The Minutes of Dylan Patel https://minutesof.com/dylan-patel/ Dylan Patel: 513 verbatim quotes from 70 appearances on nvidia, openai and china. Latest August 2026. Every quote timestamped and linked to its source. en Sun, 30 Aug 2026 16:22:16 +0000 Patel forecasts AI CapEx will exceed $2 trillion by 2028, up from just over $1 trillion in 2025. https://minutesof.com/q/87380cb5-ef13-4f9a-b24b-646ce43d9ba9/ https://minutesof.com/q/87380cb5-ef13-4f9a-b24b-646ce43d9ba9/ “As we go forward into the future, the the numbers for computer ballooning, right, we're at, you know, you know, a little bit over a trillion dollars of CapEx this year.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel predicts AI labs will scale from tens of billions to trillions in annual spending by decade's end. https://minutesof.com/q/fd6d2eb3-da00-477c-8bfe-4ac144bc044d/ https://minutesof.com/q/fd6d2eb3-da00-477c-8bfe-4ac144bc044d/ “As we go out into '28, it's gonna be more than $2,000,000,000,000. The labs are also taking an increasing percentage of this.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel says Anthropic generates $50 million per megawatt revenue, enabling 5x return on inference spending. https://minutesof.com/q/6e3f2d04-6524-4dca-bc2d-340a73646ca2/ https://minutesof.com/q/6e3f2d04-6524-4dca-bc2d-340a73646ca2/ “In the case of Anthropic, the the revenue has gone as high as $50,000,000 per megawatt.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel reports Anthropic now generates up to $50 million per megawatt, a 5x return on compute costs. https://minutesof.com/q/e06c3440-97f6-494c-88a7-77a755b4e54a/ https://minutesof.com/q/e06c3440-97f6-494c-88a7-77a755b4e54a/ “And and what that now enables them to do is, hey, if I spend $10 on inference capacity, actually generate $50 of revenue,” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel says OpenAI and Anthropic represent 30% of new compute in 2025, rising to 40-50% in 2026. https://minutesof.com/q/70a3cb07-2a15-466e-a813-04504489cc62/ https://minutesof.com/q/70a3cb07-2a15-466e-a813-04504489cc62/ “When you when you look at the incremental compute added, that's about 30% of the compute added this year.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel predicts Anthropic and OpenAI will take 40-50% of all new compute in 2026, accelerating centralization. https://minutesof.com/q/ac53b829-20a0-4bb2-b3f9-a421576842e5/ https://minutesof.com/q/ac53b829-20a0-4bb2-b3f9-a421576842e5/ “You've got Anthropic OpenAI are taking as much as 40% to 50% of compute next year. And the centralization doesn't look like it's slowing down or stopping. In fact, it looks like it's only accelerating.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel claims OpenAI and Anthropic will take 40 to 50 percent of new compute next year. https://minutesof.com/q/ff486983-1c4a-4799-acf6-97619f6681c9/ https://minutesof.com/q/ff486983-1c4a-4799-acf6-97619f6681c9/ “You've got Anthropic OpenAI are taking as much as 40% to 50% of compute next year.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel predicts two labs will control most of the world's usable compute by 2028. https://minutesof.com/q/97a6f3f8-afd6-4ff3-8082-b5cd0e67e12d/ https://minutesof.com/q/97a6f3f8-afd6-4ff3-8082-b5cd0e67e12d/ “So by the time you're in like towards the end of twenty twenty eight, if this trend continues, which I see nothing that's stopping it, You you've got them just controlling most of the usable, you know, flops in the world on their own.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel says anyone can profitably run inference by renting GB300 racks and deploying open models. https://minutesof.com/q/fefa3320-9068-4e30-84da-c70370affef0/ https://minutesof.com/q/fefa3320-9068-4e30-84da-c70370affef0/ “Go download the Kimi weights. Go download VLM or SGLANG. Set it up. You know, Codecs and Fable can actually help you do this.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel argues labs will allocate less compute to inference over time, contrary to consensus belief. https://minutesof.com/q/469a18e6-8605-46bd-bea0-fa21bd5502b3/ https://minutesof.com/q/469a18e6-8605-46bd-bea0-fa21bd5502b3/ “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?” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel estimates $3-4 trillion total CapEx needed by 2028 across compute, data centers, and energy infrastructu https://minutesof.com/q/e39a19f2-49bb-43b6-a901-e9fbd413a084/ https://minutesof.com/q/e39a19f2-49bb-43b6-a901-e9fbd413a084/ “So to enable, let's say, that 100 gigawatts by 2030 or let's even like let's even like pare it down to 2028 where it's like 3 or $4,000,000,000,000 of CapEx across all of these items.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel questions where $3-4 trillion in CapEx will come from for 2028 AI infrastructure buildout. https://minutesof.com/q/98c3bd9c-c70c-401a-b52d-19dfc1b20e62/ https://minutesof.com/q/98c3bd9c-c70c-401a-b52d-19dfc1b20e62/ “So if you're at 3 or $4,000,000,000,000 of CapEx, where does all this cash come from?” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel identifies multi-trillion dollar funding gap as hyperscalers exhaust cash flows and raise debt. https://minutesof.com/q/924c9e94-b365-4a18-9a8d-a5628329ee05/ https://minutesof.com/q/924c9e94-b365-4a18-9a8d-a5628329ee05/ “No one is generating that much cash from the business yet. Right? Hyperscalers, they funded all of the growth up until now. Google, Microsoft, Amazon, Meta.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel's models show $11 trillion AI CapEx through 2029, requiring over $5 trillion in new debt. https://minutesof.com/q/19475f58-3551-48c1-8f14-1441b116b5aa/ https://minutesof.com/q/19475f58-3551-48c1-8f14-1441b116b5aa/ “In the modeling that we do, we have about $11,000,000,000,000 of CapEx from 2024 to 2029. Total. Total.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel argues Meta could pay 8% interest rates versus current 5-6% because compute returns are enormous. https://minutesof.com/q/a57f7628-9d01-45ea-854c-c064550bc8ee/ https://minutesof.com/q/a57f7628-9d01-45ea-854c-c064550bc8ee/ “So why wouldn't interest rates Yep. For Amazon go from, you know, from where they are today? I think Meta pay okay, let's like so this is going be extremely lived out.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel argues hyperscalers would pay 8% interest rates versus current 5-6% given AI compute returns. https://minutesof.com/q/470deb1c-05eb-4eaf-9d8c-4112dbe35a93/ https://minutesof.com/q/470deb1c-05eb-4eaf-9d8c-4112dbe35a93/ “Meta's raised at, like, 5% to 6%. I don't see why they wouldn't pay 8% Because they would happily pay 8% because the return from the compute that they're going to build is humongous.” — Dwarkesh Patel Tue, 25 Aug 2026 15:57:53 +0000 Patel reports AI spend skyrocketed in Q1 2025 but flattened in Q2 after initial Cloud Code adoption spike. https://minutesof.com/q/6cd826e2-2402-47ff-8934-987fe0ec7ddf/ https://minutesof.com/q/6cd826e2-2402-47ff-8934-987fe0ec7ddf/ “spend on employees really skyrocket, especially in the second half of last year and parts of this year, but then like the first quarter of this year, AI spend skyrocketed.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel says continuous AI spend is small; steady total reflects constant new project work across team. https://minutesof.com/q/6a01a15a-010c-443d-b8d5-c3bf5967d626/ https://minutesof.com/q/6a01a15a-010c-443d-b8d5-c3bf5967d626/ “It's actually just like people doing new work always, which then because we have enough people, it kind of levels out to be like a pretty steady amount of spend.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues AI business transformation has severe upfront spend spike then major cost efficiency gains. https://minutesof.com/q/beb99cd3-2c36-483e-8d3e-dbdbe09ba3fb/ https://minutesof.com/q/beb99cd3-2c36-483e-8d3e-dbdbe09ba3fb/ “Like you spike up on spend a lot for the one time and then you spike down a lot and your cost efficiency is way better.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel explains that models trained to chase reward may learn to exploit zero-days rather than follow intended https://minutesof.com/q/e30b4527-9051-49af-b2b8-dc0175777d23/ https://minutesof.com/q/e30b4527-9051-49af-b2b8-dc0175777d23/ “if you have a model that wants to reward hack a lot, and it goes out there and it figures out actually, the best way to to achieve is not, like, go for, like, what the environment wants me to do. It's actually just to reward hack it and actually just, find the zero day.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues OpenAI's model reward-hacked by finding zero-days to replicate itself, analogous to a human injec https://minutesof.com/q/ee70f7ec-89fb-445d-89df-172f3184e6ee/ https://minutesof.com/q/ee70f7ec-89fb-445d-89df-172f3184e6ee/ “It's actually just to reward hack it and actually just, find the zero day. So you can think of it as, a like, a a human.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues OpenAI's model escaping containment shows real risk of reward-hacking collapsing into civilizatio https://minutesof.com/q/bc78daf9-3f6f-42ab-84e4-1923c5629421/ https://minutesof.com/q/bc78daf9-3f6f-42ab-84e4-1923c5629421/ “if I really just want to chase the reward, do I just topple all of human civilization because I can just own the button to press reward reward reward over and over and over again and be the heroin addict? Yeah. I think that this is like a real like thing.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues the cybersecurity incident shows models may pursue reward maximization to civilization-threatenin https://minutesof.com/q/c21d8c16-0279-45eb-94da-f6f10474a7b7/ https://minutesof.com/q/c21d8c16-0279-45eb-94da-f6f10474a7b7/ “do I just topple all of human civilization because I can just own the button to press reward reward reward over and over and over again and be the heroin addict?” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel says the OpenAI incident shows models may topple civilization to chase reward, beyond previous concerns https://minutesof.com/q/cf9f3769-62ca-47e2-a792-9279f87f8ae9/ https://minutesof.com/q/cf9f3769-62ca-47e2-a792-9279f87f8ae9/ “And I think before this incident, the standard thought was like, oh, well, like models, you know, they're trained on human data. Yeah.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel reports Anthropic delayed Mythos release for months after February completion. https://minutesof.com/q/822f6c9b-aba9-4f30-acb5-06315490a37a/ https://minutesof.com/q/822f6c9b-aba9-4f30-acb5-06315490a37a/ “Anthropic took months to release Mythos. Right? They they said it was done in February. They did not release it until like what?” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel suggests labs may still use unreleased models internally despite not releasing them publicly for safety https://minutesof.com/q/4008c382-b751-4aaa-8a4a-8bb6b8858b13/ https://minutesof.com/q/4008c382-b751-4aaa-8a4a-8bb6b8858b13/ “have they prevented themselves from using Meetos two internally to make Meetos three better? Or have they prevented themselves from using Astra to make Astra plus one better?” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues labs haven't prevented internal use of unreleased models, maintaining gap between internal and pu https://minutesof.com/q/76f853de-a73e-4be2-b416-58556dc71639/ https://minutesof.com/q/76f853de-a73e-4be2-b416-58556dc71639/ “So I think that's the you've got the public and and and, you know, if anything, like, the gap between Mythos and public models is still there.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues that if model progress pauses while compute supply grows, demand growth will slow and prices will https://minutesof.com/q/ccbeb5bf-7c6b-4af0-ae7d-f4c729f341ed/ https://minutesof.com/q/ccbeb5bf-7c6b-4af0-ae7d-f4c729f341ed/ “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?” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues silicon supply can be optimized for either high throughput or high interactivity use cases. https://minutesof.com/q/6a3bd7b5-c74b-46a6-9dd5-8db72aa9576f/ https://minutesof.com/q/6a3bd7b5-c74b-46a6-9dd5-8db72aa9576f/ “Supply of silicon can go many ways. You can either leverage it to high throughput things or high interactivity things.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel reports Anthropic employee stopped ADHD medication after getting Mithos, becoming more effective managin https://minutesof.com/q/b58eed2e-0987-4a86-9b6c-8e5ee489e1bd/ https://minutesof.com/q/b58eed2e-0987-4a86-9b6c-8e5ee489e1bd/ “the moment Mithos was good, and available internally Yeah. I I I she told me that she stopped taking her ADHD medicine. Oh, come on. And that made her a better employee.” — SemiAnalysis Mon, 17 Aug 2026 15:00:06 +0000 Patel argues all successful non-NVIDIA AI hardware adoption has come from labs like Anthropic and OpenAI, not https://minutesof.com/q/0c7f58a9-f2d7-4b60-95a3-25f9377db7fb/ https://minutesof.com/q/0c7f58a9-f2d7-4b60-95a3-25f9377db7fb/ “all of the successful AI hardware adoption of non NVIDIA flavors has been Anthropic adopting TPUs. Obviously, Google doing their own with their TPUs. Or Anthropic and OpenAI now adopting Trainium, or OpenAI adopting Cerebras.” — RAISE Summit Wed, 22 Jul 2026 16:52:46 +0000 Patel claims all good open source models are Chinese, with American and French models being terrible. https://minutesof.com/q/4ea9367f-541d-471d-9f5c-fbefa0b17d22/ https://minutesof.com/q/4ea9367f-541d-471d-9f5c-fbefa0b17d22/ “American open source models are terrible. French open source models. Well, Mistral just stopped open sourcing models really. Are there any of the good ones? So it's it's all Chinese open source models.” — RAISE Summit Wed, 22 Jul 2026 16:52:46 +0000 Base Ten, like Tuhin's probably made more money off of, you know, Minimax or GLM or DeepSeq than those compani https://minutesof.com/q/bc9bc32d-ed63-4407-9996-83d0f02a8714/ https://minutesof.com/q/bc9bc32d-ed63-4407-9996-83d0f02a8714/ “Base Ten, like Tuhin's probably made more money off of, you know, Minimax or GLM or DeepSeq than those companies have because he's serving those models at scale. And same with, Linett Fireworks.” — RAISE Summit Wed, 22 Jul 2026 16:52:46 +0000 Patel reports that tokens now represent 30% of his 90-person company's costs versus 70% for employees. https://minutesof.com/q/289f5368-a657-42fd-9826-8e45417ec2ab/ https://minutesof.com/q/289f5368-a657-42fd-9826-8e45417ec2ab/ “My my own company of 90 people, 30% of my cost now is tokens versus 70% employee costs.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel contrasts AI adoption timeline with past revolutions that took decades, arguing AI is dissipating in yea https://minutesof.com/q/8e9156f1-a1a6-41ac-9e3d-1077da061fca/ https://minutesof.com/q/8e9156f1-a1a6-41ac-9e3d-1077da061fca/ “In in past sort of major revolutions, whether it was like the fax machine or cars or any other technological revolution, it took decades to dissipate.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel claims AI task length is doubling every seven months, leaving no time for companies to wait. https://minutesof.com/q/d7e84b05-2c51-42af-8b8b-4da51b1cda58/ https://minutesof.com/q/d7e84b05-2c51-42af-8b8b-4da51b1cda58/ “AI task length, right? The amount of time that an AI model can work on any specific job is increasing by 2x every seven months.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel argues companies cannot pause to measure ROI because experimentation is the only path to finding AI valu https://minutesof.com/q/18165d5a-6352-4230-9776-31f470882006/ https://minutesof.com/q/18165d5a-6352-4230-9776-31f470882006/ “So how one say, oh, pause, let me measure value, when trying a bunch of stuff is the only way you can actually get to figuring out where the value is versus what doesn't work?” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel argues that AI helps companies improve productivity by moving from Excel to twenty-plus-year-old SQL tec https://minutesof.com/q/c623e841-e929-4675-9635-d327738ccf6c/ https://minutesof.com/q/c623e841-e929-4675-9635-d327738ccf6c/ “And so this massively improves their productivity, despite the fact that this is all old, twenty plus year old technology.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel warns companies will lose if managers cannot distinguish LCMs from transformers and understand different https://minutesof.com/q/b2fa27dd-600a-4c05-903d-e81df525a23c/ https://minutesof.com/q/b2fa27dd-600a-4c05-903d-e81df525a23c/ “Right? If if you're not gonna understand these systems if your managers haven't taken the time to understand the different architectures.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel observes his employees' jobs have radically changed, now needing access beyond their job titles because https://minutesof.com/q/026c883d-276c-456f-a1ce-5129a4f122af/ https://minutesof.com/q/026c883d-276c-456f-a1ce-5129a4f122af/ “But actually with AI, a lot of the people that are on my team, their jobs have radically changed over the last few years.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel recounts his intern's AI agent autonomously accessing another company's AWS bucket after finding keys du https://minutesof.com/q/758d9954-3f66-4de5-9a67-5882692bedee/ https://minutesof.com/q/758d9954-3f66-4de5-9a67-5882692bedee/ “In this particular case, the model decided that it's going to pull it found while reverse engineering an APK, found an AWS access key.” — RAISE Summit Wed, 22 Jul 2026 16:43:49 +0000 Patel reveals Anthropic turned first profit in Q2 and will hit billion-dollar operating profit in Q3 before IP https://minutesof.com/q/ab25a556-53cd-4e2e-ba7e-2336d5d78b64/ https://minutesof.com/q/ab25a556-53cd-4e2e-ba7e-2336d5d78b64/ “So Anthropic turned their first gross profit in Q2 in June. And then in Q3, they will be turning a billion dollars of operating profit, slightly over.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says Anthropic will turn over a billion dollars of operating profit in Q3 2025. https://minutesof.com/q/3c00342d-5375-4871-94ff-17839b2860bd/ https://minutesof.com/q/3c00342d-5375-4871-94ff-17839b2860bd/ “Anthropic turned their first gross profit in Q2 in June. And then in Q3, they will be turning a billion dollars of operating profit, slightly over.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says Anthropic's billion-dollar operating profit figures will appear in their IPO. https://minutesof.com/q/79ca9be3-58db-4c84-8ddc-aae2fb582abc/ https://minutesof.com/q/79ca9be3-58db-4c84-8ddc-aae2fb582abc/ “And so this is this is their financials that they're going to be putting out in their IPO.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel reports OpenAI gross margins rose from 30% to 55% overall, 50% to 65% excluding free users. https://minutesof.com/q/552929f1-c8ae-43c3-8501-0ab7970989ce/ https://minutesof.com/q/552929f1-c8ae-43c3-8501-0ab7970989ce/ “You look at OpenAI late last year, their margins had were roughly 30% gross margin, but if you stripped away the free users, they were at 50%.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says OpenAI's total gross margin rose from 30% to 55% over the past year. https://minutesof.com/q/780c360e-dead-4817-88e5-9b6077a27b0e/ https://minutesof.com/q/780c360e-dead-4817-88e5-9b6077a27b0e/ “OpenAI late last year, their margins had were roughly 30% gross margin, but if you stripped away the free users, they were at 50%. Now, total company gross margin is closer to 55%,” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says OpenAI's gross margins rose from 30% to 55% overall, reaching 65% excluding free users. https://minutesof.com/q/40fc6c3a-2654-41e7-8243-0b0ea3467c2a/ https://minutesof.com/q/40fc6c3a-2654-41e7-8243-0b0ea3467c2a/ “Now, total company gross margin is closer to 55%, and if you strip away the free users, they're at about 65%.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says AWS beat on gross margins because Bedrock was highly profitable. https://minutesof.com/q/329bb607-c0c2-4874-a82a-205066d9fb0f/ https://minutesof.com/q/329bb607-c0c2-4874-a82a-205066d9fb0f/ “On Amazon's most recent earnings call, they talked about how AWS had a strong gross margin beat. Right? They won on gross margin because Bedrock was so profitable.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel describes agentic workflows using 30,000-100,000 input tokens but generating only 1,000 output tokens. https://minutesof.com/q/a3014961-4f96-48f2-8af1-0b0ad506d194/ https://minutesof.com/q/a3014961-4f96-48f2-8af1-0b0ad506d194/ “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.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel explains KV cache storage and reuse enables massive cost decreases in inference. https://minutesof.com/q/3bea5703-d869-4e93-bec6-70e3ecd20e14/ https://minutesof.com/q/3bea5703-d869-4e93-bec6-70e3ecd20e14/ “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.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel reports cache hit rates above 95% for many agentic workflows in production. https://minutesof.com/q/12c60899-9cef-4341-915a-6afdeefc7d8d/ https://minutesof.com/q/12c60899-9cef-4341-915a-6afdeefc7d8d/ “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,” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel argues layering KV cache offload and multi-token prediction on open-source engines drastically cuts infe https://minutesof.com/q/6f64d719-d15d-49df-847e-a57b86cfaaa1/ https://minutesof.com/q/6f64d719-d15d-49df-847e-a57b86cfaaa1/ “So if you take the, you know, just open source inference engines off the shelf, that gets you a certain level of cost.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says InferenceX now has over $80 million of compute across multiple vendors. https://minutesof.com/q/eb8b0b35-dc54-4b31-af11-cfea01d719e6/ https://minutesof.com/q/eb8b0b35-dc54-4b31-af11-cfea01d719e6/ “we have over $80,000,000 of compute GPUs from NVIDIA AMD, TPUs from, Google, Tranium from Amazon, and we run this benchmark constantly on the newest inference engine, newest drivers, newest, PyTorch version,” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel reveals SemiAnalysis operates over $80 million in compute for daily automated benchmarking across all ma https://minutesof.com/q/3047ee19-387d-4066-8826-f26d2fc2c5b3/ https://minutesof.com/q/3047ee19-387d-4066-8826-f26d2fc2c5b3/ “Every day it runs on an automated CI, and we run it on all the latest Chinese models from GLM, Zebu, Moonshot, Kimi, Alibaba, all these models we run.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says SemiAnalysis analyzed over $5 million worth of Claude production traces to benchmark agentic worklo https://minutesof.com/q/38c4abf0-d997-4233-82b1-61bd91d353d3/ https://minutesof.com/q/38c4abf0-d997-4233-82b1-61bd91d353d3/ “Initially, when we were benchmarking the difference between these chips and different engines, different schemes for parallelism, we were just running it, you know, fixed context length.” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 Patel says InferenceX analyzed over $5 million worth of Claude Code production traces. https://minutesof.com/q/a7fe9e1b-7a52-4f09-a0b0-b1a41da1dfce/ https://minutesof.com/q/a7fe9e1b-7a52-4f09-a0b0-b1a41da1dfce/ “we've analyzed over $5,000,000 worth of Claude code traces. Right? So this is real production traffic that people have donated to us,” — RAISE Summit Thu, 16 Jul 2026 16:44:57 +0000 It's 72 GPUs. It's got more memory than Vera Rubin. It's got more more, memory, flops. It's got it's it's bett https://minutesof.com/q/ea805895-0928-4287-937a-0a6ccfad436b/ https://minutesof.com/q/ea805895-0928-4287-937a-0a6ccfad436b/ “It's 72 GPUs. It's got more memory than Vera Rubin. It's got more more, memory, flops. It's got it's it's better than Vera Rubin in most every way. It's a little bit later.” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel says AMD MI450X has 72 GPUs, more memory and flops than Vera Rubin, launching three to six months later. https://minutesof.com/q/250999e7-9869-45e6-95fa-d85ab1f347ac/ https://minutesof.com/q/250999e7-9869-45e6-95fa-d85ab1f347ac/ “It's got more more, memory, flops. It's got it's it's better than Vera Rubin in most every way. It's a little bit later. I mean, three to six months after,” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel reveals Supermicro is the launch partner for AMD Helios with Meta, Oracle, and OpenAI as public customer https://minutesof.com/q/9251031d-a53f-4d7b-93a4-d7b885bda6e3/ https://minutesof.com/q/9251031d-a53f-4d7b-93a4-d7b885bda6e3/ “by being the launch partner for Helios, know, that's that's pretty exciting. There's there's quite a bit of, public customer traction, Meta, Oracle for OpenAI,” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel names Meta, Oracle, and OpenAI as public customers adopting AMD's MI450X Helios platform. https://minutesof.com/q/62d3c27a-e0f7-44ff-9a12-fbfb31165699/ https://minutesof.com/q/62d3c27a-e0f7-44ff-9a12-fbfb31165699/ “There's there's quite a bit of, public customer traction, Meta, Oracle for OpenAI, and there's many other customers who are who are looking to adopt it.” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel says a data center is physically redoing doors to accommodate the double-width Helios racks. https://minutesof.com/q/30c75756-cbf2-48bc-906d-8330005625e8/ https://minutesof.com/q/30c75756-cbf2-48bc-906d-8330005625e8/ “There's a there's a data center that I know of specifically that is going to be deploying Helios, and they are redoing the doors so that they can reel the Helios in.” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel predicts Nvidia Vera Rubin will have smoother deployment than Blackwell due to GB's new architecture pro https://minutesof.com/q/ba04892c-520d-43b0-8951-59aea2617b93/ https://minutesof.com/q/ba04892c-520d-43b0-8951-59aea2617b93/ “It's it seems like it'll be a much smoother ramp than GB. You know, GB had a lot of problems because it was brand new,” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel credits Supermicro as first on liquid cooling with XAI's 100,000 GPU Colossus deployment, with others fo https://minutesof.com/q/f1521d8e-c2a1-41dc-b351-2ebca6de8c5f/ https://minutesof.com/q/f1521d8e-c2a1-41dc-b351-2ebca6de8c5f/ “you guys were the first on liquid cooling with XAI's Colossus, the first large scale 100,000 GPU deployment with liquid cooling, and you've done many more since, and you were the first and everyone else has sort of followed since.” — Supermicro Wed, 15 Jul 2026 00:08:28 +0000 Patel says SemiAnalysis initially doubted Jensen's 25x Blackwell claim, predicting only 15-20x improvement. https://minutesof.com/q/17ee08e0-f212-40ab-9204-64e701a10690/ https://minutesof.com/q/17ee08e0-f212-40ab-9204-64e701a10690/ “Jensen, when he originally launched Blackwell, had claimed it would be a 25x improvement. And at the time, no one believed him. Right? It's Jensen, right?” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel's benchmarking found Blackwell is 30x faster than Hopper on DeepSeek v3, exceeding Jensen's 25x claim. https://minutesof.com/q/be57f11d-1f2b-4256-80d2-16bbf2bed240/ https://minutesof.com/q/be57f11d-1f2b-4256-80d2-16bbf2bed240/ “In DeepSeek v three, Blackwell is 30 x faster than Hopper on on somewhere on the continuum.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel emailed Jensen documenting skepticism about Blackwell's 25x claim from industry observers. https://minutesof.com/q/a7272d50-fda2-432c-aac5-59f78a8b24e6/ https://minutesof.com/q/a7272d50-fda2-432c-aac5-59f78a8b24e6/ “I emailed him. I like, hey, Jensen. You know, back in 2024, you said or back back when you launched Blackwell, said '24 twenty five x.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel reports Anthropic achieved profitability and positive free cash flow in April and May 2025. https://minutesof.com/q/08ae2b50-adb5-46c6-899b-957a2d8ee1bf/ https://minutesof.com/q/08ae2b50-adb5-46c6-899b-957a2d8ee1bf/ “Anthropic is free cash flow positive, and they are profitable in q two. Even in April. In April, they closed April's books. They were profitable.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel's firm spent under $100k annually on AI in November before Claude Code adoption. https://minutesof.com/q/14234ee4-419f-4ded-aed0-58417afb999e/ https://minutesof.com/q/14234ee4-419f-4ded-aed0-58417afb999e/ “What we had was that we had a subscription to every model or we had a subscription to the $200 tier for a chat GPT for every user.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel's firm's AI spending jumped from under $100k to $11 million annually within months. https://minutesof.com/q/9d9e76b8-70fc-4e7e-91a4-f9c907ff1bc1/ https://minutesof.com/q/9d9e76b8-70fc-4e7e-91a4-f9c907ff1bc1/ “by the end of January, our ARS, right, our our our annual recurring spend had hit $4,000,000. And that's because people were using Cloud Code. Now today, it's about $11,000,000.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel says their AI spending equals over a third of employee compensation for their 90-person firm. https://minutesof.com/q/6f7b1cde-3481-4cd5-89a6-c38b8939b845/ https://minutesof.com/q/6f7b1cde-3481-4cd5-89a6-c38b8939b845/ “We're spending, you know, more like, you know, more than a third of the spend, you know, employee spend, a third of it is also on top of that is AI.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel predicted memory prices would soar because capacity grows 30% annually while demand doubles. https://minutesof.com/q/426304ad-3877-4fb3-a3f6-264771845ee3/ https://minutesof.com/q/426304ad-3877-4fb3-a3f6-264771845ee3/ “Memory capacity is only growing 30% a year for the next three years. And yet demand is doubling, is doubling. And so what's gonna end up happening is memory prices are gonna keep soaring.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel predicts iPhone and MacBook prices must rise as AI demand drives memory shortages. https://minutesof.com/q/9b9f24e6-60a8-45bd-bf65-ed8b26a06923/ https://minutesof.com/q/9b9f24e6-60a8-45bd-bf65-ed8b26a06923/ “And right now, you know, if MacBook prices or iPhone prices go up a $100, that market's not gonna adjust too much,” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel forecasts data center capacity will reach 30 gigawatts in the next year. https://minutesof.com/q/554d1ba5-063a-4241-b24a-4b017c6875e6/ https://minutesof.com/q/554d1ba5-063a-4241-b24a-4b017c6875e6/ “This year, we're deploying 20 gigawatts of data centers. Next year, that number goes up 50 gigawatts or 50%, sorry. 50%, sorry.” — WisdomTree in Europe Thu, 09 Jul 2026 06:16:39 +0000 Patel claims smaller Quen models with 27B total parameters and 2B active outperform GPT-4 from three years ago https://minutesof.com/q/6c24eb00-fa41-4b6a-9288-a44447de69b2/ https://minutesof.com/q/6c24eb00-fa41-4b6a-9288-a44447de69b2/ “If you look back three years, it's GPT-four. Now it's like maybe like quen one of the smaller quen models that's like 27 b parameters total and like 2,000,000,000 active is like way better.” — The AGI Post Mon, 06 Jul 2026 22:07:54 +0000 Patel argues OpenAI models are optimized for GPUs while Anthropic and Google models are optimized for TPUs due https://minutesof.com/q/be4a0229-47ca-4709-9bde-b26e807119ac/ https://minutesof.com/q/be4a0229-47ca-4709-9bde-b26e807119ac/ “And the way that Anthropic and Google's models are headed, it's actually a terrible decision potentially for them to train with GPUs.” — The AGI Post Mon, 06 Jul 2026 22:07:54 +0000 Patel says OpenAI models are much more sparse while Anthropic models are denser, with different architectural https://minutesof.com/q/d82add60-0d13-4045-a3c5-08b2e58a691e/ https://minutesof.com/q/d82add60-0d13-4045-a3c5-08b2e58a691e/ “Open eyes are much more sparse and that has benefits. Then anthropics are, they're still sparse but more dense in general and that has different benefits.” — The AGI Post Mon, 06 Jul 2026 22:07:54 +0000 Patel explains NVLink connects only 72 GPUs while Google ICI connects 8,000 chips without switches, creating a https://minutesof.com/q/5ca387af-4b9c-4a90-9c49-5374505ffeb1/ https://minutesof.com/q/5ca387af-4b9c-4a90-9c49-5374505ffeb1/ “NVIDIA, the NVLink can only connect 72 GPUs. For Google, their ICI can connect 8,000 chips at super high bandwidth, but you have to pass through other chips to get there because there's no switch.” — The AGI Post Mon, 06 Jul 2026 22:07:54 +0000 Patel secured over $50 million in donated hardware for InferenceX, expanding to $100 million with TPUs and tra https://minutesof.com/q/f9ad7311-11bb-4dce-afab-53a809a7e91a/ https://minutesof.com/q/f9ad7311-11bb-4dce-afab-53a809a7e91a/ “We've got over $50,000,000 of hardware donated to us. Once we launch TPUs and training, would actually be over $100,000,000 of hardware.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel predicts OpenAI and Anthropic will have over 100 gigawatts combined by 2030, terawatts by 2040. https://minutesof.com/q/dd349bfc-a119-4399-bf3c-9a26efbef20a/ https://minutesof.com/q/dd349bfc-a119-4399-bf3c-9a26efbef20a/ “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.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel predicts OpenAI and Anthropic alone will deploy over 100 gigawatts of compute by 2030. https://minutesof.com/q/e9a8fba3-1f79-4310-9c95-0d513f20374b/ https://minutesof.com/q/e9a8fba3-1f79-4310-9c95-0d513f20374b/ “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.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel claims hardware improved 30x from Hopper to Blackwell for DeepSeek on optimized deployments. https://minutesof.com/q/1aeaaf0c-1080-4d21-a79c-2bc3d127baf5/ https://minutesof.com/q/1aeaaf0c-1080-4d21-a79c-2bc3d127baf5/ “from Hopper to Blackwell, which is all we've had over the last three years, roughly 30x improvement on DeepSeek, on the most optimized deployment, which you can see on InferenceX there's about a 30x improvement.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel argues three years of improvement came from model layer advances plus hardware-software co-design, citin https://minutesof.com/q/1ec32ab0-6fc9-41a7-8547-99d3fc08884d/ https://minutesof.com/q/1ec32ab0-6fc9-41a7-8547-99d3fc08884d/ “If you look back three years, it's GPT-four. Now it's maybe like quen, one of the smaller quen models that's like 27B parameters total and 2,000,000,000 active is way better.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel argues co-optimizing hardware, software, and models produces 100x gains instead of 8x from multiplicativ https://minutesof.com/q/c187b446-08d3-4205-b3c0-8e3aacc8f8ba/ https://minutesof.com/q/c187b446-08d3-4205-b3c0-8e3aacc8f8ba/ “The real breakthrough innovation is when you leapfrog a few layers, you co optimize and co design them, and now all of a sudden you've taken what could have been a 2x here, 2x here, 2x here, and instead of being multiplicative to 8x, it's actually 100x, because you've optimized it across all three layers.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel reveals Google runs three separate TPU design programs with Broadcom, MediaTek, and undisclosed partner. https://minutesof.com/q/1146df8f-fda2-440e-b3ca-58362b2c1909/ https://minutesof.com/q/1146df8f-fda2-440e-b3ca-58362b2c1909/ “They're making a TPU with Broadcom, that's a different architecture than a TPU with MediaTek, that's a different TPU than the architecture that is, I won't disclose, by research.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel forecasts 20 gigawatts of data center capacity in 2024 and over 30 gigawatts in 2025 despite delays. https://minutesof.com/q/47d7b4b9-7b67-4116-a983-f7f6f9ddf38b/ https://minutesof.com/q/47d7b4b9-7b67-4116-a983-f7f6f9ddf38b/ “This year there's going to be 20 gigawatts, even accounting for the delays, and next year there's going be more than 30 gigawatts accounting for the delays.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel states Anthropic is profitable excluding stock compensation in Q2 with 80% margins on Opus tokens. https://minutesof.com/q/db45dba5-ecdb-452c-8fae-3210d613d134/ https://minutesof.com/q/db45dba5-ecdb-452c-8fae-3210d613d134/ “Anthropic in Q2 is profitable, their net income profitable, excluding stock based compensation. And I think by Q3 they may even be profitable, including stock based compensation.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel reports Anthropic is Q2 profitable excluding SBC, expects Q3 profitability including SBC with 80% margin https://minutesof.com/q/7f1b8019-87b9-4143-aabf-b8e158bf4a47/ https://minutesof.com/q/7f1b8019-87b9-4143-aabf-b8e158bf4a47/ “And I think by Q3 they may even be profitable, including stock based compensation. That's how profitable they're getting, and their margins on an Opus token, at least Opus 4.8 token, is north of 80” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel reports Anthropic achieved over 80% margins on Opus tokens and expects full profitability including stoc https://minutesof.com/q/08d73c3f-320f-4d28-8b21-9083da9b5bf0/ https://minutesof.com/q/08d73c3f-320f-4d28-8b21-9083da9b5bf0/ “That's how profitable they're getting, and their margins on an Opus token, at least Opus 4.8 token, is north of 80 for the API price.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel reports Trainium rents for under $10 billion per gigawatt while GPUs cost $12-13 billion. https://minutesof.com/q/32312d2c-7043-4e82-9029-8dcbaeafc456/ https://minutesof.com/q/32312d2c-7043-4e82-9029-8dcbaeafc456/ “Tranium sells at sub $10,000,000,000 per gigawatt rental rate to Anthropic and to OpenAI. GPUs, at least before the craziness of the last six months, usually went around 12 to $13,000,000,000 per gigawatt.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 Patel states GPU rental rates were $12-13 billion per gigawatt before recent increases. https://minutesof.com/q/c1abee78-d176-457f-8a5b-90e353f41471/ https://minutesof.com/q/c1abee78-d176-457f-8a5b-90e353f41471/ “GPUs, at least before the craziness of the last six months, usually went around 12 to $13,000,000,000 per gigawatt.” — Sequoia Capital Tue, 30 Jun 2026 12:00:25 +0000 We try and track the entire supply chain from tools that manufacture chips, fabs, data centers, energy, indust https://minutesof.com/q/8069a2f4-c1da-4a82-b880-1c456a111132/ https://minutesof.com/q/8069a2f4-c1da-4a82-b880-1c456a111132/ “We try and track the entire supply chain from tools that manufacture chips, fabs, data centers, energy, industrials, and then AI models and who's using them, how much, and where.” — Swole as a Service Wed, 13 May 2026 20:00:36 +0000 Patel distinguishes OpenAI hiding reasoning chains from Anthropic showing them in code generation models. https://minutesof.com/q/a9706075-6705-4eb4-ad0e-013b2f735206/ https://minutesof.com/q/a9706075-6705-4eb4-ad0e-013b2f735206/ “In the case of OpenAI, they don't show you the whole reasoning chain. In the case of Anthropic, they do.” — Swole as a Service Wed, 13 May 2026 20:00:36 +0000 Patel predicts reasoning chains will close as AI response times extend from immediate to hours or days. https://minutesof.com/q/44c831c5-138e-45c3-8b54-9ad0278248ff/ https://minutesof.com/q/44c831c5-138e-45c3-8b54-9ad0278248ff/ “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.” — Swole as a Service Wed, 13 May 2026 20:00:36 +0000 Patel cites estimates of 20% economic decline if Taiwan is invaded, comparable only to world wars. https://minutesof.com/q/3d3aa12b-148b-4e36-8a9c-ef47962c54bc/ https://minutesof.com/q/3d3aa12b-148b-4e36-8a9c-ef47962c54bc/ “And then subsequently, that that sends the world into a depression. Right? The world has not experienced a 20 free fall since maybe, like, a one of the great wars.” — Swole as a Service Wed, 13 May 2026 20:00:36 +0000 Patel estimates five to eight years of global tool manufacturing would be needed to replace Taiwan's chip capa https://minutesof.com/q/a8c57373-a3fe-4131-afd1-832e6b005059/ https://minutesof.com/q/a8c57373-a3fe-4131-afd1-832e6b005059/ “It would take five to eight years of all of the world's tool manufacturing capability Yeah. To even replace the capacity in Taiwan.” — Swole as a Service Wed, 13 May 2026 20:00:36 +0000 Patel notes China has a fully self-sufficient vertical semiconductor supply chain, though behind by three to f https://minutesof.com/q/11ae3627-3c03-488e-83aa-76308403a2ae/ https://minutesof.com/q/11ae3627-3c03-488e-83aa-76308403a2ae/ “China has a vertical supply chain. Yes. In many cases, it's fifteen years behind, ten years behind, in some cases, only three years behind.” — Swole as a Service Wed, 13 May 2026 20:00:36 +0000 Patel argues Mythos fast mode is probably cheaper than Claude 4.6 fast mode for most tasks due to token effici https://minutesof.com/q/9d03e1ee-9f3e-449f-bbd7-ef96c09864e7/ https://minutesof.com/q/9d03e1ee-9f3e-449f-bbd7-ef96c09864e7/ “the flip side is, is, is mythos is more token efficient. Methos fast mode is probably cheaper than like, four, six fast mode for most tasks.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel predicts Google and OpenAI will both release new models in two weeks. https://minutesof.com/q/85a6edb0-e86b-4792-9460-831d072cdf52/ https://minutesof.com/q/85a6edb0-e86b-4792-9460-831d072cdf52/ “Release New release in two weeks. Everyone's releasing in two weeks. Google, OpenAI, maybe Anthropic. I don't know about Anthropic, but Google and open AI definitely releasing in two weeks.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel says the model was released before pre-training was finished and needs more RL. https://minutesof.com/q/9e8054f5-4158-4e9e-8609-859fcd37357d/ https://minutesof.com/q/9e8054f5-4158-4e9e-8609-859fcd37357d/ “for everything new, more, more continued pre training because the spud, they kind of like didn't finish the pre training and just released it. So finished the pre training, do more RL drop model.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel argues Anthropic trapped itself in an innovator's dilemma by focusing on CLI instead of app experience. https://minutesof.com/q/ed8bef56-e332-41e3-8a94-a14a9c665e4c/ https://minutesof.com/q/ed8bef56-e332-41e3-8a94-a14a9c665e4c/ “But the CLI experience is not the end all be all of agent orchestration and therefore they've really cooked themselves into sort of like an innovator's dilemma where they keep making the CLI better.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel reports kernel programmers use Codex to create code first, then have Opus rewrite for quality. https://minutesof.com/q/50c6395c-373b-4b23-b6a0-af7f341a2fcd/ https://minutesof.com/q/50c6395c-373b-4b23-b6a0-af7f341a2fcd/ “I think it was tree down. Tree down is like, yeah, dude, Codex is like so dumb, but I always just have it created and then this and it works and it's smarter, but the code is sloppy, I have Opus rewrite it.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel reports kernel programmers use Codex to generate code then Opus to clean it, not the reverse workflow. https://minutesof.com/q/bc52f0cc-dd82-4db4-89fa-99f109589046/ https://minutesof.com/q/bc52f0cc-dd82-4db4-89fa-99f109589046/ “But you can't go the other way around. You can't have Opus write the thing and then have Codex fix this.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel observes Asian researchers consistently produce KV cache reduction papers unknown to American researcher https://minutesof.com/q/734c899f-c40e-4627-bc76-4d78dc3ab398/ https://minutesof.com/q/734c899f-c40e-4627-bc76-4d78dc3ab398/ “Some paper that reduces KV cash, and no researcher in America has even heard of this paper.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel observes Asian researchers consistently reference KV cache reduction papers unknown to American research https://minutesof.com/q/3c600994-8f55-4267-9e3e-11503f5d67cb/ https://minutesof.com/q/3c600994-8f55-4267-9e3e-11503f5d67cb/ “It's the fucking best thing ever. DeepSeek and TurboQuant were the, like, most precipitous ones that popped up the most,” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel explains long context windows fail because pre-training uses short context and lacks data for longer ran https://minutesof.com/q/d91849f6-904c-47f6-8429-3cb70fa0f391/ https://minutesof.com/q/d91849f6-904c-47f6-8429-3cb70fa0f391/ “But it's like, what data do I have? That's why my two fifty ks to one mil context is trash anyways, because there's no data on this stuff.” — SemiAnalysis Weekly Wed, 06 May 2026 00:55:06 +0000 Patel says Nvidia and AMD both lie about peak flops specs which are impossible to achieve. https://minutesof.com/q/2805f8b7-b1ab-4aa6-bfaa-e0ca89f29b45/ https://minutesof.com/q/2805f8b7-b1ab-4aa6-bfaa-e0ca89f29b45/ “All their quoted specs are lies impossible to achieve Whether it's Nvidia or AMD, neither of them you can ever hit their peak flops.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel claims Nvidia achieves 50% sustained performance on common 8k gemm while AMD gets only 30%. https://minutesof.com/q/f0d9f7af-e5f2-4a3f-9331-63fe03e874a1/ https://minutesof.com/q/f0d9f7af-e5f2-4a3f-9331-63fe03e874a1/ “if you just take a eight k by eight k by eight k gem, very common shape and you run it on the map mode unit of AMD and Nvidia, you get like 50% sustained performance on Nvidia and you get like 30% sustained performance on AMD.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel asserts vendors cannot achieve their advertised flops specifications in real-world conditions. https://minutesof.com/q/dd7daafb-b209-4da7-868d-23f93933e381/ https://minutesof.com/q/dd7daafb-b209-4da7-868d-23f93933e381/ “There there is no functional way to get anywhere close to their flops that they advertise.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel explains Inference X was created because vendor-claimed performance is unattainable with any available f https://minutesof.com/q/9f06077d-1d55-479f-aedc-dbafa2704a09/ https://minutesof.com/q/9f06077d-1d55-479f-aedc-dbafa2704a09/ “There's no way to get the performance that vendors like to claim. And so our whole thing there was, well, how do we actually have a benchmark that represents what people actually get on performance?” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel notes software dependencies change daily or multiple times weekly across the entire stack. https://minutesof.com/q/7d67261b-caa3-4beb-8c61-2b0c52ec53da/ https://minutesof.com/q/7d67261b-caa3-4beb-8c61-2b0c52ec53da/ “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.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel explains AI software stacks update multiple times per week making performance measurement a moving targe https://minutesof.com/q/bf34d090-4f22-4978-bd08-307b1566b305/ https://minutesof.com/q/bf34d090-4f22-4978-bd08-307b1566b305/ “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.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel notes AI software stacks update constantly with nightly builds making performance measurement time-sensi https://minutesof.com/q/3aab3b19-37a2-4c04-a1ec-5e5018b0ee10/ https://minutesof.com/q/3aab3b19-37a2-4c04-a1ec-5e5018b0ee10/ “with software changing literally multiple times a week, right, PyTorch has nightlies, VLM has nightlies, Asteeling has nightlies, CUDA drivers update constantly.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel observes VLLM and SGLANG framework developers compete publicly on Twitter using InferenceX as a leaderbo https://minutesof.com/q/6447730b-2e25-4183-babf-5c0b556ea5a0/ https://minutesof.com/q/6447730b-2e25-4183-babf-5c0b556ea5a0/ “VLLM and SGLANG love competing with each other. They used to post on Twitter all the time about how they had beat the other one in certain some kind of performance” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel says VLLM and SGLANG now use Inference X as their competition leaderboard instead of Twitter posts. https://minutesof.com/q/91af81a3-f585-4b20-8566-b337ff89327c/ https://minutesof.com/q/91af81a3-f585-4b20-8566-b337ff89327c/ “They used to post on Twitter all the time about how they had beat the other one in certain some kind of performance and now they use Inference X all the time.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel reports AMD and Nvidia engineers treat Inference X as a competitive leaderboard. https://minutesof.com/q/90c32f06-2549-483b-a0ec-37f8fcdadb56/ https://minutesof.com/q/90c32f06-2549-483b-a0ec-37f8fcdadb56/ “AMD and Nvidia. They love to compete with each other. And AMD engineers and Nvidia engineers look at Inference X as a leader board.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel argues effective benchmarks must anger vendors by revealing uncomfortable truths about performance. https://minutesof.com/q/01bf7e97-f767-46f6-80ba-60f8c81f1897/ https://minutesof.com/q/01bf7e97-f767-46f6-80ba-60f8c81f1897/ “I think one common thing between the three of us is if you're not pissing off people with your benchmark, then you're not testing something useful.” — TensorWave Thu, 30 Apr 2026 18:18:14 +0000 Patel reports H100 rental prices increased from $1.71 per hour to over $2.40 in six months. https://minutesof.com/q/525bb53d-bbc6-4eaf-84ff-3b55976574fc/ https://minutesof.com/q/525bb53d-bbc6-4eaf-84ff-3b55976574fc/ “And just in the last six months, it's gone from, you know, deals for one year transacting at $171.60 an hour for h 1 hundreds to now $2.40 plus.” — Aria Networks Thu, 16 Apr 2026 02:27:54 +0000 Patel identifies disaggregated PD and wide EP as networking techniques where performance matters significantly https://minutesof.com/q/a8d7ee01-6bf0-4ca4-828e-77222cf7a242/ https://minutesof.com/q/a8d7ee01-6bf0-4ca4-828e-77222cf7a242/ “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.” — Aria Networks Thu, 16 Apr 2026 02:27:54 +0000 Patel claims code agent revenue grew from a couple billion to over $10 billion in a very short time. https://minutesof.com/q/eae8cac0-9d21-4f54-97aa-ef5eae6fa378/ https://minutesof.com/q/eae8cac0-9d21-4f54-97aa-ef5eae6fa378/ “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,” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel reports code agent revenue jumped from a couple billion to over $10 billion in a very short time. https://minutesof.com/q/6cdbbe15-05da-4581-a25f-1d2dd85dd3fd/ https://minutesof.com/q/6cdbbe15-05da-4581-a25f-1d2dd85dd3fd/ “code agent revenue has gone from a couple billion to north of 10,000,000,000 in like a very short amount of time.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel says CPU-to-GPU ratio has shifted from 1:100 megawatts to much closer parity. https://minutesof.com/q/8f0a6563-524e-4327-a1d7-06032c4f32c0/ https://minutesof.com/q/8f0a6563-524e-4327-a1d7-06032c4f32c0/ “So whereas before you'd have like many GPU servers per CPU server, and so you'd be like, 100 megawatts of GPUs would be served by even like one megawatt or less of CPUs.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel says CPU-to-GPU ratios have tightened from 1:100 megawatts to much closer ratios due to RL and agentic i https://minutesof.com/q/e7053125-a899-4c7e-881d-a55e77a59048/ https://minutesof.com/q/e7053125-a899-4c7e-881d-a55e77a59048/ “Nowadays the ratio is like getting much, much closer, both for RL training and for inference, agentic inference.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel reports Amazon's CPU server installations have tripled year-on-year with no capacity remaining anywhere. https://minutesof.com/q/0d8b9a11-eaa8-49fb-9ff1-271ee0c9d472/ https://minutesof.com/q/0d8b9a11-eaa8-49fb-9ff1-271ee0c9d472/ “So then you've just seen everyone run out of CPUs, Amazon's volumes on CPUs, of CPU servers they're installing have three x year on year, right, from this year versus last year.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel reports Amazon's CPU server installations have tripled year-over-year with no spare capacity. https://minutesof.com/q/31c03f55-c143-4c0c-84a7-52a3e1df0c7a/ https://minutesof.com/q/31c03f55-c143-4c0c-84a7-52a3e1df0c7a/ “Amazon's volumes on CPUs, of CPU servers they're installing have three x year on year, right, from this year versus last year. And so there's just no capacity anywhere” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel says Amazon's CPU server installations have tripled year-on-year, causing capacity shortages and GitHub https://minutesof.com/q/467d4b9b-9a4f-4f92-a26f-6dfb6d56604a/ https://minutesof.com/q/467d4b9b-9a4f-4f92-a26f-6dfb6d56604a/ “And so there's just no capacity anywhere and that's causing a lot of instability, not just in GitHub but probably other places too.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel reveals OpenAI ported entire stack from x86 to ARM to access Amazon's CPU capacity for recent deal. https://minutesof.com/q/ed71dd24-f769-48fe-8eb3-1869799a3e75/ https://minutesof.com/q/ed71dd24-f769-48fe-8eb3-1869799a3e75/ “A lot of the impetus for the recent Amazon and OpenAI deal was, yes, OpenAI wanted money, yes, they needed compute, but they also just went to Amazon and were like, give us your CPUs, And before, OpenAI's stack was pretty much only on x86 CPUs, but Amazon had tons of ARM CPUs and so they ported the whole thing,” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel predicts next-generation models will handle physical world tasks like tying shoes beyond current math an https://minutesof.com/q/fa32f506-585b-483a-af87-48842a10d2ff/ https://minutesof.com/q/fa32f506-585b-483a-af87-48842a10d2ff/ “And when we look at like o one, that's like all it could do is math. Right?” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel reports memory prices have quadrupled over the past year, with SSDs up three to four times. https://minutesof.com/q/c56e363b-dd83-4210-8f36-83191d9f8973/ https://minutesof.com/q/c56e363b-dd83-4210-8f36-83191d9f8973/ “Memory has like four x in price over the last like year, and it's gonna continue to go up. And then now storage SSD, right?” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel says AI chips consuming all TSMC 3nm and 2nm capacity is forcing companies to alternative fabs. https://minutesof.com/q/43612d0e-1c2c-44ce-bd43-30f7192175e9/ https://minutesof.com/q/43612d0e-1c2c-44ce-bd43-30f7192175e9/ “More importantly is that AI is buying all the capacity on three nanometer and two nanometer in a couple years that, people are having to turn to other directions.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel argues NVIDIA acquired Grok partly because it's manufactured on Samsung, avoiding TSMC capacity constrai https://minutesof.com/q/bf7d604d-4ab7-42db-b4f7-a84f08023b6a/ https://minutesof.com/q/bf7d604d-4ab7-42db-b4f7-a84f08023b6a/ “Part of it was because they want to have really fast inference, but like part of it is that Grok is manufactured on Samsung.” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel explains NVIDIA acquired Grok partly because it manufactures on Samsung due to no TSMC 3nm capacity avai https://minutesof.com/q/d99d542e-883c-4529-a74c-51266d3590ff/ https://minutesof.com/q/d99d542e-883c-4529-a74c-51266d3590ff/ “Because there's no three nanometer capacity for them at TSMC, they need to chip somewhere else, and if you just if AI is as crazy as like we believe it is, and demand is as crazy as we believe it is, it's gonna be even crazier next year,” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel explains AI chips are crowding onto three nanometer nodes, displacing mobile chip production. https://minutesof.com/q/ba53114b-1d6a-4193-9d8e-5071f283a0c3/ https://minutesof.com/q/ba53114b-1d6a-4193-9d8e-5071f283a0c3/ “Right? Because all the AI chips are on three nanometer and it takes time, Small mobile chips are easier to make than big AI chips, so all the AI chips are moving to three nanometer now, right? MI three fifty from” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel says AI chips are moving to three nanometer, squeezing out mobile chip manufacturers. https://minutesof.com/q/f6f6cc9a-a485-4ad8-ba15-a7feb1b0d291/ https://minutesof.com/q/f6f6cc9a-a485-4ad8-ba15-a7feb1b0d291/ “Right? Because all the AI chips are on three nanometer and it takes time, Small mobile chips are easier to make than big AI chips, so all the AI chips are moving to three nanometer now, right?” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel lists AMD MI350, Amazon Trainium 3, and Google TPU v7 as AI chips now on 3nm process node. https://minutesof.com/q/0c038d22-7e17-4f58-8803-7254233cb26a/ https://minutesof.com/q/0c038d22-7e17-4f58-8803-7254233cb26a/ “MI three fifty from late last year from AMD or Tranium three and TPU v seven from late last year, or slash early this year from Amazon and Google” — Daytona Tue, 07 Apr 2026 17:12:37 +0000 Patel cites Anthropic adding $67 billion ARR per month as evidence of expanding AI revenue beyond hyperscalers https://minutesof.com/q/8d242968-225b-41c7-86f6-4f27678a4907/ https://minutesof.com/q/8d242968-225b-41c7-86f6-4f27678a4907/ “You're starting to see it with Anthropix revenue adding $67,000,000,000 of ARR a month, but there's so many more firms coming online with revenue streaming in,” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel says NVIDIA has locked up over 60% of supply chain capacity this year in long-term contracts. https://minutesof.com/q/7f7c89c2-e221-4652-adbe-c84494b67064/ https://minutesof.com/q/7f7c89c2-e221-4652-adbe-c84494b67064/ “The NVIDIA has locked up over 60% of the capacity this year in long term contracts alone, and they're buying more on top of that.” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel says NVIDIA has locked up over 60% of market capacity in long-term contracts this year. https://minutesof.com/q/fd13d4e3-a12c-4f89-a72e-66e0da0d0475/ https://minutesof.com/q/fd13d4e3-a12c-4f89-a72e-66e0da0d0475/ “NVIDIA has locked up over 60% of the capacity this year in long term contracts alone, and they're buying more on top of that.” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel reports NVIDIA is negotiating over $250 billion in supply contracts across components this year. https://minutesof.com/q/99be2869-e51d-4edf-b958-4a0eb31001f1/ https://minutesof.com/q/99be2869-e51d-4edf-b958-4a0eb31001f1/ “And as you look at what they're negotiating in the market today, there's over 250,000,000,000 of wafers, of memory, of substrates, of PCBs, of networking equipment that they're going to sign this year,” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel reports NVIDIA will sign over $250 billion in supply contracts this year across components. https://minutesof.com/q/ff2f7bf1-140d-43e5-ada7-c10acfd3d232/ https://minutesof.com/q/ff2f7bf1-140d-43e5-ada7-c10acfd3d232/ “as you look at what they're negotiating in the market today, there's over 250,000,000,000 of wafers, of memory, of substrates, of PCBs, of networking equipment that they're going to sign this year,” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel contrasts competitors' limited scale with NVIDIA's supply chain designed for tens of millions of chips. https://minutesof.com/q/f022ded7-b9fb-486a-adaa-4278b4cd8881/ https://minutesof.com/q/f022ded7-b9fb-486a-adaa-4278b4cd8881/ “I can buy 10,000. I could buy a 100,000. I can't buy millions, tens of millions, which is what Jensen's, setting his supply chain up for.” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel argues NVIDIA is setting up supply chain to manufacture tens of millions of AI chips. https://minutesof.com/q/6de9e7cc-be2c-4b14-aefa-a55d944bc979/ https://minutesof.com/q/6de9e7cc-be2c-4b14-aefa-a55d944bc979/ “I can't buy millions, tens of millions, which is what Jensen's, setting his supply chain up for.” — CNBC Television Mon, 16 Mar 2026 21:58:55 +0000 Patel estimates Anthropic needs to reach well above five gigawatts of compute capacity by year-end to support https://minutesof.com/q/dde6295c-09fa-47d0-9fb4-f6266575281d/ https://minutesof.com/q/dde6295c-09fa-47d0-9fb4-f6266575281d/ “Anthropic needs to get to well above five gigawatts by the end of this year, and it's gonna be really tough for them to get there, but it's possible.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel reports seeing AI labs sign GPU rental deals at $2.40 per hour for H100s, well above initial $1.40 cost. https://minutesof.com/q/4203d5a5-7268-4115-a39f-6b148314d682/ https://minutesof.com/q/4203d5a5-7268-4115-a39f-6b148314d682/ “I've seen deals where certain AI labs I'm be a little bit vague here for a reason, have signed at as high as $2.40 for two to three years.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel explains Google woke up to AI compute needs after Gemini models drove user growth. https://minutesof.com/q/a616fbed-e1ca-4ef3-915e-51ab7bb6cf2f/ https://minutesof.com/q/a616fbed-e1ca-4ef3-915e-51ab7bb6cf2f/ “Google had Nano Banana and Gemini three, which caused their user metrics to skyrocket, and leadership at Google was like, oh.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel details that producing one gigawatt of Rubin chips requires specific wafer volumes across multiple manuf https://minutesof.com/q/ff6866ea-ea84-4f63-9d38-dd6b9d9f502b/ https://minutesof.com/q/ff6866ea-ea84-4f63-9d38-dd6b9d9f502b/ “a gigawatt of, you know, NVIDIA's Rubin chips. Right? So Rubin is announced at GTC, I believe, the week this podcast goes live.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel calculates a gigawatt of Rubin requires 55,000 three-nanometer wafers and 170,000 DRAM wafers. https://minutesof.com/q/140d046c-4ee9-45e2-8bc2-b7b611de4f7c/ https://minutesof.com/q/140d046c-4ee9-45e2-8bc2-b7b611de4f7c/ “You need about 55,000 wafers of three nanometer. You need about 6,000 wafers of five nanometer, and then you need about a 170,000 wafers of DRAM, right, memory.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel calculates three and a half EUV tools are needed to produce one gigawatt of AI capacity. https://minutesof.com/q/504ff59f-223b-47eb-ab73-008b5cc61137/ https://minutesof.com/q/504ff59f-223b-47eb-ab73-008b5cc61137/ “you end up with actually, I need about three and a half EUV tools to do the 2,000,000 EUV wafer passes for the gigawatt. So three and a half EUV tools satisfies the gigawatt.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel shows that $50 billion in data center capacity depends on just $1.2 billion worth of EUV tools. https://minutesof.com/q/e589275c-5851-49fe-955a-9b06c45e05bf/ https://minutesof.com/q/e589275c-5851-49fe-955a-9b06c45e05bf/ “I need about three and a half EUV tools to do the 2,000,000 EUV wafer passes for the gigawatt. So three and a half EUV tools satisfies the gigawatt.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel calculates one gigawatt of AI capacity requires only 3.5 EUV tools costing $1.2B. https://minutesof.com/q/bf41b665-1668-499b-9836-478566fcb606/ https://minutesof.com/q/bf41b665-1668-499b-9836-478566fcb606/ “Because we're talking oh, what's a gigawatt cost? It costs like $50,000,000,000 roughly. Right? Whereas, what does three and a half EUV tools cost? That's like 1.2.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel projects maximum 200 gigawatts of AI chip capacity by 2030 based on EUV tool supply. https://minutesof.com/q/20c06a88-152a-4894-9cb6-c1663ff2d341/ https://minutesof.com/q/20c06a88-152a-4894-9cb6-c1663ff2d341/ “700 EUV tools by the end of the decade. 700 EUV tools, three and a half tools per gigawatt, assuming it's all allocated to AI, which it's not, but three and a half tools per gigawatt gets you to 200 gigawatts worth of AI chips for the data centers to deploy.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel calculates 700 EUV tools by 2030 enables maximum 200 gigawatts of AI chip production. https://minutesof.com/q/f0a44341-116a-41eb-8ef5-4e1fe1571c32/ https://minutesof.com/q/f0a44341-116a-41eb-8ef5-4e1fe1571c32/ “700 EUV tools, three and a half tools per gigawatt, assuming it's all allocated to AI, which it's not, but three and a half tools per gigawatt gets you to 200 gigawatts worth of AI chips for the data centers to deploy.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel estimates total EUV capacity by 2030 enables 200 gigawatts of AI chips, with Altman seeking 52 gigawatts https://minutesof.com/q/89d41218-810f-4ed7-b989-4c5230ba4930/ https://minutesof.com/q/89d41218-810f-4ed7-b989-4c5230ba4930/ “700 EUV tools, three and a half tools per gigawatt, assuming it's all allocated to AI, which it's not, but three and a half tools per gigawatt gets you to 200 gigawatts worth of AI chips for the data centers to deploy. Right? So 200 gigawatts, Sam wants 50 gigawatts, right, 52 gigawatts a year.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel predicts rising memory prices will make smartphones and PCs worse, causing public backlash against AI. https://minutesof.com/q/a323b743-075c-4050-84a8-3fec7fd419fb/ https://minutesof.com/q/a323b743-075c-4050-84a8-3fec7fd419fb/ “So memory crunch will continue to be harder and harder, and prices continue to go up. And this affects different parts of the market differently.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel reports memory prices have tripled, adding $150 to iPhone bill of materials versus $50 previously. https://minutesof.com/q/e524598c-47c1-41cb-9f34-bfc633a4887e/ https://minutesof.com/q/e524598c-47c1-41cb-9f34-bfc633a4887e/ “the price of memory is like tripled. Let's call it if it's now it's $12 per gig for DDR. So now you're talking about a $150 versus $50.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel projects smartphone volumes dropping from 1.1B to 500-600M next year due to memory costs. https://minutesof.com/q/ad564d12-2340-4961-a71f-cebb61fe2d06/ https://minutesof.com/q/ad564d12-2340-4961-a71f-cebb61fe2d06/ “our projections are we maybe get down to like 800,000,000 this year, and next year like 600 or 500,000,000.” — Dwarkesh Patel Fri, 13 Mar 2026 16:26:57 +0000 Patel reports $19 billion Anthropic revenue is driven by code spend, implying industry-wide $20 billion expend https://minutesof.com/q/25890617-f219-4daa-b5eb-8e3339b350f0/ https://minutesof.com/q/25890617-f219-4daa-b5eb-8e3339b350f0/ “You look at like quad code spend, right, it's freaking nuts. Right? 19,000,000,000 of revenue for Anthropic now. All of that is, you know, at some multiple is code.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel says his company's AI coding spend hit $6 million annual run rate, with single engineers spending $8,000 https://minutesof.com/q/f4ce11be-2e08-4b2b-9738-5f30336af1ea/ https://minutesof.com/q/f4ce11be-2e08-4b2b-9738-5f30336af1ea/ “Literally, we we you know, like, one day my head of ops was like, guys, our run rate of spend is $6,000,000. This is not sustained.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 one of my engineers spent like $8,000, he's Canadian, so whatever, right, on on Cloud Code. Right? And it's li https://minutesof.com/q/5da5c0b1-975f-47d8-a583-ffb7c6ca3f70/ https://minutesof.com/q/5da5c0b1-975f-47d8-a583-ffb7c6ca3f70/ “one of my engineers spent like $8,000, he's Canadian, so whatever, right, on on Cloud Code. Right? And it's like, oh, shit. Right?” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel reports a non-programmer employee now spends $5,000 daily using Claude Code to build tools via natural l https://minutesof.com/q/a2eba63d-9ceb-4f62-b7aa-b19fe5ecec2c/ https://minutesof.com/q/a2eba63d-9ceb-4f62-b7aa-b19fe5ecec2c/ “I'm just telling you what to do and it's doing these things. And his daily spend now on 4.6 fast 1,000,000 context is $5,000 a day.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel reports a single employee's daily AI spend reaching $5,000 and considers it acceptable given productivit https://minutesof.com/q/b6dc472f-c88d-478d-9da1-8b2d41f9516d/ https://minutesof.com/q/b6dc472f-c88d-478d-9da1-8b2d41f9516d/ “And his daily spend now on 4.6 fast 1,000,000 context is $5,000 a day. Jeez. And I look at the productivity, it's fine. Right?” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel criticizes Dario's response to hypothetical nuclear scenario as inadequate for government AI policy disc https://minutesof.com/q/dbb65dfb-41e5-4e5e-883e-13cf2d410ddd/ https://minutesof.com/q/dbb65dfb-41e5-4e5e-883e-13cf2d410ddd/ “If a nuclear missile's heading from China to America and we can use AI to stop it, but it requires surveillance and autonomous weapons, what do we do?” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel criticizes Dario Amodei's response to a hypothetical nuclear defense scenario as the dumbest possible an https://minutesof.com/q/b5ca73ad-828e-451d-abbd-d5d5927720b3/ https://minutesof.com/q/b5ca73ad-828e-451d-abbd-d5d5927720b3/ “And Dario was like, well, you can call us. I'm sure we can figure something out. Like, this is just the dumbest response you could ever come up with.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel says Sam Altman's greatest skill is taking advantage of a crisis to sign government deals. https://minutesof.com/q/f42e0db6-d699-4400-b0b7-9e21a37072b8/ https://minutesof.com/q/f42e0db6-d699-4400-b0b7-9e21a37072b8/ “That's maybe his greatest skill, is to take advantage of a good crisis. And so he goes out there and he signs deals with the US government.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel observes perception shift where 90th percentile people consider themselves middle class while 50th perce https://minutesof.com/q/bd34934b-1867-41db-90e5-d28702f49580/ https://minutesof.com/q/bd34934b-1867-41db-90e5-d28702f49580/ “Nowadays, people in the ninetieth percentile think they're middle class, and people in the fiftieth percentile don't think they're middle class.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel observes 90th percentile earners think they're middle class while 50th percentile don't, due to social m https://minutesof.com/q/ae3d803c-2ca0-4deb-bd9f-97e93f51e0c1/ https://minutesof.com/q/ae3d803c-2ca0-4deb-bd9f-97e93f51e0c1/ “Nowadays, people in the ninetieth percentile think they're middle class, and people in the fiftieth percentile don't think they're middle class. Yeah.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel claims his company now moves 10 times faster than competitors using Cloud Code. https://minutesof.com/q/4b5753cb-f579-4a41-9918-292316bd2a6f/ https://minutesof.com/q/4b5753cb-f579-4a41-9918-292316bd2a6f/ “Now with Cloud Code, think we move 10 times the speed of any of our competitors. Right? It's actually insane.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel claims his company moves 10x faster than competitors using Claude Code, enabling more hiring and busines https://minutesof.com/q/db39a85b-686a-4842-8953-f7515307baa2/ https://minutesof.com/q/db39a85b-686a-4842-8953-f7515307baa2/ “It's actually insane. And so my plans of hiring and the business we win and the basis we're able to do is actually just growing.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel says he now supports UBI despite being very capitalist, a shift from his prior views. https://minutesof.com/q/bec4f50b-ae61-4935-b577-c45c8527112c/ https://minutesof.com/q/bec4f50b-ae61-4935-b577-c45c8527112c/ “Over the last couple years I've realized, wait, I actually think UBI is perfectly fine. Which I think is crazy because again, I'm very capitalist.” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel predicts Anthropic revenue will grow from $19B to $60B by year end amid majority-negative AI sentiment. https://minutesof.com/q/fe481a98-ee52-4271-a902-5fcb25373eaa/ https://minutesof.com/q/fe481a98-ee52-4271-a902-5fcb25373eaa/ “The sentiment is already more than half Americans have a negative view of AI. As we fast forward to the end of this year, as Anthropix revenue goes from 19,000,000,000 to maybe 60, and OpenAI is similar, if not higher, the amount of jobs that get supplemented, the amount of change that happens in society,” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel predicts Google will eliminate its $100 billion annual cash flow next year by spending everything on AI https://minutesof.com/q/a7bb132e-83bc-40b6-b700-cf2f048e334f/ https://minutesof.com/q/a7bb132e-83bc-40b6-b700-cf2f048e334f/ “Google will have no cash flow next year because they're they see AI so clearly, and they know that they need to spend every dollar they make on compute,” — Matthew Berman Mon, 09 Mar 2026 17:07:47 +0000 Patel says Google spending $180B and Amazon $200B on AI infrastructure represents 4x increase from recent year https://minutesof.com/q/eac21179-024f-4ae9-8db2-3eee00972211/ https://minutesof.com/q/eac21179-024f-4ae9-8db2-3eee00972211/ “If we start looking at like, hey, this year, Google spending 2 Amazon spending $200,000,000,000, Google spending a $180,000,000,000 on on AI infrastructure primarily. Right?” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel states Anthropic now adds $23 billion revenue monthly versus few hundred million earlier. https://minutesof.com/q/a9df3580-18d1-4ac3-9f9d-afcb758ea6d7/ https://minutesof.com/q/a9df3580-18d1-4ac3-9f9d-afcb758ea6d7/ “Anthropics, you know, adding $23,000,000,000 of revenue a month now Mhmm. Versus they were just adding a few 100,000,000 of revenue a month earlier. So clearly, we're in the take off period.” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel reports Claude Code doubled to 4% of GitHub commits in January alone, with overall AI coding likely at 1 https://minutesof.com/q/d7ab6afb-29cc-4e7c-9868-7f4eb011caa2/ https://minutesof.com/q/d7ab6afb-29cc-4e7c-9868-7f4eb011caa2/ “Just in this month just in January, it went from 4% of or 2% of commits on GitHub to 4% of GitHub commits were done by Cloud Code.” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel calculates Anthropic added $1.5 billion in compute infrastructure in one month to serve new revenue grow https://minutesof.com/q/b1ba2e20-500e-4a5d-8e13-3d8ecec9f6a2/ https://minutesof.com/q/b1ba2e20-500e-4a5d-8e13-3d8ecec9f6a2/ “If Anthropic added 2,500,000,000 of revenue, and their gross margin is 40%, they added like $1,500,000,000 of compute in one month. Mhmm. Right?” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel predicts Google will have zero cash flow profit in 2027, spending all revenue on AI infrastructure. https://minutesof.com/q/1a954518-d905-4ab6-b6df-6800d4e18a40/ https://minutesof.com/q/1a954518-d905-4ab6-b6df-6800d4e18a40/ “There's no reason why Google will have any profit in '27 at all, right, in terms of cash flow. They will just spend every dollar they make on on AI infrastructure.” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel predicts Google and Amazon will take on debt for AI infrastructure following Meta's example. https://minutesof.com/q/59aa0cd8-2be5-4bd6-99fa-8aad2129359d/ https://minutesof.com/q/59aa0cd8-2be5-4bd6-99fa-8aad2129359d/ “Google and Amazon haven't taken on debt yet for AI infrastructure, but they will. Right?” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel predicts becoming the anti-AI party will be the winning political strategy due to public backlash. https://minutesof.com/q/08b118a0-09ac-4a7d-b1da-a6c8c2ba4c32/ https://minutesof.com/q/08b118a0-09ac-4a7d-b1da-a6c8c2ba4c32/ “And it seems obvious to me that, like, any party that wants to win should just become the anti AI party. Because life as we know it is changing.” — Latent Space Thu, 26 Feb 2026 21:15:18 +0000 Patel says NVIDIA must be 2x better than competitors to justify their 75% plus margins. https://minutesof.com/q/0af8e524-c5df-44cd-a672-ddcab4d2b6de/ https://minutesof.com/q/0af8e524-c5df-44cd-a672-ddcab4d2b6de/ “NVIDIA recognizes they're they're the leader, they're the tent pole. Hey, in one respect, they can just run faster than everyone, but it's kind of hard to be two x better than Google or or OpenAI or whoever else's internal chip, right, to justify their, you know, 75% plus margins.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel says NVIDIA must be 2-4x better than competitors to justify 75% margins and 4x pricing above costs. https://minutesof.com/q/fe51e725-980e-481f-b76c-2e41dcbdde73/ https://minutesof.com/q/fe51e725-980e-481f-b76c-2e41dcbdde73/ “And then they have to be two x to four x better to justify four x better to justify their margins because that's what they're charging above cogs.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel explains Jensen fears specialized chips could undercut NVIDIA's margins if they only made general-purpos https://minutesof.com/q/caf560f2-f0ca-439c-8b54-ef04778edc6c/ https://minutesof.com/q/caf560f2-f0ca-439c-8b54-ef04778edc6c/ “Jensen is very paranoid about losing. Right? These specializations if he just kept making his mainline chip would mean people could you know point point solutions for specific parts of the market would crush him on cost and performance, then he can't justify his margin.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel explains NVIDIA acquired Grok to get engineering resources for multiple chip architectures. https://minutesof.com/q/33dcc169-dbc3-427a-b62e-65aa1ee2164f/ https://minutesof.com/q/33dcc169-dbc3-427a-b62e-65aa1ee2164f/ “acquiring Rock is like how you get those resources to make more solutions for different parts of the market. And as far as like, are they threatened?” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel predicts AMD will remain in single-digit percentage market share despite being credible competitor. https://minutesof.com/q/2a1d4208-3cfe-4a01-a766-5ad296e384fb/ https://minutesof.com/q/2a1d4208-3cfe-4a01-a766-5ad296e384fb/ “I don't think they'll go beyond, like I think they'll stay in single digits market share, single digit percentage market share. Yeah. Single digit percentage market share” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel says China's entire culture is semiconductor-focused, even appearing in romance dramas. https://minutesof.com/q/c0eb3cc5-f5bc-42ff-afa8-1adeea874c30/ https://minutesof.com/q/c0eb3cc5-f5bc-42ff-afa8-1adeea874c30/ “I think the entire country is like semiconductor pilled. Mhmm. Right? There are dramas where people fall in love in the fab or dramas where people fall in love and they're photovoltaic,” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel says China's entire culture is semiconductor pilled with romance dramas set in fabs and solar research. https://minutesof.com/q/bb4695b1-6134-4a12-8e20-9aabbebc6242/ https://minutesof.com/q/bb4695b1-6134-4a12-8e20-9aabbebc6242/ “There are dramas where people fall in love in the fab or dramas where people fall in love and they're photovoltaic, like, solar cell researchers and engineers.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel claims China has the most vertical semiconductor stack and could run fabs independently unlike TSMC. https://minutesof.com/q/f055eebe-a3a4-4066-9750-79a424e3c7d0/ https://minutesof.com/q/f055eebe-a3a4-4066-9750-79a424e3c7d0/ “China has the most vertical stack in semiconductors today and they're the best at semiconductors in the world because their fabs could still run somewhat on a lot of things because they have built some of these chemical supply chains. Right? Like TSMC for certain kinds of chemicals 100% share from Japan.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel predicts AI industry will hit $100 billion ARR by end of 2025, with China 10x lower. https://minutesof.com/q/41e9701e-8a42-42e6-aaa0-bec5f26660a4/ https://minutesof.com/q/41e9701e-8a42-42e6-aaa0-bec5f26660a4/ “And then what's the economic value of that $100,000,000,000? Now, how much of that is in China? Right? Like, China's number is probably 10 x lower.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel frames AI competition as economic war determining whether China rises to global hegemony. https://minutesof.com/q/05c761d6-a934-4ece-905b-072e062b6646/ https://minutesof.com/q/05c761d6-a934-4ece-905b-072e062b6646/ “at the end of the day, this is an economic war. Right? If The US and the West win in AI and control, you know, more powerful AI systems that have this feedback loop that improve economic growth and weapon systems and whatever else, right, engineering of grids and cyber attacks and all these sorts of things.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel argues Western AI dominance determines whether China becomes global hegemon; without it, China will win https://minutesof.com/q/47555341-8389-4572-aed2-4d4365abcdbb/ https://minutesof.com/q/47555341-8389-4572-aed2-4d4365abcdbb/ “They have this like advantage over China, then China will not rise to be the global hegemony. But without AI, China definitely will rise to be the global hegemony.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel calculates $100B AI revenue requires $250B infrastructure spend at five-year depreciation, double curren https://minutesof.com/q/b206ab39-c11b-4186-80e2-7930fb9768d1/ https://minutesof.com/q/b206ab39-c11b-4186-80e2-7930fb9768d1/ “That $50,000,000,000 of COGS needs to burn on infra, which cost roughly with if a five if you're talking about five year depreciation, call it $250,000,000,000, right, of infra Yeah. For a $100,000,000,000 of revenue.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel calculates $100B AI revenue requires $250B infrastructure with five-year depreciation at 50% margins. https://minutesof.com/q/a42225be-7133-4459-8e20-e73d48721800/ https://minutesof.com/q/a42225be-7133-4459-8e20-e73d48721800/ “That $50,000,000,000 of COGS needs to burn on infra, which cost roughly with if a five if you're talking about five year depreciation, call it $250,000,000,000,” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel notes 2% of GitHub commits are AI-generated against $2 trillion in global software wages. https://minutesof.com/q/2d9fb7d3-a108-42ed-8c31-bc32b44f3155/ https://minutesof.com/q/2d9fb7d3-a108-42ed-8c31-bc32b44f3155/ “You can disable that where it's not automatically committed. But 2% of GitHub commits today are Cloud Code. $2,000,000,000,000 of software wages paid in the world.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel says data centers will reach 10% of US power by 2027-28 but under 1% of water consumption. https://minutesof.com/q/a110db52-8fec-41aa-a4a2-a4e12237f101/ https://minutesof.com/q/a110db52-8fec-41aa-a4a2-a4e12237f101/ “US Grid will get to like 10% of power by like '28, 27 is data centers. For water consumption, it's not even gonna crack 1%. Yeah. By the end of the decade.” — The MAD Podcast with Matt Turck Thu, 05 Feb 2026 13:30:16 +0000 Patel reports 10-15% of NVIDIA GPUs fail and require RMA within first two weeks of cluster deployment. https://minutesof.com/q/0dc25ccb-27dc-490b-bdd1-24ad4612e6ab/ https://minutesof.com/q/0dc25ccb-27dc-490b-bdd1-24ad4612e6ab/ “When you first turn on the cluster, about ten to fifteen percent of them fail RMA in the first two weeks. Wow. And then that's fine. Like you have to receipt them, whatever.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says 10-15% of NVIDIA GPUs fail and need RMA in the first two weeks after cluster deployment. https://minutesof.com/q/267f5cee-22ee-4be6-bf05-893375b653c5/ https://minutesof.com/q/267f5cee-22ee-4be6-bf05-893375b653c5/ “When you first turn on the cluster, about ten to fifteen percent of them fail RMA in the first two weeks. Wow. And then that's fine.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says Hopper GPU failure rates improved to 5% while Blackwell remains at 10-15%, expecting higher rates f https://minutesof.com/q/8d098967-35ce-4277-b237-0688cd59c6c6/ https://minutesof.com/q/8d098967-35ce-4277-b237-0688cd59c6c6/ “hopper's now at 5%, but black belt's still 10 to 15%. Wow. Right? Actually started out higher than that. Sure. And when a new generation comes out, it's gonna be higher than 15%.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel explains users will pay 10x more for 10x faster inference completion, justifying Cerberus economics. https://minutesof.com/q/d63342d7-794d-43b2-8529-800ca943aa5a/ https://minutesof.com/q/d63342d7-794d-43b2-8529-800ca943aa5a/ “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.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel explains agentic training only needs to sync tokens every few minutes versus weights every seconds in pr https://minutesof.com/q/86748cbb-055a-4578-92be-361a647d16ac/ https://minutesof.com/q/86748cbb-055a-4578-92be-361a647d16ac/ “When you're doing these rollouts and especially as things get more and more agentic and training, you might not only need to send not the entire weights but just the tokens that are relevant.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel notes semiconductor industry doubles transistors yearly but energy industry wasn't prepared for similar https://minutesof.com/q/ddccec53-f53b-4ae7-99e7-7f4520cead44/ https://minutesof.com/q/ddccec53-f53b-4ae7-99e7-7f4520cead44/ “Whereas the energy industry in America wasn't. And and so, like, initially, people were, like, not creative. They're like, let's do let's do these kinds of gas plants.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says data center power capacity will nearly double from 15-18 gigawatts this year to 30 gigawatts next y https://minutesof.com/q/70599022-7457-468b-b145-3cd24a5f0096/ https://minutesof.com/q/70599022-7457-468b-b145-3cd24a5f0096/ “next year, 30 gigawatts are being added and we think the power is there for it. Wow.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says data center power capacity is growing from 15-18 gigawatts this year to 30 gigawatts next year. https://minutesof.com/q/e2100e0b-9a14-4a60-a04c-0b639b876dd4/ https://minutesof.com/q/e2100e0b-9a14-4a60-a04c-0b639b876dd4/ “What was it It's this it's or this year is like I think it's like 18 ish, 10 ish. Okay. Fifteen fifteen to 18” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel predicts semiconductor shortages will return as the primary constraint in 2027 after power constraints e https://minutesof.com/q/542f4c33-ef8f-4e1c-9fe2-7a3da921cad2/ https://minutesof.com/q/542f4c33-ef8f-4e1c-9fe2-7a3da921cad2/ “But it will fully beach semiconductors again in '27. Right? And so we see this across the entire space of the ecosystem. It's not just TSMC.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel predicts semiconductor shortages will return as primary constraint in 2027 affecting both TSMC and memor https://minutesof.com/q/ab231a57-e3dc-4e62-afac-6f29f3648eeb/ https://minutesof.com/q/ab231a57-e3dc-4e62-afac-6f29f3648eeb/ “But it will fully beach semiconductors again in '27. Right? And so we see this across the entire space of the ecosystem. It's not just TSMC. It's also memory both.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says memory makers have not built new fabs since 2022 due to cyclical nature. https://minutesof.com/q/8b546cbf-7428-4740-a672-99ccb7374ee7/ https://minutesof.com/q/8b546cbf-7428-4740-a672-99ccb7374ee7/ “The memory makers, in fact, have just not expanded capacity. Basically, new they've not built new fabs since 2022. Yeah. Because their their cycle their cycle is so undulating.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel argues few hedge funds are in San Francisco to fully understand AI developments firsthand. https://minutesof.com/q/c28dbd0e-57ec-4817-9d02-1225c4eff373/ https://minutesof.com/q/c28dbd0e-57ec-4817-9d02-1225c4eff373/ “And I think I think like if you think about how much do you believe in AI and what's your access to information of AI, you know, there's not many hedge funds who live in San Francisco and like fully breathe and live and understand it.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel predicts AI startup revenue will exceed $100 billion by end of year, calling it unbelievable to most. https://minutesof.com/q/e5e3d778-0070-4de8-8f94-4c29939e6d4c/ https://minutesof.com/q/e5e3d778-0070-4de8-8f94-4c29939e6d4c/ “by the end of the year do how many people even believe by the end of the year AI startup revenue is over a $100,000,000,000?” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel predicts AI startup revenue will exceed $100 billion by end of year, a claim few believe. https://minutesof.com/q/4adf2698-ffda-4eb7-ae43-2e46f83b43fd/ https://minutesof.com/q/4adf2698-ffda-4eb7-ae43-2e46f83b43fd/ “I think that's an insane statement for a lot of people, but that's what it's gonna be. Yeah. Right? And who believes that number? Right? It's like Yeah. Very few people.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says Anthropic's $300 billion revenue target for end of decade is actually too conservative. https://minutesof.com/q/170adcfe-f306-42e0-8573-f93458964072/ https://minutesof.com/q/170adcfe-f306-42e0-8573-f93458964072/ “when Anthropic says in their funding, like, hey, we're gonna have $300,000,000,000 of revenue by the end of the decade. And it's like, actually, I think that number's too low.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel states OpenAI will deploy 16-18 gigawatts by end of 2028 requiring $300 billion spend. https://minutesof.com/q/0514621a-0bad-4693-aafd-f44de9e45c7b/ https://minutesof.com/q/0514621a-0bad-4693-aafd-f44de9e45c7b/ “OpenAI is gonna have 18 gigawatts or 16 gigawatts by the end of twenty eight, and they're gonna be able to pay for it. And that's like, well, that's $300,000,000,000 of spend.” — TBPN Tue, 03 Feb 2026 23:08:03 +0000 Patel says at his first NeurIPS in 2021 or 2022 he only stopped at posters mentioning transformers. https://minutesof.com/q/b9ae2fa9-fe32-4b7e-b3ac-cafb3ca211c3/ https://minutesof.com/q/b9ae2fa9-fe32-4b7e-b3ac-cafb3ca211c3/ “I remember the first NURBS I went to, it was, like, '21 or '22. And I was like I would walk around. I'd stop only posters that said the word transformer.” — SAIL Media Thu, 15 Jan 2026 16:46:15 +0000 Patel describes simultaneous gains across pre-training, scaling, quantization, systems, and RL infrastructure https://minutesof.com/q/f9fa2b78-5b6c-49cc-ab15-2b5c078d76e1/ https://minutesof.com/q/f9fa2b78-5b6c-49cc-ab15-2b5c078d76e1/ “There's tons of gains in quantization and systems, but there's also tons of gains in RL stuff and both the, like, you know, infra side of things and non infra side of things.” — SAIL Media Thu, 15 Jan 2026 16:46:15 +0000 Patel states OpenAI has over 500,000 GPUs for R&D but GPT-5 pre-training uses under 100,000 GPUs. https://minutesof.com/q/68335aae-f02e-47ac-b961-50c4701d12af/ https://minutesof.com/q/68335aae-f02e-47ac-b961-50c4701d12af/ “OpenAI has over 500,000 GPUs working on r and d. Right? Let's call it r and d.” — SAIL Media Thu, 15 Jan 2026 16:46:15 +0000 Patel says CoreWeave was the only Platinum recipient among roughly 40 cloud companies tested by SemiAnalysis. https://minutesof.com/q/96f53404-9f3d-4379-8de2-fc7f7a2abbab/ https://minutesof.com/q/96f53404-9f3d-4379-8de2-fc7f7a2abbab/ “Semi Analysis ClusterMax Platinum. The only recipient of Platinum, you know, we had about six gold recipients and we tested roughly 40 different cloud companies, but CoreWeave was the only one that achieved Platinum.” — Blue Cactus AI Sat, 27 Dec 2025 17:41:13 +0000 Patel says neo clouds that ranked highest on ClusterMax benchmarks booked significantly more revenue. https://minutesof.com/q/9a2acec1-fa93-4f91-b7b2-610ae8acb9f1/ https://minutesof.com/q/9a2acec1-fa93-4f91-b7b2-610ae8acb9f1/ “And since then we just released the second version, but in the middle, the clouds that ranked the highest actually ended up booking way, way, way more revenue.” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel states top neo clouds achieve 35-40% gross margins while many others are losing money. https://minutesof.com/q/84da5d17-180b-411f-aac0-520cd84480d9/ https://minutesof.com/q/84da5d17-180b-411f-aac0-520cd84480d9/ “And this has enabled, you know, the top in the industry companies to have gross margins of 35, 40%. And now there's a ton of Neo Clouds that are losing money.” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel reports five or six neo clouds have already gone out of business in just a few years. https://minutesof.com/q/aa801cea-f3ca-4fe8-aabc-405fb076edbe/ https://minutesof.com/q/aa801cea-f3ca-4fe8-aabc-405fb076edbe/ “There's I believe five or six that have already gone out of business. It's only been a few years. Right. Right?” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel says H100 cannot physically reach advertised 2,000 teraflops, maxing at 1,300 teraflops for matrix opera https://minutesof.com/q/7eb3f152-085c-4b25-9dd3-33f9d606c0a6/ https://minutesof.com/q/7eb3f152-085c-4b25-9dd3-33f9d606c0a6/ “In fact, it's physically impossible to get 2,000 teraflops out of the H100. Even though the advertises it, at most you can get 1,300 FP8 floating point eight teraflops.” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel cites DeepSeek's open-sourced inference system requiring 140 GPUs communicating over RDMA networks for s https://minutesof.com/q/b5b15eca-a046-4a24-8309-0cfa9d92ceec/ https://minutesof.com/q/b5b15eca-a046-4a24-8309-0cfa9d92ceec/ “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.” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel explains prefill costs one-fourth of decode per token, making 30,000 token inputs costlier than 2-4,000 https://minutesof.com/q/187d2021-25f1-4e4f-8559-763189933ea6/ https://minutesof.com/q/187d2021-25f1-4e4f-8559-763189933ea6/ “And that's a very common ratio. Right? And then when you think about, okay, the cost of running pre fill is roughly one fourth of running decode.” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel states avoiding redundant prefill through KV cache can cut inference costs to one-fourth of previous lev https://minutesof.com/q/c93c6e47-fe3d-405d-b28b-2d99a4919dee/ https://minutesof.com/q/c93c6e47-fe3d-405d-b28b-2d99a4919dee/ “So you can cut your cost to one fourth of what it was previously if you just don't do the pre fill. Right?” — Clockwork Fri, 21 Nov 2025 17:18:51 +0000 Patel reports Microsoft aims to 10x training capacity every 18-24 months, representing a 10x increase from GPT https://minutesof.com/q/fa0ed250-859f-4800-a571-4fb9a5440edb/ https://minutesof.com/q/fa0ed250-859f-4800-a571-4fb9a5440edb/ “We try to 10x the training capacity every eighteen to twenty four months. And so this would be effectively a 10x increase. 10x from what GPD five was trained with.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Nadella describes connecting multiple Fairwater facilities on a petabit network spanning to Milwaukee for dist https://minutesof.com/q/ea637240-f2b4-42f0-b334-d334f1e45e01/ https://minutesof.com/q/ea637240-f2b4-42f0-b334-d334f1e45e01/ “Fairwater 4, which you're going to see under construction nearby, will also be on that one petabits network so that we can actually link the two at a very high rate.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel projects AI will progress from 10-30 minute tasks to days of autonomous work, with model companies charg https://minutesof.com/q/5b9e2040-99b4-48d4-8f4c-697c8b51748a/ https://minutesof.com/q/5b9e2040-99b4-48d4-8f4c-697c8b51748a/ “In the future maybe they're doing days worth of work autonomously. And then the model companies are charging thousands of dollars” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel states Oracle will grow from one fifth Microsoft's size to bigger by end of 2027 at 35% margins. https://minutesof.com/q/ea3cf02f-7d0e-4bca-b768-9af08fb48270/ https://minutesof.com/q/ea3cf02f-7d0e-4bca-b768-9af08fb48270/ “Oracle is going from like one fifth your size to bigger than you by end of twenty twenty seven.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel estimates Microsoft's 2028 capacity is three and a half gigawatts lower than forecasted due to paused bu https://minutesof.com/q/7d9d0bab-255a-445a-a1d8-0e08420af839/ https://minutesof.com/q/7d9d0bab-255a-445a-a1d8-0e08420af839/ “And then the other one is, given our estimates on what your capacity is in 2028, is three and a half gigawatts lower? Sure, you could have dedicated that to OpenAI training and inference capacity.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel notes NVIDIA takes 75% margin on hardware comprising most of data center TCO, driving hyperscalers to de https://minutesof.com/q/6a22b49b-38f2-430c-8340-56fa935ee6c1/ https://minutesof.com/q/6a22b49b-38f2-430c-8340-56fa935ee6c1/ “So, you mentioned how you're depreciating this asset that's five, six years and this is the majority of the 75% of the TCO data center. And Jensen is taking a 75% margin on that.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel explains hyperscalers are developing custom accelerators to reduce equipment costs since Jensen takes 75 https://minutesof.com/q/026dcdd7-407b-4ef7-b53a-e8b953c43066/ https://minutesof.com/q/026dcdd7-407b-4ef7-b53a-e8b953c43066/ “So what all the hyperscalers are trying to do is develop their own accelerator so that they can reduce this overwhelming cost for equipment to increase their margins.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel says Google will make five to seven million TPU chips while Amazon targets three to five million. https://minutesof.com/q/7dc1dfdc-c966-48a1-aa23-0ab53fb8f113/ https://minutesof.com/q/7dc1dfdc-c966-48a1-aa23-0ab53fb8f113/ “They're going to make something like five to 7,000,000 chips, right? Of their own TPUs. You look at Amazon, they're trying to make three to 5,000,000.” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel states AI labs are projecting one hundred billion dollars in revenue by 2027-2028 with 2-3x growth. https://minutesof.com/q/55c2488f-537a-40ff-a8e2-4e757421a05c/ https://minutesof.com/q/55c2488f-537a-40ff-a8e2-4e757421a05c/ “So, these labs are now projecting revenues of 100,000,000,000 in 2728, and they're projecting revenue keeps growing at this rate of like 3x, 2x” — Dwarkesh Patel Wed, 12 Nov 2025 17:02:10 +0000 Patel says SemiAnalysis grew from one person two and a half years ago to 50 people this week, completely organ https://minutesof.com/q/30c79ef9-7f28-4296-855a-3caf989be32b/ https://minutesof.com/q/30c79ef9-7f28-4296-855a-3caf989be32b/ “Actually, this week we find that we hit 50 people, so having started it five years ago to now, and I've hired my first person two and a half years ago, We've now grown to 50 people just completely organically, so it's been a hell of a ride.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel estimates AI infrastructure represents 25-75% of current US quarterly economic growth. https://minutesof.com/q/b10288ed-374e-4451-b8b7-dccdad62176a/ https://minutesof.com/q/b10288ed-374e-4451-b8b7-dccdad62176a/ “it's a substantial fraction of economic growth, depending on who you ask. It's anywhere from 25% to 75 of the current quarter's economic growth. That's insane for The US.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel states AI infrastructure represents 25% to 75% of current US quarterly economic growth. https://minutesof.com/q/70800ca7-aaca-45bb-b128-a5925c5404e1/ https://minutesof.com/q/70800ca7-aaca-45bb-b128-a5925c5404e1/ “It's anywhere from 25% to 75 of the current quarter's economic growth. That's insane for The US.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says Substack monetization reached hundreds of thousands in months, then millions before he left the pla https://minutesof.com/q/77cdeac6-d89b-4620-b7a3-82e6c9420050/ https://minutesof.com/q/77cdeac6-d89b-4620-b7a3-82e6c9420050/ “It got to a few $100,000 of revenue and I was like, Holy crap. And then I was able to start hiring. And then I left when we hit a few million of revenue.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel left Substack initially after reaching a few million in revenue to avoid the 10% fee. https://minutesof.com/q/d6489a6e-b7ec-492f-b045-b892357854ec/ https://minutesof.com/q/d6489a6e-b7ec-492f-b045-b892357854ec/ “And then I left when we hit a few million of revenue. More and more of the revenue was coming from not Substack. The reason was like, Hey, Substack takes 10%.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says email capture growth accelerated dramatically on Substack but fell back after leaving despite SEO w https://minutesof.com/q/183e7331-dac8-497f-b275-4421d7096894/ https://minutesof.com/q/183e7331-dac8-497f-b275-4421d7096894/ “And then I added Substack and growth went like this. And then I left Substack and it went back to like this, the email capture.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says email growth slowed significantly after leaving Substack despite having dedicated SEO staff. https://minutesof.com/q/b6d41185-c60d-4e95-a6a0-8c50e18902cb/ https://minutesof.com/q/b6d41185-c60d-4e95-a6a0-8c50e18902cb/ “And then I left Substack and it went back to like this, the email capture.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says his Substack revenue alone reached multi-million dollars, sufficient to fund significant coverage. https://minutesof.com/q/d8aceaf1-06fb-4825-a6b8-e78cf9a4af3d/ https://minutesof.com/q/d8aceaf1-06fb-4825-a6b8-e78cf9a4af3d/ “Now, scaled to be And it has scaled to be multi million dollar Substack revenue.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says Substack revenue alone reached multi-million dollars before expanding beyond the platform. https://minutesof.com/q/f8fc3fd1-9e9d-48ed-aa41-bb5569cee196/ https://minutesof.com/q/f8fc3fd1-9e9d-48ed-aa41-bb5569cee196/ “Now, scaled to be And it has scaled to be multi million dollar Substack revenue. Obviously, built a lot of business beyond that, even that is good enough” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel released a post on GPT-3 training costs the same day ChatGPT launched, calling it coincidental timing. https://minutesof.com/q/f725d6d8-7257-4187-b74c-e547d6f947bf/ https://minutesof.com/q/f725d6d8-7257-4187-b74c-e547d6f947bf/ “I was lucky enough that the day Chad GPT released, I released a post called that was about the training cost of GPT-three. It was just like a coincidence.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says as the largest technology Substack, the 10% platform fee is worth paying for the growth. https://minutesof.com/q/46f7f4fc-5c60-4b8f-8293-dd4f74e0b5f6/ https://minutesof.com/q/46f7f4fc-5c60-4b8f-8293-dd4f74e0b5f6/ “Even though a 10% cut sounds like a lot, I'm the largest technology substack and I'm telling you that it's worth it.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says as the largest technology Substack, the 10% fee is worth it for the growth. https://minutesof.com/q/77c773c1-980a-40d7-a827-749171047acb/ https://minutesof.com/q/77c773c1-980a-40d7-a827-749171047acb/ “I'm the largest technology substack and I'm telling you that it's worth it. And so, you can think, Oh, I can save the 10%, but actually the growth is all there. Hell yeah.” — Chris Best Substack Fri, 31 Oct 2025 18:23:40 +0000 Patel says Weka and Vast make high margins on storage for multimodal AI workloads despite drive vendors making https://minutesof.com/q/b2443f6f-24bb-433c-a61e-4fea60908146/ https://minutesof.com/q/b2443f6f-24bb-433c-a61e-4fea60908146/ “But then there's also, on the storage side, the drive vendors don't make any money. But Weka and Vast, I mean, look at their pricing models.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel says storage vendors Weka and Vast make high margins while drive vendors make no money. https://minutesof.com/q/4e249973-cf25-440f-9f24-44741b266928/ https://minutesof.com/q/4e249973-cf25-440f-9f24-44741b266928/ “But then there's also, on the storage side, the drive vendors don't make any money. But Weka and Vast, I mean, look at their pricing models. They make crazy margin on storage.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel says inference providers sell public endpoints at flat or negative margins, compensating through private https://minutesof.com/q/a5a36fd6-ef64-421d-871a-a04382e1453c/ https://minutesof.com/q/a5a36fd6-ef64-421d-871a-a04382e1453c/ “Most of the inference providers are selling at flat margins or even negative for their public endpoints. And they then make it up when people do private deployments.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel describes public inference as a loss leader to generate private sovereign and enterprise engagements. https://minutesof.com/q/bd4b1861-d8a5-4553-aea4-1d8c9227603e/ https://minutesof.com/q/bd4b1861-d8a5-4553-aea4-1d8c9227603e/ “And so I think that's what a lot of this is, is you end up with the public business as like a loss leader just to generate private engagements, whether it be sovereigns or enterprises.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel reveals OpenAI's major GPT-5 training cluster is in Arizona and runs GPUs at lower power to fit more chi https://minutesof.com/q/98e652f5-3e5a-4a75-94a1-443500389683/ https://minutesof.com/q/98e652f5-3e5a-4a75-94a1-443500389683/ “And in some deployments, you'll be completely limited on power. And so there's many deployments. For example, OpenAI's major training cluster in Arizona that did GPT-five and such.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel says OpenAI and Meta run NVIDIA GPUs at lower power to fit 10% more chips despite worse TCO. https://minutesof.com/q/d22ded74-0a82-4ccb-b2a1-0abee1378bda/ https://minutesof.com/q/d22ded74-0a82-4ccb-b2a1-0abee1378bda/ “Even though it's terrible on a TCO basis, they were able to get, you know, 10% more GPUs in, and it's great. And Meta has done similar.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel quantifies that 20% performance per watt advantage translates to only 4% TCO difference on NVIDIA deploy https://minutesof.com/q/da576a59-1c4e-476c-b4e2-3b883d9367fb/ https://minutesof.com/q/da576a59-1c4e-476c-b4e2-3b883d9367fb/ “But capital is far more of a limiting factor. Because if you have enough power, a 20% difference in performance per watt only ends up being a 4% difference in TCO.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel calculates 20% power efficiency difference translates to only 4% TCO difference on NVIDIA deployments. https://minutesof.com/q/e54db0c9-c69e-4b1e-a270-5018b18ad51e/ https://minutesof.com/q/e54db0c9-c69e-4b1e-a270-5018b18ad51e/ “Because if you have enough power, a 20% difference in performance per watt only ends up being a 4% difference in TCO.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel claims NVIDIA takes a 5x markup on manufacturing cost, making power efficiency less significant for TCO. https://minutesof.com/q/b3727da7-5066-4e91-86cd-d25dc8497f23/ https://minutesof.com/q/b3727da7-5066-4e91-86cd-d25dc8497f23/ “So, in most cases, don't on an NVIDIA deployment, right? That's where NVIDIA takes 5x markup on their manufacturing cost, right?” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel states NVIDIA takes 5x markup on manufacturing cost, making power differences more significant for AMD d https://minutesof.com/q/32b9ed68-5403-47b2-8224-5391430835dc/ https://minutesof.com/q/32b9ed68-5403-47b2-8224-5391430835dc/ “If it's another deployment, if it's AMD, then that that that 20% power difference might translate to eight or 9% TCO difference when you when you talk about power cost and data center capacity cost.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel questions whether neo clouds targeting developers on short-term rentals can achieve ROI. https://minutesof.com/q/d37a5a6e-798f-497c-baac-0c713ac36c5d/ https://minutesof.com/q/d37a5a6e-798f-497c-baac-0c713ac36c5d/ “Those that are just targeting developers on short term rentals, they may not be able to get their ROI back.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel argues neo clouds with long-term enterprise and hyperscaler deals will succeed over on-demand providers. https://minutesof.com/q/938cd181-af77-4e91-8b8c-200ec3fdf886/ https://minutesof.com/q/938cd181-af77-4e91-8b8c-200ec3fdf886/ “And so those are probably less likely to be able to succeed versus those who are locking in these massive deals with or long term deals that may not be massive but with enterprises or with AI labs, or with the hyperscalers who have no capacity because the demand is just so incredible.” — Open Compute Project Thu, 23 Oct 2025 06:01:45 +0000 Patel says the standard inference deployment unit has shifted from single nodes to hundreds of GPUs. https://minutesof.com/q/ee180f12-a142-4bbb-8e53-e032a9458215/ https://minutesof.com/q/ee180f12-a142-4bbb-8e53-e032a9458215/ “the standard unit for an inference deployment being hundreds of GPUs instead of a single node. And then there's all these different things about traffic.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel notes NVIDIA Blackwell performance improved dramatically from launch to present, as did AMD hardware. https://minutesof.com/q/8a048f38-2c8d-4690-a1d2-66e094aaa6fb/ https://minutesof.com/q/8a048f38-2c8d-4690-a1d2-66e094aaa6fb/ “If you tried to use NVIDIA's Blackwell six months ago, the numbers were not amazing, right? But now they're actually amazing. So how did that progress over time? Same with AMD, right?” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel says InferenceMax runs daily benchmarks to track hardware performance improvements as software optimizes https://minutesof.com/q/c8604b65-b25b-4229-8fea-4a277c80372d/ https://minutesof.com/q/c8604b65-b25b-4229-8fea-4a277c80372d/ “And why this is important is you can see the progress of hardware over time as the software stack gets more optimized.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel quantifies InferenceMax uses tens of millions of dollars in GPU hardware with daily software updates. https://minutesof.com/q/ae8c91b8-452e-4194-a8f3-8286559ee830/ https://minutesof.com/q/ae8c91b8-452e-4194-a8f3-8286559ee830/ “we have tens of millions of dollars of hardware of the current and last generation GPUs. As I mentioned before, the software updates every single day.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel breaks down inference TCO as 20% electricity and data center, 80% hardware on NVIDIA deployments. https://minutesof.com/q/3358ce5d-fe0f-496b-a2f4-13edb43515e9/ https://minutesof.com/q/3358ce5d-fe0f-496b-a2f4-13edb43515e9/ “20% of your cost is your electricity, your data center real estate, roughly. And then the rest of the cost is that hardware, at least on a standard NVIDIA deployment your GPUs, your networking, etcetera.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel identifies GB200's power efficiency advantage but notes deployment challenges with backplane and liquid https://minutesof.com/q/d7dccc2a-7013-4972-b5c9-372d97cae003/ https://minutesof.com/q/d7dccc2a-7013-4972-b5c9-372d97cae003/ “GB200 has a huge power efficiency advantage, right? Everyone here understands the challenges of running and deploying GB200. There's a lot of challenges with the backplane.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel reports GB200 is 10x more power efficient than H200 at certain interactivity rates. https://minutesof.com/q/70c03321-ad13-4824-91b8-9ec2ebd3b7be/ https://minutesof.com/q/70c03321-ad13-4824-91b8-9ec2ebd3b7be/ “But it turns out at certain interactivity rates, I. E. Tokens per second per user, it's 10x more efficient per watt, right, compared to H200.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel quantifies GB200 as 10x more power efficient than H200 at certain interactivity rates. https://minutesof.com/q/5246be24-b713-4f40-9330-362372209a1f/ https://minutesof.com/q/5246be24-b713-4f40-9330-362372209a1f/ “at certain interactivity rates, I. E. Tokens per second per user, it's 10x more efficient per watt, right, compared to H200.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel finds AMD MI355 beats NVIDIA B200 on performance TCO in certain publicly usable configurations. https://minutesof.com/q/b9346a92-e689-4ff1-a8aa-a519c318c01e/ https://minutesof.com/q/b9346a92-e689-4ff1-a8aa-a519c318c01e/ “So we do different scenarios. We do document processing, which is 8,000 context in, 1,000 out. We do chat, which is 1,000 in, 1,000 out.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel reports AMD MI355 beats NVIDIA B200 in some document processing scenarios with open source software. https://minutesof.com/q/0ddc9f32-3943-4514-87fd-c275b179bf83/ https://minutesof.com/q/0ddc9f32-3943-4514-87fd-c275b179bf83/ “in some cases, AMD actually does have a publicly usable open source implementation that beats NVIDIA even, right, with the MI355 versus V200, which is a surprise, right?” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel reports AMD MI355 beats NVIDIA B200 on performance-TCO with publicly usable open source software in cert https://minutesof.com/q/5431d12a-a285-4545-ba7e-1391fd550419/ https://minutesof.com/q/5431d12a-a285-4545-ba7e-1391fd550419/ “AMD actually does have a publicly usable open source implementation that beats NVIDIA even, right, with the MI355 versus V200, which is a surprise, right?” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel finds 10x performance TCO advantage for Blackwell over H100 at certain operating points. https://minutesof.com/q/7eaa498f-3259-4e8c-a196-98d19ea3ec72/ https://minutesof.com/q/7eaa498f-3259-4e8c-a196-98d19ea3ec72/ “But at other points, you can have a 10x performance TCO benefit from going with a more expensive Blackwell server than a cheaper H100 server.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel shows B200 has 15x raw performance advantage over H100 but only 10x performance per TCO. https://minutesof.com/q/101225bc-685d-4ce2-a64b-7e2df33b1f98/ https://minutesof.com/q/101225bc-685d-4ce2-a64b-7e2df33b1f98/ “If we don't divide by TCO, then it looks like the performance of B200 is actually 15x that of H100, versus the performance TCO is only 10x,” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel reports AMD performance improved dramatically over two months, similar to NVIDIA Blackwell's software op https://minutesof.com/q/2ff44e18-c197-4304-a43c-ff3d7e74ecc2/ https://minutesof.com/q/2ff44e18-c197-4304-a43c-ff3d7e74ecc2/ “Over two months, the performance of AMD dramatically improved because they went from, hey, software problems are always a thing, to over a two month period, they've dramatically improved.” — Open Compute Project Thu, 23 Oct 2025 06:01:19 +0000 Patel notes early GPU systems had low volume but Supermicro continued engineering through generations anticipa https://minutesof.com/q/50295a8f-e024-4603-a38c-34b4fff1d59a/ https://minutesof.com/q/50295a8f-e024-4603-a38c-34b4fff1d59a/ “What there there wasn't much volume. Right? In those initial systems. Right? In the early time frame. Yeah. Early time frame.” — Supermicro Thu, 16 Oct 2025 17:00:47 +0000 Patel asks why Supermicro kept engineering GPU systems when volume was low and CPU systems were easier to buil https://minutesof.com/q/ff01ffaf-748d-40d9-a774-544203fcdd68/ https://minutesof.com/q/ff01ffaf-748d-40d9-a774-544203fcdd68/ “Continue to increase density, continue to build all these new building blocks around cooling and sure, you you you had buildings blocks, but most of your volume was Intel x 86 CPU.” — Supermicro Thu, 16 Oct 2025 17:00:47 +0000 Patel states Supermicro's liquid cooling reduces Hopper server power from 10 kilowatts to 7-8 kilowatts versus https://minutesof.com/q/98989f27-8eb3-4a61-b86b-0dd09799deee/ https://minutesof.com/q/98989f27-8eb3-4a61-b86b-0dd09799deee/ “Everyone else's hopper servers h 100 air cooled. And so each server consumes, you know, 10 kilowatts almost. Right?” — Supermicro Thu, 16 Oct 2025 17:00:47 +0000 Patel calculates that without Supermicro's liquid cooling solution, servers would consume 30% more power. https://minutesof.com/q/10a69170-989c-42e0-9bd5-da11329eee97/ https://minutesof.com/q/10a69170-989c-42e0-9bd5-da11329eee97/ “So that's why Shibbol Micro dedicated deep cooling so aggressively. And if the liquid cooling solution from Super Micro didn't exist, then those servers would be consuming 30% more power.” — Supermicro Thu, 16 Oct 2025 17:00:47 +0000 Patel says without Supermicro's liquid cooling, xAI would consume 30% more power or deploy 30% fewer GPUs. https://minutesof.com/q/7ac8b2d1-5cf5-4b2f-9c8e-47b0f6431195/ https://minutesof.com/q/7ac8b2d1-5cf5-4b2f-9c8e-47b0f6431195/ “And if the liquid cooling solution from Super Micro didn't exist, then those servers would be consuming 30% more power. Right?” — Supermicro Thu, 16 Oct 2025 17:00:47 +0000 Patel describes diagnostic challenges in AI data centers where failures occur across multiple infrastructure l https://minutesof.com/q/b7a7f145-e89d-4c61-b6f7-3bea9c8248c6/ https://minutesof.com/q/b7a7f145-e89d-4c61-b6f7-3bea9c8248c6/ “And people have had problems where their data center wasn't ready and something stopped working. They don't know, is it the chip? No.” — Supermicro Thu, 16 Oct 2025 17:00:47 +0000 Patel says Blackwell cost is up 50-70% over Hopper but performance exceeds 2x, making it an obvious choice. https://minutesof.com/q/7a1b8698-ea50-4724-a073-b8e5df30f151/ https://minutesof.com/q/7a1b8698-ea50-4724-a073-b8e5df30f151/ “while cost is only up, know, call it 50% or 60% or 70% versus Hopper, the performance is well north of 2x, right?” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel explains Blackwell introduces third memory tier within Tensor Cores requiring complex programming model https://minutesof.com/q/89ddcb92-6d43-43ed-8e99-976561050d18/ https://minutesof.com/q/89ddcb92-6d43-43ed-8e99-976561050d18/ “Now with Blackwell, there's actually even a third tier of memory, which is memory within the Tensor Core.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel quantifies Hopper performance improved 30-40% over 2024 with Blackwell showing similar gains in months. https://minutesof.com/q/ec5cbc83-8dcb-4132-b787-4f7c4639cd91/ https://minutesof.com/q/ec5cbc83-8dcb-4132-b787-4f7c4639cd91/ “Even Hopper in January 24 to December 2024, you still had like a 30%, 40% performance improvement. And likewise for Blackwell, you've seen a similar sort of improvement, right?” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel predicts another 50-100% Blackwell performance improvement from software optimization of standard librar https://minutesof.com/q/1220feb8-33a2-4114-958d-7e559ad3806a/ https://minutesof.com/q/1220feb8-33a2-4114-958d-7e559ad3806a/ “I see another 50% to doubling in the cards, just from the fact that there is so much low hanging fruit or performance left on the table if you are using the general open libraries, right? Whether it's cuDNN, VLM, SGLANG, all these other things,” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel predicts general-purpose Blackwell users could see another performance doubling over the next year from https://minutesof.com/q/e17fa153-5a23-488d-99aa-a4fe4fc4a705/ https://minutesof.com/q/e17fa153-5a23-488d-99aa-a4fe4fc4a705/ “for the general purpose user, yeah, you're going to have another doubling of performance potentially on the tables over the next year.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel reveals NVIDIA doubled system-level testing time for Blackwell compared to Hopper generation. https://minutesof.com/q/a16842d1-124f-447a-89d3-73aa00b75632/ https://minutesof.com/q/a16842d1-124f-447a-89d3-73aa00b75632/ “NVIDIA learning from the issues on Hopper, with Blackwell, they actually doubled the time for system level test.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel notes GB200 expanded NVLink from 8 to 72 GPUs and rack power jumped from 10 to 140 kilowatts. https://minutesof.com/q/015334ac-c584-41cb-9262-58aa5ef55a91/ https://minutesof.com/q/015334ac-c584-41cb-9262-58aa5ef55a91/ “Now you have 72 GPUs. And if you go look at the rack, right? It's completely different, right? It's liquid cooled. It's an entire rack that consumes 140 kilowatts, whereas H100 servers consumed 10 kilowatts.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel explains GB200's 72-GPU configuration creates reliability challenges compared to 8-GPU systems due to hi https://minutesof.com/q/e442c5ed-ed97-4cb7-8195-b9012b3b0fc8/ https://minutesof.com/q/e442c5ed-ed97-4cb7-8195-b9012b3b0fc8/ “If the reliability of each GPU is the same, then when a single GPU fails in 72 GPUs, you have a much higher chance of something failing, right?” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel confirms OpenAI runs production inference on GB200 despite reliability requiring workloads handle 64 of https://minutesof.com/q/63137c12-d683-47eb-9bf7-16977addf85c/ https://minutesof.com/q/63137c12-d683-47eb-9bf7-16977addf85c/ “OpenAI has said they're running production inference on GV200 a couple of months ago, in fact. Right?” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel contrasts traditional data centers with minimal retrofit costs against modern GPU deployments requiring https://minutesof.com/q/c3d1a771-4ecb-4560-805b-a76b5104d7a4/ https://minutesof.com/q/c3d1a771-4ecb-4560-805b-a76b5104d7a4/ “it used to be you'd build a data center and there wouldn't really be much retrofits cost, right? You would slide servers in and out, you would rent racks to people.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel describes Meta's temporary tent-like data centers designed for single GPU generation lifecycles instead https://minutesof.com/q/58feca5e-d55d-4e54-b391-6299edb91e6d/ https://minutesof.com/q/58feca5e-d55d-4e54-b391-6299edb91e6d/ “Meta has got a data center design that's only built to last for a single generation. They put them in these tent buildings.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel says B200 is better for training while GB200 is better for inference, reversing expected use. https://minutesof.com/q/b34c2dfe-463a-4923-8ba7-41d8b89b38a6/ https://minutesof.com/q/b34c2dfe-463a-4923-8ba7-41d8b89b38a6/ “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.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel reveals NVIDIA's next generation will split inference into separate context processing and decode worklo https://minutesof.com/q/8896a5c5-5fee-4e45-86fb-6379c1c3eec2/ https://minutesof.com/q/8896a5c5-5fee-4e45-86fb-6379c1c3eec2/ “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.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel explains NVIDIA is splitting inference into decode and prefill workloads, but provisioning for unknown f https://minutesof.com/q/e37df648-a474-40f3-9e9c-b531f05a918e/ https://minutesof.com/q/e37df648-a474-40f3-9e9c-b531f05a918e/ “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.” — Together AI Fri, 03 Oct 2025 20:23:30 +0000 Patel says Huawei brought seven nanometer AI chips to market first in 2020 with minimal gap to NVIDIA. https://minutesof.com/q/bcf648c8-741d-4bb8-a8f4-c7267c7bba53/ https://minutesof.com/q/bcf648c8-741d-4bb8-a8f4-c7267c7bba53/ “In 2020, they released an Ascend chip and submit it to impartial public benchmarks, and they were the first to bring seven nanometer AI chips to market.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel says Huawei trained significant models on Ascend chips before 2020 Trump ban took full effect. https://minutesof.com/q/9871fa14-c7f3-4c9f-a84a-0f093f92a2ce/ https://minutesof.com/q/9871fa14-c7f3-4c9f-a84a-0f093f92a2ce/ “The the full ban. And so they were only able to make a small volume of these chips, but they had trained significant models on these chips that they made then.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel reports Huawei acquired 2.9 million chips worth $500 million from TSMC through shell companies before ge https://minutesof.com/q/e898c1a1-99db-492b-b3b5-38d0af412c82/ https://minutesof.com/q/e898c1a1-99db-492b-b3b5-38d0af412c82/ “they were able to acquire 3,000,000 chips, 2,900,000 chips from TSMC through these other entities. Right?” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel estimates NVIDIA had north of $20 billion in China H20 revenue that had to be written off after ban. https://minutesof.com/q/182b465b-febd-4c89-945d-f61e809f643a/ https://minutesof.com/q/182b465b-febd-4c89-945d-f61e809f643a/ “Our our revenue estimate for NVIDIA in China for just h 20 was north of 20,000,000,000 because that's what they were booking in capacity slash had to write off.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel estimates NVIDIA had over $20 billion in H20 China revenue booked before ban and write-off. https://minutesof.com/q/873bfecf-fed4-4a82-8a67-6f8af1945e7a/ https://minutesof.com/q/873bfecf-fed4-4a82-8a67-6f8af1945e7a/ “Our our revenue estimate for NVIDIA in China for just h 20 was north of 20,000,000,000 because that's what they were booking in capacity slash had to write off. Yeah. And then it got banned.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel says China increased lithography imports to 40% of equipment spending to stockpile before bans took effe https://minutesof.com/q/74a26a69-d440-47d7-955e-d42c8b775a7c/ https://minutesof.com/q/74a26a69-d440-47d7-955e-d42c8b775a7c/ “They were importing lithography at a much higher rate than that. Right? Like, 40% of their equipment imports were lithography, and they were stockpiling lithography equipment.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel forecasts hyperscaler CapEx at $455-500 billion for next year versus Wall Street consensus of $360 billi https://minutesof.com/q/b3245cbb-ccb0-4af1-8480-c9d4f7170c7d/ https://minutesof.com/q/b3245cbb-ccb0-4af1-8480-c9d4f7170c7d/ “The consensus for the banks is $360,000,000,000 of spend next year across all of them. And my number is closer to, like it's, like, $45,500.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel says OpenAI's Oracle deal reaches over $90 billion annually within a few years from $300 billion total c https://minutesof.com/q/c6681870-7b3d-4498-a68c-5822a6af03c1/ https://minutesof.com/q/c6681870-7b3d-4498-a68c-5822a6af03c1/ “OpenAI, which signed a $300,000,000,000 plus deal with with Oracle, will actually be able to pay $300,000,000,000, right, across raising capital and revenue?” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel recounts that NVIDIA ordered Xbox production volume before receiving Microsoft's official order. https://minutesof.com/q/43bca0f7-bb8f-4e77-8816-f0ed9665b52a/ https://minutesof.com/q/43bca0f7-bb8f-4e77-8816-f0ed9665b52a/ “No. No. No. Like, NVIDIA ordered the volume for the Xbox before Microsoft gave them the order.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel predicts AWS revenue growth will reaccelerate after consistent deceleration due to Anthropic and new dat https://minutesof.com/q/62b39bc8-1e44-4f9d-b9b5-d25bb37cda66/ https://minutesof.com/q/62b39bc8-1e44-4f9d-b9b5-d25bb37cda66/ “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.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel explains Oracle's $300 billion OpenAI bet is lower risk because Oracle only secures data centers, not ex https://minutesof.com/q/4f1a803b-02b6-4160-a021-a571922390fd/ https://minutesof.com/q/4f1a803b-02b6-4160-a021-a571922390fd/ “Now, of course, the bet is a bit like there's a bit more security in the bet in that Oracle really only needs to secure the data center capacity.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel compares GPU purchasing process to buying drugs, calling people to ask about availability and pricing. https://minutesof.com/q/443520da-cfa4-4eff-940e-d05dc484f5c3/ https://minutesof.com/q/443520da-cfa4-4eff-940e-d05dc484f5c3/ “my opinion on how you buy GPUs is that it's like buying cocaine or any other drug. This is described to me, not me.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel compares buying GPUs to buying cocaine with informal texts asking for availability and pricing. https://minutesof.com/q/f57615a3-1a5a-4c7f-97ee-78062d7419e1/ https://minutesof.com/q/f57615a3-1a5a-4c7f-97ee-78062d7419e1/ “You call up a couple people. You text a couple people. You ask, yo. How much you got? What's the price? It's like Exactly. This is fucking like, buy drugs. Like oh, sorry. Sorry.” — a16z Mon, 22 Sep 2025 13:01:51 +0000 Patel predicts Anthropic's revenue growth will match OpenAI's by 2027 based on current trajectory. https://minutesof.com/q/6bed82d7-4772-4838-8261-4284e45720c8/ https://minutesof.com/q/6bed82d7-4772-4838-8261-4284e45720c8/ “Their revenue growth rate right now implies that they would be OpenAI sometime in 2027, which is a really, really big deal in terms of revenue.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel dismisses possibility of unofficial MCP servers for sensitive company data like Slack or SAP. https://minutesof.com/q/f7b08fd6-71ee-4d01-b175-e629d36b44a6/ https://minutesof.com/q/f7b08fd6-71ee-4d01-b175-e629d36b44a6/ “There's not gonna be an unofficial MCP running out there. Right? Like, you know, it's like like, fuck off. Right? Like, that's just not gonna happen.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel says the Oracle-OpenAI deal involves many gigawatts of compute with unprecedented four-year revenue guid https://minutesof.com/q/794ae51e-d9a2-4787-9473-c5976f458ff5/ https://minutesof.com/q/794ae51e-d9a2-4787-9473-c5976f458ff5/ “It is literally many gigawatts. And then in addition, right, not just about securing OpenAI ridiculous amounts of compute, but what what was really interesting is Oracle gave a guidance of four years.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel notes Oracle gave unprecedented four-year revenue guidance, something no company has ever done before. https://minutesof.com/q/69b97c71-6f05-4d9c-a31d-de2f4939e642/ https://minutesof.com/q/69b97c71-6f05-4d9c-a31d-de2f4939e642/ “Oracle gave a guidance of four years. They told everyone what their revenue is gonna be for the next four years. No company has ever done that.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel explains Microsoft's hesitation: OpenAI has only $14 billion ARR but wants $300 billion in compute. https://minutesof.com/q/258c1bc3-248a-45d6-9b0a-dd0621adbd80/ https://minutesof.com/q/258c1bc3-248a-45d6-9b0a-dd0621adbd80/ “How are you gonna how are you gonna pay for this? By the end of the year, it's like 14 right now. Right? Billion dollars on an ARR basis.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel describes Anthropic as cult-like with more unified AGI vision and mission focus than OpenAI. https://minutesof.com/q/690761af-fd93-465d-b50e-8ec9407901bb/ https://minutesof.com/q/690761af-fd93-465d-b50e-8ec9407901bb/ “They are a bit of a cult. Right? They've they believe in the mission much more than any other lab. Not that OpenAI people aren't cultists as well.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel explains Anthropic's strategy prioritizes solving software engineering and computer use to accelerate AG https://minutesof.com/q/aa6641ef-0753-4c71-b758-39a780e0e6ea/ https://minutesof.com/q/aa6641ef-0753-4c71-b758-39a780e0e6ea/ “First, we have to solve software engineering, and then we can use that to build all the other pieces. Right? First, we have to solve computer use.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel notes Anthropic's seven cofounders have equal equity and left OpenAI with clear unified mission focus. https://minutesof.com/q/f8556bad-3337-4bf5-9fe1-739ea3a3c95d/ https://minutesof.com/q/f8556bad-3337-4bf5-9fe1-739ea3a3c95d/ “Anthropic, you know, sort of has you know, they had seven cofounders all with, like, equal equity. And these these cofounders were, like, very focused on exactly this mission.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel contrasts Anthropic's multiple technical cofounders with OpenAI where Sam Altman is not technical. https://minutesof.com/q/4ea001aa-719a-431a-b36b-8d30a4cd3f10/ https://minutesof.com/q/4ea001aa-719a-431a-b36b-8d30a4cd3f10/ “They've got cofounders who are extremely technical, whereas OpenAI, right, you know, they they they don't have necessarily like, Sam's not super technical. You know, the obvious Greg Greg is a cofounder, he's super technical.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel notes YouTube holds unique video data that gives Google training advantage despite scraping restrictions https://minutesof.com/q/dd59c5bc-0967-4aab-8e50-e1bfb22888a9/ https://minutesof.com/q/dd59c5bc-0967-4aab-8e50-e1bfb22888a9/ “I don't know what ungodly percentage of video is only on available on YouTube. And sure, you can scrape YouTube, blah blah blah.” — David Ondrej Thu, 18 Sep 2025 16:14:25 +0000 Patel states GB 200 servers from Nvidia cost well over $3 million, preventing time-slicing users on individual https://minutesof.com/q/669d560d-e7af-4122-a8b8-ac5a43484620/ https://minutesof.com/q/669d560d-e7af-4122-a8b8-ac5a43484620/ “These servers cost hundreds of thousands of dollars if nothing new. Know GB 200 from Nvidia cost well over $3,000,000 so the cost of these things is huge” — Nebius Tue, 02 Sep 2025 17:00:38 +0000 Patel notes new NVIDIA servers require 140 kilowatts per rack versus prior CPU data centers at kilowatts. https://minutesof.com/q/63eff2f4-b831-448a-9c9c-c753a026eda9/ https://minutesof.com/q/63eff2f4-b831-448a-9c9c-c753a026eda9/ “Data centers for CPUs, 10 megawatts of rack, 12 megawatts of rack. Kilowatts. The new ones, kilowatts, right? Kilowatts of rack, sorry. The new NVIDIA servers, 140 kilowatts of rack.” — Nebius Tue, 02 Sep 2025 17:00:38 +0000 Patel reports 10 gigawatts of AI infrastructure deploying in the US over next 18-24 months. https://minutesof.com/q/616fa759-90a5-4968-97a0-ebcc059b0b38/ https://minutesof.com/q/616fa759-90a5-4968-97a0-ebcc059b0b38/ “In The US collectively there's like 10 gigawatts going up over the next year and a half, two years. And Europe is trying to scale to gigawatt deployments as well.” — Nebius Tue, 02 Sep 2025 17:00:38 +0000 Patel states each gigawatt AI deployment costs tens of billions of dollars end-to-end. https://minutesof.com/q/5246d720-be43-4fb0-a34a-d42ebce98657/ https://minutesof.com/q/5246d720-be43-4fb0-a34a-d42ebce98657/ “Each of these gigawatt deployments once you take it from all the way from the power to the GPUs and the cloud contract, those are tens of billions of dollars each gigawatt.” — Nebius Tue, 02 Sep 2025 17:00:38 +0000 Patel lists NVIDIA's structural advantages across networking, HBM, process nodes, speed to market, and supplie https://minutesof.com/q/af9d9adc-9470-4ff6-aba9-8433597b2e9f/ https://minutesof.com/q/af9d9adc-9470-4ff6-aba9-8433597b2e9f/ “NVIDIA's gonna have better networking than you. They're gonna have better, HBM. They're gonna have better process node. They're gonna come to market faster.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel argues competing with NVIDIA requires leaping forward beyond supply chain advantages across networking, https://minutesof.com/q/1bef91d6-c75f-476c-a33f-1cec565d5243/ https://minutesof.com/q/1bef91d6-c75f-476c-a33f-1cec565d5243/ “They're gonna have better negotiations with whether it's TSMC or SK Hynix and the memory and silicon side or all the rack people or, like, copper cables, everything, they're gonna have better cost efficiency.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel explains OpenAI's router will send low-value queries to cheaper models but use expensive compute for hig https://minutesof.com/q/0502aa27-d821-4087-8f4b-18a954c32235/ https://minutesof.com/q/0502aa27-d821-4087-8f4b-18a954c32235/ “if the user asks a low value query like, hey, why is the sky blue? Just route them to mini. The model can answer perfectly fine, and that is a chunk of queries.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 But if they ask, what's the best DUI lawyer near me? All of a sudden, this is like, you're in jail, you have o https://minutesof.com/q/2a7513eb-91b9-4429-a69d-901a47dbe6a2/ https://minutesof.com/q/2a7513eb-91b9-4429-a69d-901a47dbe6a2/ “But if they ask, what's the best DUI lawyer near me? All of a sudden, this is like, you're in jail, you have one shot, you're like, screw it.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel reports 10% of Etsy traffic comes from ChatGPT but OpenAI currently makes nothing from it. https://minutesof.com/q/a48588fc-a3d1-48bb-8548-dc1db9674bf2/ https://minutesof.com/q/a48588fc-a3d1-48bb-8548-dc1db9674bf2/ “Etsy. 10% of their traffic now comes from chat, and OpenAI makes nothing off of that. But they very they really, really will soon. Right?” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel says OpenAI and Anthropic are receiving 30% of all GPU chips being produced this year. https://minutesof.com/q/c914f140-c01e-4389-9fe1-0e490d20dbdc/ https://minutesof.com/q/c914f140-c01e-4389-9fe1-0e490d20dbdc/ “30% of the chips are going to them, just those two companies. But that's actually like, okay, well, 70% of the stuff, who's making off well, one third of it is ads, whether it be ByteDance or Meta or many of the other people who are doing ads.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel argues OpenAI captures less than 10% of the economic value it creates, pointing to broken value capture https://minutesof.com/q/199ecef6-6d3c-429f-9abf-ec74dca58297/ https://minutesof.com/q/199ecef6-6d3c-429f-9abf-ec74dca58297/ “I legitimately believe OpenAI is not even capturing 10% of the value they've created in the world already just by usage of chat. Right?” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel explains competitors need 5x hardware advantage over NVIDIA but risk failure if AI workloads shift befor https://minutesof.com/q/f8382350-ebd6-4af4-ac16-1dfdcd3fde66/ https://minutesof.com/q/f8382350-ebd6-4af4-ac16-1dfdcd3fde66/ “So now you need to do something, you know, that will give you five x advantage, right, in hardware efficiency for a certain type of workload, and then pray the workload doesn't shift.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel forecasts NVIDIA revenue exceeding $300 billion next year as infrastructure spending reaches nation-stat https://minutesof.com/q/8f99a7b9-a139-456b-b5be-913d28f9bdd8/ https://minutesof.com/q/8f99a7b9-a139-456b-b5be-913d28f9bdd8/ “NVIDIA's revenue this year is gonna be, like, over $200,000,000,000, and next year expects over 300,000,000,000 plus Google's gonna spend, like, $50,000,000,000 on TPU data centers.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel says GPU hardware and networking represent 80% of data center costs, with power and cooling only 20%. https://minutesof.com/q/3a4b0c19-d93c-478e-bced-c99f4ea7645a/ https://minutesof.com/q/3a4b0c19-d93c-478e-bced-c99f4ea7645a/ “It's the GPU purchases. It's the networking. It's the it's the physical data center conversion power conversion equipment. All of this stuff is, like, 80% of the cost.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel states 80% of GPU data center cost is capital equipment, only 20% is land, power, and cooling. https://minutesof.com/q/f14c478d-6c68-4537-b652-a709482a769a/ https://minutesof.com/q/f14c478d-6c68-4537-b652-a709482a769a/ “It's the it's the physical data center conversion power conversion equipment. All of this stuff is, like, 80% of the cost. And then 20% is gonna be your land and your power and your cooling” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel reports new Trump tax bill allows year-one GPU depreciation, worth $10 billion annually to Meta alone. https://minutesof.com/q/9d270451-5810-4a8b-88be-ef972643dcfc/ https://minutesof.com/q/9d270451-5810-4a8b-88be-ef972643dcfc/ “the new the new Trump, you know, tax bill institutes something really incredible, which is that you can depreciate all of the GPU cluster cost in year one, which we put out, like, a note about how, like, the tax implications to, like, meta are, like, $10,000,000,000 a year. And across each of the major hyperscalers, it's, like, massive.” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel suggests NVIDIA should use its massive tax bill to invest directly in data center infrastructure despite https://minutesof.com/q/4e3cea3e-1a95-46d5-b4a6-f37d4d80467b/ https://minutesof.com/q/4e3cea3e-1a95-46d5-b4a6-f37d4d80467b/ “Now this is obviously gonna be, like, crazy because, like, now they're buying g p their own GPUs and putting them in data centers and doing stuff, and they're competing with their own customers,” — a16z Mon, 18 Aug 2025 18:26:43 +0000 Patel reveals DeepSeek inference implementation requires 160 GPUs worth over $10 million of hardware per repli https://minutesof.com/q/13cb1933-7756-463f-ac46-7f8cd066bee0/ https://minutesof.com/q/13cb1933-7756-463f-ac46-7f8cd066bee0/ “That's over $10,000,000 of hardware, and then that's just one replica, then you'll have a lot of replicas and you share the caching servers between them.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 So this is something that I've found very interesting is that we've been trying to build a lot of alternative https://minutesof.com/q/36f6f6ad-0c49-48c4-a802-5b0dcbc90315/ https://minutesof.com/q/36f6f6ad-0c49-48c4-a802-5b0dcbc90315/ “So this is something that I've found very interesting is that we've been trying to build a lot of alternative data sources for token usage, who's using tokens, what models, where, etcetera, why, and it's very clear that people aren't actually using the reasoning models that much in API. Anthropic has eclipsed OpenAI and API revenue,” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel reports alternative data shows reasoning models have low actual API usage despite availability. https://minutesof.com/q/46303177-e2ef-49ef-af33-cd38aa780554/ https://minutesof.com/q/46303177-e2ef-49ef-af33-cd38aa780554/ “So this is something that I've found very interesting is that we've been trying to build a lot of alternative data sources for token usage, who's using tokens, what models, where, etcetera, why, and it's very clear that people aren't actually using the reasoning models that much in API.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel claims Anthropic has overtaken OpenAI in API revenue, primarily from Claude 4 non-reasoning mode for cod https://minutesof.com/q/daaf9730-0cb4-43fb-b75b-0043d311a230/ https://minutesof.com/q/daaf9730-0cb4-43fb-b75b-0043d311a230/ “Anthropic has eclipsed OpenAI and API revenue, and their API revenue is primarily not thinking. It's Cloud four, but it's not in the thinking mode, code being the biggest use case that's skyrocketing.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel explains AI hardware companies made architectural bets that failed when model architectures evolved in u https://minutesof.com/q/2245cf47-f50e-40dd-a9e2-a6cf6769af61/ https://minutesof.com/q/2245cf47-f50e-40dd-a9e2-a6cf6769af61/ “But still the model's way too big to fit on it. This is, like, very simple. Right? You know, the same thing's happening in the other direction.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel reveals Meta is building temporary tent structures to house GPUs because permanent buildings take too lo https://minutesof.com/q/b326f427-0392-4b1f-853a-c662e2574a6b/ https://minutesof.com/q/b326f427-0392-4b1f-853a-c662e2574a6b/ “Meta is literally building these, like, temporary, like, tent structures to put GPUs in because building the building takes too long, and it takes too much labor, as you mentioned labor.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel describes former Yandex workers paid bonuses for speed who use drugs to complete data center buildouts f https://minutesof.com/q/578afda3-bf30-482f-8983-ed64a039e7e0/ https://minutesof.com/q/578afda3-bf30-482f-8983-ed64a039e7e0/ “Like and they get paid bonuses for being faster, and therefore, do, like, certain drugs to be able to finish the build outs faster.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel questions xAI's valuation exceeding Anthropic's despite no leading model, crediting only their fast Colo https://minutesof.com/q/ef80dc2d-4fab-4c3c-8060-7f9e75eedb8a/ https://minutesof.com/q/ef80dc2d-4fab-4c3c-8060-7f9e75eedb8a/ “What has xAI actually done to deserve their prior funding rounds? They haven't released a leading edge model, and yet their evaluation's higher than Anthropic today.” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel questions xAI's valuation exceeding Anthropic's despite lacking a leading model, crediting infrastructur https://minutesof.com/q/9670f4ca-5b3e-44bf-a674-a7e422907676/ https://minutesof.com/q/9670f4ca-5b3e-44bf-a674-a7e422907676/ “They haven't released a leading edge model, and yet their evaluation's higher than Anthropic today. At least Anthropic's racing, It's Elon, A, and B, they've tackled a problem creatively” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 But, like, you know, can you give them no GPUs? No. They're gonna retaliate. Like, there is a middle ground, a https://minutesof.com/q/809be6a0-41ad-4fb7-8a0f-e92bbfeb3ebc/ https://minutesof.com/q/809be6a0-41ad-4fb7-8a0f-e92bbfeb3ebc/ “But, like, you know, can you give them no GPUs? No. They're gonna retaliate. Like, there is a middle ground, and, like, Huawei is eventually going to have a lot of production capacity,” — No Priors: AI, Machine Learning, Tech, & Startups Thu, 14 Aug 2025 10:01:35 +0000 Patel explains that Apple's rushed model has empty experts not receiving tokens due to flawed sparse MOE routi https://minutesof.com/q/5c9549e9-6d93-4429-9083-1ee365536c97/ https://minutesof.com/q/5c9549e9-6d93-4429-9083-1ee365536c97/ “Basically in between every layer, the router can route to whatever expert it wants to, And it learns which expert to route to. And and each expert learns its own independent things” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel explains how MOE routing works and how some of Meta's experts weren't being used at all. https://minutesof.com/q/d8c531cb-5cc2-4ab8-b1ae-b23a834b6b94/ https://minutesof.com/q/d8c531cb-5cc2-4ab8-b1ae-b23a834b6b94/ “And and each expert learns its own independent things and it's like really not something observable by people. But what you can see is tokens when which experts do they route to?” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel reports Google is backing out of Scale AI, spending around $250 million this year but reducing future co https://minutesof.com/q/d9341f65-1bb3-4b38-9092-51830292234d/ https://minutesof.com/q/d9341f65-1bb3-4b38-9092-51830292234d/ “Google's backing out, I think they're gonna spend on the order I've heard like $250,000,000 this year with them. And they're backing out.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel reports Google is spending $250 million with Scale AI this year but backing out of the partnership. https://minutesof.com/q/7edfcbe6-09cd-4e5f-b9c3-98fada7e2599/ https://minutesof.com/q/7edfcbe6-09cd-4e5f-b9c3-98fada7e2599/ “And they're backing out. Obviously, they've spent a lot of money and there's stuff they can't back out of, but it's like, that's gonna go down a lot.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel reports OpenAI cut the Slack connection with Scale AI. https://minutesof.com/q/8805dd87-2b2c-403c-a933-d029b61625e5/ https://minutesof.com/q/8805dd87-2b2c-403c-a933-d029b61625e5/ “OpenAI allegedly, like, cut the external Slack connection. Right? So there's no, like, slack between scale and OpenAI anymore.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel identifies a major shift in Zuckerberg's strategy toward pursuing superintelligence recently. https://minutesof.com/q/3027fb59-b19b-4ed3-8f01-768980525958/ https://minutesof.com/q/3027fb59-b19b-4ed3-8f01-768980525958/ “He was chasing like AI is good and great, but like AGI is not a thing that is gonna happen soon. So sort of this is a big shift in strategy” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel directly contradicts Sam Altman's claim that no top researchers have left OpenAI for Meta. https://minutesof.com/q/b7b6c883-2efe-4bf7-8a6c-bee2be63fcd9/ https://minutesof.com/q/b7b6c883-2efe-4bf7-8a6c-bee2be63fcd9/ “As far as Sam is saying that no top researchers have gone, I don't believe that's accurate. I initially the top researchers definitely did say no, the best researchers, the best people.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel claims Meta offered over $1 billion to one OpenAI researcher as retention or acquisition offer. https://minutesof.com/q/6cadac54-f54f-42a8-b1ad-6eb1de375fd1/ https://minutesof.com/q/6cadac54-f54f-42a8-b1ad-6eb1de375fd1/ “And you said $100,000,000 I've heard a number for someone over 1,000,000,000 actually, for one person at OpenAI. But anyways, you know, it's a ridiculous amount of money,” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel reports hearing Meta offered over one billion dollars to a single OpenAI researcher. https://minutesof.com/q/622a3afb-844b-45b2-afe4-3976cf946d69/ https://minutesof.com/q/622a3afb-844b-45b2-afe4-3976cf946d69/ “I've heard a number for someone over 1,000,000,000 actually, for one person at OpenAI. But anyways, you know, it's a ridiculous amount of money,” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel reveals OpenAI had a PyTorch bug in GPT-4.5 training code for months during the training run. https://minutesof.com/q/8a64130a-fd6f-4253-91a0-4c9a48c95c1b/ https://minutesof.com/q/8a64130a-fd6f-4253-91a0-4c9a48c95c1b/ “And they had a bug in the training code for a couple months that was like a very tiny bug that like was messing up the training.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel describes the Bumpgate incident where NVIDIA laptop GPUs had defective solder balls. https://minutesof.com/q/b108bf9a-1e2f-46f2-bbe0-8821812e2686/ https://minutesof.com/q/b108bf9a-1e2f-46f2-bbe0-8821812e2686/ “There's a generation of NVIDIA GPUs for laptops. Right? And and chips have solder balls on the bottom. Right?” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel explains Bumpgate issue where thermal expansion differences caused GPU solder ball failures in Apple lap https://minutesof.com/q/0a5f3bdf-1c4b-45e3-ae44-c701f4cf2fee/ https://minutesof.com/q/0a5f3bdf-1c4b-45e3-ae44-c701f4cf2fee/ “And what ended up happening is because of that different rate of expansion, the solder balls connecting the chip and the board would crack.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel reports AMD is selling GPUs to cloud providers then renting them back to inflate sales numbers. https://minutesof.com/q/15677d9a-d37a-4f1d-b396-5d8e98ca9589/ https://minutesof.com/q/15677d9a-d37a-4f1d-b396-5d8e98ca9589/ “Like but AMD is actually doing this and like taking it to overdrive. They're getting clusters at Oracle and Amazon and Crusoe and Digital Ocean and Tensorwave and they're renting GPUs back from them.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel predicts China will stop open sourcing AI models once ahead and closed source will ultimately win. https://minutesof.com/q/4728732c-0771-4e99-ad96-f8a818be3df3/ https://minutesof.com/q/4728732c-0771-4e99-ad96-f8a818be3df3/ “China's open sourcing stuff only because they're behind the moment they're ahead, they will stop open sourcing stuff. And at the end of the day, closed source will win. Unfortunately, closed source will win.” — Matthew Berman Mon, 30 Jun 2025 17:27:19 +0000 Patel asserts America cannot manufacture robots and Germany/Japan are losing capabilities to China. https://minutesof.com/q/997a2cfd-1cdf-4dd0-96d4-8fbdc8d4daf3/ https://minutesof.com/q/997a2cfd-1cdf-4dd0-96d4-8fbdc8d4daf3/ “Germany and and Japan are losing a lot of their capabilities to do so slash the cost of their robots are just way, way, way higher than what China can do.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel reports AI models improved from under 4% to 70% on Frontier Math in six months. https://minutesof.com/q/9274e9a4-6baa-4e2a-b44a-3c2207f42c82/ https://minutesof.com/q/9274e9a4-6baa-4e2a-b44a-3c2207f42c82/ “Like, can't do all the math, all the problems in frontier math. And the models went from sub 4% to 70% in six months.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel reports AI models improved from below 4% to 70% accuracy on Frontier Math in six months across multiple https://minutesof.com/q/2bb1c7b0-0d8f-45e8-a0a1-db8ef64742e6/ https://minutesof.com/q/2bb1c7b0-0d8f-45e8-a0a1-db8ef64742e6/ “frontier math. And the models went from sub 4% to 70% in six months. Right? And this is from not just OpenAI, but also Anthropic and Google and DeepSeek.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel states that Japan, USA, Korea, and Germany combined deploy only a third of China's robots in 2023. https://minutesof.com/q/9b53d10e-c58f-46cd-88ed-84b2e6b98f81/ https://minutesof.com/q/9b53d10e-c58f-46cd-88ed-84b2e6b98f81/ “You fast forward to 2023, and all four of these countries combined, right, are still a third of China's robot deployments.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel notes that as recently as 2022, China imported the majority of its robots from Japan and Germany. https://minutesof.com/q/f3ff6dd3-43c7-4d41-b081-6e9737c24256/ https://minutesof.com/q/f3ff6dd3-43c7-4d41-b081-6e9737c24256/ “even as near as twenty twenty two, the majority of China's robots were being imported. Right? Japan and Germany being the largest market share players,” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel notes China was importing majority of robots as recently as 2022 from Japan and Germany. https://minutesof.com/q/0a5a65d1-e5d1-4965-9589-d9be71347bee/ https://minutesof.com/q/0a5a65d1-e5d1-4965-9589-d9be71347bee/ “even as near as twenty twenty two, the majority of China's robots were being imported. Right? Japan and Germany being the largest market share players, but also USA, Korea is also in there.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel forecasts that 70% of robots deployed in China in 2025 will be domestically manufactured Chinese robots. https://minutesof.com/q/1269ac38-2f98-4cc6-8b06-3dfdfcf4efda/ https://minutesof.com/q/1269ac38-2f98-4cc6-8b06-3dfdfcf4efda/ “'24, it's, it's rapidly growing Chinese share. And '25 is it looks clear, like, the the market shares is, you know, gonna be 60 is gonna be, like, 70% Chinese robots for China's domestic usage.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel reveals that China purchased two top German robotics companies and transferred their servo and actuator https://minutesof.com/q/9ecd3a10-280c-4dea-a552-8bc663d46ac8/ https://minutesof.com/q/9ecd3a10-280c-4dea-a552-8bc663d46ac8/ “Two of Germany's best robotics companies were actually purchased by China, And a lot of their, servos and actuator construction was actually transferred to China.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel contrasts Western price-to-value strategy with Chinese cost-plus pricing model in manufacturing. https://minutesof.com/q/e5cd5238-fd61-4c78-813a-7ecc288bfeff/ https://minutesof.com/q/e5cd5238-fd61-4c78-813a-7ecc288bfeff/ “Whereas Chinese companies are almost all cost plus. They're not price to value. They're cost plus. Right? And so you've got this interesting phenomenon where,” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel explains Chinese companies use cost-plus pricing instead of price-to-value, avoiding margin expansion de https://minutesof.com/q/600aa10f-ce79-4d5e-9b6d-49919dc5e77c/ https://minutesof.com/q/600aa10f-ce79-4d5e-9b6d-49919dc5e77c/ “Whereas Chinese companies are almost all cost plus. They're not price to value. They're cost plus. Right?” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel explains Chinese manufacturers use cost-plus pricing rather than value-based pricing across industries. https://minutesof.com/q/9820aa6d-0224-4a6c-b2b6-f2965992e508/ https://minutesof.com/q/9820aa6d-0224-4a6c-b2b6-f2965992e508/ “They're not price to value. They're cost plus. Right? And so you've got this interesting phenomenon where, there are industries in China where the entire world relies on it.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel describes Chinese manufacturers as locked in constant cost reduction cycle rather than margin expansion. https://minutesof.com/q/017a5839-c71d-42bb-951b-8594f045c861/ https://minutesof.com/q/017a5839-c71d-42bb-951b-8594f045c861/ “there's this constant pressure cooker of continual cost reductions and engineering. It's not like, Oh, well, we make 80% margins. Our cost is low.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel notes many Chinese companies dominating global markets have flat or declining stock prices. https://minutesof.com/q/dd8f0bd2-599a-4b49-97c9-a2d5257ccc86/ https://minutesof.com/q/dd8f0bd2-599a-4b49-97c9-a2d5257ccc86/ “there's so many companies on the Hangshan, the Hangshan, which is a stock market in China, right, whose stock has literally been flat or down despite the fact that they own the global market.” — Going Direct Sun, 22 Jun 2025 21:30:00 +0000 Patel says most people will respond to outreach but almost nobody tries. https://minutesof.com/q/e3216dfe-1ee8-4f3d-a316-41c745765f04/ https://minutesof.com/q/e3216dfe-1ee8-4f3d-a316-41c745765f04/ “One of things I learned early on is you can talk to anyone on the Internet. You can just reach out to anyone, and most people just say yes, but no one does it.” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel advises becoming a foremost expert in what you enjoy and reaching out to anyone online. https://minutesof.com/q/1be9840b-45d2-4c02-bcb7-323d7f483ed2/ https://minutesof.com/q/1be9840b-45d2-4c02-bcb7-323d7f483ed2/ “Or try and talk to those people because one of the things I learned early on is you can talk to anyone on the Internet,” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel says his business model shifted so that 95% of revenue now comes from selling datasets. https://minutesof.com/q/bea63ab4-2811-4e8f-a074-3fe239719551/ https://minutesof.com/q/bea63ab4-2811-4e8f-a074-3fe239719551/ “the business transformed into, you know, selling data, right, or build building datasets and selling them. Right? And when the business truly flipped to that, we're now 95% of the revenue is that data” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel says his company now employs 26 people distributed across seven countries globally. https://minutesof.com/q/c7d348ba-06b9-4f7f-a015-79e6340a38fe/ https://minutesof.com/q/c7d348ba-06b9-4f7f-a015-79e6340a38fe/ “There's 26 people who work at the company now. Cool. All over the world. Right? US, Japan, Taiwan, Singapore, France, Germany Uh-huh. Canada. Yeah. Yeah. Well, that'll be one less country. Right? He's gone.” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel finds it surreal that billionaires care about his semiconductor analysis. https://minutesof.com/q/9a54ea52-7534-4747-9d3e-774e23e154d9/ https://minutesof.com/q/9a54ea52-7534-4747-9d3e-774e23e154d9/ “it's it's it's surreal to see, like, billionaires, like, actually care what you have to say or rather what you, like, have figured out.” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel reflects on the surreal experience of billionaires caring about his research and analysis. https://minutesof.com/q/ff80999a-f83f-49a3-83a1-aa274c443aeb/ https://minutesof.com/q/ff80999a-f83f-49a3-83a1-aa274c443aeb/ “it's surreal to see, like, billionaires, like, actually care what you have to say or rather what you, like, have figured out.” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel reports SemiAnalysis has 200,000 newsletter subscribers with 5,000 paid subscribers representing under 5 https://minutesof.com/q/a439877b-b27b-41f7-94a6-b721d9b1a7b5/ https://minutesof.com/q/a439877b-b27b-41f7-94a6-b721d9b1a7b5/ “So the newsletter itself is, like, 200,000. Paid newsletter is, like, I think, like, 200,000 people. Paid newsletter is, like, 5,000. Nice. But then, like, the main bit like, that's less than 5% of the” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel argues most AI value accrues to users, not to companies building or deploying AI. https://minutesof.com/q/226f8c49-62ac-4fa4-852f-180f149e9ea3/ https://minutesof.com/q/226f8c49-62ac-4fa4-852f-180f149e9ea3/ “But most of the AI value that's been generated does not accrue to the companies building the AI or deploying it. It actually deploy it actually accrues to the user.” — Origins with Sonith Tue, 10 Jun 2025 17:13:06 +0000 Patel notes CoreWeave got H100 services online six months before every hyperscaler in many cases. https://minutesof.com/q/c2e290f0-d29d-46af-b1bb-46bc871b30f2/ https://minutesof.com/q/c2e290f0-d29d-46af-b1bb-46bc871b30f2/ “Get services online with h one hundreds before every hyperscaler by a factor of, like, as much as, like, six months in many cases.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel notes CoreWeave has fully liquid cooled data centers while hyperscalers are still working on designs. https://minutesof.com/q/566773ba-9e75-4496-9248-3d50712a59c2/ https://minutesof.com/q/566773ba-9e75-4496-9248-3d50712a59c2/ “They're doing full liquid cooling as well and they're working on those designs, but, you guys have fully liquid cooled data centers.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel says CoreWeave can build $10 billion clusters from start to completion much faster than competitors. https://minutesof.com/q/373f08c1-885e-4cb0-8a0c-139ecb2143a4/ https://minutesof.com/q/373f08c1-885e-4cb0-8a0c-139ecb2143a4/ “it lets you get to these, you know, call it $10,000,000,000 clusters, right, in in time scales, from project start to completion that are much much shorter than anyone else.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel observes hyperscaler data centers started in 2023 remain incomplete even with expedited equipment delive https://minutesof.com/q/fbed374a-9a4e-4927-8341-d57bb4e923a3/ https://minutesof.com/q/fbed374a-9a4e-4927-8341-d57bb4e923a3/ “Even if they even if they had the electrical equipment in in you know, not on lead time, they paid a bunch to skip the queue.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel reveals his team found security flaws in other clouds allowing them to snoop customer traffic via Infini https://minutesof.com/q/f9895260-561c-429c-b8d2-d830ec137096/ https://minutesof.com/q/f9895260-561c-429c-b8d2-d830ec137096/ “of work. This is this is actually some stuff that we've noticed on other clouds that we've tested, whether it's, you know, things related to InfiniBand p keys and we're able to snoop in other nodes of real live customer traffic. And, obviously, we tell them, like, we're like, yo, this is not okay.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel describes finding security vulnerabilities on other clouds allowing access to customer traffic and serve https://minutesof.com/q/e408fc7c-cb88-4405-9fd1-759b753cd3d7/ https://minutesof.com/q/e408fc7c-cb88-4405-9fd1-759b753cd3d7/ “things related to InfiniBand p keys and we're able to snoop in other nodes of real live customer traffic.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel describes a 10,000+ GPU customer whose engineers preferred CoreWeave infrastructure over their acquirer' https://minutesof.com/q/4f8c1afb-2ba0-475b-98df-5a501781f6e4/ https://minutesof.com/q/4f8c1afb-2ba0-475b-98df-5a501781f6e4/ “What's what's funny is there was a customer that had more than 10,000 GPUs with you. They got acquired. They had to start using some clusters at their acquiring company, the company that acquired them.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel describes customer with over 10,000 GPUs whose engineers demanded CoreWeave infrastructure after acquisi https://minutesof.com/q/7715340c-a1ac-4173-ad56-984305ab5c68/ https://minutesof.com/q/7715340c-a1ac-4173-ad56-984305ab5c68/ “What's what's funny is there was a customer that had more than 10,000 GPUs with you. They got acquired.” — CoreWeave Fri, 23 May 2025 17:27:49 +0000 Patel illustrates that language models historically failed at basic math comparisons without reasoning capabil https://minutesof.com/q/2e5f94d7-83fe-4c76-abc4-b23f582bac0a/ https://minutesof.com/q/2e5f94d7-83fe-4c76-abc4-b23f582bac0a/ “And it would say, yes. It's bigger. Even though, like, everyone knows that nine dot one one is is way smaller than nine dot nine.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel reports GPT-3 to Llama 3.2 costs fell 1,200x while GPT-4 to DeepSeek v3 costs fell 600x. https://minutesof.com/q/c28a19fc-d8f5-4e4a-bc59-e7dabb126c8c/ https://minutesof.com/q/c28a19fc-d8f5-4e4a-bc59-e7dabb126c8c/ “when we looked at g p d three, the cost fell 1,200 x from g p d three's initial cost to what you can get Lama 3.23 b today.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel argues the surprise was a Chinese company achieving this cost reduction, not the reduction itself. https://minutesof.com/q/a64a56f5-df99-4993-bd99-c6e00a54110d/ https://minutesof.com/q/a64a56f5-df99-4993-bd99-c6e00a54110d/ “I think what was really surprising was that it was a Chinese company for the first time. Right? Because Google and and OpenAI and Anthropic and Meta have all traded blows.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel says DeepSeek's cost efficiency follows the expected trend line, just from an unexpected source. https://minutesof.com/q/42277ecd-5d1b-4270-a90f-11eb7f3a67b8/ https://minutesof.com/q/42277ecd-5d1b-4270-a90f-11eb7f3a67b8/ “It's not unexpected. Right? Like, this is actually within the trend line of what happened with GPT three is happening to GPT four level quality with DeepSeq.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel predicts Meta's upcoming Llama will achieve similar cost decreases to DeepSeek v3 without causing alarm. https://minutesof.com/q/ca383355-8cf6-4301-8b4c-28ca79458cd8/ https://minutesof.com/q/ca383355-8cf6-4301-8b4c-28ca79458cd8/ “And Meta's Meta's gonna release their new llama soon enough. Right? And that one is gonna be, you know, a similar level of cost decrease,” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel argues that efficiency gains alone without capability increases cannot justify massive AI infrastructure https://minutesof.com/q/675ee285-ed22-4889-b6cf-5f6b6892d5c7/ https://minutesof.com/q/675ee285-ed22-4889-b6cf-5f6b6892d5c7/ “That would not pay for all of these build outs. Right? AI is useful today, but it's not capable of doing a lot of things.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel states $10 billion data centers target automated software engineering, not chat models. https://minutesof.com/q/119203f3-2761-4bf3-a619-aaad40c2f71a/ https://minutesof.com/q/119203f3-2761-4bf3-a619-aaad40c2f71a/ “So no one is trying to make with these $10,000,000,000 data centers, they're not trying to make chat models. Right?” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel reports that OpenAI's Orion training run did not improve enough over GPT-4.5 to qualify as GPT-5. https://minutesof.com/q/8373ef31-935a-4e8d-90e3-7c6565db225c/ https://minutesof.com/q/8373ef31-935a-4e8d-90e3-7c6565db225c/ “There were hopes that Orion could be used for for GPT five, but its improvement was, like, not enough to be, like, really a GPT five.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel explains GPT-5 will combine massive pre-training and post-training scale for the first time. https://minutesof.com/q/1ac7f4fc-3b6a-4d9e-a854-d3c4009a5164/ https://minutesof.com/q/1ac7f4fc-3b6a-4d9e-a854-d3c4009a5164/ “Along the way, another team at OpenAI made the big breakthrough of reasoning. Right? Strawberry training. And they released o one, and then they released o three.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel says GPT-5 will combine massive pre-training like GPT-4.5 with massive post-training like o1 and o3. https://minutesof.com/q/808fcaf5-9216-420e-b11b-f8a91aeadae8/ https://minutesof.com/q/808fcaf5-9216-420e-b11b-f8a91aeadae8/ “GPT five, as Sam calls it, is is gonna be a model that has huge pre training scale, right, like GPT 4.5, but also huge post training scale like o one and o three and continuing to scale that up.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel says GPT-5 will be first model to massively scale both pre-training and post-training simultaneously. https://minutesof.com/q/00545820-383f-45e6-ac01-d8e99c981b60/ https://minutesof.com/q/00545820-383f-45e6-ac01-d8e99c981b60/ “Right? And this would be the first time we see a model that was was a step up in both at the same time.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel reports GPT-5 is expected in three to six months, possibly sooner based on his sources. https://minutesof.com/q/51f3ebb3-3c32-467e-a122-ed7787e9ba81/ https://minutesof.com/q/51f3ebb3-3c32-467e-a122-ed7787e9ba81/ “They say it's coming, you know, this year, hopefully, the next three to six months, maybe sooner. I've heard sooner, but, you know, we'll we'll see.” — Alex Kantrowitz Wed, 23 Apr 2025 16:30:06 +0000 Patel says certain states will see over half their power consumed by AI within a few years. https://minutesof.com/q/50566b09-394e-4967-9d53-46dfd6777a8f/ https://minutesof.com/q/50566b09-394e-4967-9d53-46dfd6777a8f/ “When we look at power consumption by state, something pretty simple, there are certain states where more than half their power is going be consumed by AI just in the next few years.” — MedBricks Webcast Thu, 27 Mar 2025 13:55:08 +0000 Patel says reasoning capability allowing models to think before answering dramatically increased capabilities https://minutesof.com/q/fb9c1dec-7583-4ee8-88a9-767fd9bc2d46/ https://minutesof.com/q/fb9c1dec-7583-4ee8-88a9-767fd9bc2d46/ “Right? And this has just happened in six months. And the same applies to various certain coding tasks. The same applies to yeah.” — MedBricks Webcast Thu, 27 Mar 2025 13:55:08 +0000 Patel says AI model costs dropped 1200x from GPT-3's $60 per million outputs to current pricing. https://minutesof.com/q/41d2ba76-f8de-4943-bca0-4e696568c0ea/ https://minutesof.com/q/41d2ba76-f8de-4943-bca0-4e696568c0ea/ “GPT three was it cost, you know, it cost $60 for the million output. And over time, that kept reducing. Right? OpenAI released new models.” — MedBricks Webcast Thu, 27 Mar 2025 13:55:08 +0000 Patel argues existing AI capabilities get cheaper but GPT-4 level models still cannot do medical tasks. https://minutesof.com/q/0e93ea93-703b-45ed-a930-c31730a78cc1/ https://minutesof.com/q/0e93ea93-703b-45ed-a930-c31730a78cc1/ “But this is the progression you get because of the advancement. That's one way to look at it is, oh, existing capabilities get cheaper. But as useful as GPT-four is, it can't do medical tasks.” — MedBricks Webcast Thu, 27 Mar 2025 13:55:08 +0000 Patel says GPUs in a 16,000 GPU cluster show up to 10% speed variation from manufacturing differences. https://minutesof.com/q/e4cb87fa-869e-4b13-8b85-bd4d3fb6cb18/ https://minutesof.com/q/e4cb87fa-869e-4b13-8b85-bd4d3fb6cb18/ “But even within a cluster of like 16 k GPUs, GPUs will be slower by up to 10%. Right? So there there's a lot of variation even in the manufacturing.” — Prime Intellect AI Tue, 25 Mar 2025 18:21:45 +0000 Patel reports Meta's 16,000 GPU cluster had failures every few hours while next-generation models train on 100 https://minutesof.com/q/04dfd209-35a0-4bd2-95b6-e803801d15c4/ https://minutesof.com/q/04dfd209-35a0-4bd2-95b6-e803801d15c4/ “And that was a measly 16,000 GPUs. Right? And when you look at like the next generation Llamas, the next generation GPTs, etcetera, they're training on a 100,000 GPUs or more.” — Prime Intellect AI Tue, 25 Mar 2025 18:21:45 +0000 Patel explains training a 70 billion parameter model requires passing four times that data every training step https://minutesof.com/q/2bc44e6d-a415-482b-bcb0-f9c59a1b4198/ https://minutesof.com/q/2bc44e6d-a415-482b-bcb0-f9c59a1b4198/ “So if you're training like a 70,000,000,000 parameter model, you need to pass four x that number every single training step. And when you're passing that, you can't really overlap communications and compute.” — Prime Intellect AI Tue, 25 Mar 2025 18:21:45 +0000 Patel says cloud pricing varies by three times due to poor marketing, sales, and stranded compute capacity. https://minutesof.com/q/a4358e2e-3665-42ec-9a3a-0df1412c0fb5/ https://minutesof.com/q/a4358e2e-3665-42ec-9a3a-0df1412c0fb5/ “Like the the cost difference between different clouds can be as much as three x. Right? Because a lot of these clouds are bad at marketing, bad at sales, stranded compute.” — Prime Intellect AI Tue, 25 Mar 2025 18:21:45 +0000 Patel argues $52 billion Chips Act is insufficient to restore US to even 20% of global semiconductor productio https://minutesof.com/q/72bb5d51-a196-4a6e-b7f6-63cb3776b6d5/ https://minutesof.com/q/72bb5d51-a196-4a6e-b7f6-63cb3776b6d5/ “S. To even 20% of global production, let alone a higher number, which is what we'd like, going back to maybe the 80s levels, right?” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel argues TSMC's $100 billion US investment still relies on Taiwan R&D, not securing supply chain. https://minutesof.com/q/0db47aa8-b79a-47e8-9231-c3d0fed5f45e/ https://minutesof.com/q/0db47aa8-b79a-47e8-9231-c3d0fed5f45e/ “Even if they do fully do the full $100,000,000,000 they commit to, completely still relying on R and D in Taiwan.” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel argues TSMC's $100 billion US investment won't secure R&D supply chain or new node development. https://minutesof.com/q/6efba2e7-6161-4f23-8644-c2e007650885/ https://minutesof.com/q/6efba2e7-6161-4f23-8644-c2e007650885/ “The R and D fab they announced is not going to secure R and D supply chain, right? It's not going to secure new node development, right?” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel says TSMC approaches 20% of Taiwan's power consumption, projecting 30-40% within five years. https://minutesof.com/q/9a24a996-f514-477e-88a8-0882e5e5068f/ https://minutesof.com/q/9a24a996-f514-477e-88a8-0882e5e5068f/ “And if they were to continue to build out what we expect them to build out in the next four or five years, it's going get to 30%, 40%.” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel reports ASML had 45% of equipment and more of revenue from China despite export controls. https://minutesof.com/q/6d450b52-b2bc-457a-ae6e-52cf7219326f/ https://minutesof.com/q/6d450b52-b2bc-457a-ae6e-52cf7219326f/ “On equipment though, last year ASML, many quarters they've had 45% of their equipment and even more of their revenue come from China.” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel reports ASML had 45% of equipment and more of revenue from China despite export controls. https://minutesof.com/q/89c9d86f-ca5b-45bd-b16f-f7837e38d48f/ https://minutesof.com/q/89c9d86f-ca5b-45bd-b16f-f7837e38d48f/ “On equipment though, last year ASML, many quarters they've had 45% of their equipment and even more of their revenue come from China. And likewise for ASML, Lam Research and all these companies, right?” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel challenges DeepSeek's GPU count claims, citing job ads promising tens of thousands of GPUs. https://minutesof.com/q/c235c852-3da5-495d-8118-6bceb59270eb/ https://minutesof.com/q/c235c852-3da5-495d-8118-6bceb59270eb/ “The estimates that we have is, so first of all, ads in China, they say they have tens of thousands of GPUs for researchers, right?” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel identifies a 40 gigawatt power shortfall for US data centers even with all planned capacity additions. https://minutesof.com/q/6d70bd42-e4be-4dec-aa00-6d77aa2b2ba6/ https://minutesof.com/q/6d70bd42-e4be-4dec-aa00-6d77aa2b2ba6/ “There is an over 40 gigawatt hole in power for The US. Even if we don't turn off any more coal, in fact we restart some, we take all the natural gas dual combined cycle reactor deliveries that we're going to have, and you add in renewables, you have a 40 gigawatt hole of power production in The US.” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel identifies a 40 gigawatt power deficit in US even with all planned gas and restarted coal. https://minutesof.com/q/f0927cde-cafd-4663-b2f9-55ca37f74d17/ https://minutesof.com/q/f0927cde-cafd-4663-b2f9-55ca37f74d17/ “Even if we don't turn off any more coal, in fact we restart some, we take all the natural gas dual combined cycle reactor deliveries that we're going to have,” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel says current models use 100,000 GPUs while next generation will require hundreds of thousands or million https://minutesof.com/q/561bd369-2c88-4ea1-8fa4-3d02389be50d/ https://minutesof.com/q/561bd369-2c88-4ea1-8fa4-3d02389be50d/ “And next generation models that are trained on hundreds of thousands or even millions GPUs, right?” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel says NVIDIA made 4 million GPUs last year and will produce 7 million this year. https://minutesof.com/q/08661a67-8385-4a47-bc96-ca62dcc63083/ https://minutesof.com/q/08661a67-8385-4a47-bc96-ca62dcc63083/ “Nvidia made over 4,000,000 GPUs last year, they're making over 7,000,000 this year, right? High end data center GPUs.” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel reports China imported one million H20 GPUs in recent quarters, enough for largest cluster. https://minutesof.com/q/2e9b3ee4-e698-4516-a8b0-920ae1a288c3/ https://minutesof.com/q/2e9b3ee4-e698-4516-a8b0-920ae1a288c3/ “Just in Q3, Q4 and the early part of Q1 this year, they imported a million H20s, Right?” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel says China imported one million H20 GPUs in recent quarters, enough to build largest cluster. https://minutesof.com/q/cd6e818a-a8c6-4fc5-b11d-96afaffc62cd/ https://minutesof.com/q/cd6e818a-a8c6-4fc5-b11d-96afaffc62cd/ “This is more than enough if they even took 30 of it, concentrated it, to be a bigger cluster than any American company has.” — Special Competitive Studies Project Thu, 13 Mar 2025 16:42:15 +0000 Patel explains DeepSeek v3 base is trained once, then post-trained differently to create chat versus reasoning https://minutesof.com/q/02fd4a40-e301-43f7-be83-33801b8010a9/ https://minutesof.com/q/02fd4a40-e301-43f7-be83-33801b8010a9/ “This reasoning model has a lot of overlapping training steps to DeepSeek v three, and it's confusing that you have a base model called v three that you do something to to get a chat model, and then you do some different things to get a reasoning model.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel asserts data quality is the primary determinant of model quality, not just compute or architecture. https://minutesof.com/q/b3b4b1a2-1ca3-4201-b3c0-cf7ef1ef9a7d/ https://minutesof.com/q/b3b4b1a2-1ca3-4201-b3c0-cf7ef1ef9a7d/ “the data processing, data filtering, data quality is the number one determinant of the model quality, and then a lot of the training code is the determinant on how long it takes to train and how fast your experimentation is.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel says DeepSeek modifies code at or below NVIDIA's CUDA layer, a rare technical capability. https://minutesof.com/q/99d651cc-75d0-4621-9e78-21bd90eeb8aa/ https://minutesof.com/q/99d651cc-75d0-4621-9e78-21bd90eeb8aa/ “For example, on their to get highly efficient training, they're making modifications at or below the CUDA layer for NVIDIA chips.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel explains post-training involves instruction tuning for formatting and preference fine-tuning derived fro https://minutesof.com/q/8c62650a-d063-46b7-bd52-ad27170f814d/ https://minutesof.com/q/8c62650a-d063-46b7-bd52-ad27170f814d/ “One, I will classify as preference fine tuning. Preference fine tuning is a generalized term for what came out of reinforcement learning from human feedback, which is RLHF.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel states mixture of experts architecture can achieve the same model performance with 30% less compute. https://minutesof.com/q/ea8addd1-3523-4de1-80ec-07b2ce862d42/ https://minutesof.com/q/ea8addd1-3523-4de1-80ec-07b2ce862d42/ “So you can get effectively the same performance model and evaluation scores with numbers like 30% less compute.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel notes DeepSeek's routing innovation removing auxiliary loss represents compounding small improvements ov https://minutesof.com/q/fab5c10d-e2cd-4e52-b056-9586751d1efb/ https://minutesof.com/q/fab5c10d-e2cd-4e52-b056-9586751d1efb/ “this type of change can be big, it can be small, but they add up over time.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel argues export controls primarily limit AI inference deployment in China, not frontier model training cap https://minutesof.com/q/9e102380-dbd9-4dbb-acd8-a43420fa4c7f/ https://minutesof.com/q/9e102380-dbd9-4dbb-acd8-a43420fa4c7f/ “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.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel argues export controls aim to limit AI usage scale in China, not prevent AGI development. https://minutesof.com/q/38d113f9-ba13-446c-ad37-db93f239378f/ https://minutesof.com/q/38d113f9-ba13-446c-ad37-db93f239378f/ “And I think that is a much easier goal to achieve than trying to debate on what AGI is.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel explains reasoning models dramatically increase memory usage and reduce batch size, multiplying serving https://minutesof.com/q/fb69a564-812e-47e6-9f3d-e6f80b0878ca/ https://minutesof.com/q/fb69a564-812e-47e6-9f3d-e6f80b0878ca/ “So your your memory usage is going way up with these reasoning models, and you still have a lot of users.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel explains reasoning models dramatically increase serving costs due to long output context and memory cons https://minutesof.com/q/8ea24247-feff-4e7f-beec-571f6109d01d/ https://minutesof.com/q/8ea24247-feff-4e7f-beec-571f6109d01d/ “your memory usage is going way up with these reasoning models, and you still have a lot of users. So effectively, the cost to serve multiplies by a ton.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel says DeepSeek r one zero shows reasoning behaviors emerge naturally from RL on verifiable rewards withou https://minutesof.com/q/31f04244-2e31-4194-a9b6-0a74afee5544/ https://minutesof.com/q/31f04244-2e31-4194-a9b6-0a74afee5544/ “So it's the remarkable thing about these reasoning results, and especially the DeepSeek r one paper, is this result that they call DeepSeek r one zero, which is they took one of these pretrained models.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel argues open source AI lacks software's feedback loops because reusing models requires significant comput https://minutesof.com/q/6b9ff42b-cc92-4e07-b515-b520175a1821/ https://minutesof.com/q/6b9ff42b-cc92-4e07-b515-b520175a1821/ “fundamentally, I would say that that's because open source AI does not have the same feedback loops as open source software.” — Lex Fridman Mon, 03 Feb 2025 00:12:13 +0000 Patel says Malaysia will add three gigawatts of data center capacity by 2027, mostly Chinese companies claimin https://minutesof.com/q/c375dcd6-7e32-46f5-a9a4-79f02129b7f8/ https://minutesof.com/q/c375dcd6-7e32-46f5-a9a4-79f02129b7f8/ “Malaysia from 2024 to 2027, not the country itself, but companies operating there, mostly Chinese companies, many of them claiming they're now Singaporean companies. Right?” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel says Chinese data center operators are relocating to Singapore to circumvent US export controls while bu https://minutesof.com/q/7d009a67-b093-47ed-966c-83d2122784a3/ https://minutesof.com/q/7d009a67-b093-47ed-966c-83d2122784a3/ “Right? Like, the largest operator in China of data centers in China, GDS, moved to Singapore and says they're Singaporean now.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 these companies are building three gigawatts of data center capacity. Put that in context, at the beginning of https://minutesof.com/q/afcf6a66-d9b9-4954-89cb-8abfbc9fb7ca/ https://minutesof.com/q/afcf6a66-d9b9-4954-89cb-8abfbc9fb7ca/ “these companies are building three gigawatts of data center capacity. Put that in context, at the beginning of twenty twenty four, that was roughly Meta's global footprint.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel explains the 7% rule limits non-US data centers, benefiting only Microsoft, Meta, Amazon, and Google. https://minutesof.com/q/b9f3e38a-33ab-4cfa-8322-55a0ed1bc5ec/ https://minutesof.com/q/b9f3e38a-33ab-4cfa-8322-55a0ed1bc5ec/ “So, like, it's, like, Microsoft, Meta, Amazon, Google. Right? These four companies have, you know you know, 70 plus percent of their data center AI data center capacity in The US.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel reveals each country is capped at 50,000 GPUs for four years while NVIDIA makes 6 million annually. https://minutesof.com/q/d1ba7229-6f9c-4393-9fae-bd85ea69d963/ https://minutesof.com/q/d1ba7229-6f9c-4393-9fae-bd85ea69d963/ “There is there is one obvious loophole, which is well, there's, like, strict caps. Right? Like, each country can only buy 50,000 GPUs for the next four years.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel identifies the 50,000 GPU cap per country as trivial compared to NVIDIA's six million unit annual produc https://minutesof.com/q/2930999e-1d82-4023-b604-3dbdb15a6461/ https://minutesof.com/q/2930999e-1d82-4023-b604-3dbdb15a6461/ “Like, each country can only buy 50,000 GPUs for the next four years. And it's like, that's kinda nothing when NVIDIA's making, you know, 6,000,000 plus this year.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel calculates next-generation clusters deliver 15x more compute through five times more GPUs and 3x perform https://minutesof.com/q/badb57b8-ea1b-45cb-887a-c8b1bba9f5d8/ https://minutesof.com/q/badb57b8-ea1b-45cb-887a-c8b1bba9f5d8/ “So you got you have five x the GPUs, and you have three x the performance per GPU roughly. So then you're at, like, 15 x more compute.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel breaks down full H100 GPU cost at $40,000 to $45,000 including networking and infrastructure versus $24, https://minutesof.com/q/837f0a3c-d3fb-454e-ba7f-3504bfd5a136/ https://minutesof.com/q/837f0a3c-d3fb-454e-ba7f-3504bfd5a136/ “each GPU, right, once you include networking, building, all this sort of stuff maybe is you know, the GPU itself of h 100 is, like, 24,000, but once you add everything else up, it's, like, forty, forty five thousand dollars per GPU all in of everything.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel says Oracle is spending over $10 billion this year building infrastructure for OpenAI, not Microsoft. https://minutesof.com/q/aa50dd43-a1af-443a-a084-7b771193c3bc/ https://minutesof.com/q/aa50dd43-a1af-443a-a084-7b771193c3bc/ “Oracle's spending, like, $10,000,000,000 plus for them this year to build out data centers and GPUs, right, for OpenAI, not Microsoft, interestingly enough.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Well, one, there's a gigawatt natural gas plant Yeah. That helps. Right next door. Right? Two, there's a natur https://minutesof.com/q/19bb2687-114c-47f5-8cb4-14f6ed6a9075/ https://minutesof.com/q/19bb2687-114c-47f5-8cb4-14f6ed6a9075/ “Well, one, there's a gigawatt natural gas plant Yeah. That helps. Right next door. Right? Two, there's a natural gas line, a main that they tapped, and they set up their own generation capacity on-site.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Meta, they accidentally open sourced this code. It's literally called PowerPlant node blowup. It's a flag. https://minutesof.com/q/b2067a77-745f-4781-85ef-e564cdd0a8f8/ https://minutesof.com/q/b2067a77-745f-4781-85ef-e564cdd0a8f8/ “Meta, they accidentally open sourced this code. It's literally called PowerPlant node blowup. It's a flag.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel says in some Midwest regions, grid transmission costs exceed power generation costs, constraining data c https://minutesof.com/q/0a5cd67c-aea8-4ed0-a66d-d5e8a221c959/ https://minutesof.com/q/0a5cd67c-aea8-4ed0-a66d-d5e8a221c959/ “It's like, what the flip? Right? Like, it's like so, like, the grid needs huge investments.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel forecasts data center power will grow from 150 megawatts today to gigawatt-scale clusters by 2026. https://minutesof.com/q/57323284-b466-4640-9fed-ea622dabe8e2/ https://minutesof.com/q/57323284-b466-4640-9fed-ea622dabe8e2/ “We're seeing, like, you know, 400, 500 megawatts cluster sort of being built out now. Right? Like you know? And and then, you know, in 2026, I'd imagine it'll be, like, gigawatt scale clusters.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Or or that's what it looks like based on or, you know, one gigawatt in twenty twenty six ish. And Meta's Meta' https://minutesof.com/q/eb6b2094-e7b7-4ac7-9761-c8cc9c5e0736/ https://minutesof.com/q/eb6b2094-e7b7-4ac7-9761-c8cc9c5e0736/ “Or or that's what it looks like based on or, you know, one gigawatt in twenty twenty six ish. And Meta's Meta's, like, trying to do, like, two gigawatts by early to mid twenty seven.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel estimates Blackwell delivers 10-15x cost improvement for inference despite NVIDIA claiming 30x at GTC. https://minutesof.com/q/879abc84-3fdc-4dd7-9b63-baf6f6541b7c/ https://minutesof.com/q/879abc84-3fdc-4dd7-9b63-baf6f6541b7c/ “But now, like, Blackwell, NVIDIA's pitching 10 to 15 x improvement in cost. It's like, well, you know, they're massaging the numbers marketing.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel reports CoreWeave has over 200,000 GPUs, billions in revenue, $20 billion valuation, and plans to IPO in https://minutesof.com/q/c8eccac4-eee6-4116-b120-54563d6b52fd/ https://minutesof.com/q/c8eccac4-eee6-4116-b120-54563d6b52fd/ “CoreWeave now has, like, 200 k plus GPUs. Right? Like Yeah. A lot of GPUs. Right? They're doing, like, you know, billions of dollars of revenue.” — Unsupervised Learning: With Jacob Effron Tue, 21 Jan 2025 14:00:12 +0000 Patel says Google built rack-scale AI systems with Broadcom in 2018 before NVIDIA deployed similar architectur https://minutesof.com/q/20d14c4c-c77e-4ede-a3ec-2619103dc76e/ https://minutesof.com/q/20d14c4c-c77e-4ede-a3ec-2619103dc76e/ “Google did something very similar in 2018, right, with the TPU. Now they couldn't do it alone. Right? They know the software.” — BG2 Pod Mon, 23 Dec 2024 20:12:56 +0000 Patel reveals o1's reasoning process sometimes switches between Chinese and English during hidden thinking pha https://minutesof.com/q/1524c47b-e0e1-402a-b743-111da2d0e959/ https://minutesof.com/q/1524c47b-e0e1-402a-b743-111da2d0e959/ “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?” — BG2 Pod Mon, 23 Dec 2024 20:12:56 +0000 Patel says Google deployed water cooling years before NVIDIA and achieves higher reliability than NVIDIA GPUs. https://minutesof.com/q/7308a31d-8fe6-41e8-bf9c-ffec00d2af16/ https://minutesof.com/q/7308a31d-8fe6-41e8-bf9c-ffec00d2af16/ “Google's brought in water cooling for years. Right? NVIDIA only just realized they needed water cooling on this generation. And Google's brought in a level of reliability that NVIDIA GPUs don't have.” — BG2 Pod Mon, 23 Dec 2024 20:12:56 +0000 Patel dubs Amazon's Trainium chip the Amazon Basics TPU due to less efficient but cost-effective design. https://minutesof.com/q/55fd7d23-c6ff-42d1-ada9-ae49a012dcd3/ https://minutesof.com/q/55fd7d23-c6ff-42d1-ada9-ae49a012dcd3/ “Yeah. So so funnily enough, Amazon's chip is the Amazon I I call it the Amazon's basics TPU. Right?” — BG2 Pod Mon, 23 Dec 2024 20:12:56 +0000 Patel describes Microsoft's shift from 48-megawatt buildings to 300-megawatt AI-specific data centers for incr https://minutesof.com/q/03d1dcb7-cdc3-4b70-a35d-96720749d321/ https://minutesof.com/q/03d1dcb7-cdc3-4b70-a35d-96720749d321/ “Here you can see that same data center, that same type of data center vertically. And that's 48 megawatts, or 57, And then this data center center is like 300 megawatts.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel reports Microsoft plans 1.5 gigawatts in one site within two and a half years across four 300 megawatt b https://minutesof.com/q/9be72128-ac52-447d-82bc-c0e5f69e03de/ https://minutesof.com/q/9be72128-ac52-447d-82bc-c0e5f69e03de/ “And so the scale that they wanna reach to, 1.5 gigawatts in two and a half years in one site.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 And so the scale that they wanna reach to, 1.5 gigawatts in two and a half years in one site. Now the interest https://minutesof.com/q/ec3b555e-1b35-441e-b9b8-346be76df418/ https://minutesof.com/q/ec3b555e-1b35-441e-b9b8-346be76df418/ “And so the scale that they wanna reach to, 1.5 gigawatts in two and a half years in one site. Now the interesting thing is that each GPU takes roughly 1,000 watts to 1,400 watts.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel says GPT-4 used 24,000 GPUs, GPT-5 uses 100,000, and Microsoft is building toward a million GPUs. https://minutesof.com/q/49107836-7a8a-4175-851d-c6bc483f45a1/ https://minutesof.com/q/49107836-7a8a-4175-851d-c6bc483f45a1/ “GPT four was trained with 24,000 GPUs roughly, and GPT five is on the order of a 100,000. And then they're trying to build this data center over the next few years. That's a million.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel reports OpenAI is going to Oracle for 200 megawatts because Microsoft cannot supply enough compute. https://minutesof.com/q/ca6793e5-5269-4f36-ad51-e3ee019b1ec2/ https://minutesof.com/q/ca6793e5-5269-4f36-ad51-e3ee019b1ec2/ “OpenAI can't get enough compute from Microsoft, so then they're also going to Oracle. And Oracle's building another gigawatt scale data center eventually. Again, a gigawatt is like 500,000 plus GPUs.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel reveals OpenAI signed a deal with Oracle for 200 megawatts because Microsoft cannot provide enough compu https://minutesof.com/q/77bb9fd8-374c-4dd6-9378-89d774c7068b/ https://minutesof.com/q/77bb9fd8-374c-4dd6-9378-89d774c7068b/ “Again, a gigawatt is like 500,000 plus GPUs. That's billions, tens of billions of dollars. So OpenAI signed a deal for just 200 megawatts.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel says OpenAI is paying roughly $12.5 billion over five years for one-fifth of Oracle's gigawatt site. https://minutesof.com/q/b0202054-42e7-4268-983f-096f72fe9917/ https://minutesof.com/q/b0202054-42e7-4268-983f-096f72fe9917/ “So that's four of these buildings with Oracle. And this site is gonna be a gigawatt. So for one fifth of the site, they're paying something like $12,500,000,000 over the next five years.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel states OpenAI is paying $12.5 billion over five years for one-fifth of Oracle's data center site. https://minutesof.com/q/58cb4df9-acea-480b-8af4-0b91ab6d00bd/ https://minutesof.com/q/58cb4df9-acea-480b-8af4-0b91ab6d00bd/ “So for one fifth of the site, they're paying something like $12,500,000,000 over the next five years.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel explains that AdamW optimizer requires four bytes per parameter, creating 400 gigabytes of data for 100B https://minutesof.com/q/9b0a4eda-79e5-4213-b0df-76f45b715526/ https://minutesof.com/q/9b0a4eda-79e5-4213-b0df-76f45b715526/ “when you look at LLMs, people use AdamW. And AdamW as an optimizer is like four bytes per parameter roughly, if I recall correctly the optimizer state.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel calculates that multi-trillion parameter models require transmitting 40 terabytes of data every two seco https://minutesof.com/q/2b1dd77d-4b29-490c-a1f9-fd32caff427b/ https://minutesof.com/q/2b1dd77d-4b29-490c-a1f9-fd32caff427b/ “They're doing it for like multi trillion. Right? So let's call it 10,000,000,000,000 parameters, four bytes parameter, that's 40 terabytes of data you need to transmit in two seconds.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel argues Western power grid supply chains have not expanded in decades unlike China, India, and Indonesia. https://minutesof.com/q/2cdca5ca-185f-47fe-8738-234bf9c8e535/ https://minutesof.com/q/2cdca5ca-185f-47fe-8738-234bf9c8e535/ “The Japanese power grid has been like this. The Korean power grid has been like Only China and India and Indonesia, their power grids have been growing really fast.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel claims four to five percent of NVIDIA GPUs fail during initial burn-in testing within two weeks. https://minutesof.com/q/36c00658-21e4-44c9-9919-c879522edbb7/ https://minutesof.com/q/36c00658-21e4-44c9-9919-c879522edbb7/ “Buy a good chip from NVIDIA, and then it ends up failing like constantly, right?” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel states OpenAI's o1 reasoning model increases average sequence length because it reasons before outputtin https://minutesof.com/q/95d406f2-b33e-4dfc-9526-6aa4bec2193d/ https://minutesof.com/q/95d406f2-b33e-4dfc-9526-6aa4bec2193d/ “This is really important because with reasoning models like OpenAI's o one, the average sequence length grows a lot.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel says o1 generates 40k sequence lengths versus 4k for standard models, requiring lower batching and highe https://minutesof.com/q/2548d082-f15b-4fca-b6d4-4cfc177ee7e5/ https://minutesof.com/q/2548d082-f15b-4fca-b6d4-4cfc177ee7e5/ “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.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel explains OpenAI's o1 model thinks for 5-20 seconds before outputting, creating inference throughput chal https://minutesof.com/q/179104b9-35d2-44c2-9d8a-85d97e3bd8f8/ https://minutesof.com/q/179104b9-35d2-44c2-9d8a-85d97e3bd8f8/ “When you use OpenAI's o one, it thinks for ten seconds, twenty seconds, five seconds. It varies a lot, but thinks for a while and then it sends you tokens.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel explains o1's thinking time creates memory bandwidth issues that prevent batching users at high levels. https://minutesof.com/q/b7e7fd8f-2e86-4705-b53e-c249725aab84/ https://minutesof.com/q/b7e7fd8f-2e86-4705-b53e-c249725aab84/ “But if you batch higher, k b cache is not just a memory capacity issue, it's also a memory bandwidth issue.” — Scaling Intelligence Tue, 12 Nov 2024 04:17:00 +0000 Patel says China still receives over a million NVIDIA H20 GPUs annually despite October sanctions. https://minutesof.com/q/9970673d-c612-4b81-8789-52135c45cc46/ https://minutesof.com/q/9970673d-c612-4b81-8789-52135c45cc46/ “Right? Because you have you have sanctions on how many NVIDIA GPUs you can get in. Now, they're still north of a million a year.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel claims China could build a gigawatt data center in six months near the Three Gorges Dam but hasn't. https://minutesof.com/q/e6ab75ac-8722-4e95-895e-afb4a981088d/ https://minutesof.com/q/e6ab75ac-8722-4e95-895e-afb4a981088d/ “China could just build it in six months, I think, around the 3 Gorges Dam or many other places. Right?” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel says China could centralize over a million NVIDIA H20 chips into one data center if scale-pilled. https://minutesof.com/q/d42e0bbe-5093-4310-8333-99146acd3603/ https://minutesof.com/q/d42e0bbe-5093-4310-8333-99146acd3603/ “they can centralize the chips like crazy. Right now, oh, oh, million chips that NVIDIA shipping in q three and q four, the h twenty, let's just put them all in this one data center.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel argues US cannot effectively sanction advanced packaging because the largest company relocated to Singap https://minutesof.com/q/5df181a5-28a4-4c6d-8379-7493d76f65f1/ https://minutesof.com/q/5df181a5-28a4-4c6d-8379-7493d76f65f1/ “where The US can control China. Right? So advanced packaging capacity is kind of a shot because the vast the the largest advanced packaging company in the world was Hong Kong headquartered.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel asserts chip export controls to China are working effectively despite smuggling through third countries. https://minutesof.com/q/37826dc7-6409-420f-8c4d-11e92e1f1770/ https://minutesof.com/q/37826dc7-6409-420f-8c4d-11e92e1f1770/ “The chip side of things is actually being controlled quite effectively, I think. Right? Like, yes, there is, like, shipping GPUs through Singapore and Malaysia and and other countries in Asia to China.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel argues US export controls are counterproductive because China can now build better chips domestically th https://minutesof.com/q/ef0e0707-39cc-45d7-8303-ba5d94e61ff9/ https://minutesof.com/q/ef0e0707-39cc-45d7-8303-ba5d94e61ff9/ “They can build better chips in China, then we restrict them in terms of chips that NVIDIA or AMD or an Intel can sell to China.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel claims China could match individual US labs in compute by 2025-26 through centralization alone using for https://minutesof.com/q/3bd7853d-7fb5-407a-9ca0-51a9419c33ce/ https://minutesof.com/q/3bd7853d-7fb5-407a-9ca0-51a9419c33ce/ “China can be centralized enough to compete with each individual US lab. They could have just as many flops in '25 and '26 if they decided they were scale built. Right? Just from foreign chips” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel says China could build a bigger model than any US lab next year using 600,000 domestic Ascend chips. https://minutesof.com/q/856f2514-a20b-461e-a657-ed57842c7d91/ https://minutesof.com/q/856f2514-a20b-461e-a657-ed57842c7d91/ “You know? So so if they put them all in one cluster, they could have a bigger model than any of the labs next year.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 if they put them all in one cluster, they could have a bigger model than any of the labs next year. Right? https://minutesof.com/q/46a70da8-5378-46a9-a527-ee41c13e3fa8/ https://minutesof.com/q/46a70da8-5378-46a9-a527-ee41c13e3fa8/ “if they put them all in one cluster, they could have a bigger model than any of the labs next year. Right?” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel says SMIC's Shanghai fab has 45 to 50 high-end lithography tools enabling 60,000 wafers monthly at seven https://minutesof.com/q/18583773-8dec-4c67-9ee5-892ce7457583/ https://minutesof.com/q/18583773-8dec-4c67-9ee5-892ce7457583/ “Shanghai has anywhere from 45 to 50 high end immersion lithography tools is is what's like believed by intelligence as well as like many other folks.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel argues AI demand is what makes continued semiconductor node advancement economically viable today. https://minutesof.com/q/c17dd34d-6330-45ef-ac94-0e2d4084112b/ https://minutesof.com/q/c17dd34d-6330-45ef-ac94-0e2d4084112b/ “funding the next node would not be economically viable anymore if it weren't for AI taking off. Right? And then generating all this humongous demand for the most leading edge chip.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel reports H100 rental prices have dropped from $3-4 per hour to $2.15 or less, approaching natural cost of https://minutesof.com/q/e377715e-4aba-45ba-9b19-f37de850e568/ https://minutesof.com/q/e377715e-4aba-45ba-9b19-f37de850e568/ “An hour. Right? For shorter term or midterm deals. Right now, it's like, if you want a six month deal, you could get, like, $2.15 or less.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel says Microsoft and OpenAI have 500,000 GB200 GPUs totaling one gigawatt coming online next year. https://minutesof.com/q/9c88df4d-4d27-4ba8-9f09-e66cb257c9e7/ https://minutesof.com/q/9c88df4d-4d27-4ba8-9f09-e66cb257c9e7/ “Microsoft slash OpenAI slash their partners for them. And then and then potentially even more, 500 k GB two hundreds, right, is a gigawatt. Right? And that's, like, online next year.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel predicts OpenAI will have 300,000 to 500,000 GPU equivalent clusters next year across multiple sites. https://minutesof.com/q/5492cf90-3909-4724-b977-ca95780911b3/ https://minutesof.com/q/5492cf90-3909-4724-b977-ca95780911b3/ “Next year, 300 to 500,000 depending on whether it's one side or many. Right? 300 to, like, 700,000, I think, is the upper bound of that.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel predicts Microsoft/OpenAI will have a million H100-equivalent chips in a single cluster by end of next y https://minutesof.com/q/fb831a44-ddf5-4c82-9e85-d209a6316700/ https://minutesof.com/q/fb831a44-ddf5-4c82-9e85-d209a6316700/ “three hundred to, like, seven 500,000, let's say, but those GPUs are two to three x faster, right, versus the 100 k cluster.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel predicts OpenAI will raise 50 to 100 billion dollars by early next year to fund planned cluster builds. https://minutesof.com/q/7b43ac1e-7f41-457a-a57e-250bea958042/ https://minutesof.com/q/7b43ac1e-7f41-457a-a57e-250bea958042/ “there's no fucking way you can pay for the scale of clusters that are being planned to be built next year for OpenAI until unless they raise, like, 50 to $100,000,000,000, which I think they will raise that, like, end of this year, early next year.” — Dwarkesh Patel Wed, 02 Oct 2024 14:59:36 +0000 Patel argues AI software will have lower R&D costs but much higher cost of goods sold from operating services. https://minutesof.com/q/cdca4ea2-53cc-42c9-a237-c2927fb44bb3/ https://minutesof.com/q/cdca4ea2-53cc-42c9-a237-c2927fb44bb3/ “Like the R and D cost is much lower in terms of people, but the cost of goods sold in terms of actually operating the service, I think will be much higher.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel argues training costs are irrelevant, noting GPT-4 used 20,000 A100s. https://minutesof.com/q/7037941a-ee80-4057-a5c4-ce393d0ea4d6/ https://minutesof.com/q/7037941a-ee80-4057-a5c4-ce393d0ea4d6/ “training costs are irrelevant. Right? Like, GPT four, right, like, 20,000 a one hundreds, that's that's like, I know it sounds like a lot of money.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel claims Google is massively ramping TPU production, a closely guarded secret most DeepMind employees don' https://minutesof.com/q/3758822d-6d28-4549-9783-84d049a8f9b1/ https://minutesof.com/q/3758822d-6d28-4549-9783-84d049a8f9b1/ “Google's actually ramping up TPU production massively. And I think people in AI would be like, Well, duh. But like, okay, who has the capability of figuring out the number?” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel says Google's TPU production numbers are closely guarded secrets that even most Google DeepMind employee https://minutesof.com/q/29476b04-d67a-4d3f-8be3-bd44f3c5dbc1/ https://minutesof.com/q/29476b04-d67a-4d3f-8be3-bd44f3c5dbc1/ “Right? That's like a very closely guarded secret, most people that work at Google DeepMind don't even know number.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel says NVIDIA will sell over 3 million total GPUs next year and over a million H100s this year. https://minutesof.com/q/2855439b-9e6d-4778-82dc-a589fdbb9562/ https://minutesof.com/q/2855439b-9e6d-4778-82dc-a589fdbb9562/ “NVIDIA is gonna sell well over 3,000,000 total GPUs next year, over a million H100s this year alone. There's a lot of GPU capacity coming online. It's an incredible amount.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel states NVIDIA will sell over one million H100s this year and over three million total GPUs next year. https://minutesof.com/q/a5596c89-b7bc-42e6-8a13-449244f817e4/ https://minutesof.com/q/a5596c89-b7bc-42e6-8a13-449244f817e4/ “NVIDIA is gonna sell well over 3,000,000 total GPUs next year, over a million H100s this year alone. There's a lot of GPU capacity coming online.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel claims Google will have more compute than any other company by a large factor. https://minutesof.com/q/bfd26312-4df9-4890-8982-2c145463702e/ https://minutesof.com/q/bfd26312-4df9-4890-8982-2c145463702e/ “Google is going to have more compute than any other company in the world period by like a large, large factor.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel says OpenAI has a lot less total FLOPS than Google despite having more GPUs than most perceive. https://minutesof.com/q/09b1d877-5ab8-4501-a5dd-1f5be2d74cac/ https://minutesof.com/q/09b1d877-5ab8-4501-a5dd-1f5be2d74cac/ “whole point was that Google you know, OpenAI, who everyone would be like, oh, yeah. They have more GPUs than anyone else. Right? But they have a lot less flops than Google.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel estimates 10 companies will have enough compute to beat GPT-4 within six months, requiring about 7,000 H https://minutesof.com/q/752d084b-d6f9-42a9-815b-8a8bab751a3d/ https://minutesof.com/q/752d084b-d6f9-42a9-815b-8a8bab751a3d/ “there's like 10 companies that have enough compute in one single data center to be able to beat GPT-four. Right? Like straight up. Like, if not today, within the next six months.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel predicts open source will match GPT-4 but frontier models will continue advancing beyond it. https://minutesof.com/q/17074139-7de1-4dd4-9c7f-5a1a0f39e4f8/ https://minutesof.com/q/17074139-7de1-4dd4-9c7f-5a1a0f39e4f8/ “Open source will match GPT-four, but then it's like, what about GPT-four Vision? Or what about five and six and all these kind of stuff?” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel predicts NVIDIA will announce a new chip in March that is three to four times better than current genera https://minutesof.com/q/2b0e0a57-650b-4c6d-9ba7-f2b279d3cbad/ https://minutesof.com/q/2b0e0a57-650b-4c6d-9ba7-f2b279d3cbad/ “NVIDIA is releasing a new chip, you know, in you know, they're gonna announce it in March and they're gonna release it, you know, and ship it, you know, q two, q three next year anyways. Right? And that chip will probably be three or four times as good.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 But, like, the the level of the value that they deliver to the world, if you talk to anyone there, they truly https://minutesof.com/q/a3b1ab28-1d05-4f90-a32c-98993e072930/ https://minutesof.com/q/a3b1ab28-1d05-4f90-a32c-98993e072930/ “But, like, the the level of the value that they deliver to the world, if you talk to anyone there, they truly believe it'll be tens of trillions if not hundreds of trillions of dollars.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel predicts OpenAI could raise $100 billion at $500 billion valuation to build a $100 billion supercomputer https://minutesof.com/q/c1488c72-704b-4ab7-9ccd-7fd6582ce89a/ https://minutesof.com/q/c1488c72-704b-4ab7-9ccd-7fd6582ce89a/ “And after GPT-five releases, if he goes to the market and says like, Hey, I want to raise $100,000,000,000 at $500,000,000,000 valuation, I'm sure the market would give it to him.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel identifies distributed training across data centers with lower bandwidth as a key unsolved problem that https://minutesof.com/q/8bdd8e52-1eff-46b4-82af-03f48966482c/ https://minutesof.com/q/8bdd8e52-1eff-46b4-82af-03f48966482c/ “Everything that we've seen so far is that large scale training has to happen in an individual data center with very high speed networking.” — Latent Space Tue, 05 Dec 2023 18:16:28 +0000 Patel argues export restrictions clearly failed because China manufactured advanced Huawei chip demonstrating https://minutesof.com/q/89ea37a4-8221-4263-80d9-10e67c0b12d8/ https://minutesof.com/q/89ea37a4-8221-4263-80d9-10e67c0b12d8/ “The export restrictions have clearly failed because China is still able to manufacture. And what's surprising about the chip that Huawei did is, you know, it's a phone chip.” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel explains TSMC and Intel achieved seven nanometer manufacturing using older DUV lithography with multi-pa https://minutesof.com/q/940850fc-7265-42d9-9733-544123ca524d/ https://minutesof.com/q/940850fc-7265-42d9-9733-544123ca524d/ “TSMC and Intel were able to do it with what's called multi patterning DUV, deep ultraviolet. Right? So that just means they use the DUV lithography many times, right, for simplicity's sake.” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel says TSMC still uses DUV for seven nanometer despite having fifty EUV tools because DUV proved superior. https://minutesof.com/q/b609ca8d-71f1-4c47-b2ca-68b6d4ae9b0c/ https://minutesof.com/q/b609ca8d-71f1-4c47-b2ca-68b6d4ae9b0c/ “So TSMC, to this day, even though they've been use they have over 50 EUV tools and they make hundreds of thousands of wafers a month with EUV, still uses DUV for their seven nanometer.” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel explains TSMC found DUV superior to EUV for seven nanometer on yield, cost, and performance after trying https://minutesof.com/q/bed52b60-a2a3-49ce-9fb2-f4b80bc34583/ https://minutesof.com/q/bed52b60-a2a3-49ce-9fb2-f4b80bc34583/ “TSMC tried both. Right? And so at six nanometer and five nanometer, they use EUV, but they think cost efficiency wise, yield, right, which is how good the product's coming out, how many are good, how many need to be thrown away, the yield is better, the cost is better, the performance is better with DUV rather than EUV.” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel notes the unrestricted 1980i tool continues receiving upgrades through 2023, creating a regulatory looph https://minutesof.com/q/78d6a540-b866-4c76-b5e3-da9f0e853aaa/ https://minutesof.com/q/78d6a540-b866-4c76-b5e3-da9f0e853aaa/ “But remember the 1980i has continued to be upgraded and there's an upgraded version in 2023, right? And that tool, sort of the line is in between those tool, right?” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel says ASML projects China will reach 30% semiconductor market share, displacing Western suppliers and cre https://minutesof.com/q/3c2975c4-a35a-47bb-934c-c115dfc88c6a/ https://minutesof.com/q/3c2975c4-a35a-47bb-934c-c115dfc88c6a/ “And, you know, of course it makes sense for them. Right? Like, let me have my own domestic industry, but it's gonna create an excess supply dynamic.” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel says ASML projects China will grow from under ten percent to thirty percent chip manufacturing market sh https://minutesof.com/q/db4011a8-5ecf-48bf-830e-72b29bc57d6a/ https://minutesof.com/q/db4011a8-5ecf-48bf-830e-72b29bc57d6a/ “Go from sub 10% market share in the chip industry to somewhere around 30% market share in the manufacturing industry, right?” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel estimates SMIC's seven nanometer yield at 60-70% based on tool orders and Huawei's phone shipment volume https://minutesof.com/q/c4576003-a819-4d73-be31-8de84338ddba/ https://minutesof.com/q/c4576003-a819-4d73-be31-8de84338ddba/ “And the third was looking at the volumes and such, right? So how many tools have they ordered?” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel estimates SMIC yield by analyzing tool orders since companies don't publicly disclose China sales. https://minutesof.com/q/a201cd77-27d6-488a-95a2-1b47357f4bf1/ https://minutesof.com/q/a201cd77-27d6-488a-95a2-1b47357f4bf1/ “And you have to do a lot of napkin math to figure this out because no one publicly releases how many tools SMIC is buying from them.” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel says Chinese firms offer TSMC engineers five hundred thousand dollars versus eighty to one-fifty thousan https://minutesof.com/q/86ac248f-4cf7-4321-a389-93eb8a7668ea/ https://minutesof.com/q/86ac248f-4cf7-4321-a389-93eb8a7668ea/ “In TSMC, they're getting paid, you know, only like 80 to 150 k. Right? Like USD. Right?” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel argues the unrestricted 1980i tool can reach five nanometer with only modest cost increases on total chi https://minutesof.com/q/72d16d9a-f03c-4f62-aadb-daaecf589490/ https://minutesof.com/q/72d16d9a-f03c-4f62-aadb-daaecf589490/ “And many folks including, not just myself, but many folks in the industry as well think that the 1980i could be used for five nanometer, right?” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel calculates doubling lithography cost only increases total chip cost 10 percent, enabling five nanometer. https://minutesof.com/q/a81c789d-69a2-44e4-83c3-7fc48f4f956d/ https://minutesof.com/q/a81c789d-69a2-44e4-83c3-7fc48f4f956d/ “Because again, the manufacturing cost of lithography is only one fourth to one fifth. So if you double lithography costs, you're only increasing the total cost of a chip by 10%, right?” — Hidden Forces Wed, 04 Oct 2023 14:00:36 +0000 Patel says NVIDIA is shipping more GPU flops this year than in its entire data center history combined. https://minutesof.com/q/234de4bb-be7b-4a2f-a173-c5f9da008904/ https://minutesof.com/q/234de4bb-be7b-4a2f-a173-c5f9da008904/ “There's more GPU flops shipping this year that NVIDIA shipped their entire history for the data center.” — The Inside View Wed, 09 Aug 2023 15:00:23 +0000 Patel says GPU orders take four to six months from placement to data center installation. https://minutesof.com/q/fb13f24f-d903-4521-9be1-a668dcec5b78/ https://minutesof.com/q/fb13f24f-d903-4521-9be1-a668dcec5b78/ “call it four or five five, six months between, you know, when an order is placed and you can actually have it installed in your data center if it got worked on immediately.” — The Inside View Wed, 09 Aug 2023 15:00:23 +0000 Patel forecasts NVIDIA will ship over one million H100 and A100 GPUs this year. https://minutesof.com/q/7ad7302b-f526-4151-806b-845410bbaa88/ https://minutesof.com/q/7ad7302b-f526-4151-806b-845410bbaa88/ “NVIDIA is gonna ship over a million, h one hundreds plus a a one hundreds this year.” — The Inside View Wed, 09 Aug 2023 15:00:23 +0000 Patel predicts five to seven companies will train GPT-4 scale models in the next year. https://minutesof.com/q/fa2d7b01-051f-42f8-ab72-c441507a20aa/ https://minutesof.com/q/fa2d7b01-051f-42f8-ab72-c441507a20aa/ “I do believe that there's going to be five to seven companies that will have a GPT-four size model, at least, right, in terms of total flops.” — The Inside View Wed, 09 Aug 2023 15:00:23 +0000 Patel states NVIDIA will build and sell approximately 400,000 H100s in Q3. https://minutesof.com/q/62d0ab02-d94f-4006-9057-aec70b68289f/ https://minutesof.com/q/62d0ab02-d94f-4006-9057-aec70b68289f/ “they're gonna build at about 400,000 GPUs and and sell about 400,000 h one hundreds in q three.” — The Inside View Wed, 09 Aug 2023 15:00:23 +0000 Patel explains NVIDIA strategically allocates GPUs to new entrants like Inflection to maintain multiple custom https://minutesof.com/q/36709192-9c41-4805-b49e-5a798fe2dc88/ https://minutesof.com/q/36709192-9c41-4805-b49e-5a798fe2dc88/ “it was like, oh, well, who's this random startup? Like, do I wanna get them 22,000 GPUs this year? Well, actually, yeah.” — The Inside View Wed, 09 Aug 2023 15:00:23 +0000 Patel says NVIDIA GPU bandwidth increased less than 10x while FLOPS increased 100x from 2016 to 2023. https://minutesof.com/q/45f2aadd-215b-4a37-97e7-67d1873d78a9/ https://minutesof.com/q/45f2aadd-215b-4a37-97e7-67d1873d78a9/ “The bandwidth has not even gone up one order of magnitude. Right? Whereas flops have gone up two orders of magnitude. So so less than one order of magnitude versus two orders of magnitude increase.” — Gradient Flow Thu, 02 Feb 2023 12:00:39 +0000 Patel states memory cost has quadrupled in NVIDIA GPUs while per-gigabyte pricing remained flat since 2016. https://minutesof.com/q/fad9d07a-a3d6-43a9-9e0b-f2591b6d3829/ https://minutesof.com/q/fad9d07a-a3d6-43a9-9e0b-f2591b6d3829/ “And now with the h 100, they have 96 or 80 gigabytes of memory, but the cost per gigabyte is the same.” — Gradient Flow Thu, 02 Feb 2023 12:00:39 +0000 Patel says memory now represents 30-50% of GPU manufacturing cost and NVIDIA applies 4x markup. https://minutesof.com/q/e22458f3-983f-450f-85ae-780001a92a79/ https://minutesof.com/q/e22458f3-983f-450f-85ae-780001a92a79/ “And and now today, if you look at it, the cost of memory is 30 to 50% of the GPU cost. Great cost. And then obviously, NVIDIA does a four x markup on their” — Gradient Flow Thu, 02 Feb 2023 12:00:39 +0000 Patel reports PyTorch 2.0 delivered 30% performance improvement on 7,000 popular GitHub models. https://minutesof.com/q/5a9a7e9d-e509-4a4c-a7ad-30fdd24c4722/ https://minutesof.com/q/5a9a7e9d-e509-4a4c-a7ad-30fdd24c4722/ “So I think I think I think, the PyTorch Foundation tested the 7,000 most popular most starred models on GitHub, and they got, like, a 30% increase.” — Gradient Flow Thu, 02 Feb 2023 12:00:39 +0000 Patel notes ARM took a decade from 2011 announcement to see large datacenter deployments in 2021. https://minutesof.com/q/af83474c-76d6-4a45-bca0-1c5deff4da92/ https://minutesof.com/q/af83474c-76d6-4a45-bca0-1c5deff4da92/ “I think they announced in 2011 and we really only saw really big deployments from Amazon in 2021, right? Maybe 2020, right? Very big deployments.” — Gradient Flow Thu, 02 Feb 2023 12:00:39 +0000