Who is on this subjectordered by the date of their first quote here
1 of 5 lane rests on fewer than 5 quotes and is marked thin. Offsets are days from the middle first-quote date, 25 Sep 2024 — a date, and nothing else. It is not a claim about who reached a view first.
-
Gavin Baker
40
26 Nov 2019 – 4 Aug 2026
-1765 d
-
Brad Gerstner
7
2 Feb 2024 – 15 May 2026
-236 d
-
Bill Gurley
3
thin
25 Sep 2024 – 21 Nov 2024
+0 d
-
Jensen Huang
5
13 Oct 2024 – 5 Feb 2026
+18 d
- Dylan Patel 32 12 Nov 2024 – 25 Aug 2026 +48 d
The chronologysourced and dated, oldest first
-
-
Gavin Baker
Baker argues AI revolution stems from cloud computing power and mobile-generated data, not algorithmic advances.
“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.”
-
Gavin Baker
Baker states data quantity is the single most predictive element of AI quality, not algorithms or infrastructure.
“The single most predictive element of knowledge about AI quality is the quantity of data used to train the algorithm.”
-
Gavin Baker
Baker cites research showing every 10x increase in training data doubles AI quality.
“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.”
-
-
-
Brad Gerstner
Gerstner says Microsoft, Amazon, and Meta all signaled massive inference and infrastructure investments driving NVIDIA higher.
“It's up because Amazon told us the same thing. It's up because Susan Lee from Meta just told us yesterday that they're going to make huge future investments in infrastructure, read NVIDIA,”
-
Gavin Baker
Baker predicts Tesla FSD will achieve 100x improvement quickly as compute scales to GPT-4.5 levels.
“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.”
-
Bill Gurley
Gurley explains Strawberry's logarithmic scaling means 10X to 100X inference costs for linear improvement requiring selective application.
“The implication being in order to get linear improvement, you have to do maybe 10X the amount of processing. And this is all inference.”
-
Bill Gurley
Gurley explains Strawberry requires 10X processing for linear improvement, meaning 10-100X inference costs for better solutions.
“in order to get linear improvement, you have to do maybe 10X the amount of processing. And this is all inference. So what are the implications of that?”
-
Jensen Huang
Huang says models may run 10,000 internal inferences before answering, including tree search and reinforcement learning.
“The idea that a model before it answers your answer had already done internal inference 10,000 times, it's probably not unreasonable. And it's probably done tree search, it's probably done reinforcement learning on that,”
-
Jensen Huang
Huang says models may run 10,000 internal inferences before answering, including tree search and reinforcement learning.
“The idea that a model before it answers your answer had already done internal inference 10,000 times, it's probably not unreasonable. And it's probably done tree search, it's probably done reinforcement learning on that,”
-
Brad Gerstner
Gerstner reports Jensen predicts inference will scale 100x to billion-x with 40% of NVIDIA revenue already from inference
“he said as a consequence of that, inference is going to a 100 x, thousand x, a million x, maybe even a billion x.”
-
Brad Gerstner
Gerstner says inference costs fell 90% last year and will drop another 90% in coming months.
“What Sunny has told us and other people in the field have told us that inference is going to drop by another 90% over the cost of the next period of months.”
-
Dylan Patel
Patel says o1 generates 40k sequence lengths versus 4k for standard models, requiring lower batching and higher prices.
“If you ask it to, like, generate a web scraper, in the standard one it'll just start outputting code, and it'll be maybe like a four k sequence length.”
-
Dylan Patel
Patel explains o1's thinking time creates memory bandwidth issues that prevent batching users at high levels.
“But if you batch higher, k b cache is not just a memory capacity issue, it's also a memory bandwidth issue.”
-
Bill Gurley
Gurley questions hardware implications if LLM differentiation shifts from pre-training to inference focus.
“if the world shifts to inference sooner or faster, and if the path of differentiation for the language model companies becomes more inference based than pre training based, what does that mean for the hardware systems that people are using?”
-
Brad Gerstner
Gerstner states inference time reasoning requires way more compute than scaling pretraining.
“inference time reasoning is way more compute intensive than simply scaling pretraining. Why don't you walk us through? We have a really interesting graph here about why that's the case”
-
Dylan Patel
Patel reveals o1's reasoning process sometimes switches between Chinese and English during hidden thinking phase.
“It generates tons of things. It's like it it sometimes switches between Chinese and English. Right? Like, whatever it is. It's thinking. Right?”
-
Dylan Patel
Patel calculates o1 reduces concurrent users per server to one-fourth or one-fifth versus standard models.
“E. Concurrent users I can have is a fraction of that. One fourth to one fifth the number of users can currently use the server.”
-
-
-
Gavin Baker
Baker predicts big businesses will pay a million dollars for AI to think for six weeks on key questions.
“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.”
-
Gavin Baker
Baker predicts inference compute will be the derivative winner and compute will run out in 2025.
“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.”
-
Gavin Baker
Baker predicts inference compute will be the derivative winner and compute will run out in 2025.
“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.”
-
Gavin Baker
Baker identifies three axes of AI scaling: pretraining, inference time compute, and now reasoning as the third multiplicative dimension.
“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.”
-
Dylan Patel
Patel argues export controls primarily limit AI inference deployment in China, not frontier model training capabilities.
“A large part of export controls, if they work, is just that the amount of AI that can be run-in China is going to be much lower.”
-
Dylan Patel
Patel notes OpenAI is releasing custom kernels with model weights so everyone has optimized inference day one.
“Make your own inference implementation, But OpenAI is, like, actually, like, dropping the model weights and, like, all these custom kernels for people to implement in inference.”
-
Brad Gerstner
Gerstner says Google's monthly token generation exploded 100x in a year, from 9 trillion to 980 trillion tokens.
“Today, it's 980,000,000,000,000 tokens. So from 9,000,000,000,000 to nine eighty, it's a 100 x increase in a year.”
-
Dylan Patel
Patel predicts AWS revenue growth will reaccelerate after consistent deceleration due to Anthropic and new data centers.
“AWS has been decelerating revenue. Year on year revenue has been falling consistently. And and our big call is that it's actually going to start reaccelerating.”
-
Dylan Patel
Patel confirms OpenAI runs production inference on GB200 despite reliability requiring workloads handle 64 of 72 GPUs.
“OpenAI has said they're running production inference on GV200 a couple of months ago, in fact. Right?”
-
Dylan Patel
Patel says B200 is better for training while GB200 is better for inference, reversing expected use.
“And so you've sort of Which is the exact opposite of what you would have expected. Oh, use the big thing for training and use the small thing for inference.”
-
Dylan Patel
Patel reveals NVIDIA's next generation will split inference into separate context processing and decode workloads, not training versus inference.
“NVIDIA's next generation actually has something very different. They're not saying, Hey, there's a training GPU and an inference GPU, right? Because either is fine.”
-
Dylan Patel
Patel explains NVIDIA is splitting inference into decode and prefill workloads, but provisioning for unknown future ratios is challenging.
“what sort of NVIDIA's pitching is like a split of inference into two workloads, and we'll see if they're successful. There's a lot of challenges on the infrastructure side.”
-
Dylan Patel
Patel announces InferenceMax is completely open source under Apache 2.0 license for full transparency.
“InferenceMax is aiming to change all of that in the benchmarking space. So it's completely transparent in that it's completely open source, Apache 2.0 license.”
-
Dylan Patel
Patel cites DeepSeek's open-sourced inference system requiring 140 GPUs communicating over RDMA networks for single replica.
“One example is one that DeepSeek open sourced over December of last yearJanuary, February of this year, where a single replica of inference for a single model is going to be like 140 GPUs.”
-
Dylan Patel
Patel says avoiding repeated pre-fill on 30,000-token inputs can reduce inference costs by 75% through KV cache.
“So then that 30,000 tokens, that pre fill, that 30,000 tokens still cost a lot more than that output.”
-
-
-
Brad Gerstner
Gerstner says the world will be supply constrained on inference compute indefinitely, citing NVIDIA's Grok acquisition.
“The world is going to be supply constrained on inference compute as far as the eye can see. It was part of the inspiration. We were also investors in Grok.”
-
Dylan Patel
Patel explains users will pay 10x more for 10x faster inference completion, justifying Cerberus economics.
“for a lot of people, I'm fine to spend 10x the price on something that completes 10x faster. So Cerberus sort of just makes a ton of sense there.”
-
Jensen Huang
Huang says AI creates abundance of intelligence by orders of magnitude, compressing year-long work into hours or real-time.
“AI reduces the cost of intelligence or create the abundance of intelligence by orders of magnitude.”
-
Jensen Huang
Huang describes PhysicsNEMO as a physics-aware AI framework combining principled simulation with AI prediction.
“PhysicsNEMO is essentially a physics aware AI model simulation system and AI framework that allows us to create these AI models that are either trained by principled simulators or work alongside principled simulators”
-
Jensen Huang
Huang claims PhysicsNEMO can predict physics simulations ten thousand times faster than traditional methods.
“so it's grounded in the laws of physics, but able to predict 10,000 times faster.”
-
Dylan Patel
Patel claims code agent revenue grew from a couple billion to over $10 billion in a very short time.
“code agent revenue has gone from a couple billion to north of 10,000,000,000 in like a very short amount of time. And these, the horizon of these has also increased dramatically,”
-
Dylan Patel
Patel identifies disaggregated PD and wide EP as networking techniques where performance matters significantly.
“When someone has really good networking with these leading edge techniques to run a model such as KIMI or DeepSeq called disaggregated PD or wide EP, these are two different techniques, network performance is a humongous factor.”
-
Dylan Patel
Patel claims network performance drives 5x to 10x difference in inference for models using disaggregated PD or wide EP techniques.
“And so if you look into inference x, it turns out network performance drives a not just a, hey, 30%. It's actually multiple, like, x's.”
-
Dylan Patel
Patel notes software dependencies change daily or multiple times weekly across the entire stack.
“And furthermore, with software changing literally multiple times a week, right, PyTorch has nightlies, VLM has nightlies, Asteeling has nightlies, CUDA drivers update constantly. You just go You go through the whole list.”
-
Dylan Patel
Patel explains AI software stacks update multiple times per week making performance measurement a moving target.
“with software changing literally multiple times a week, right, PyTorch has nightlies, VLM has nightlies, Asteeling has nightlies, CUDA drivers update constantly. You just go You go through the whole list.”
-
Dylan Patel
Patel distinguishes OpenAI hiding reasoning chains from Anthropic showing them in code generation models.
“In the case of OpenAI, they don't show you the whole reasoning chain. In the case of Anthropic, they do.”
-
Dylan Patel
Patel predicts reasoning chains will close as AI response times extend from immediate to hours or days.
“And then as the horizon of AI models gets longer, right, rather than a question answer immediately, the question answer becomes ten minutes, hours, days, the value of all the reasoning is gone.”
-
Gavin Baker
Baker says AI models shifting to usage-based pricing with overage reveals no ceiling on spending yet.
“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.”
-
Gavin Baker
Baker emphasizes only 0.1% of the world uses AI models properly yet there's massive shortage despite trillions spent.
“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.”
-
Gavin Baker
Baker predicts Trainium will be to 2026 what TPUs were to 2025 as Trainium 3 ramps.
“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.”
-
Gavin Baker
Baker predicts Trainium will dominate 2026 like TPUs did in 2025, with Trainium 3 ramping in second half.
“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.”
-
Gavin Baker
Baker argues Trainium is most underestimated because frontier mixture-of-expert models require switched scale-up networks for inference.
“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.”
-
Gavin Baker
Baker states only NVIDIA and Amazon Trainium have functioning switched scale-up networks for inference today.
“And the only two functioning switched scale up networks in the world today are the ones that power NVIDIA GPUs and Amazon's Trainiums.”
-
Gavin Baker
Baker reveals Atreides could have invested over $50 million in CoreWeave at $1.1 billion valuation but was conflicted out.
“I could've Atreides could've invested over $50,000,000 in the round at 1,100,000,000, And I was conflicted out by Crusoe,”
-
Brad Gerstner
Gerstner argues that token production and consumption is the fundamental basis of all AI intelligence.
“There is no intelligence. There's no consumer chat GPT. There's no enterprise intelligence. There's no clawed code without the production of tokens.”
-
Gavin Baker
Baker estimates Anthropic would be doing $100-150B ARR if not compute-constrained, versus current $50B.
“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.”
-
Gavin Baker
Baker believes Anthropic is likely already generating cash or will start this year.
“I think Anthropic probably starts generating cash this year if they are not already generating cash, which I think is probably the case.”
-
Gavin Baker
Baker argues America will consume all available compute, reducing edge AI bear case concerns.
“And I just think the same is true of compute. It's why I'm probably less worried about like an edge AI bear case than I was.”
-
Gavin Baker
Baker argues America will consume all available compute, making him less worried about edge AI bear cases.
“It's why I'm probably less worried about like an edge AI bear case than I was. We're going to consume as much compute as we can.”
-
Gavin Baker
Baker explains harness engineering matters significantly and harnesses are increasingly co-developed with models.
“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.”
-
Gavin Baker
Baker says understanding frontier AI now requires enterprise usage-based plans, not $250 monthly subscriptions which are rate-limited.
“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.”
-
Gavin Baker
Baker says understanding frontier AI now requires enterprise usage-based plans, not consumer subscriptions, due to rate limiting.
“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.”
-
Gavin Baker
Baker says AI shifting from flat pricing to usage-based is extremely bullish as people consume more AI.
“AI is just shifting from all you can eat to pay by the drink. Then it turns out people really like to talk to their friends long distance.”
-
Gavin Baker
Baker predicts OpenAI and Anthropic will exceed $200B ARR this year due to shift to usage-based pricing.
“So I think the shift to usage based pricing is probably why you will see OpenAI and Anthropic exceed well over $200,000,000,000 in ARR this year.”
-
Gavin Baker
Baker predicts OpenAI and Anthropic will exceed $200B ARR this year due to usage-based pricing shift.
“I think the shift to usage based pricing is probably why you will see OpenAI and Anthropic exceed well over $200,000,000,000 in ARR this year.”
-
Gavin Baker
Baker explains continual learning as models dynamically updating weights in real-time, unlike current reinforcement learning approaches.
“Continual learning is a model that dynamically adjusts its weights or adjusts in some way in real time.”
-
Gavin Baker
Baker predicts GPUs will have 10-15 year useful lives due to prefill-inference disaggregation, extending older chips' value.
“The disaggregation of inference means that I think these GPUs are going to have ten or fifteen year lives.”
-
Gavin Baker
Baker predicts GPUs will have 10-15 year useful lives due to prefill-inference disaggregation, extending older chips' value.
“The disaggregation of inference means that I think these GPUs are going to have ten or fifteen year lives.”
-
Gavin Baker
Baker argues GPU useful lives will extend to 10-15 years due to inference disaggregation, contradicting AI skeptics.
“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.”
-
Gavin Baker
Baker argues GPU useful lives will extend to 10-15 years due to inference disaggregation, contradicting AI skeptics.
“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.”
-
Gavin Baker
Baker argues GPU useful lives will extend to 10-15 years due to prefill/decode disaggregation, not 1-2 years.
“The useful life of GPU is only a year or two. The useful life of CPU is only four years because the rapid technological change.”
-
Dylan Patel
Patel explains SemiAnalysis publishes public research on GPU benchmarking to build scale beyond institutional client work.
“Let's figure out how to figure it out and just post it publicly, and that begets more and more scale. So we've done this with a lot of GPU benchmarking and testing”
-
Dylan Patel
Patel notes AI inference libraries update twice weekly, causing constant performance changes.
“In fact, the update cycle for most of these libraries is twice a week, so you basically have the software updating all the time and therefore performance changing.”
-
Dylan Patel
Patel predicts OpenAI and Anthropic will have over 100 gigawatts combined by 2030, terawatts by 2040.
“I think by 2030, just OpenAI and Anthropic will have over 100 gigawatts combined, and then you'll add Meta and Google and so on and so forth.”
-
Dylan Patel
Patel predicts OpenAI and Anthropic alone will deploy over 100 gigawatts of compute by 2030.
“by 2030, just OpenAI and Anthropic will have over 100 gigawatts combined, and then you'll add Meta and Google and so on and so forth.”
-
Gavin Baker
Baker explains AI recomputes answers probabilistically each time, enabling superhuman capabilities but requiring extreme computational expense.
“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.”
-
Gavin Baker
Baker calculates rendering Monopoly Go with VEO3 would cost over 100x the game's revenue.
“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.”
-
Dylan Patel
Patel describes agentic workflows using 30,000-100,000 input tokens but generating only 1,000 output tokens.
“And and that makes, you know, initially, that'd be like, okay, well now I need a ton ton of compute to calculate all the prefilled tokens.”
-
Dylan Patel
Patel explains KV cache storage and reuse enables massive cost decreases in inference.
“you calculate that once, you store it off in memory, whether it be system memory or storage, and then you pull it back in when you run the turn.”
-
Dylan Patel
Patel reports cache hit rates above 95% for many agentic workflows in production.
“we're seeing cache hit rates above 95% for many AgenTeq workflows, which which means the cost for a cache hit is it's not free,”
-
Gavin Baker
Baker argues contracted compute trades at massive discount to spot, repricing will accelerate cash flows and answer ROI questions.
“And so, essentially, you have the contracted base of installed compute trading at a massive discount to the current spot market.”
-
Gavin Baker
Baker reports GPU rental prices doubled from mid-$2 to nearly $4 per hour over seven months for identical clusters.
“And they had rented a cluster of several thousand black wells, and we'll just call it somewhere in the mid $2 per GPU hour.”
-
Gavin Baker
Baker reports inference cloud expects to pay 100% more for Blackwells when contracts expire, showing hyperscalers are underearning.
“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.”
-
Gavin Baker
Baker cites analysis showing compute margins, quantity, and inference margins all rising simultaneously, driving lab acceleration.
“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,”
-
Gavin Baker
Baker cites analysis showing compute margins, quantity, and inference margins all rising simultaneously, driving lab acceleration.
“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,”
-
Gavin Baker
Baker estimates SpaceX monetizes compute at $50B per gigawatt versus $73B consensus, with Grok and Cursor hitting $10B ARR quickly.
“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.”
-
Dylan Patel
Patel argues that if model progress pauses while compute supply grows, demand growth will slow and prices will collapse.
“If model progress at the labs pause, then more compute comes online. It has to slow down. Right? Sort of right now we have supply demand, right?”
-
Dylan Patel
Patel says anyone can profitably run inference by renting GB300 racks and deploying open models.
“Go download the Kimi weights. Go download VLM or SGLANG. Set it up. You know, Codecs and Fable can actually help you do this.”
-
Dylan Patel
Patel argues labs will allocate less compute to inference over time, contrary to consensus belief.
“So, this is sort of what I personally believe that the labs are going to allocate less and less compute to inference over time, I think is very non consensus, right?”
-
Where they overlapsame programme, same subject
- Brad Gerstner and Dylan Patel — BG2 Pod, 1 appearance together
- BG2 Pod also carries Bill Gurley and Jensen Huang on this subject, at different times.