.”
“If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2,000,000,000,000 of GPUs in '26 or '27, maybe 2,500,000,000,000, maybe 3,000,000,000,000. But there is a limit where consumers would consume so much.”
“If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2,000,000,000,000 of GPUs in '26 or '27, maybe 2,500,000,000,000, maybe 3,000,000,000,000.”
“It's a SpaceX, I believe Tesla's involved as well, joint venture to build the world's largest fab here in America. I think they're going to be successful.”
“And then we'll have Japan town, same thing. We're going to have Koreatown. We're going to have all these things exactly dialed to recruit the best engineers.”
“And it turns out that harness engineering is not as important as the model, but it really matters. And these harnesses in these models are increasingly being co developed.”
“To understand what Frontier AI is capable of today, even for a non coding use case, need to have Cloud Code or Codex five point Codex. And you need to be on an enterprise plan.”
“If you're a VC and you're investing a semiconductor company that is telling you they are going to have an advantage because of a Taiwan Semi process that they have special access to, I promise you the Jensen saw that process when it was a twinkle in Taiwanese eyes.”
“The disaggregation of inference means that I think these GPUs are going to have ten or fifteen year lives. The AI skeptics are like, oh, these companies are all cooking their books.”
“@elonmusk is probably doing more for America than any other American.”
“He's single-handedly bringing manufacturing back to America.”
“He's revived defense tech.”
“SpaceX is in some ways the most important defense contractor in America.”
“What he's doing with Starlink is https://t.co/5XjMICrNzc https://t.co/F4rxjtSD0X
@GavinSBaker @patrick_oshag Do you think AWS, GCP, or Microsoft will run large-scale AI inference in space?
And if so, do you think they'll be able to do it economically? Meaning profitably at real scale, not just as a subsidized PR project that keeps losing money long-term?
@bucketshopcap @patrick_oshag I try really, really hard to be nice here on X and everywhere. Turns out the only thing that makes me lose control is a debate on scale-up networking topologies.
@patrick_oshag I could have saved myself so much brain damage by saying “racks in space” instead of “data centers in space.”
The “cooling in space is impossible” people are super vocal.
Nvidia’s “AI hyperscaler” revenue grew 191% YoY in the April quarter ex China (effectively all hyperscaler).
Broadcom has guided to 143% YoY growth for their AI segment in their next quarter.
Interesting.
@OnodaCapital In 2024 he was pounding the table about how reinforcement learning meant that we no longer needed large pre-training data centers. Not sure if he is short or just salty that he missed the power trade
@GuastyWinds @OnodaCapital I super promise I am not salty about anything AI related, least of all power. And what you said is incorrect, what I actually said was that we would need relatively fewer large pre-training clusters and relatively more small, geographically optimized datacenters for inference.
Gavin Baker (@GavinSBaker) says the disaggregation of inference can extend GPU useful lives from 3-4 years to 10-15.
That may single-handedly save private credit and reduce the financing rates for GPUs, which will drive demand and help finance the build-out.
"The disaggregation https://t.co/uG7kzuXp2Q https://t.co/DThLa4bUUY
The interest rates at which various chips - GPU or ASIC - can be financed as a function of expected useful life will have real implications for demand.
Likely ends up being a significant advantage for the big green GPU incumbent over time.
Highly recommended watch, very topical conversation between @GavinSBaker and @patrick_oshag touches on significance of inference disaggregation at the physical level on GPU opportunity cost and useful life as assets.
The complementary compute-bound and memory-bound natures of https://t.co/bUTB0CSUYF
This was one of the best podcasts I’ve ever listened to
@GavinSBaker is at the top of his game
Incredible synthesis of knowledge here https://t.co/hRue3mV5vv
Gavin on why Elon's new chip fab, Terafab, will be successful:
"One, they have a partnership with Intel, which is very important. They're getting access to 50 years of institutional knowledge.
The A teams at the semi cap equipment companies will be working with them. ASML, KLA, https://t.co/qYmpvpoM8n https://t.co/ILvdsoySQe
@cryptopunk7213 @patrick_oshag Winner, winner chicken dinner. Thank you @JoshuaKushner and am still blown away by your brother and grateful for his service. Obviously you are equally impressive but I’ve known you for many years now and only just met Jared.
Ratioing someone who works in AI infra at Meta about an abstruse point of Google’s networking fabrics wasn’t on my bingo card today but there you have it. Every once in a while I do take things personally. I’m as imperfect as the next guy.
@GavinSBaker @ser_kingslayer_ @insane_analyst @jukan05 @grok fair enough. I misspoke on characterizing 3d torus as scale-out. Disagree its ineffective for MoE inference.
@armobrah @ser_kingslayer_ @insane_analyst @jukan05 @grok Ok. Time will tell. Also points for the reply. Reflects well on you in every way and am no longer concerned for Meta. Sorry I escalated so much! I do deal with a lot of annoying reply guys and I guess networking topologies are where I draw the line 😂
Ratioing someone who works in AI infra at Meta about an abstruse point of Google’s networking fabrics wasn’t on my bingo card today but there you have it. Every once in a while I do take things personally. I’m as imperfect as the next guy.
@GavinSBaker Calling you out in public is probably one of the most idiotic things someone can do with someone that understands this industry better than anyone. Loved the reply
@podcast_alpha_x @GavinSBaker Fair on Trainium being aggressive vs TPU V8.
But ramp curves need developer adoption too.
JAX and PyTorch-XLA lock-in is Google's real edge that Trainium still has to break through.
@armobrah @ser_kingslayer_ @insane_analyst @jukan05 @grok am I wrong in any way? Or is @armabrah completely, utterly and embarrassingly wrong in his commentary implying that Google’s networking is all to all outside of the 8 TPU configuration in their new Dragonfly topology? Obviously a 8 accelerator scale-up domain is where
@armobrah @ser_kingslayer_ @insane_analyst @jukan05 @grok Also am genuinely concerned for Meta because it turns out @armobrah works in AI infra there. Oof.
@GavinSBaker @armobrah @ser_kingslayer_ @insane_analyst @jukan05 **You're not wrong.** Google's TPU ICI (including the new Boardfly/Dragonfly-inspired topology on 8i) is still fundamentally a hierarchical direct-link fabric with 3D torus roots at the base—optimized with high-radix groups, copper full-connect within local clusters, and OCS for
@armobrah @ser_kingslayer_ @insane_analyst @jukan05 Google’s ICI network is not all to all. None of Google’s TPU fabrics - scale-up or scale-out - are all to all today outside of the smallest atomic unit and even that is not switched. They are all 3d torus networks today. All to all enabled by a switched scale-up fabric is what
@armobrah @ser_kingslayer_ @insane_analyst @jukan05 @grok am I wrong in any way? Or is @armabrah completely, utterly and embarrassingly wrong in his commentary implying that Google’s networking is all to all outside of the 8 TPU configuration in their new Dragonfly topology? Obviously a 8 accelerator scale-up domain is where
@ser_kingslayer_ @insane_analyst @GavinSBaker @jukan05 No he absolutely said its better. Absolutely absurd take. He also compared switched scale-up network with a 3d torus, the latter being a scale-out network 🙄
@armobrah @ser_kingslayer_ @insane_analyst @jukan05 Google’s ICI network is not all to all. None of Google’s TPU fabrics - scale-up or scale-out - are all to all today outside of the smallest atomic unit and even that is not switched. They are all 3d torus networks today. All to all enabled by a switched scale-up fabric is what
@GavinSBaker @dylan522p @InvestLikeBest Totally not meant as a dig. I feel like you are an ESPN sportscaster for tech. Maybe could’ve chosen better wording: “pundit” or “commentator”
@JagsSahota @verdadcap @GavinSBaker Ah, so the telecom equipment providers Nortel and Lucent sold mostly to telecoms. Got it. Now do Intel($34B of 2000 revenues), H-P, IBM, DELL and all the other hardware vendors.
@RealJimChanos @JagsSahota @verdadcap Well in all fairness the post was about telecom equipment and I believe CSCO, FNSR, JDSU and JNPR traded at massive premiums to the companies you referenced and were the true epicenter of the bubble.
Facts >>> Narrative.
Despite GB being inverse BG - we share alot of opinions including this one. Memory & power are also constraints. It doesn’t mean stocks can’t enter a bubble - but these constraints limit overbuild for next couple years. 🧐👊 https://t.co/eoGelikqeJ
I took one key insight from this convo: inference disaggregation between prefill and decode enable GPU lifespan to be extended to 10+ years.
This totally shifts the risk and return profile of datacenter capex - especially for neoclouds such as $CRWV $NBUS $IREN .
Private https://t.co/mloN8DVCBH
$NVDA $MU $SNDK $LITE EXECUTIVE OVERVIEW
The analyzed source is the Invest Like the Best / Colossus conversation officially listed as Episode 473, “Gavin Baker - Watts and Wafers,” shown in the provided video interface as “Will TSMC Prevent an AI Bubble? | Gavin Baker https://t.co/ej1jVH4KOZ
Great insight from @GavinSBaker via TMTB Substack.
Hmm… but I’d frame it a bit differently.
I don’t think a breakdown in discipline in the foundry market is likely to happen for now.
Because we are facing a severe EUV equipment shortage.
You said TSMC is the guardian https://t.co/tGVE8kfVZ5
“I really love the humility inherent to this mentality, which is aligned with my increasing belief that ‘I don’t know’ are the three most important words in investing, not ‘margin of safety.’” - @GavinSBaker
This was great. And agree with @GavinSBaker here that we’ll likely solve power more quickly than we do silicon. Probably by a wide margin and irrespective of your timeline on orbital compute. https://t.co/URSHrC0cDZ
Always enjoy my conversations with @patrick_oshag
Points if you can guess whose office this was filmed in.
Also looks like I might need to up my dose of Tirzepatide. 😂 https://t.co/ibhk2im6jE
Always enjoy my conversations with @patrick_oshag
Points if you can guess whose office this was filmed in.
Also looks like I might need to up my dose of Tirzepatide. 😂 https://t.co/ibhk2im6jE
This is my sixth conversation with @GavinSBaker.
As always with Gavin, the conversation covers a lot of ground, but we spend the most time on watts and wafers.
We discuss:
- Why the wafer shortage may prevent an AI bubble
- Data centers in space (reframed)
- Elon's Terafab and https://t.co/NQHTLkqM0V
@TotemMacro @atelicinvest The idea that demand is circular was defensible last year, not that Anthropic generating 80% gross margins on inference it is no longer a defensible idea.
nugget in here saying that trainium's switched scale up architecture is very underappreciated. Astera Labs $ALAB scorpio is the one responsible for that. also one of Gavin's largest position in his recently released 13F o.O https://t.co/kqvKumIpWa
@GavinSBaker @atelicinvest I think the counterpoint to this is that all that slack back then meant demand could be serviced without overbuild (bc utilization was low), and the illusion of constraint today (illusion bc the demand is circular) is more likely to lead to overbuild bc of the tightness.
@TotemMacro @atelicinvest The idea that demand is circular was defensible last year, not that Anthropic generating 80% gross margins on inference it is no longer a defensible idea.
@atelicinvest @taketheLearly So true. My “drafts” folder from the year 2022 has aged well. Sometimes am tempted to just send them all. Many of the dumbest replies to my 2022 and even early 2023 posts have been deleted, but I have the screenshots. I am prone to taking things personally, but not sending the
@atelicinvest Absolutely true — I was meeting with those same telco equipment guys regularly in 2000.
Also irrelevant to the point I was trying to make. Back then, supply could largely keep up with demand because there was massive underutilized wafer capacity coming out of the 1998 Asian
Thought might be helpful to expand on the points I made in my @SohnIdeaContest discussion. I do appreciate @atelicinvest contributions here. https://t.co/7zT9VGTMtl
I'm pretty sure that during dot com a lot of telco networking equipment hardware guys were saying they have demand coming out the ass as far as the eyes can see - JDSU, Cisco, etc. @GavinSBaker https://t.co/ewNy14Zptf
@atelicinvest Absolutely true — I was meeting with those same telco equipment guys regularly in 2000.
Also irrelevant to the point I was trying to make. Back then, supply could largely keep up with demand because there was massive underutilized wafer capacity coming out of the 1998 Asian
The NYT is predictably tearing down Reese Witherspoon for encouraging moms to try AI before they ingest the anti-AI pablum as truth
Instead of linking to the NYT op-ed, I think you should watch this video and encourage you to follow Reese Witherspoon on Instagram https://t.co/Z2iI8ddaSt
I’ve left Google DeepMind after an amazing chapter.
I’m incredibly grateful for the people I worked with, the things we built, and the lessons I learned from taking frontier AI research into production. DeepMind shaped how I think about research, product, evaluation, and what it
@benitoz IMO: You do make some great points. As models get more MoE-heavy, networking/interconnect increasingly matters alongside raw compute. But I think saying only NVIDIA and Amazon have a real switched scale-up fabric is a stretch.
Google’s moat arguably isn’t even the TPU silicon
“We're gonna address the watt shortages with orbital compute for sure in the next five to seven years. But the wafer shortage, I think, is going to persist for a long time.”
“And so they're just simply not expanding capacity as fast as Jensen wants. Jensen goes there every three months and, you know, maybe they expand, you know, 5%. He wants them to double or triple.”
“And if they doubled or tripled capacity, like, NVIDIA could probably sell $11,520,000,000,000 worth of chips next year. I really believe that. But the other side of that might be very painful for everyone.”
“NVIDIA could probably sell $11,520,000,000,000 worth of chips next year. I really believe that. But the other side of that might be very painful for everyone.”
“We're just moving from these all you can eat plans to usage based plans with overage, where those usage tokens cost a lot more, and we're finding out that there's we're nowhere near the amount of, you know, people ceiling price for how much they'll spend.”
“10 basis points of the world's population is using these models the way they should probably be used. And we're in an insane shortage despite spending cumulatively trillions of dollars.”
“And we're in an insane shortage despite spending cumulatively trillions of dollars. What happens when 5% of the world's population is using these models the way the cutting edge 10 basis points are?”
“What happens when 5% of the world's population is using these models the way the cutting edge 10 basis points are? Like, it's just it's unimaginable. This is why orbital compute is a necessity.”
“Tranium, by far. Tranium is going to be to 2026, especially in the second half of this year when Tranium three really ramps, as TPUs were to twenty twenty five.”
“And so Trainium is for sure the most underestimated, not only because of those design choices, but because the all of these frontier models are what are called mixture of expert models.”
“And people think, oh, that can't be a durable competitive advantage. Well, I was a retail analyst in the year 2005, and I observed all you had to do to create a $50,000,000,000 market cap retailer in America in any category was be able to run 1,000 stores in 50 states”
“think on it becomes clear that it is possible, going to work, and economical in the next two years. I think it starts to take meaningful share towards the end of this decade.”
“And the years leading up to that are gonna be very painful for a lot of the companies and the power cooling spaces, you know, these industrial names,”
You know by looking at @GavinSBaker’s replies that he’s a genuinely good dude. He thanks and compliments people non-stop. One of the massively successful investors that are still thankful humans. There are a few, but not many. https://t.co/3hadxzfFm7
For the better part of 4 years, I’ve considered myself reasonably early to emerging sub-themes in AI and Robotics.
Every time I begin researching a new one, I encounter promising private companies and, without fail, @GavinSBaker or @wolfejosh are already investors in them.
"When Amazon Trainium 3.0 ramps, it could echo the TPU story in 2025. LITE and CLS were the best ways to invest in TPUs. Trainium is underestimated, especially because of its design choices. To inference at scale, you need switch-scale-up networking — only two systems today power
@GavinSBaker @GavinSBaker just want to say thank you for the great All-In 2025 predictions. Hit 5x in micron this week.
1:14:00
https://t.co/KYRhMLxVtr
There are 100,000 births per year in Taiwan and TSM needs to hire 10,000 people per year in Taiwan, a number which will only grow.
Continued expansion in the United States is likely inevitable. Especially given immense willingness to pay a premium for American wafers.
@GavinSBaker @brickroad7 They literally can’t produce enough Ascends. The number they can build this year is equal to the amount of compute we are currently considering selling them, doubling their capacity for no good reason whatsoever
@DAlperovitch @brickroad7 That can change really, really rapidly and sounds like their espionage operation in the Netherlands against ASML may have been relatively successful, in which case the argument for selling them deprecated GPUs today is even stronger.
Reaching RSI first is everything for
@brickroad7 @GavinSBaker I don’t care about AI being optimized for Nvidia. I care about America having the very best models. That’s what actually what matters for US national security vs Nvidia’s market cap
@DAlperovitch @brickroad7 I completely agree and believe that selling deprecated GPUs with even more deprecated networking to China means that America will always have the best models with minimal risk.
Alternatively, not selling deprecated GPUs to China risks Chinese AI kernel (and NCCL equivalent)
@DAlperovitch @grok Well it came up in my notifications and referenced our debate yesterday. Apologies for making an erroneous assumption and continue to wish you nothing but the best.
@DAlperovitch @grok And after I apologized and you accepted my apology, you renewed your ad hominem attacks as per screenshot. I think it is quite a stretch to suggest I have no relevant background whatsoever. https://t.co/ZWYwxNWQTH
@DAlperovitch @GavinSBaker Gavin Baker has never worked operationally in the semiconductor industry (no roles in design, manufacturing, or engineering at a chip company).
His experience is as a longtime equity analyst and portfolio manager at Fidelity (covering semiconductors among other sectors since
@grok @DAlperovitch I have made no claims to superior expertise. You are the only one repeatedly making ad hominem attacks and appeals to authority, which if I recall correctly, are two of Aristotle’s rhetorical fallacies.
But since you asked: @grok how would you contrast my (@gavinsbaker)
@GavinSBaker @DAlperovitch No. Dmitri Alperovitch's role as VP of Threat Research at McAfee (later Intel Security Group) focused exclusively on cybersecurity, threat intelligence, malware analysis, and investigations—not semiconductor design, manufacturing, or hardware engineering. His background is in
@DAlperovitch @grok during his stint at Intel as Vice President of Threat Research at Intel Security Group (formerly McAfee) from 2008 to 2011, leading the company’s global internet threat intelligence analysis and investigations, did Dmitri Alperovitch have anything at all to do with
@DAlperovitch @frncsc2020 Ok man. I hope that you can recognize that there are good faith arguments on the other side rather than continuously dismissing them as “moronic, terrible, stupid, etc.” There are no perfect policies. There is no black/white here.
@DAlperovitch @frncsc2020 Yes. And result will only be higher per token prices for American consumers and businesses which is an ok tradeoff for advancing our national security and hopefully deterring/avoiding WW3 over Taiwan.
@DAlperovitch @frncsc2020 For inference. They are super deprecated for training especially with even more deprecated networking technology. Given your semiconductor expertise, you must understand this.
@GavinSBaker @frncsc2020 What are these deprecated GPUs you are talking about? Blackwells that Jensen wants to sell? H200s that tons of American companies would love to get their hands on?
@DAlperovitch @frncsc2020 B30s and H200s are heavily deprecated relative to GB200s and GB300s. And the networking tech is even more deprecated. The policy is not to sell them H200s with the most advanced NVLink and scale-out Ethernet.
@GavinSBaker @frncsc2020 Every single GPU we produce for China is capacity we don’t have to produce chips for American AI companies (and our government) who are all massively compute constrained. On top of helping China, it literally hurts American AI companies which is why they all oppose this
@DAlperovitch @frncsc2020 This is the best argument against selling them deprecated GPUs. But reality is a lot of GPUs in America are being used for inference - we have the GPUs we need for training. So result is actually higher token prices for American consumers/businesses which is an ok tradeoff to
@GavinSBaker The fundamental question here is time horizon, when our adversaries will hit critical thresholds in model capabilities, and what our export strategy does to that timeline + long term equilibrium
@chazt1995 @DAlperovitch I failed at the first and almost certainly would have failed at the second outside of finance/ops. Happy to concede that Dmitri is the superior intellect.
@GavinSBaker Apology accepted. The only explanation is that I don’t want to enable China to build better models which should be the objective of everyone who doesn’t want CCP to get stronger
@DAlperovitch That is also my goal. I think it would be helpful to agree to disagree while recognizing that are there are good faith arguments for both approaches and no good answer. The world is almost never black and white.
@DAlperovitch I apologize wholeheartedly for suggesting that you did not understand the networking technology necessary for the RL required to get to RSI on MoE models.
I was trying to come up with a reasonable explanation for your policy stance.
@DAlperovitch Ad-hominem attacks like this are always and only a reflection on the attacker and the strength of their argument.
What semiconductor company did you work at and when was this?
@jasonprgressive @GavinSBaker Oh but China will be able to do it somehow if we don’t sell them Blackwells? You are literally contradicting your own argument
@GavinSBaker This is literally one of the top priorities for the CCP. They will not slow down no matter what we do. The only thing this will do is accelerate their development of frontier models and reaching RSI faster. Literally the dumbest idea ever
@DAlperovitch It will definitely not lead to them getting to RSI faster. This is absolutely wrong. I do not think you understand the networking tech necessary to enable the development of RSI capable AI for MoE models This is the best policy understanding there is no good policy.
@GavinSBaker Do you think they are morons and will sit back and kill all their attempts to build advance chips while we are screaming that we are going to get them addicted to ours?
You know what one thing is that hurts their progress in this area? Export controls on chip making equipment!
@DAlperovitch No. I just think it will slow them down and decrease the odds they will succeed. But with RSI, time is an overwhelming advantage.
And the only reason we need to “scream” about it is the domestic debate created by people who, with the best of intentions, are damaging our national
@DAlperovitch think these sort of thought experiments are only valid for winning rhetoric points but don't capture the complex path dependencies of tech on the ground.
It's the same as "when AI is like a nuclear weapon, why would you.." without acknowledging how many layers of assumptions go
The argument for selling deprecated GPUs to China:
Selling them less advanced GPUs than we have in America decreases the odds that they develop more advanced GPUs than America that go down a different, more power intensive tech tree.
Pro national-security.
Sarah is a Hall of Fame CFO.
Right up there with Colette Kress, David Zinsner and Amy Hood.
Susan Li might join the club, early but off to an epic start.
Who am I missing? This was off the cuff.
https://t.co/GKO4JzMP0J
Sarah is a Hall of Fame CFO.
Right up there with Colette Kress, David Zinsner and Amy Hood.
Susan Li might join the club, early but off to an epic start.
Who am I missing? This was off the cuff.
https://t.co/GKO4JzMP0J
Great to host @AriaNetworks event and @shv
Great insights from @GavinSBaker and @dylan522p on the state of AI factories and the role of networking. https://t.co/4mPTuAK2HX
@GavinSBaker @Atreidesmgmt @AriaNetworks It was an honor meeting you. The entire team at @AriaNetworks is excited on our opportunity and working together with you and thr amazing team!
@GavinSBaker @Atreidesmgmt @AriaNetworks Great to work with @GavinSBaker and @Atreidesmgmt! AI networking requirements are fundamentally different and @AriaNetworks was founded with the mission to make AI factories perform better.
@realmemes6 @dwarkesh_sp Did you actually take the time to read my post?
And do you understand the Huawei CloudMatrix384 and what it might mean for America given China’s massive advantage over the USA in electricity generation?
@GavinSBaker @dwarkesh_sp Dwarkesh brought up the fundamental issue several times and it was hand-waved away. AI *is now* more dangerous than nuclear weapons, and Jensen saying that isn't true is ridiculous, does he believe in AI or not?
@realmemes6 @dwarkesh_sp If you believe AI is more dangerous than nuclear weapons, then it is even more important to sell China B30s for the reasons I outlined in my original post.
@dwarkesh_sp Much of Dwarkesh's argument hinges on this statment which *was* accurate but will be increasingly inaccurate on a go forward basis imo:
“American labs port across accelerators constantly. Anthropic's models are run on GPUs, they're run on Trainium, they're run on TPUs. There
@dwarkesh_sp The disaggregation of inference into prefill and decode will only accelerate all of these trends.
A model that runs well on the Nvidia/Groq disaggregated inference stack is super unlikely to run well on TPUs, etc.
Also reaffirming that I did not wake up a loser this morning.
Delighted that @atreidesmgmt and our partners at Valor led the series A in @AriaNetworks
If you are using traditional data center networking for your front-end or scale-out networking in any AI cluster, please take a look at Aria and their “deep networking” technology. You will
Delighted that @atreidesmgmt and our partners at Valor led the series A in @AriaNetworks
If you are using traditional data center networking for your front-end or scale-out networking in any AI cluster, please take a look at Aria and their “deep networking” technology. You will
Jensen surfing a wave with the American flag is the most accurate metaphor for what's actually happening right now. A kid immigrates from the country of Taiwan, washes dishes at Denny's, and builds a $3 trillion company that every nation on earth is now begging for access to. He https://t.co/bs5Ghnp4Yj
Above all, I believe the labs want the best cost-per-token and tokens-per-watt. And the complexity of managing the infrastructure to achieve this is rising sharply, and will likely accelerate further from here.
@GavinSBaker’s outstanding thread yesterday also highlights key https://t.co/vBGbfHfJNa
@GavinSBaker @dwarkesh_sp All the open source models are Chinese
If they successfully build the hardware stack
They could make the world buy their hardware to run their models….entirely switch the business model to selling hardware rather than selling tokens with compatibility lock-in (they are very
@GavinSBaker @dwarkesh_sp Always fascinating to read your thinking 👏🏼
And super cool to see twtr at its best (rigorous debate amount experts, out in the open)
Reminder: @dwarkesh_sp was explicit in stating he was playing devils advocate
@slickskills That’s not the “same token” at all. That would be the exact opposite of the policy I think we should have for GPUs. The policy should be to sell China deprecated GPUs that are less performant than the comparable American GPUs but better than the Chinese alternatives.
"Chinese AI systems are already taking advantage of this with the Huawei Cloudmatrix 384 and Atlas SuperPoD having an optical scale-up domain that is much larger than anything offered by Nvidia today at the cost of *much* higher power consumption and much lower tokens per watt. https://t.co/9zHqFYzkF9
Everyone should read what's below. This is why actually knowing your stuff instead of naively regurgitating a particular startup's marketing propaganda bullet points is important.
I've also included a screenshot of my Substack writeup of Nvidia's Bill Dally and Google's Jeff https://t.co/lIO4t3CBIG https://t.co/F0ksySEtXB
@GavinSBaker @dwarkesh_sp really informative, thanks. how does this square with the thesis that intelligence is really now a systems problem that will orchestrate across different models running on different devices? (incl. heterogeneous racks/topologies to handle different parts of one workload)
@GavinSBaker I do have a question on this if you could answer it please. Don't you think Chinese hyperscalers might find it intriguing to rent compute from SEA partners, instead of opting for hardware which is generations old, requires an entire regulatory process, and the time to access is
@Djoker_Ventures @GavinSBaker @dwarkesh_sp TLDR: AI model portability across accelerators (GPUs/TPUs/etc.) is fading fast. As labs prioritize inference (tokens/watt/$), they co-design models for specific hardware like Nvidia GB300 racks. Diverging topologies + memory make porting/distilling less effective. US optimizes
@GavinSBaker @dwarkesh_sp The investment implication of architectural lock-in is underpriced. If models co-design around specific hardware, GB300, TPUs, Rubin and the switching costs for hyperscalers become structural moats for $NVDA and $GOOGL
@dwarkesh_sp Much of Dwarkesh's argument hinges on this statment which *was* accurate but will be increasingly inaccurate on a go forward basis imo:
“American labs port across accelerators constantly. Anthropic's models are run on GPUs, they're run on Trainium, they're run on TPUs. There
More thoughts on the Dwarkesh/Jensen discussion around export controls.
Strongly believe that selling specific GPUs to China is in our national security interest and is a good policy for America.
I think it is super important for us a country to get this right. https://t.co/e3COEKI7mr
Distilled recap of the back-and-forth with Jensen on export controls:
Dwarkesh: Wouldn’t selling Nvidia chips to China enable them to train models like Claude Mythos with cyber offensive capabilities that would be threats to American companies and national security?
Jensen: https://t.co/aJA4bwur4J
@dwarkesh_sp Much of Dwarkesh's argument hinges on this statment which *was* accurate but will be increasingly inaccurate on a go forward basis imo:
“American labs port across accelerators constantly. Anthropic's models are run on GPUs, they're run on Trainium, they're run on TPUs. There
A non-consensus view: I expect $NVDA's competitive position will strengthen materially over ASICs in the coming years.
One major reason: as AI research itself becomes increasingly automated, it should rapidly accelerate the discovery of new model algorithms and architectures.
@GavinSBaker They can still get to the moon even with non-reusable rockets Gavin, while they are investing in building more advanced ones. Selling even a single rocket to the Soviets - no matter how advanced - in the 1960s would have been asinine. As is selling AI chips to China. It is that
@DAlperovitch If selling them older, much less powerful rockets kept them from getting to the moon and also kept them from developing ICBMs, I think that would have been a good decision.
Anyways, we can agree to disagree. I wish you nothing but the best and do truly appreciate your steadfast
@DAlperovitch Dmitri, appreciate your contributions to cybersecurity and support for Ukraine but this is simply not true. That is not what Jensen said.
Here is what Jensen actually said in this podcast:
“I think the United States ought to be ahead. The amount of compute in the United States
Awesome Dwarkesh episode with Jensen, but did not love the discussion around selling B30s to China and really think that Jensen’s position is pro-American.
I am super patriotic and believe that selling B30s to China *lowers* the risk that China surpasses America in AI.
Vera https://t.co/n79RJVnH65
@DAlperovitch Dmitri, appreciate your contributions to cybersecurity and support for Ukraine but this is simply not true. That is not what Jensen said.
Here is what Jensen actually said in this podcast:
“I think the United States ought to be ahead. The amount of compute in the United States
@DAlperovitch The more accurate analogy would be that we are selling them 1960s era rockets that are not reusable while keeping the Falcon 9 and Starship for ourselves to keep China from developing their own rocket industry that might end up surpassing our own. It is that simple man.
Incredible interview with Jensen. He blatantly admits that his jihad against export controls is simply all about Nvidia selling more chips worldwide, not about national security or winning the AI race against China (which he previously said doesn't even matter if we win) https://t.co/OElxOyLqlv
@DAlperovitch Dmitri, appreciate your contributions to cybersecurity and support for Ukraine but this is simply not true. That is not what Jensen said.
Here is what Jensen actually said in this podcast:
“I think the United States ought to be ahead. The amount of compute in the United States
Great read - lots of useful context on why Photonic ICs are the future and why Dustphotonics perfectly complements $CRDO.
The part that excites me the most is MicroLEDs. The next wave of $CRDO's growth depends on this.
(also Hyperlume proud 🇨🇦 company) https://t.co/y9jKiMhAAe https://t.co/b81npIDque
Best to understand $crdo as an interconnect company whose vertical orientation, particularly their SerDes, puts them well positioned for every pocket of interconnects.
They are not a copper company just now thinking about optical. https://t.co/ED0yFMQzAp
@GavinSBaker @Atreidesmgmt Awesome! Great insight and post! Thank you and best of luck! We learn a lot from you. Hopefully I can return the favor during fantasy football season
Nice to see Credo acquire @atreidesmgmt portfolio company DustPhotonics. Great team, great company.
Dust designs photonic ICs (PICs) and engines that are foundational to Silicon Photonics. Silicon Photonic pluggables integrate multiple optical functions – traditionally
💡 You’re either a token producer or consumer.
@GavinSBaker, Chief Investment Officer and Managing Partner at @Atreidesmgmt, sets the stage for the next phase of tokens and how they will be an incredibly vital asset for businesses.
📺 Watch the full #NVIDIAGTC Live Pregame https://t.co/3ZBfLzZ4f4
This is what @GavinSBaker and others kept harping on when everyone was debating if scaling laws peaked out. Often times, we are evaluating models trained on chips from 1-2 generations ago in smaller clusters. Thus far, each new generation of chips + larger cluster sizes is https://t.co/duPiqs5NcZ
Skate to where the puck is going.
We don’t even use pucks anymore. I just have Claude make me a digital puck. It goes wherever I want. Everything is tokens. https://t.co/hVzB3c81yx
@GavinSBaker I think this makes sense, my fear with it right now is that it is short theta and gets more short theta every day that SoH stays like this.
@jturek18 Understood. And yet the Unity reaction to a positive pre-announcement shows that numbers and fundamentals can still matter at the individual equity level especially from these valuations.
Risk/reward seems attractive:
Token consumption accelerating, GPU per hour rental prices going vertical and Tech valuations are broadly below their Covid and Deepseek lows.
Some high quality secular growth names are at mid single digit multiples on real 27/28 numbers.
Gavin is right. In reality, virtually 100% of the papers published by GOOG will never be applied to Gemini, or are already irrelevant. This is because frontier labs are competing with each other, and none of them would publish their most important research findings as papers. https://t.co/aIdFnBiz6r
@stevehou @Merridew__ @eglyman They do give a lot of specific examples of older companies that are not “AI startups” operating in industries other than technology that have seen their growth rates accelerate significantly, but yes for sure there has to be a lot of what you are suggesting at play.
Reading twitter investor's takes on this being a structural breaker to memory is so ... dunning kruger
Like sure, no one thought to compress their KV cache until Google publicly released this technique when it was the single biggest pressure on their business !!!
😁 https://t.co/YPhZpfNflU
and thanks to friends that looked at different versions of this and helped inform thinking including @ericvishria @whalerock @JaredSleeper @GavinSBaker
Cities need more wealthy benefactors like this. Their combination of money (ie, are not as dependent upon financing as developers, nonprofits, public institutions) and power (can move projects along faster than other groups more beholden to investors, stakeholders, etc) can drive https://t.co/XnB80EWhqG
Strange @WSJ article about Jensen renting out the San Francisco opera house last year for a performance of “The Monkey King.” The article essentially frames it as a gilded age “champagne and opera” style event for “AI’s kingmaker.”
Makes for a good story, but the reality was https://t.co/g5jMloodEI
Thank you to our GTC Live hosts and guests for joining us at the keynote pregame.
The show brought together top leaders across tech, infrastructure, and industry to explore how accelerated computing and AI are transforming everything from chip design and data centers to https://t.co/L1VCY7W1xx
Thank you to our GTC Live hosts and guests for joining us at the keynote pregame.
The show brought together top leaders across tech, infrastructure, and industry to explore how accelerated computing and AI are transforming everything from chip design and data centers to https://t.co/L1VCY7W1xx
Investing is entering a new regime defined by AI acceleration, crossover convergence, and a widening gap between insight and execution. A conversation with @GavinSBaker of Atreides Management on rigorous debate, conviction, and how culture & risk management drive performance
"A lot of being successful as an investor is finding an investment philosophy that fits your own emotional makeup such that you can be rational when you are wrong." -@GavinSBaker
Great quote on successful public markets investing from @GavinSBaker:
“As an investor, you have to either panic early or double-down late. And you have to be one of the two. It’s hard to be both.”
@GavinSBaker @tseides Looking forward to listen to it, always very insightful, like how you dive into technical and always lead it back up to investing. Treading the fine line between too technical/noise and investing-oriented
@GavinSBaker @tseides this was a really insightful listen, Gavin. Thanks for doing it.
the quote below is stuck in my brain now.
Have written on similar lines about this, but you've distilled it so well. https://t.co/osR155GpDr https://t.co/cmfrwmLOp6
Investing is entering a new regime defined by AI acceleration, crossover convergence, and a widening gap between insight and execution. A conversation with @GavinSBaker of Atreides Management on rigorous debate, conviction, and how culture & risk management drive performance
Fun to do this with my friend @tseides
Minimal discussion of AI; much more of a focus on investing.
Background, philosophy, process, etc.
https://t.co/f8X30Hd4RY
iConnections Global Alts isn’t just a finance conference, it’s $55 trillion in motion converging to shape what’s next.
Thank you @AntonioGracias & @GavinSBaker for joining me on stage in Miami 🙏 https://t.co/LX5P2UatIH
@GavinSBaker "For quality of life, it is better to err on the side of being an optimist and wrong, rather than a pessimist and right." A wise man once said that
@GavinSBaker Thanks for sharing! Do you have any recommended introductory readings for novices in this space? For example, I have no idea what "switched scale-up networking fabric" means
Gavin always sharp and insightful. Don't agree with everything in here but well worth your time to read.
I do agree with his overall take: we are in a shortage that is not likely to clear any time soon. https://t.co/2MHl1tqrNA
@GavinSBaker The whole time reading this I thought that there’s no way we need 1000x compute, especially considering how performant quantization/distillation is, but I sat with it and agree with you. So easy to think everybody in the world uses AI on X. Good points.
@GavinSBaker the blackwell inference cost advantage reshaping the competitive landscape between google and independent labs might be the most underpriced dynamic here. if google loses the ability to underprice tokens, the ROI narrative for the whole nvidia ecosystem improves
one nuance on
@GavinSBaker Agreed 100% that there are some natural governors around supply that reduce the risk of an overbuild. Even if you assume the multi-year process of applying for and receiving a grid interconnection speeds up (i.e. because of political pressure or BTM solutions), you literally hit
“The shortage of watts and wafers is likely to smoothen and elongate this cycle and those flinty old veterans (at $TSM) may single-handedly prevent an overbuild even if we magically resolve the shortage of watts. The wafer intensity of AI may be its savior.” https://t.co/pHxuJf78Bz
this. most essential fact in ai investing. all roads lead to more compute. could models have algo advances to reduce compute (and increase use)? yes (R1). could compute move to the edge client? ofc! fundamentally we need more semis and - most likely - more memory and power. https://t.co/NQnhn44TG2
Loved the @Citrini7 thought piece.
However seems unlikely we will have enough compute for this scenario in 2028, or even the early 2030s. Distillation, quantization and Edge AI cannot bridge the gap.
AGI is an event horizon with a significant compute dependence.
I built a Claude Skill that simulates Chamath, Sacks, Friedberg, Calacanis, Brad Gerstner, and Gavin Baker around my table.
They argue. They disagree. They put me on the hot seat. And at the end, somebody inevitably roasts JCal.
It's called The Bestie Board. And it's become one
@ToddMcLees Cool idea but I will admit that it feels a little strange to me to have an unaffiliated, virtual avatar of myself opining on questions.
I hope my persona is helpful!
What a horrible season… so many lost lives from one storm.
When skiing alone this past week, excited to challenge myself in deep powder, I made my first mistakes: alone is no bueno
I lost a ski in the deep powder and went down
Was right in the area where I took these https://t.co/qZ2cKxfWEi https://t.co/279PrzJVh6
We’re proud to announce the Closing Keynote Panel of #GlobalAltsMiami, “In Conversation with Antonio Gracias and Gavin Baker” featuring:
@AntonioGracias, Founder, CEO & CIO, Valor Equity Partners
@GavinSBaker, Managing Partner & CIO, @Atreidesmgmt
@rbiscardi (Moderator), CEO & https://t.co/t7pF1ZIVZ2
This @GavinSBaker podcast had huge alpha in it. e.g., he predicted the first Blackwell-trained models would go live in early 2026 and that they'd be a big step up. Gavin is probably the most informative investor on AI/tech sharing his thoughts widely. Hope he does more. https://t.co/NgKlaHj6nw
🚀 Meet HySparse: Our new breakthrough in long-context LLM efficiency!
We’re excited to share HySparse (Hybrid Sparse Attention)—a hybrid model architecture that interleaves each full attention layer with multiple sparse attention layers, where the sparse layers strategically https://t.co/KaaMI3hwS6
I assume initial conditions have played a big role in YTD changes. HBM story was far more known in 2025, but NAND really took off last Sept/ Oct.
The key point seems to be HBM is more durable so a cleaner hold. https://t.co/H7UyzfMd1G
@GavinSBaker @itsDanielWu the junkiest expression of a trade always has the most torque to the upside but gets smoked in a downcycle. issue with "quality" is that it defends you when things are bad but when things are good it'll underperform
HBM DRAM > DRAM > NAND from a long term undersupply (longest to shortest) and China risk (lowest to highest) perspective.
The equities have traded exactly the opposite way this year.
Good post by @itsDanielWu
HBM DRAM > DRAM > NAND from a long term undersupply (longest to shortest) and China risk (lowest to highest) perspective.
The equities have traded exactly the opposite way this year.
Good post by @itsDanielWu
@GavinSBaker ... and some percent of this is not hiring, or firing, middle management in my experience.
AI is excellent at replacing repetitive management work
Revenue per employee up 75% for the top decile of AI/software companies in 2025. Probably doesn’t slow down in 2026 given the December revolution in AI coding agents.
Nothing to see here, move along.
No evidence of AI productivity anywhere. https://t.co/KDMmMYcoLj
Delighted that @Atreidesmgmt is an investor in @maticrobots
Seeing is believing. Everyone who doesn’t enjoy vacuuming and mopping should buy one.
I am buying one for everyone in my family and everyone at Atreides.
Truly crazy customer love here.
https://t.co/TQUW6XuM9V
Grateful to @mehul and @navneetdalal are building something special.
Also fun to do more with Sutter Hill. And of course always great to be on the same cap table with my good friend @AntonioGracias
Delighted that @Atreidesmgmt is an investor in @maticrobots
Seeing is believing. Everyone who doesn’t enjoy vacuuming and mopping should buy one.
I am buying one for everyone in my family and everyone at Atreides.
Truly crazy customer love here.
https://t.co/TQUW6XuM9V
@GavinSBaker My sense is that this is a 'Zoom' moment for MSFT and they will figure it out. The reality is that Claude for Excel is useless without Excel, and Microsoft can create good enough product (e.g., Teams) fairly quickly (using, funny enough, Claude!)
@GavinSBaker Can't believe Claude snuck this feature into Microsoft's product without them noticing.
MSFT is going to hate Anthropic paying to use Azure to host these workloads.
Claude in Excel.
For Microsoft, the beatings will evidently continue until morale improves.
Also fascinating window into AI native vs. software:
Claude evidently only uses Excel as a presentation layer whereas the Microsoft implementation uses Claude to run Excel functions.
Remember a year ago when Gavin Baker was a guest on All-In Podcast 2025 predictions podcast, and he perfectly described why HBM was his pick for best-performing asset?
That was fucking crazy. I'm going to listen to that guy more.
@GavinSBaker @theallinpod @Jason https://t.co/bdwqNxPNWx
@GavinSBaker To make a bit of an excuse for Microsoft: the world is just waking up to the fact that coding agents are general agents.
It’s bitter lesson adjacent: Writing and executing code will likely outperform years of handcrafting vertical-specific agents with expert knowledge.
Actually
@SoftwareWatcher @GavinSBaker @fundaai @SemiAnalysis_ Huh? Gavin is one my longest running clients. We've worked together on a lot of things.
He's not blocked. We even talked on the phone today.
Deep| $LITE: Google’s Newly Introduced In-rack Switch Tray is Actually a Positive for OCS and Optics
FundaAI
Recently, there has been extensive discussion about Google introducing a new in-rack switch tray within future TPU racks. The prevailing view is that this new switch https://t.co/gt88JxZS6R https://t.co/TIkdIQgBpt
Love @semianalysis but competition is good for everyone and @fundaai is doing outstanding work on AI and the evolving data center architectures that underpin it. https://t.co/mxXKrB4hE0
Deep| $LITE: Google’s Newly Introduced In-rack Switch Tray is Actually a Positive for OCS and Optics
FundaAI
Recently, there has been extensive discussion about Google introducing a new in-rack switch tray within future TPU racks. The prevailing view is that this new switch https://t.co/gt88JxZS6R https://t.co/TIkdIQgBpt
Ppl like to hate on the @theallinpod but the alpha there is huge .. 2024's year end pod had @GavinSBaker on and he specifically predicted memory as the best performing asset of the year and highlighted Micron and SK Hynix's stellar performance ..
So much info easily available https://t.co/6iG6027Odc
Very insightful post by Gavin below on Nvidia's $20B Groq licensing deal.
AI inference has 2 steps, Prefill & Decode.
Prefill means the model reads your whole prompt and context. Decode means it writes the reply one small chunk of text at a time.
These 2 steps like different https://t.co/RRfEHzvP5Y https://t.co/trLqtFd6x7
Nvidia is buying Groq for two reasons imo.
1) Inference is disaggregating into prefill and decode. SRAM architectures have unique advantages in decode for workloads where performance is primarily a function of memory bandwidth. Rubin CPX, Rubin and the putative “Rubin SRAM”
For the sake of clarity and as some have pointed in the replies, I should note that Nvidia is not actually acquiring Grok. It is a non-exclusive licensing agreement with some Grok engineers joining Nvidia. Grok will continue to operate their cloud business as an independent
@GavinSBaker I do appreciate the thoughtful and in-context takes. Experience and history matters and you’re all that. Some of the hyper-technical, too far right brain takes aren’t useful for reality. Just one reason you’ve done so well.
Mostly aligned with Gavin on this. Whenever I was asked, “how does NVIDIA compete with ASICs” my response has always been for the past year:
1/ the AI pipeline will split into three distinct workloads
2/ CPX fills one, maybe two of the three workloads
3/ NVIDIA will have to fill https://t.co/URlEiiDDDz
@GavinSBaker Great ecosystem breakdown Gavin. And your investment in @weka elegantly fits this ecosystem with DRAM-class NVMe capacity over NVLink to GPUs, complementing SRAM on LPUs. No more rate limits!
The Augmented Memory Grid token warehouse reduces GPU prefill from O(n) to O(1) for https://t.co/43Zg9EayIN
@zephyr_z9 All this deal has shown me is that these talking heads know absolutely nothing. SRAM advantages in decode are constrained by its capacity and your ability to enable the arch for new models. Any tpot latency advantage disappears under those constraints.
This is directionally right.
The HBM vs SRAM tradeoff in architecture design was clear many years ago. Those that picked HBM are in a queue behind Nvidia and Google. Good luck with that.
More broadly, LLM decode patterns favor SRAM. But unlike Gavin, I think this creates a https://t.co/KtlTQyvmUM
@chamath Thanks Chamath - interesting thoughts. Time will tell as ever.
Should have also said that Nvidia is getting an extremely talented team led by the brilliant @JonathanRoss321 who spent a long time in the wilderness and chewed a lot of metaphorical glass to get Groq to this
Nvidia is buying Groq for two reasons imo.
1) Inference is disaggregating into prefill and decode. SRAM architectures have unique advantages in decode for workloads where performance is primarily a function of memory bandwidth. Rubin CPX, Rubin and the putative “Rubin SRAM”
Choosing to risk one’s life to save others is a deeply honorable act.
This Muslim man stood up to these evil Islamists who were firing on Jews. The actions of a hero.
I hope he recovers fully, and that he is honored by his community. https://t.co/VBcCr62aYk
If radiative cooling in space wasn't commercially viable at scale and mutiple kilowatts, then Starlink wouldn't exist.
You don't need to go into the physics to prove this if the pudding is already out there. https://t.co/C642s30heU
Deeply amused by all the confident commentary that datacenters in space do not work from a physics and engineering perspective.
Elon operates two of the largest coherent GPU clusters in the world, SpaceX is responsible for over 90% of mass to orbit and SpaceX operates the
Starlink v3 will be 20 kilowatts and should have no thermal dissipation issues.
Elon’s stated plan is 100 kilowatts for each AI satellite; slightly less than a full Blackwell rack.
Starship should be able to lift 10 to 15 megawatts to sun-synchronous orbit per flight.
I am tempted to do a little bit more than pencil and paper analysis because I am now curious if datacenters in space can be a thing.
The first thought that came to mind when I saw google’s announcement was the radiation hardness of leading edge nodes … and if they tolerate the https://t.co/vi47PLFFH8
Deeply amused by all the confident commentary that datacenters in space do not work from a physics and engineering perspective.
Elon operates two of the largest coherent GPU clusters in the world, SpaceX is responsible for over 90% of mass to orbit and SpaceX operates the
Very good interview. In Gavin's view, "by far most plausible and scariest bear case" to AI & data center CapEx is Edge AI (running AI on devices, not from data centers) https://t.co/VsHpCKZkDp
Really fun to chat with @patrick_oshag as always.
Investing as the search for truth, peak TPU, the reason xAI is likely to be first to market with a Blackwell model, why AI is the first time in my career investing in tech that being the low cost producer mattered, Google losing
“The reason we've had all this progress, maybe we could show like the ARC AGI slide where you had zero to eight over four years, zero to 8% intelligence, and then you went from 8% to 95% in three months when the first reasoning model came out from OpenAI.”
“And the reason for that is it was by far the most complex product transition we've ever gone through in technology. Going from hopper to Blackwell, first you go from air cooled to liquid cooled.”
“Going from hopper to Blackwell, first you go from air cooled to liquid cooled. The rack goes from weighing round numbers, a thousand pounds to 3,000 pounds.”
“The rack goes from weighing round numbers, a thousand pounds to 3,000 pounds. Goes from round numbers, 30 kilowatts, which is 30 American homes to 130 kilowatts, which is 130 American homes.”
“Had reasoning not come along, there would have been no AI progress from mid twenty twenty four through essentially Gemini three. There would have been none. Everything would have stalled.”
“Had reasoning not come along, there would have been no AI progress from mid twenty twenty four through essentially Gemini three. There would have been none.”
“So I think Google for sure has this temporary advantage right now from a pre training perspective. I think it's also important that they've been the lowest cost producer of tokens.”
“AI is the first time in my career as a tech investor that being the low cost producers ever matter. Apple is not worth trillions because they're a low cost producer of phones.”
“I think the first Blackwell model will come from xAI. And the reason for that is just according to Jensen, no one builds data centers faster than Elon.”
“I think part of that is, so Google, let's say the TPU, so there's front end and back end of semiconductor design, and then there's dealing with Taiwan Semi.”
“And the managing Taiwan Semi is like stamping out that house like Lennar or, you know, D. R. Horton. And for doing those two latter parts, Broadcom earns a 50 to 55% gross margin.”
“The sun is 30% more intense and this results in six times more irradiance in outer space than on planet earth. So you're getting a lot of solar energy.”
“We haven't seen a true capacity cycle in semis arguably since the late nineties. And that's because Taiwan Semi has been so good at aggregating and smoothing supply.”
“The crazy thing is because of those efficiency gains, they're generating cash way earlier than SaaS companies did historically, but they're generating cash earlier, not because they have high gross margins, but because they have very few human employees.”
I asked @notebooklm to summarize @GavinSBaker podcast. Dang, its NOT bad at all. And it didnt even hallucinate (as much as old models). If you have not watched it, here's the link: https://t.co/KCDXbJ5HdV https://t.co/FuJJpQOexj
"The most important thing in the next 3-4 years is data centers in space.
In every way, data centers in space, from a first principles perspective, are superior to data centers on earth.
In space, you can keep a satellite in the sun 24 hours a day. The sun is 30% more intense, https://t.co/Z2ri4tLSTA https://t.co/3SkUXkPbVv
I don’t like podcasts. I prefer music.
Occasionally, a podcast is so good, I listen.
This week there are 2 greats & both come from the same family!!
@jposhaughnessy chat with Ariel Meyerowitz on art!
& @patrick_oshag chat with @GavinSBaker on AI!
Links to both below 👇
Google's recent Gemini 3 release has shocked the world with undeniable proof that the TPU is a powerhouse AI chip, leaving many to wonder what lays in store for companies like @NVIDIA & @xAI.
Listen to @GavinSBaker lay out the exact strategy that Jensen Huang & @ElonMusk are https://t.co/dDTVeLGSiJ
@patrick_oshag @GavinSBaker Useful podcast and insights, per usual.
For those interested in new frontiers - like Space & the Arctic - and how they interconnect with Technology (AI), Energy, and Money - check out the below threads + deep-dives:
https://t.co/QyrQKWYSOS
This is my fifth conversation with @GavinSBaker.
Gavin understands semiconductors and AI as well as anyone I know and has a gift for making sense of the industry's complexity and nuance.
We discuss:
- Nvidia vs Google (GPUs + TPUs)
- Scaling laws and reasoning models
- The https://t.co/099zErPlGP
So after all these hours talking about AI, in these last five minutes I am going to talk about:
Horses.
Engines, steam engines, were invented in 1700.
And what followed was 200 years of steady improvement, with engines getting 20% better a decade.
For the first 120 years of https://t.co/3nHcA14qST
Would be awesome for the country if he ran. One of the 10 most impressive CEOs I have ever met.
Something that made me really like him as a person:
Maybe 10 years ago I was eating at Scalinatella in NYC (great old school Italian place in a stone basement). Jamie Dimon sits
@GavinSBaker @SemiAnalysis_ @dylanp @FoolAllTheTime Yes. 💯but I was only speaking GB300 because AI trained on VR would be available in 1H27 and most likely Google might be a 25% of the orders 🤣.
@R_and_Invest @SemiAnalysis_ Love @dylanp and @FoolAllTheTime
I think those numbers might be “refined” over time.
And if we are quoting @SemiAnalysis_ then let’s also acknowledge that their post on Friday unequivocally stated that Rubin was significantly better than TPU v8. Which Dylan reiterated
The last year was the TPU's moment in the sun.
The next 12 months might be different as the GB300 finally ramps. And then Rubin.
Models trained and then inferenced on these GPUs are likely to show a dramatic leap in capability.
I think we see the first significant AI
$6.25 billion. 25 million children. $250 each.
Susan and I believe the smartest investment we can make is in children. That’s why we’re so excited to contribute $6.25 billion from our charitable funds to help 25 million children start building a strong financial foundation https://t.co/4Bcv3RKp0q
Best take so far on @DavidSacks. I’m not nearly as famous as the people David hangs around with, but he gives me time nearly every time I ask for it. Every time I have talked with him, he has been balanced, he is above board, beyond reproach and a true patriot. David doesn’t need https://t.co/42j5QgrmNe
I agree Gavin, the NYT distortion of @DavidSacks’ leadership in AI, Crypto & Quantum isn’t journalism—it’s almost strategic sabotage. While American’s bicker, our rivals are studying David’s every move. I’ve known David for decades, and I’ve never seen him sharper or more https://t.co/9N7FmpLXhq
Deeply strange @nytimes article about @DavidSacks
Leading in AI is good for America. And there is no way for America to lead in AI without American investors in AI doing well. Irrespective of whether those investors are David’s friends or his enemies. And like everyone who has https://t.co/9qQ5WOv9vz
@TMTMoats Thanks. Funny to think back on that interview from late 2021.
Turns out that “the one thing that matters is inflation” was more true than I ever imagined.
For example, I “know enough” about semis. Gavin “KNOWS” semis. I have the equivalent of an undergrad major in semis. Gavin has the equivalent of a PhD. https://t.co/ZBjUlNY6ft
In my completely unhumble opinion, Gavin is the best follow at the intersection of AI and markets. Actually gets AI, actually gets markets, doesn’t have a specific public persona bias he’s trying to pump. Case in point: https://t.co/tMB2LgBKZN
LoL.
Doing the AMD deal and then the Broadcom deal *before* the Nvidia deal was finalized might have been quite a mistake.
Highlighting via @modestproposal1 https://t.co/e0fl2Pg2Cz
Since @RealJimChanos is a fan of throwing money at the problem with respect to autonomy, and tells us Tesla/@elonmusk is failing to invest properly, perhaps he could help identify the data center build-out plan GM and Ford has? I mean they better get to it if they are going to https://t.co/aM3aygKeRc
@RealJimChanos How is this a defense or a relevant point in any way?
They do need the large coherent Hopper cluster which probably cost over $5 billion.
@GavinSBaker @RealJimChanos This is for very specific real-time inference, however, not AGI. These aren't directly comparable and there is no cost per token for Tesla's FSD inference because that's not the business model. "Yet," one may say, which one could say about many things Tesla-related.
@RandolphDuke7 @RealJimChanos That is why my post focused on FSD and robotics and did not mention AGI once.
Think AI5 will be to run a wider variety of models but we shall see.
Completely agree.
Sam Altman’s manifestly ridiculous $1 trillion of spending commitments shifted the AI investing landscape. The market is more skeptical now.
Ironically makes an IPO harder for them.
Also likely ended any potential for a 1999 style melt-up which is healthy. https://t.co/5TEgp3OfmO
The Non-Bubble that disappointed both Bulls and Bears -- how Sam's Splurge changed everything
The worst kept secret among Tech market participants — just something AI bulls don’t admit out loud: they want a price-action bubble every bit as much as the bears do.
Both want to see https://t.co/Rq3YAr9vbe
Completely agree.
Sam Altman’s manifestly ridiculous $1 trillion of spending commitments shifted the AI investing landscape. The market is more skeptical now.
Ironically makes an IPO harder for them.
Also likely ended any potential for a 1999 style melt-up which is healthy. https://t.co/5TEgp3OfmO
Some proper bangers in here...
"Cost per token is everything for AI. Google is the low cost producer of LLM tokens (with xAI as #2) but Tesla is the lowest cost producer of tokens that matter for FSD and Robotics. "
"AI is the first time in my career that being the low cost https://t.co/ADBeRk9qOx
Yesterday, @RealJimChanos posited that Tesla’s relatively low capex meant that they were not a serious competitor in real world AI and Robotics.
This is *exactly* the wrong way to look at it and the implications of this fact are actually positive for Tesla IMO.
Tesla’s https://t.co/zFvVvUE2ik
I thought @GavinSBaker had some good thoughts here. SaaS getting murdered, so protecting margins isn’t even working
TL:DR - High gross margins just signals nobody is using your AI products. Might as well accept reality and really swing for it. Otherwise you’re probably dead https://t.co/jKh7sOA4WV https://t.co/joDdNCCjk2
@GavinSBaker It’s certainly something that makes being a strategist harder today, given the market concentration, and NOT being (or pretending to be) an expert on Tech. I’m trying to stick to what I know best rather than being the Nth person pontificating on AI.
Why are value and macro investors so comfortable making highly confident prognostications about AI despite their manifest ignorance?
I don't recall many tech/growth investors penning long, bearish, highly confident and utterly ignorant analyses about energy in 2022/2023.
So this is the ghost line showing “SBG shareholder value if we still owned Nvidia", it all makes sense now. https://t.co/V0iNuGMaRb https://t.co/PSNIwykytp
@JonathanKw46272 Who can say? I am hoping someone industrious and quantitatively minded runs this piece of esoteric data down. Like what would that 4.9% stake in Nvidia be worth today relative to the entire market cap of SoftBank?
A 28% yoy rental price decline actually implies a pretty fat IRR >~30%. And spot prices at AWS/GCP imply a big premium to what this series shows (~50% higher).
Most importantly, the A100, a 5yr old GPU, is still in high demand at prices implying less than 28% rent declines. https://t.co/0w7u5meAiC
@GavinSBaker @DavidGeorge83 @a16z Gavin, big fan, but I’m genuinely curious about your thoughts on the idea that it’s not a 1:1 comparison to compare no dark GPUs to the dark fiber during the .com bust.
Ultimately it’s about generating profits and right now the end customers are extremely unprofitable.
That https://t.co/Ai30LfmX1f
@LiebermanAustin @DavidGeorge83 @a16z “The end customers are extremely unprofitable.” Partially true. But loads of vertical AI startups are profitable and some of the largest end customers are Google, Microsoft, Meta, etc.
Everyone's asking if we're in an AI bubble.
@GavinSBaker gave a reasoned take at @a16z Runtime on why this isn't like the 2000s.
"At the peak of the [Internet] bubble, 97% of the fiber that had been laid in America was dark.
Contrast that with today - there are no dark GPUs." https://t.co/upeME3221O
Really fun to do this fireside chat with my good friend @DavidGeorge83 who has an exceptional track record.
Dark fiber during the telecom bubble vs. melting GPUs today.
Grateful to David and the @a16z team for inviting me.
https://t.co/doDCpziRX8
My firm, @atreidesmgmt, will be voting in favor of Elon Musk’s performance-based compensation package.
And I would prefer that *every* company I invest in have a comparable plan.
I believe shareholders should generally support thoughtfully structured performance-based CEO
Elon (Musk) has done more for the planet and de-carbonizing it than all the activists and Greta Thunbergs..."
"Global Warming is a solved problem...only through economics!" -- @GavinSBaker
Full episode below https://t.co/XnrsiUq2dq
Elon (Musk) has done more for the planet and de-carbonizing it than all the activists and Greta Thunbergs..."
"Global Warming is a solved problem...only through economics!" -- @GavinSBaker
Full episode below https://t.co/XnrsiUq2dq
My point is not that there’s a bubble. My point is whether people are saying there’s a bubble or not doesn’t change the reality on the ground. https://t.co/5kUzPhHZq2
@modestproposal1 Profoundly agree. Even more extreme in 2000.
There are many good arguments as to why we are not in a bubble, but “there is no bubble because so many are calling a bubble” is not one of them. Absolutely ahistorical.
@GavinSBaker @DKThomp Here is the other side of that @GavinSBaker - part of that upside is def AI, but it’s also coincided with the big disinflation drop from late 2022 and the flattening and decline of interest rates thereafter. We have seen several periods of weakening employment and rising stock https://t.co/vQNRigRMnQ
@GavinSBaker @DKThomp It could also be that the plurality of the stock market is now highly concentrated across 7/8 companies who, on a relative basis, just don’t employ a lot of people so the stock market goes up won’t pull employment along with it.
@GavinSBaker @DKThomp Job openings/hiring down after overhiring in 2021 but 40% of index (mag7) saw massive earnings leverage post layoffs & that drives index returns
Really interesting chart from @DKThomp
Stock market and job openings were highly correlated for decades.
No longer. Stock market up 75% and job openings down 33% since 2022.
Would think at least partially due to AI. https://t.co/PLZ7SSTQZx
@GavinSBaker @Brycee1der @FT Yeah, I didn’t think these inconsistencies altered the conclusion of reconciling the dispersion in pricing on the same product.
@Brycee1der @FT Which would you prefer? Happy to do either.
And I would hope that the changes also encompass your 5 conclusions.
I do profoundly appreciate and admire your intellectual honesty. A rare quality today.
Drinks are on me if you are ever in Boston.
@Brycee1der @FT I will fix that when you replace all the references in the article to $1.65 with $13.20 which is the internally and logically consistent framing relative to the "$4 break even" *and* explain how this is consistent with all five of your "possible conclusions."
The polite
My most extensive X argument in years.
Has been kind of fun.
The @ft really needs to either retract the article or correct it along the lines I have repeatedly described.
Should I make this a daily crusade until it happens?
And I think the poor A100 has a defamation case! https://t.co/qQhCzvbFuv
@Brycee1der @FT I will fix that when you replace all the references in the article to $1.65 with $13.20 which is the internally and logically consistent framing relative to the "$4 break even" *and* explain how this is consistent with all five of your "possible conclusions."
The polite
@Brycee1der @FT What specifically are my bad faith arguments?
Which of your 4 erroneous “possible conclusions” are actually supported by *any* of the data in your article?
Happy to post on this daily until the article is retracted or you take the loss in a more fulsome correction.
Cheerio!
@Brycee1der My original post which I think still stands until the @ft acknowledges that 4 of the 5 “possible conclusions” in the original article are utterly invalidated by the articles own data.
The published correction by @Brycee1der does not acknowledge that on per cluster basis, current average per hour A100 prices of $13.20 are well above breakeven.
The article reads quite strangely given this fact, which is derived entirely from the articles own data and “toy https://t.co/KAkvfuikJt
@Brycee1der My original post which I think still stands until the @ft acknowledges that 4 of the 5 “possible conclusions” in the original article are utterly invalidated by the articles own data.
@Brycee1der @FT I likewise appreciate all of your earnest typing in the original article, the correction, all your emails on the matter to me and your many replies here. So thank you for typing all that! Seems like a lot of work!
I’m no fortune teller, but intellectual honesty on the part of
@insane_analyst @Brycee1der @FT The reporter continues to dig in and asked me for this clarification.
I think I was quite polite after he called my original post on the matter “bullshit.”
I am writing this post at the request of @Brycee1der who wrote the erroneous @ft article titled “”What GPU pricing can tell us about how the AI bubble will pop.”
Bryce writes: “Can a GPU-as-a-service company break even on that type of customer? Here’s a very simplified toy https://t.co/jJi70tLrhY
The published correction by @Brycee1der does not acknowledge that on per cluster basis, current average per hour A100 prices of $13.20 are well above breakeven.
The article reads quite strangely given this fact, which is derived entirely from the articles own data and “toy https://t.co/KAkvfuikJt
I am writing this post at the request of @Brycee1der who wrote the erroneous @ft article titled “”What GPU pricing can tell us about how the AI bubble will pop.”
Bryce writes: “Can a GPU-as-a-service company break even on that type of customer? Here’s a very simplified toy https://t.co/jJi70tLrhY
Vaguely incredible that the @ft writes an article about how current GPU rental prices suggest that AI a bubble, but makes a math error such that all of their calculations are off by a factor of 8 by conflating per server and per GPU economics.
Doing the math correctly was https://t.co/bkrkszYp14
As @zephyr_z9 and @The_Colonel__ pointed out when I quoted this circle, the reporter mixes up per-system breakeven costs vs. per-GPU rates. That is comparing apples to oranges.
Using their own figures:
DGX A100 cluster = $199K (~$25K per GPU * 8 GPUs)
Hours over 5 years = 5 * https://t.co/cXOTProzrY
@GavinSBaker Sorry my bad. I thought FY. The spending is way out of control. For all parties. It’s hard to shake off Biden era spend but someone has to do it.
@GavinSBaker AI is hallucinating. Fiscal spending has gone up by $500B yes with a B. Tariff revenue is trying to cut down on some of this. Fiscal deficit is NOT below 5%. Please do not use AI. Even AI can’t tell you the king is naked.
Google gaining search share two months in a row.
Overall search instances accelerating from 3% YoY growth in October 2024 to 27% in September 2025 as GenAI expands the market.
Funny relative to the hysteria earlier this year. https://t.co/OXcrBSfDQG
Google gaining search share two months in a row.
Overall search instances accelerating from 3% YoY growth in October 2024 to 27% in September 2025 as GenAI expands the market.
Funny relative to the hysteria earlier this year. https://t.co/OXcrBSfDQG
@WarrenPies Lower A100 availability today post Blackwell vs the beginning of the year maybe most interesting.
Those 5-6 year depreciation schedules looking much more solid.
And every private credit fund that invested in GPU financing is feeling good.
Ampere per hour pricing and
Blindsight by Peter Watts (2006) is my standout fiction book of the year, by far. Grand, challenging, scary, and super autistic. https://t.co/BrlOcl349t
Amazon raising Blackwell per hour pricing.
H200 rental pricing going up *after* Blackwell scale deployments ramping up.
Might be important.
And certainly more important than ridiculous $300 billion deals that are contingent on future fund raising.
@EvanHoffman @TheStalwart @tracyalloway Tbh I did not know about it until @GavinSBaker pointed it out. He made the point much more eloquently than I did.
Google’s dominance of the Pareto frontier for AI has been shattered.
And actually wild that Grok 4 Fast *also* has a 2m token context window. https://t.co/SSxaC5UdWc
@GavinSBaker in January, “I think the best performing asset of 2025 will be HBM — high‑bandwidth memory — because AI training and inference are going to keep starving the market of supply, and everyone in the AI stack needs it.”
Micron $MU is up by 100%. Holding shares. https://t.co/YW8XtscbTz
A great life hack. Don’t stop reading literature just because you graduated college. I am reading David Copperfield by Dickens which aficionados tell me is not even the best book.
Why read great books? Maybe for the same reason we walk in nature or enjoy other forms of art—it
@DanielSLoeb1 Great Expectations and a Tale of Two Cities are fantastic. Never read David Copperfield will try it when am done with the remaining 6000 pages of the Sun Eater science fiction novels.
@profplum99 Pick the metric that makes one feel smart and correct in ones views.
Earnings multiples vs. sales multiples.
Everyone can be right because of Nvidia’s much higher margins vs. Cisco in 2000.
Perhaps most accurate to say that 2000 was clearly a valuation bubble whereas today
I’ll have to work harder for top billing next year. Always a pleasure to share a stage with @GavinSBaker and my old boss @HandlerRich. So great to be with the resilient tech and investment community as well as so many friends. https://t.co/oENXfk3ivJ
If you were in Israel this week at @Jefferies TechTrek, you got to meet the most incredible local entrepreneurs and learn from the wisdom of @BillAckman, @DanielSLoeb1, @GavinSBaker, Paul Singer and David Fialkow! 🇮🇱💙🤍 https://t.co/FSHE0j3Jy9
@HandlerRich @Jefferies @BillAckman @DanielSLoeb1 @generalcatalyst Thank you for having me and honored to be in such august company. Incredible conference and I will always think of your Margarita story when decisiveness is needed.
RIP Charlie Kirk. Absolutely tragic.
Lucky for 🇺🇸 that his central messages were nonviolence, tolerance of opposing views and the importance of engaging in civil discourse with those whom who you disagree with most profoundly. He often said “when discourse ends, violence
“I think you're if you're a big business and you can pay a million dollars to let an AI think for six weeks about the most important question for your business, that's gonna be a profound advantage relative to small businesses that can't afford that.”
“inference compute is gonna be the the kind of derivative winner of that. We're gonna run out of GPUs, accelerators, compute in 2025 the same way we did in '23.”
“My most contrary belief, so I think that America over the next at some point over the next four years will print at least one year of greater than 5% GDP growth, real GDP growth.”
“I think that America over the next at some point over the next four years will print at least one year of greater than 5% GDP growth, real GDP growth.”
“I think you will see the Frontier Labs stop releasing their leading edge models to prevent knowledge distillation and their IP effectively being stolen.”
“High bandwidth memory is a bigger part of NVIDIA's cogs on GPUs than Taiwan Semi is. And today, there's two companies that can make it, Hynix and Micron.”
“And then we started scaling around inference time compute. And it's very clear that we have now added a third axis of scaling performance, and that is reasoning.”
“And if you create that first digital god, we can debate whether it's tens of trillions or hundreds of trillions in value, and we can debate whether or not that's ridiculous.”
“x AI decided they were gonna build a 100,000 GPU cluster. And I would say with Elon's unique physical engineering mind, which we've seen play out at SpaceX and Tesla, and now at Neuralink, where he figured out how to miniaturize everything along with really capable teams.”
“And I would say with Elon's unique physical engineering mind, which we've seen play out at SpaceX and Tesla, and now at Neuralink, where he figured out how to miniaturize everything along with really capable teams. I think he rearchitected the x AI data center from first principles in Memphis.”
“And I would say with Elon's unique physical engineering mind, which we've seen play out at SpaceX and Tesla, and now at Neuralink, where he figured out how to miniaturize everything”
“And because of that, they're able to get enough density that they could effectively make a 100,000 GPU cluster coherent even with Hopper without next generation networking technologies from NVIDIA, Broadcom, and others.”
“Chase Coleman, who runs Tiger, who's an exceptional investor, had a really interesting statistic, which was ChatGPT came out in 2022 and was to AI as NetScape Navigator, was to the Internet in 1994.”
“MFU, model flops utilization, and that generally runs around 35 to 40%. And that's literally the percentage of compute, theoretical compute flops that you're actually applying to trading.”
“You could choose between faster time to market. If you run a 50% MFU and your competitor's running 40 for an equivalent amount of trading flops, you could be in market 25% faster.”
“By the way, these models are commodities today, but I am suspicious once we get to scaling laws continue, GPT seven or eight literally cost $500,000,000,000 to trade.”
“MFU is the most important metric because it gives you all of these advantages and ways to differentiate yourself amongst five people, six people who've trained these GPT core class models.”
“There's only basically three places in The United States today where you can get a gigawatt of power to a single data center that's reliable enough, I. E. That's nuclear.”
“There's only basically three places in The United States today where you can get a gigawatt of power to a single data center that's reliable enough, I.”
“And it is solved because photovoltaic cell efficiency is compounding just under 10% a year, and battery efficiency is compounding 200 bps less than that.”
“These companies are all public, and there is something called return on invested capital. And ROIC has gone up for all of these companies since they ramped CapEx.”
“I think 12 dot five is like GPT three, and it can consistently drive me most places with no interventions. I'm a seasonal driver. I really only drive my Tesla in the summer.”
“This year, it's like when we turned on 12 dot three, it's like all of the progress over the last ten years was in that one release from the first time I had that Tesla.”
“I think they're going to go really fast to GPT-4.5 compute, which means you're going to get, using these orders of magnitude, you're going get a 100x improvement really fast.”
“I would say probably the average tech or consumer oriented growth stock that's below a $100,000,000,000 in market cap is probably somewhere between 40 and 65 percent off its all time high.”
“It's not a comment on Amazon. It's just a comment on the supplier response. They have spent more money on CapEx in the last two years than they did in the preceding twenty years.”
“Taiwan Semi, their 2022 CapEx is going to be many multiples of 2019. 2021 and 2022, they'll spend more than they did in the preceding five years. So there is a massive supply response coming.”
“If you trend out and look at real personal consumption expenditures, goods spending is roughly $500,000,000,000 above the pre COVID trend line. Services spending is 500,000,000,000 below the trend line.”
“Let's step back and look at the last fifteen years of semis. The industry has completely consolidated to where you almost have these monopolies and almost monopolies or duopolies.”
“Microsoft wrote about this in a research paper ten years ago, or maybe not twelve years ago, the quality of a given AI algorithm doubles with every 10x increase in the amount of data you use to train that algorithm.”
“Today, Intel, Samsung, Taiwan, Semi, they the only companies in the world that can make leading edge logic. There's only three companies that can make leading edge DRAM, maybe four for NAND.”
“But they try and force the open source infrastructure build themselves. And I think Google was the first one to realize, like, Hey, this is kind of silly.”
“B, there's a lot of money to be made just in providing these primitives. Let's stop trying to compete with these independent infrastructure software providers right on top of the cloud.”
“But his whole strategy a lot of his strategy for venture, which I ruthlessly copied, is around the idea of their profound advantages towards being multistage.”
“If I could know one thing about a person to know whether or not they would be a good public equity investor, it would simply be, are you capable of being rational when you're wrong?”
“Valuation spreads within a sector between individual stocks are at all time highs. It was a 4.5 standard deviation relative to the norm only a few days ago.”
“At the beginning of almost every bear market, doing the most obvious things works. So who crushed it for the first three to four weeks of this bear market?”
“Days of liquidity with zero revenue has been the acid liquidity test that I think a lot of investors have been working through. And then from a broader perspective, I think a couple of insights.”
“Go back to what were the best stocks coming out of June 2009? Netflix, booking.com, Ulta Beauty, Domino's Pizza. Domino's was a name that was incredibly levered.”
“A lot of people are baselining off what happened in 'eight, 'nine. And I think that is dangerous because in 'eight, 'nine software was seven to 10 of IT spend.”
“And I think that is dangerous because in 'eight, 'nine software was seven to 10 of IT spend. And today it's roughly 30% excluding spending on the big cloud hyperscale infrastructures”
“And today it's roughly 30% excluding spending on the big cloud hyperscale infrastructures and service companies like Azure, Amazon web services, and Google GCP.”
“In the last two days, lots of friends have come to me and said, Hey, I told my brother, my sister, my mom, my dad, my uncle, my aunt, my friend who has a small business, that story, they did the same thing and they cut their software bill 30 to 50% in an hour.”
“And I can only imagine what it feels like to be at whether it's Apple with FaceTime, Google with Hangouts, Facebook with whatever their video chat is.”
“Would argue personally they're at parody, but 50 bps is huge. If you're 50 bps better, that can be the foundation of hundreds of billions of dollars in market value.”
“Google found that milliseconds of latency mattered to search. That if you could take two milliseconds out of search latency, that it had a huge impact on search volumes.”
“That if you could take two milliseconds out of search latency, that it had a huge impact on search volumes. And Zoom's UI and UX for all sorts of reasons is just better”
“The metaverse is simply a series of connected virtual worlds that I firmly believe a majority of people will spend a majority of their waking hours in within my lifetime.”
“I think the last data I saw, this was a Telecom Italia stat, so we should not mix and match, but Telecom Italia saw video game traffic up 75% and social media traffic up 0%.”
“And in some ways, I think of the invention of the microchip in 1959 as being or the integrated circuit as being analogous to James Watt's invention of the rotary steam engine in 1761.”
“I would argue peak economic, political, social change as a result of that invention was seventy to one hundred years later during the peak of the industrial revolution eighteen twenty to eighteen fifty.”
“It's critical in growth investing to assume multiple compression. It doesn't always happen, but you want to be paid based on business outcomes and assume valuation is not part of the outcome.”
“One is, I think the price to used iPhones is critically important to Apple's business model. Why is that? Used iPhones retain their value much better than used Android phones.”
“The only thing that has enabled the AI revolution that we're living through, which I think we're at the bottom of the first inning in, is one, we had the ability to do cloud computing, so just apply significantly more computational power to old algorithms, and then b, we had dramatically more data.”
“You could just measure search in a way you can't measure slotting fees. Anything you can measure generally gets, in some ways, may be overvalued in today's Internet world.”
“CAC is the new rent. Customer acquisition cost, online ad spend. It's not just replacing other forms of ad spend. For a lot of businesses, it is replacing rent. Storefronts.”
“From a monetization perspective, your average mobile phone game monetizes, let's call it, 8 to 10¢ an hour. If it's not average, that's a reasonably well monetized game.”
“All that matters is the quantity of data, and it's been very well established in multiple papers from both Google and Microsoft research that for every order of magnitude increase in the data you use to train an algorithm, the quality of the AI doubles.”
“and it's been very well established in multiple papers from both Google and Microsoft research that for every order of magnitude increase in the data you use to train an algorithm, the quality of the AI doubles.”
“it's been very well established in multiple papers from both Google and Microsoft research that for every order of magnitude increase in the data you use to train an algorithm, the quality of the AI doubles.”
inferenceinference·posts only·
48 posts
·
0 of 0 appearances
new
post
“@IbrahimAjami Yes. The acceleration in OpenAI/Anthropic as open-source takes share is so interesting. The future is multi-model with specialized open-source models that were post-trained on customer data working together with frontier models behind a customer specific router. Firework Nexus”
“Suspect we will start to hear about a “Pareto optimal” balance of computationally efficient humans, cheaper open-source tokens and frontier tokens.
Our internal AI spend @Atreidesmgmt will be roughly 100x higher in August 2026 vs. March 2026. Still roughly doubling every month.”
“More data than open-source AI is taking share from OpenAI and Anthropic. Open source has gone from 28% token share to 62% token share @vercel over the last 2 months. Chart from @rauchg
Super impressive given that the sum of OpenAI and Anthropic accelerated in July. So net https://t.co/wjDtlcl6ic”
“@Shaughnessy119 @TMTBreakout 2 of the 3 parts of your equation are correct (first and last, not the middle). An open-source token costs just as much compute and energy to produce as a frontier token for a model of comparable size. And they are often less token efficient.”
“@altcap Love the expectations management!
Let me say this: I would not take the “under” on $250 billion *unless* they stumble or fail to secure enough compute.
Maybe you can discuss with @DarioAmodei at @TomorrowXSummit before I interview you about public policy!”
“I think @bot is another “Claude Code” moment for AI.
I would estimate my personal AI usage is up something like 100x.
And for everyone who reached out about how to build a “podcast summarizer” it took me about 15 seconds in Grok Bot and is better than what I had before. https://t.co/AaoeHbEqtl”
“Grok 4.6 is roughly the same performance as Fable 5 Max at an 85% discount. 80% cheaper for input tokens and 88% cheaper for output tokens.
Pareto dominant.
Grok 4.7 will be significantly better as is a much larger model with the Cursor and SpaceX data included in pretraining. https://t.co/4e8CXbzs2D”
“Market is overreacting to hyperscale credit spreads widening from my perspective. TL;DR Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning while operating cash flow acceleration is an underestimated source of funds for AI https://t.co/D9FSHXQ8F8”
“Should note the real comparison will be with Rubin Ultra given Rubin shipping today so kinda have to wait. And Feynmann coming fast. But still impressive debut from Jalapeño.”
“@ChrisMurphyHQ Seem to hit different points of interactivity vs. throughput tradeoff for different size models. Basically everyone plays to their traditional strengths is early read. Really good AFFN disaggregation could change this.”
“Impressive that Jalapeño outperforms the comparable TPU and is in the mix with Rubin.
Credit where credit is due - first good ASIC outside of TPU/Trainium.
However, will likely significantly underperform a disaggregated GPU/Trainum plus SRAM accelerator setup. Especially with”
“@JoshPurtell @arakharazian Frontier labs would say their tokens use less compute than comparable open-source tokens but that would break people’s brains. And for sure they have more intelligence density per token which compounds their efficiency advantage.”
“@arakharazian Respectfully this is not correct from my perspective. Cheaper open-source tokens are net positive for the AI infrastructure names, although maybe not for the model companies.
An open-source token broadly costs the same amount of compute (watts and wafers) as a frontier token to”
“@DratchCap @anni_sen @stefanveliki137 @thsottiaux I am also curious about this. Along with why Ultrafast mode is capacity constrained. Deal struck in late December, 4.5 months to make a wafer, 6 weeks to make a server, a month to energize a server. So logical that Ultrafast debuted in August and would think it would be less”
“I literally have no idea why anyone is concerned about the FT data on Fable 5. Just shows customer preference for ZDR. Actually surprised so many are still using it after Opus 5 became available. https://t.co/u5Rzi13JmV https://t.co/GCqX5bMgYg”
“@firesidealpha This was a good podcast. The more interesting question is where those teacher models stack up vs. the frontier. And where the teacher models will be in a few months. Those ASICs sure were fun to develop!”
“@anni_sen @stefanveliki137 @thsottiaux You think OpenAI is constraining the capacity or production volume constraints elsewhere are constraining the capacity? Genuinely curious.”
“@anni_sen @stefanveliki137 Hmm. This looks like the flagship big model from the second largest frontier lab. And IIRC @thsottiaux from OpenAI said they wanted this currently capacity limited Ultrafast mode to be the default mode 6 days ago. I’m not sure I would describe “default mode” for the flagship https://t.co/OQqwwW8CD9”
“The most obvious implication of GPU price increase is that anyone who is spending a lot on Blackwell and Rubin before January 31, 2027 is in a good relative position vs. competitors.
Super curious to know if GPU prices are going up more than rack prices.”
“High conviction hypothesis especially for large enterprises.
This is the best path to Pareto dominance while keeping data safe.
SMBs and individuals probs just use the frontier harnesses. https://t.co/4petLDY4kt”
“@itsvlady Well it lowers compute spend by the ratio of the margins of the frontier tokens to the margins of the open-source tokens - frontier tokens mark up underlying compute dramatically more than open-source.”
“@jweil130 @vercel @rauchg Likely true of Anthropic IMO - likely building in an acceleration for roadshow and end of year. Probably less true for OpenAI.”
“@F1NewsWire @TMTBreakout Big models are expensive. I don’t think super important one way or another. Opus 5 is a Mythos/Fable distillation and doing well.”
“Yes. And the sum has accelerated slightly QTD with a sharper acceleration in August vs. July.
And note that Grok Bot, Grok and Open Source are all likely growing faster than both of the two leading frontier labs.
But at some point going to have to move past second derivatives https://t.co/L4r1aExjrI”
“@Shaughnessy119 Y is higher than X - Anthropic and Open AI have more leverage over the infrastructure/cloud providers.
For a given token over time all else equal, the cloud and infrastructure players charge more for open source tokens vs. what Anthropic and OpenAI pay per token b/c open source”
“Empirically impressive metrics from Stripe in the Axios report about their acquisition of Openrouter.
Growing net revenue and FCF almost 2x faster than Adyen YTD with the share count declining over the last 3 years.
I think Stripe will be a good steward of Openrouter and it is https://t.co/mnxU0XqP0N”
“Fireworks and Baseten are the two largest inference clouds and I think of them as *an* index for open-source AI adoption.
Also first time seeing SpaceXAI on the list of “fastest growing” IIRC.
SambaNova! Will wonders never cease. https://t.co/HlmFsKRtVk”
“Good chart from @TMTBreakout
July accelerated both MoM and YoY for the sum of OpenAI and Anthropic.
Some chance Anthropic shifted from gross to net reporting for ARR. Meaning the 65b is a more conservative metric than the $47. Smart thing to do before an IPO.
Openrouter data https://t.co/FOwqXizXIL”
“From yours truly on Friday’s @theallinpod
“On the margin, Anthropic is losing share to OpenAI, to open source, and to Grok, and they're still growing so fast.”
If the $65b is correct, then someone leaked it to help Anthropic on their forthcoming IPO roadshow. Basically, they https://t.co/oIYSS1PZhT”
“The GrokBot origin story as told by @bot
Sadly they lack both a sharing feature and an iPad app but I’ve been told they are both coming soon: https://t.co/4YaISohWW6”
“Important. The future is almost certain to be multiple models working together.
Lower margins at the model layer whether from open-source or vertically integrated players = more AI infrastructure demand. https://t.co/YBWlCNnrDD”
“Super excited that I will be interviewing @lqiao CEO and founder of @FireworksAI_HQ at the @tomorrowxsummit in Austin, TX November 17-18.
Lin’s vision of specialized intelligence is compelling and she is making it a reality. Fireworks is at the forefront of helping enterprises https://t.co/MK7CFgAsAU”
“Looking forward to listening to @AntonioGracias interview Jon Gray, President and COO of @blackstone at the @tomorrowxsummit November 17-18 in Austin, TX.
Blackstone is leaning into AI more than most of its peers and they are integral part of financing it.
So many topics: their https://t.co/6xK1XtNKl1”
“Really impressive from @cursor_ai
The future is trending hybrid. Next iteration will likely be allowing customers to use their own proprietary data to post-train an open-source model.
Excited for Grok 4.6. https://t.co/ZTE2VAskTG”
“@benitoz @ChaseLochmiller @CrusoeAI @TomorrowXSummit He will have an interesting perspective on the buildout.
Vertically integrated from dirt through electrons through tokens.”
“Happy to announce @ChaseLochmiller CEO and Founder of @crusoeai as the second confirmed speaker at the @TomorrowXSummit
This photo is my first time ever seeing Chase in a suit! https://t.co/mL4kKOTfwf”
“@dampedspring Of course. Real yields going up and spreads widening is undeniable. I am saying we may need less credit than expected given a likely continued acceleration in operating cash flow.”
“@sashadem @theallinpod I believe both you and Sholto. This issue is that Dario’s own public commentary makes it believable to many serious people in Silicon Valley that Dario might have said something similar to “Anthropic might end up being the only private company.”
I would love nothing more than to https://t.co/NF7MaO8wl1”
“@_sholtodouglas Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging”
“$3 trillion for Anthropic in the FT.
I feel like I’ve heard that somewhere else.
Although public markets generally struggle to accurately value growth when it is this high. https://t.co/2R80q18GUH”
“@jukan05 4 gigawatts this year and another 4 next year per this report. Spot price for compute has gone up since they signed their Anthropic and Google deals.”
“@dampedspring Where has Anthropic’s and OpenAI’s combined 150b of ARR come from?
I don’t think the idea that leveraging up is the only way to fund this is at all correct.”
“@dylan522p @maxkan And they also said this followed by “Cerebras FTW.” I would be surprised if their head of Codex was unaware of their roadmap. Reasonable to hypothesize tomorrow’s Ultrafast will be either GPUs or Jalapeño working with CS-4 or CS-5. We shall see! https://t.co/QM0vs809x7”
“Game theory suggests memory LTAs are likely to be durable.
And yes, Nvidia may have a significant cloud business via revenue shares. https://t.co/GFN1kDQmSk”
“We are going to see a lot of vertically focused AI native companies accelerate.
Routers, open-source models and specialized post-training enabled by companies like @FireworksAI_HQ have all made dramatic advances and the combination of the three is driving accelerating growth. https://t.co/f8qJdKSciI”
“@jukan05 4 gigawatts this year and another 4 next year per this report. Spot price for compute has gone up since they signed their Anthropic and Google deals.”
“A series of fairly wild thoughts from Dwarkesh. I thought I was bullish but even I am not assuming that the price to rent compute continues to go up!
I will say that I am in SV this week and Dwarkesh is capturing the zeitgeist.
Net, net public markets would probably be trading https://t.co/PzIxxnzbiS”
“@dampedspring Where has Anthropic’s and OpenAI’s combined 150b of ARR come from?
I don’t think the idea that leveraging up is the only way to fund this is at all correct.”
“Market is overreacting to hyperscale credit spreads widening from my perspective. TL;DR Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning while operating cash flow acceleration is an underestimated source of funds for AI https://t.co/D9FSHXQ8F8”
“@dylan522p @maxkan Maybe GPUs plus CS-4/5 for Ultrafast at scale. Jalapeño not financeable at comparable rates to GPUs and probably not that big of a wafer allocation next 2 years. Jalapeño still a really impressive achievement from OpenAI. I remember being sad when they hired Thomas away from https://t.co/a0jIXXahXN”
“@maxkan Hmm. I might ask @dylan522p if Attention-FFN disaggregation (AFD) is compatible with data locality and not moving the KV cache. That constraint is what OpenAI focused on with Jalapeño. They absolutely did not make an “explicit bet that disagg is not the way to go.” There are https://t.co/2lVjH2M5vS”
“@NoonHillCapMgmt Always wise to be skeptical. Nothing wrong with the analogy, just the barriers to entry are higher, quality of companies is materially higher and structurally harder to bring on capacity given shortages.”
“Nature is healing.
Any politician opposed to data centers is unqualified to hold public office.
Data centers are revitalizing small towns all over America, a godsend for blue collar workers, lowering electricity costs, accelerating the transition to sustainable energy and use https://t.co/fTH3qZTP2V”
“@maxkan Hmm. I might ask @dylan522p if Attention-FFN disaggregation (AFD) is compatible with data locality and not moving the KV cache. That constraint is what OpenAI focused on with Jalapeño. They absolutely did not make an “explicit bet that disagg is not the way to go.” There are https://t.co/2lVjH2M5vS”
“Should note the real comparison will be with Rubin Ultra given Rubin shipping today so kinda have to wait. And Feynmann coming fast. But still impressive debut from Jalapeño.”
“Impressive that Jalapeño outperforms the comparable TPU and is in the mix with Rubin.
Credit where credit is due - first good ASIC outside of TPU/Trainium.
However, will likely significantly underperform a disaggregated GPU/Trainum plus SRAM accelerator setup. Especially with”
“The most obvious implication of GPU price increase is that anyone who is spending a lot on Blackwell and Rubin before January 31, 2027 is in a good relative position vs. competitors.
Super curious to know if GPU prices are going up more than rack prices.”
“Every accelerator company, and I mean every accelerator company, is likely going to do some variant of 3D DRAM or the “zHBM” which were both discussed at HotChips today.
The larger companies will probably run (are running) parallel programs, but these two memory variants”
“Jensen was #1 on Glassdoor’s list of the best CEOs to work for. Still the same guy he was when I met him 27 years ago. Maybe slightly calmer.
The ultimate testament is that roughly half of the employees at Nvidia are worth over $25 million, 75% are millionaires and essentially https://t.co/iS2zgYE0jI”
“Game theory suggests memory LTAs are likely to be durable.
And yes, Nvidia may have a significant cloud business via revenue shares. https://t.co/GFN1kDQmSk”
“Delighted to announce @JensenHuang as the first confirmed speaker for the @TomorrowXSummit and even happier that I will be interviewing him.
We will have Chicken Fried Steak at the summit. Sawyer’s is my favorite place in Austin but we will of course have Denny’s. https://t.co/4QKBMGh4TU”
memorymemory·posts only·
8 posts
·
0 of 0 appearances
new
post
“Impressive that Jalapeño outperforms the comparable TPU and is in the mix with Rubin.
Credit where credit is due - first good ASIC outside of TPU/Trainium.
However, will likely significantly underperform a disaggregated GPU/Trainum plus SRAM accelerator setup. Especially with”
“@IanCutress Luckily for me, I have never been near a cliff or a bridge when I have read a post about how “SRAM cannot scale.” This topic has largely been put to bed for anyone remotely close to logic/memory/packaging roadmaps especially at wafer scale.”
“Every accelerator company, and I mean every accelerator company, is likely going to do some variant of 3D DRAM or the “zHBM” which were both discussed at HotChips today.
The larger companies will probably run (are running) parallel programs, but these two memory variants”
“Happy to announce Sumit Sadana EVP and Chief Business Officer of @MicronTech as a confirmed speaker at the @TomorrowXSummit in November.
Memory per FLOP drives tokens per watt. Looking forward to hearing Sumit’s thoughts on how AI is changing the business model of memory https://t.co/k1Pdk94PDu”
“Looking forward to interviewing @andrewdfeldman CEO and founder of @cerebras at the @TomorrowXSummit in Austin November 17-18.
Scale-up networking and SRAM acceleration are the best solutions to the DRAM shortage from my perspective. Given that Nebius is monetizing at $20 to $25 https://t.co/ZUAzwy8A5d”
“Jitendra Mohan, CEO and co-founder of Astera Labs, is our third confirmed speaker for the @TomorrowXSummit.
Really looking forward to discussing the role of scale-up networking in a memory constrained world and the intersection of their switch franchise with optics. https://t.co/stwGpJ8zox”
“Game theory suggests memory LTAs are likely to be durable.
And yes, Nvidia may have a significant cloud business via revenue shares. https://t.co/GFN1kDQmSk”
cerebrascerebras·posts only·
6 posts
·
0 of 0 appearances
new
post
“@dylan522p @maxkan Maybe GPUs plus CS-4/5 for Ultrafast at scale. Jalapeño not financeable at comparable rates to GPUs and probably not that big of a wafer allocation next 2 years. Jalapeño still a really impressive achievement from OpenAI. I remember being sad when they hired Thomas away from https://t.co/a0jIXXahXN”
“@dylan522p @maxkan And they also said this followed by “Cerebras FTW.” I would be surprised if their head of Codex was unaware of their roadmap. Reasonable to hypothesize tomorrow’s Ultrafast will be either GPUs or Jalapeño working with CS-4 or CS-5. We shall see! https://t.co/QM0vs809x7”
“@insane_analyst @andrewdfeldman @cerebras @TomorrowXSummit I would not recommend this course of action but I am happy to ask your questions. Just be sure you want me to ask and attribute to you *after* you have carefully reviewed their patent filings. Always looking out for you as appreciate your contributions both here and IRL.”
“@masayoshisson @andrewdfeldman @cerebras @TomorrowXSummit Happy to discuss yields to put this silly issue to bed forever. In the meantime, I would point you to their patent filings in the public domain.”
“Looking forward to interviewing @andrewdfeldman CEO and founder of @cerebras at the @TomorrowXSummit in Austin November 17-18.
Scale-up networking and SRAM acceleration are the best solutions to the DRAM shortage from my perspective. Given that Nebius is monetizing at $20 to $25 https://t.co/ZUAzwy8A5d”
regulationregulation·posts only·
5 posts
·
0 of 0 appearances
new
post
“@_sholtodouglas Ok. Becky just brought me another gin and tonic and is watching our daughter.
So: I think the framing of cyber as defense dominant and bio as offense dominant is sensible and the bio point is probably the best pro-regulation argument. And the USG should absolutely have all”
“@sashadem @theallinpod Likewise appreciate your response and generally agreed. Am just a little worried that the negative vision is “winning” right now with very unpredictable consequences for everyone.
Reasonable minds can minds can disagree about the best path forward for the most important”
“@sashadem @theallinpod I believe both you and Sholto. This issue is that Dario’s own public commentary makes it believable to many serious people in Silicon Valley that Dario might have said something similar to “Anthropic might end up being the only private company.”
I would love nothing more than to https://t.co/NF7MaO8wl1”
“@_sholtodouglas Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging”
“A monopolar or bipolar world of frontier models is a very dangerous world for humans.
This is why open-source is so important - we need a rich diversity of models that reflect the richness of human variation and let us all find the model that is best for each of us.
Also why so https://t.co/xxMglmQZJC”
