This is a long conversation centered around AI asset pricing, compute supply and demand, capital expenditures, open-source models, and market structure. This article is compiled from the video interview "Why the Markets Are Pricing AI Wrong" with Gavin Baker on "Invest Like the Best".
The full text is divided into two parts: Part 1 extracts the most noteworthy core judgments from this conversation, helping to quickly understand Gavin Baker's main views on the current AI market; Part 2 retains the complete Q&A and context, striving to restore the original discussion logic for further reading and verification of specific viewpoints.
About Gavin Baker
Gavin Baker is the founder, managing partner, and chief investment officer of the investment firm Atreides Management, primarily focusing on technology and high-growth company investments. Before founding Atreides, he worked at Fidelity Investments for nearly 20 years, serving as the portfolio manager for the Fidelity OTC Portfolio from 2009 to 2017, while also participating in Fidelity's venture capital business. His earlier research and investment experience covers multiple industries, including semiconductors, telecommunications, pharmaceuticals, and retail.
Having long tracked the technology industry and the semiconductor sector, the focus of this interview is not merely "whether AI stocks will rise or fall," but rather judging whether the market is mispricing AI from the perspectives of GPU supply and demand, cloud providers' capital expenditures, model economics, the open-source ecosystem, and cash flow realization.
Part 1 | Conversation Summary: The 8 Most Memorable Judgments
This conversation does not discuss an abstract question of "whether there is an AI bubble," but rather a more specific pricing divergence: what assumptions the public markets are using to price the AI industry chain, and whether these assumptions align with frontline compute prices, token demand, and cash flow changes.
1. The Market Is Trading on "AI Earnings Nearing a Cycle Peak," While Gavin Sees Fundamentals Still Accelerating
In July, AI-related stocks plummeted, and the market began trading on narratives of compute oversupply, peaking capital expenditures, and open source eroding business models. However, the quantitative signals Gavin sees in Silicon Valley have not deteriorated synchronously: GPUs remain scarce, renewal prices for some older GPU clusters have risen significantly, and DRAM and token demand remain strong. His core divergence lies in this: stock prices are being revised down rapidly, but most observable frontline industry indicators are not following suit. If this divergence persists, the market may be pricing AI infrastructure with overly pessimistic assumptions.
2. What Truly Determines the Viability of This Round of AI Investment Is Whether Operating Cash Flow Can Catch Up with Capital Expenditures
The most worrying risk for the market is not "AI demand suddenly disappearing," but that this round of infrastructure construction ultimately turns into a capital cycle highly dependent on debt financing. Once over-leveraged, even a slight mismatch in supply and demand could cause the entire system to reverse rapidly. Gavin's counterargument is that much of the existing GPU compute is still locked in low-priced long-term contracts signed years ago. As these contracts expire and are repriced at higher market prices, the revenue and operating cash flow of hyperscale cloud providers may continue to accelerate. If this cash flow materializes, a large portion of capital expenditures originally expected to require credit support could be covered directly by internal cash flow. Therefore, whether the bullish logic for this round of AI holds up ultimately comes down to cash flow, rather than just looking at the scale of capital expenditures.
3. Open-Source Models Are Not "Disappearing Compute Demand," but Rather a Redistribution of Profits in the AI Industry Chain
The market tends to interpret users shifting from expensive frontier models to cheaper open-source models as a decline in AI demand. However, Gavin emphasizes that for the infrastructure, a token is a token: whether the token comes from a closed-source frontier model or an open-source model, it still requires FLOPs, memory, power, cooling, and data centers. What truly changes is the profit margin and profit attribution. Open source may compress the gross margin at the frontier model layer, shifting some profits to inference clouds, infrastructure, and application layers; meanwhile, lower token prices could also stimulate higher usage volumes. Therefore, while open source may pose a challenge to the business models of certain model companies, it does not necessarily constitute a bearish factor for underlying compute demand.
4. The Next Phase of AI May Not Be "One Super Model Consuming All Demand," but Multi-Model Layering and Routing
Gavin leans towards a future of multi-model coexistence: cheaper open-source models handle a massive volume of routine tasks, while the most powerful frontier models are responsible for planning, checking, and the most difficult problems; enterprises then train specialized models with their own data, distributing tasks among different models via a router.
This will give AI-native applications like Harvey and Cursor the opportunity to form their own model layers and data moats, rather than simply relying on a single frontier model provider. An interesting counterintuitive point is: the emergence of a large number of cheap "120 IQ" models does not necessarily reduce the value of "160 IQ" frontier models; instead, it may make the highest-level intelligence responsible for orchestrating these models even more valuable.
5. AI Spending Must Ultimately Find Its Source in the Real Economy, and the Most Optimistic Answer Is Not Layoffs, but Growth
If token spending continues to expand, this money can ultimately only come from two places: substituting labor costs, or AI driving faster economic growth.
Gavin mentions that token spending at some highly AI-integrated companies already accounts for 20% to 30% of payroll costs, yet many founder-led tech companies have not conducted layoffs on a corresponding scale as a result. This makes him lean towards the second possibility: AI is not just replacing people with tokens, but enabling the same people to generate more revenue. Therefore, the strongest true bullish logic for AI is productivity and economic growth, not mere labor substitution.
6. The Indicators That Could Truly Falsify the AI Bullish Logic Are Actually Very Clear
Gavin's risk list is not complicated: operating cash flow stops accelerating; total demand from Anthropic, OpenAI, Grok, Cursor, and the open-source ecosystem slows down significantly; GPU leasing prices continue to drop sharply; GPUs shift from being scarce to easily accessible; and AI infrastructure becomes increasingly reliant on debt financing. On the technical side, he pays particular attention to continuous learning and sample-efficient learning. If models can in the future be trained with significantly less data and continue to learn efficiently after deployment, training compute demand could be affected. However, in the long run, he still believes inference will gradually become a larger source of compute demand.
7. AI Is Not Just Being Priced by the Market; It Is Also Beginning to Change the Market's Own Pricing Mechanism
An increasing number of public market investors are feeding news to models like Claude for immediate analysis. When a large number of investors use similar models and receive similar explanations, the cognitive diversity of the market may decline. The result is that the same narratives are traded much faster; industry cycles that used to take years to play out can now complete their rise, crowding, and collapse within weeks. Gavin views this phenomenon as an important new variable in understanding the recent violent volatility in AI stocks.
8. What This Conversation Ultimately Leaves Us With Is a Framework of Judgments That Can Be Continuously Verified
Gavin does not frame his conclusion as "AI stocks will definitely go up." On the contrary, he repeatedly emphasizes that he could be wrong and tries to find negative data that could overturn his own judgments. But after experiencing a round of significant declines, his observation remains: stock prices have deteriorated significantly, while most industry quantitative indicators have not deteriorated synchronously, and some indicators have even continued to improve. Therefore, there is no need to continue debating the grand "AI bubble" narrative; we just need to keep an eye on a few things: token usage, GPU leasing prices and availability, repricing of existing compute contracts, and the operating cash flow of hyperscale cloud providers.
Part 2 | Full Video & Conversation
Below is the organized full interview. For easier reading on the official account, ad reads and uninformative verbal fillers have been removed, and corrections have been made to subtitle phrasing, obvious translationese, and confirmed English transcription/Chinese translation errors; the substantive viewpoints, numbers, cases, follow-up questions, and divergences in the interview have been retained as completely as possible.
July Was Like "Condensing 2022 into a Single Month," Yet AI Fundamentals Are Actually Improving
Patrick: Gavin, it's only been two months. The model release cycle is getting shorter and shorter, and the interval between our podcast episodes is also shrinking; we're now almost recording shows in sync with the model release rhythm. Someone previously said that our podcast always happens to appear exactly at local market peaks. After this time, no one should be able to say that anymore.
This month has been absolutely crazy. What do you most want to talk about right now?
Gavin: I would describe July as "condensing 2022 into a single month." There are indeed some negative factors in the fundamentals that we should discuss, but overall, I think the fundamentals are actually improving significantly.
Many AI-related stocks have fallen 50% to 60% from their highs. More accurately, they have dropped almost in a straight line by 40% to 60% within a month.
Before we started recording, I even asked you: you've been here all summer, have you heard of a single negative quantitative indicator regarding AI?
Patrick: No. In fact, almost every indicator I see is accelerating. And this isn't because everyone is blindly optimistic about AI; they can indeed pull out data from their respective positions to prove this.
Gavin: Completely agree. Whether looking at GPU availability, GPU leasing prices, DRAM spot prices this month, or token growth, all indicators are accelerating.
I think the market has a big problem right now: it cannot see the real operating conditions of Anthropic and OpenAI. Additionally, the market cannot see these U.S. open-source inference clouds monetizing through inference, such as Fireworks, Baseten, Modal, and Together. Once these private companies are included in the observation, the entire picture will be completely different.
Open source is accelerating significantly due to models like GLM-5.2 and Kimi K3; Nemotron is also advancing steadily, and the U.S. has seen some good small open-source models emerge. Meanwhile, OpenAI's growth is accelerating, Anthropic continues to grow strongly, and is almost certainly generating quite substantial free cash flow.
There is a chart in the market that everyone likes to look at: the cash flow of semiconductor companies is going straight up, while the free cash flow of hyperscale cloud providers is heading in the other direction. But this chart leaves out Anthropic, OpenAI, and these inference clouds, and also misses another very important change.
Back in 2024 and 2025, even if you were extremely bullish on AI, you would probably assume that GPU leasing prices would decline slowly; if you were bearish, you would expect prices to drop rapidly. But I don't think anyone at the time would have imagined that by 2026, the prices of older GPUs would actually still be rising in a straight line.
So at the time, everyone thought they were being smart: go sign long-term contracts. Many new cloud providers actually had to do this too, because only by securing long-term offtake agreements first could they finance their GPUs.
The result is: the existing compute that has already been contracted is priced at a huge discount relative to the spot market. As these contracts expire one after another, the compute will be repriced at higher prices. Even if spot prices decline in the future, contracted compute could still be revalued upwards. By then, you will see a clear acceleration in revenue and cash flow, and many doubts about AI return on investment will be answered as a result.
Signs have already started to appear this quarter. Note that I am talking about operating cash flow, not free cash flow. The operating cash flow growth rates disclosed by Microsoft, Meta, and Amazon accelerated from 28% to 32%. Moreover, there were quite a few large one-off items this quarter; for example, these hyperscalers often have billions of dollars in legal expenses, many of which come from EU fines. If these one-off items are excluded, the growth rate actually increased from 28% to about 35%. On such a large base, this is a very obvious acceleration.
More importantly, this is happening before Rubin comes online on a large scale and is priced at a significant premium, and also before a large number of old contracts are repriced.
What Is the Market Actually Afraid of This Month? From Meta and Open-Source Models to DUV Lithography
Patrick: This month has indeed been tough. Should we review chronologically how the market got here step by step?
Gavin: Sure. The first thing was Meta saying it would lease out some of its compute. The market immediately interpreted this as a major negative: Meta has excess compute, it's going to cut capital expenditures, and AI construction is going to have problems.
But the actual situation was completely different. Meta's subsequent earnings report has proven that it did not cut capital expenditures, nor did other indicators worsen; instead, it continued to increase investment.
A closer-to-reality explanation at the time was: Meta saw that there were people in the market who owned large, already-deployed compute clusters and could lease these clusters out at prices far higher than early contract prices. So the market speculated that Meta might want to take out a small portion of capacity first to prove that this compute could achieve a very high internal rate of return, and then go raise equity capital to further increase capital expenditures.
Later, the facts didn't necessarily look like this exact plan, but regardless, the market initially interpreted this event as "insufficient demand and excess capacity," and I was very certain at the time that this interpretation was wrong.
Meta's subsequent response was not a contraction either. It quickly released Muse Spark 1.1, one of the best models in a long time. This model was later overshadowed by the hype around Grok 4.5, but it is a very good model in its own right, far stronger than many of Meta's products from the past two years. So from this perspective, Meta has absolutely no reason to take its foot off the gas.
Immediately after, the market started to panic over open-source models. At the same time, Silicon Data's token index stabilized after declining; these two things are actually connected.
This token index does not cover all tokens, and the impact of models like GLM-5.2 and Kimi was only incorporated after some time. What is really happening is: the structure of tokens is shifting from higher-priced frontier models to cheaper open-source models.
The inference gross margin for frontier models might be 80%, 90%, or even 95%—the exact numbers are debatable; whereas the corresponding token gross margin for open-source models is much lower. Seeing the revenue structure shift from high-margin frontier tokens to low-margin open-source tokens, the market assumes this means demand is worsening.
But for the infrastructure, a token is a token. The FLOPs, memory, and power consumption required to generate a token do not fundamentally change just because it is an open-source model or a closed-source frontier model.
So open source taking market share is more about shifting profits away from the frontier model layer. Low prices create demand elasticity; users will use more tokens, thus requiring more compute. Both Anthropic and open-source models ultimately run on the underlying cloud infrastructure and both have to pay for compute. In other words, this is more like profits shifting from the frontier model layer to the AI infrastructure layer, rather than AI infrastructure demand disappearing.
Patrick: Jensen Huang can almost be called the world's biggest open-source advocate. If open source were truly detrimental to NVIDIA's business, would he still support it so much?
Gavin: He is certainly a very idealistic person and a very patriotic American; I believe if something is right for the world, he might support it even if he doesn't make money. But if open source were clearly detrimental to NVIDIA, it probably wouldn't be such a core and high-frequency topic for him.
Moreover, I think open source itself is very important for society. A world where only one or two frontier models dominate and earn gross margins close to 90% for a long time is not necessarily a good thing for humanity and society. We need many models, as we've discussed before.
The market had barely started to digest and accept the open-source situation when news broke that China already has DUV lithography machines. Because everyone's positions are concentrated in similar baskets of stocks, this news triggered a massive sell-off in semiconductor equipment stocks.
But among these things, I don't think what's truly worth worrying about is Meta or open source, but rather rising real yields and widening credit spreads.
We are currently investing a massive amount of capital to support AI infrastructure construction. Although the vast majority of the funds still come from operating cash flow, the proportion of credit in it is indeed rising. Real yields are rising, and credit spreads are widening. Meta issued a bond last week, and its pricing was not as easy as traditional Meta bonds; the CDS spreads of many companies are also widening.
Some very smart private capital folks would say this is just banks hedging financing they've already committed to, which would happen anyway. But regardless, it certainly doesn't look good on the surface, and these are undeniable facts.
If we really have to rely on debt to complete this round of AI infrastructure construction, it would be terrifying. And precisely because of this, the "huge gap between existing compute contract prices and spot prices" I mentioned earlier is so critical.
Is Debt the Real Risk?
If Operating Cash Flow Can Accelerate
The $700 Billion Credit Demand Might Not Even Materialize
Patrick: So one of the most important questions for the next six months is exactly how much this round of construction will need to rely on credit, right? If it ultimately turns into a debt-driven capital cycle, that's a classic danger mode.
Gavin: Absolutely correct. Once you start over-leveraging, things become dangerous. Debt-driven expansion requires rapid debt repayment; if there is even a slight mismatch in supply and demand, the entire system could reverse very quickly. That's exactly how it was during the dot-com bubble.
After going through this month, I am much more open to all my judgments and have been stress-testing them. I have studied credit especially seriously because this is a real risk. If AI construction ultimately must rely on credit, that is a major negative factor.
But after modeling the numbers, the conclusion I reached is not that pessimistic.
If you look at the planned gigawatts coming online in the future, and then look at the market's consensus forecasts for hyperscale cloud providers, these new compute additions are mainly Blackwell and Rubin. Rubin is NVIDIA's next-generation chip, and Blackwell is the current generation. Yet in their models, the market is actually estimating the monetization capability of this compute roughly based on Ampere—which is already two generations behind, not even Hopper, but the older Ampere.
Under this assumption, the future operating cash flow of hyperscale cloud providers is about $1.3 trillion to $1.4 trillion.
But I think it is highly unlikely that Blackwell and Rubin will ultimately only monetize at Ampere levels. The real quantitative demand signals we see on the front lines in Silicon Valley do not support this assumption either.
Even if you just conservatively assume their monetization level is lower than the current Blackwell level, the operating cash flow would be closer to $2 trillion. In this way, the approximately $700 billion credit demand simply disappears.
Moreover, the stronger the operating cash flow, the more all credit indicators will improve. After existing compute is repriced, operating cash flow will continue to accelerate. The market's current models instead assume growth will decelerate; I think this assumption itself is questionable.
Once credit indicators improve, even if companies still want to use debt financing, financing will become easier. Whether they ultimately use it is another matter.
We've discussed a risk before: a so-called "demand air gap" could occur with Blackwell. You invest hundreds of billions of dollars to buy Blackwell; initially, a large number of chips are used for training, and training itself won't generate revenue and returns immediately, so for a period, there could be a situation where capital expenditures go up first while cash flow hasn't caught up yet.
We did see similar signs in the first quarter. But two months ago, I started to gradually worry less about this risk, and an important reason is that what Anthropic demonstrated at the time was just astonishing. The market in April, May, and June was also willing to temporarily look past this air gap and look forward.
By July, due to a series of overlapping events, the market suddenly became unwilling to look forward anymore—precisely when operating cash flow had already started to truly accelerate. Microsoft brought online a massive amount of new capacity in June, much of which hasn't even been reflected in the second-quarter data yet.
So at the end of the day, what you need to judge is: do you believe the quantitative demand signals being sent out by these private companies in Silicon Valley will persist? If they do, then after old contracts expire, existing compute will be repriced at higher prices, operating cash flow will rise, and most of the construction can be supported directly by operating cash flow, or potentially even entirely by operating cash flow.
If repriced at current price levels, it might be able to fund all construction over the next few years.
Gavin's Mission in Silicon Valley: Please Tell Me a Real Negative Data Point
Gavin: This is why I feel this market action is very unusual.
Back in 2022, you knew exactly what the market was worried about: recession, rising interest rates, inflation. During the DeepSeek shock, you also knew what the market was afraid of; "Liberation Day" was the same. The risks were very clear, which paradoxically gave a strange sense of comfort.
This time, we have certainly discussed many specific concerns, but apart from credit, many other narratives seem to me unable to withstand factual verification. Yet stocks are still falling. Technical analysts would tell you this is actually a bit scary, because what really hits you is often the bullet you didn't see.
I've said before that I think the three most important words in investing might not be "margin of safety," but "I don't know." So my main mission coming to Silicon Valley this week is to keep asking people: okay, tell me something negative. Give me a real negative quantitative indicator.
This morning I just talked to a very hot startup that everyone wants to work with. It rented a cluster of several thousand Blackwell GPUs. A few months ago, the price it rented this batch of B200s was around $2 per GPU hour; now, for an almost identical cluster, when renewing the lease seven months later, it expects to pay close to $4.
In other words, originally, even if GPU rents just declined moderately, I would have considered that a positive; but the reality is that over six or seven months, prices have actually risen by 50% to 60%.
There are many similar examples. I remember one inference cloud, maybe Baseten—I'm not 100% sure—they said directly on a podcast: when the Blackwell contracts expire, we expect to pay 100% more.
This actually means that almost all hyperscale cloud providers are now "earning less"—because they are selling compute at early contract prices far below the spot market.
Patrick: So did you actually find any negative indicators?
Gavin: I asked a lot of people. The only thing I heard repeatedly is that third-party data for Anthropic seems to show its growth curve has slightly deviated from its original trajectory. I think this is very likely true.
But at the same time, both OpenAI and open-source models are accelerating significantly. If you aggregate all the demand, what I still see is net acceleration, perhaps even faster than before.
Open source is a bit like "dark matter" in the public markets. It's hard for public markets to measure it. Much of the information you can get is just statements from inference cloud leaders on podcasts and at conferences, not audited financial data, so you have to remain cautious. But from what various companies have revealed, demand is clearly accelerating, which also aligns with the capability leaps brought by GLM-5.2 and Kimi K3.
I think NVIDIA will also keep pushing Nemotron closer to frontier models.
This has been a very humbling month. I have stress-tested almost every assumption, but in the end, what I come back to see is: fundamentals are improving.
Meanwhile, NVIDIA's forward P/E ratio has dropped to near its lowest point in the past 10 years. Semiconductor stocks were only cheaper than now during "Liberation Day" and the DeepSeek shock, and both of those were very typical V-shaped bottoms.
Patrick: Does this mean the market actually believes these companies' current earnings are far above sustainable levels?
Gavin: One hundred percent. That's exactly what the market thinks.
We must be humble; maybe the market is right, maybe these companies are indeed over-earning. But I came here this week specifically to find negative data that could support this judgment, and I've tried as hard as possible to find it. Usually when you come to Silicon Valley, you always hear some positive things, some negative things, and overall it leans positive because technology creates value in the long run. But this time, I have hardly found any real quantitative negative indicators, apart from the Anthropic third-party data just mentioned.
Moreover, Anthropic's shareholders are highly controversial about this. You can clearly feel that, on the one hand, they are eager to tell you what they know, and on the other hand, they are worried that the company will find out the news leaked from them, affecting future IPO allocations. Their attitude is very clear: the third-party data does not fully reflect the real situation.
I also find it hard to believe that the acceleration of open source and OpenAI would be large enough to completely explain all the changes, but Anthropic is still clearly in the lead. Third-party data also shows Grok and Cursor growing, especially after the release of Grok 4.5 and Grok Build, there was a very obvious change in July.
So, this is really a very difficult month to process.
Is AI Making the Stock Market Itself More "Homogeneous"?
Gavin: A friend of mine at Fidelity said that the most effective market strategy over the past three years has been to make the simplest, most superficial trades as quickly as possible, and then rotate rapidly.
Patrick: So what is the simplest trade right now?
Gavin: De-risking. Throughout July, many people have been de-risking around these narratives. But apart from credit, many narratives do not hold up on a factual level. Our research actually makes me increasingly believe that credit will not be the decisive issue.
And now a very interesting new mechanism has emerged in the market.
Suppose we do need credit to build the compute required for the future, but the credit suddenly doesn't materialize, then existing compute will only become more valuable. This reminds me of a theory Michael Mauboussin talked about before: when cognitive diversity in the market disappears, bubbles and crashes both become more likely to occur.
Today, public market investors I know, whether individuals or institutions, almost all throw every piece of news into Claude immediately. Some use Claude, some use Claude Code, and some even directly run Claude Agents.
But Claude is essentially a probabilistic model. For the same piece of news, the explanations received by a large number of users often won't differ much. This has never happened in the history of the stock market.
People used to say that in the traditional media era, there was an almost singular "voice of authority" like Walter Cronkite, and later the media fragmented. But now, Claude is to some extent becoming the Walter Cronkite of the stock market: many people default to believing the explanations it provides.
Claude is smart, but it is not always right. The problem is that the stock market is essentially pricing the future in a Bayesian way. Now it's increasingly looking like: news comes out, Claude gives an explanation, and then a large group of people trade based on that same explanation.
This homogenization has started to create very violent, very short market cycles.
There's a guy named TBU in the anonymous "semiconductor circle" who posted a very astonishing chart of Japanese capacitor stocks, saying the capacitor industry completed an entire cycle in six weeks. Those stocks went up almost in a straight line, some possibly doubling, tripling, quadrupling, and then quickly fell back down.
The most absurd thing is that the real fundamentals haven't even had time to materialize. A cycle that might have taken three years to play out in the past is now completed in six weeks.
If Continuous Learning Breaks Through, Could It Actually Reduce Training Compute Demand?
Patrick: You've been in touch with many companies here lately. I'm particularly curious, regarding efficiency improvements around inference, training, and the entire model serving chain, what long-term innovations are worth paying attention to now? Is there any technical direction that makes you particularly excited, or makes you particularly vigilant that it might change the public market's judgment on compute demand?
Gavin: There is one direction I am paying very close attention to: continuous learning and sample-efficient learning. Many people seem to think these two things are close to a real breakthrough.
We've discussed this before. If continuous learning and sample efficiency are truly solved, could there be some kind of cliff in training demand in the short term? That's a reasonable question.
Someone gave me a very vivid analogy: humans might be equivalent to being "trained" with only about 20 billion tokens, whereas today's large model training data has reached the scale of 300 trillion tokens. If in the future a model only needs 10 trillion tokens to be trained, and can continue to learn efficiently with very few samples after release, then purely from the perspective of training demand, this is obviously not a positive.
However, from a long-term structural perspective, training will inherently account for an increasingly smaller portion of semiconductor compute demand—approaching a very small, but not zero, proportion. The truly massive demand is ultimately inference.
So I think continuous learning and sample-efficient learning are among the most noteworthy scientific or technical variables right now. No one knows whether its impact will be long-term or short-term. SSI said they will release a model in August, and many new labs of this generation are also focusing on this direction.
It must be emphasized: if these problems are truly solved, it would of course be a very good thing for the world. We all hope it happens. It's just that I still find it hard to believe it will ultimately truly reduce the total demand for AI infrastructure.
After going through this month, I have been trying hard to keep myself very open and very humble. If you absolutely must ask me to find something on the technical side that could change my judgment, I would put continuous learning and sample efficiency very high on the list.
And NVIDIA is actually actively involved in many such startups, so it can see these changes itself.
Patrick: If you absolutely had to list a set of conditions that would truly change your AI bullish judgment, or even make you scared, what would you list? Is it that operating cash flow doesn't materialize, and ultimately the entire round of construction has to rely on debt financing?
Gavin: Yes. First, if operating cash flow cannot continue to accelerate, that is a clear negative signal. This largely depends on the total demand from Anthropic, OpenAI, Grok, Cursor, and the open-source ecosystem.
Second, if GPU leasing prices continue to drop significantly, the market will reflect it very quickly.
Third, if GPUs become very easy to obtain, that would also worry me.
Patrick: Have you heard anyone say "I have too many GPUs" right now?
Gavin: Not a single one. In fact, quite the opposite.
This even sounds a bit like a drug market, but reality is just that exaggerated. The list of truly obvious risks is actually very clear: if the aggregated demand growth from labs like Anthropic and OpenAI stalls or even declines, that is very negative—unless it's simply because open-source tokens are taking their market share while expanding the overall market.
I increasingly believe the future will be multi-model coexistence. Especially AI-native companies, they will be increasingly willing to use open-source models. Many inference clouds are now very mature in supervised fine-tuning and reinforcement learning; you can customize an open-source model with your own data and put it behind a router.
When a request comes in, it can first be handed to your own model for processing, and then frontier models like Claude and Grok can do the checking or handle the most difficult parts. In many cases, the final result is even slightly better, but the overall cost is only half of what it was.
Many people hear "costs cut in half" and think this is bad for AI infrastructure demand. It's actually not.
The price the user ultimately pays depends on the gross margin of different token tiers. You are just shifting demand from expensive frontier tokens with gross margins possibly close to 90%, to low-priced tokens with gross margins possibly only 30%. The cost reduction comes from profit margin redistribution, not from the disappearance of compute needed to generate tokens.
The compute consumed behind these tokens is basically the same.
Moreover, different companies are in completely different AI adoption cycles. Many large public companies will say: "My AI spending grew 20 times in three months, and the budget was instantly spent." So they deploy a router and hand more tasks to cheaper models; on the surface, AI spending stabilizes or even drops.
But the actual situation might be that token generation volume has actually increased, and GPU compute hours are still rising. The company just became smarter: choosing different models for different tasks.
This kind of thing is happening in some of the most cutting-edge public companies. At the same time, there is a whole wave of AI-native companies that are investing in AI extremely aggressively, not hiring a massive number of human employees, but putting more and more money directly into tokens.
Looking further out, many companies on the U.S. East Coast have hardly truly adopted AI yet; a large number of enterprises in the U.S. interior are also far from being that cutting-edge overall; Europe is even still discussing how to regulate AI, with many places considering how to manage it before actually using AI.
So, there are actually many waves of AI adoption happening simultaneously right now, just at completely different stages.
Where Does the Customers' Money Come From? The $5 Trillion in Token Spending, Preferably from Growth Rather Than Layoffs
Gavin: There is a number I can never forget. Globally, the number of people truly using Agentic AI at a high frequency right now is maybe only 500,000, or even just 250,000. But at such a low penetration rate, we already have a very severe compute shortage.
There are 7 to 8 billion people on Earth. What will happen when the number of people using Agentic AI goes from 500,000 to 100 million, and then to 500 million?
Patrick: This is also the counterpoint many people have to your view. Suppose we accept the judgment that "hyperscale cloud providers' current earnings are below what they should be," and also accept that after existing compute is repriced, their operating cash flow will accelerate, even enough to support capital expenditures.
But where does this operating cash flow ultimately come from? Who are the customers?
Gavin: By definition, there are only two possibilities.
Just like Satya said: either the economy will start growing at a faster pace because of AI, say 10%, or it won't. The other source is labor substitution.
In many AI-native companies, we have indeed seen some form of labor substitution, but it doesn't necessarily manifest as layoffs; rather, the company simply isn't hiring as many people as it used to.
You can look at "gross profit generated per full-time employee." Many firms like a16z and ICONIQ have done similar research. Compared to past generations of startups, this metric for AI-native companies is rising almost in a straight line.
Patrick: Have you surveyed the ratio of token spending to human spending in your portfolio companies?
Gavin: Yes. Usually looking at the ratio of token spending to total payroll costs. For those companies that are truly deeply "AI-ified," this ratio can already reach 20% or 25%.
Patrick: Our friend Dylan Patel is a staunch believer in ASI (Artificial Super Intelligence), and his company is already at 30%.
Gavin: That might be one of the highest I've ever heard.
Patrick: I've even heard of 50%.
Gavin: Global knowledge work is roughly a $25 trillion economic size. If you calculate using the 20% you just mentioned, that's $5 trillion.
This $5 trillion must come either from substituting labor spending or from faster economic growth. As humans, we certainly very much hope it comes primarily from economic growth.
I just heard an interesting thing from a very senior tech CEO who has founded multiple companies this morning: if you look at companies that are still founder-led and controlled, and strip out the corrections caused by over-hiring during the pandemic, there are actually almost no companies doing true massive layoffs.
Logically, these founders should be the most willing to quickly adopt AI and improve efficiency. Yet apart from some moves by Zuckerberg, they are not universally conducting giant layoffs.
This might indicate that the opportunity they see is not simply "replacing people with tokens," but that people and tokens can create more things together.
Patrick: So the true bullish logic is growth, not labor substitution.
Gavin: Yes, it's growth.
You've already seen the data released by companies like Cognition, Ramp, and Stripe: the companies with the highest AI spending are also growing significantly faster.
Patrick: I really like the Cognition Index.
Gavin: That index is indeed astonishing. Of course, skeptics would reasonably point out that it doesn't fully control for industry differences. But if you drill further into the data, you'll find that even some very traditional companies—I can't remember if the example was plumbers or HVAC contractors—can also clearly benefit from AI.
And blue-collar workers as a whole are actually benefiting right now too. We will return to this point later when we talk about data centers.
Why Is Everyone Signing Long-Term Agreements? Memory and Compute Are Turning into a "Game of Thrones"
Gavin: There is another issue I think is very important: everyone is signing long-term agreements, or LTAs, right now.
Almost everything is in a state of shortage right now. If anything looks a bit weak, it's mainly because we can't power up these gigawatt-scale data centers fast enough, not because demand is insufficient.
These gigawatts will eventually be powered up. Policy directions are improving; turbine manufacturers and diesel generator set manufacturers are all accelerating capacity expansion; some are even taking turbines off old planes, refurbishing them, and reusing them. Many things that sounded crazy before are happening. Capitalism is very good at solving supply bottlenecks.
I had a judgment wrong before: I thought memory companies would continue to pursue massive short-term earnings beats. But now they are clearly shifting to another strategy—sacrificing some short-term upside in exchange for long-term supply chain agreements.
These LTAs have many structures. Some require advance payments from customers, some set price floors and ceilings, and some lock in supply volumes.
This is a bit like "labor hoarding" after the pandemic. Many companies laid off people too harshly back then, and later became increasingly unwilling to lay off easily. Now a similar psychology has emerged in the supply chain: after experiencing severe shortages, parties are more willing to sacrifice short-term price advantages in exchange for long-term certainty.
But what's really interesting is: if you want to tear up these LTAs in the future, game theory will be very ruthless.
There are probably four types of players that truly matter in scale right now: Amazon has its own Trainium, Google has TPUs, AMD is an important player, and then NVIDIA, whose scale is far greater than the sum of all other players. Beyond that, there are many startups.
Suppose it's 2027 or 2028. You must recognize the importance of memory: the more memory allocated per FLOP, the more tokens output per unit of compute. This is one of the most important ways to increase token output and directly lowers unit costs. This is also why demand currently shows almost no obvious negative elasticity to price changes—performance and output improvements outweigh the price itself.
In a sense, this has already become a large-scale "Game of Thrones" among these companies.
Suppose it's 2028 or 2029, you think the market is a bit oversupplied, so you start thinking about tearing up previously signed long-term agreements to try to get lower prices.
But in the coming years, market share will largely depend on how much supply chain quota you can get and how much product you have locked in advance.
So you have to ask yourself: if you tear up the LTA today, and for any reason in the next two or three years, bargaining power shifts back to the memory suppliers, what will happen? The answer might be that you are out of the game entirely.
For example, Google tears up long-term agreements today because of oversupply. After oversupply, prices drop, and capacity naturally contracts. But this is a cyclical industry; after oversupply, it usually moves towards undersupply again. When the next round of shortage comes, how much quota can Google get? Who will suppliers prioritize?
You might destroy the competitive foundation of the entire business just to lower the procurement price a bit in the short term.
It wasn't like this when Apple did procurement in the past. Apple had almost no true peers as competitors and was the largest buyer. Even if it squeezed SK Hynix very hard, Micron would be very happy to take Apple's orders, so Apple long knew it had immense bargaining power.
Today's AI supply chain is completely different. There are at least four large competitors, plus a large batch of startups. If you tear up the price agreement, the other party can completely say: "Okay, then we'll also tear up the supply volume agreement and give this inventory to your competitors." You could instantly lose market share.
Why Is the Current Environment Reinforcing NVIDIA's Advantages Almost Everywhere?
Gavin: This is also one of the reasons I find it hard to understand why NVIDIA is trading at such a low P/E ratio.
If you have to finance chips—and everyone needs to right now—there is no asset easier to finance than NVIDIA GPUs.
If you also need land and power, NVIDIA is also very good at coordinating within the ecosystem, helping connect resources and partners for all parties.
They have also recently been developing a new business model I think is very smart, which I would call "credit packaging with revenue sharing." Simply put, if the actual leasing price of GPUs is above a certain floor, NVIDIA can get a continuous revenue share in addition to chip revenue.
This could allow NVIDIA to very quickly build a massive cloud-related business through a mechanism similar to royalty revenue, and also alleviate the cash flow duration mismatch problem across the industry.
Patrick: Does this count as vendor financing?
Gavin: Strictly speaking, not entirely. NVIDIA is not directly lending money to the companies buying GPUs; rather, other capital providers are lending money to the GPU buyers. NVIDIA still does equity investments, but it's not simply investing in a company and having them take that money back to buy NVIDIA chips.
NVIDIA has also specifically stated that its equity investment agreements will specify that the investment funds cannot be used directly to purchase NVIDIA chips. Of course, money itself is fungible, which is why this matter often sparks controversy.
Patrick: If you were the CEO of SK Hynix, what would you do?
Gavin: I would follow NVIDIA's line of thinking.
Patrick: Specifically how?
Gavin: I would go to the buyers of GPUs, Trainium, and other chips, and say to them: I can also participate in this kind of credit packaging.
The memory business itself is not as stable or predictable as NVIDIA's, but you can definitely design some kind of structure. For example, putting up some cash upfront to reduce the risk borne by debt capital providers, acting as a sort of guarantee; at the same time, demanding a continuous share of revenue in the future.
I'm just brainstorming a structure on the spot right now, but the direction is very natural. Memory companies have a lot of cash right now, while the credit market has started to show resistance. Firms like Blackstone and Apollo, I believe, will definitely design various similar schemes for memory companies.
Essentially, this is a further extension of the LTA logic. Traditional LTAs are suppliers using some price upside in exchange for business stability and continuity; if you overlay revenue sharing, you can even continuously receive royalty-like returns from the customer's future recurring revenue.
This is exactly what NVIDIA is doing, and I think the market has severely misunderstood this.
It would be very helpful for NVIDIA to explain this clearly. First, they are indeed extremely bullish on AI. In fact, almost every time in the past that NVIDIA didn't get equity in an important AI company, it looked like a mistake in hindsight. They have invested in almost all the important companies; they just didn't invest in memory companies for a long time; they didn't invest in Anthropic before, but now they have taken a stake.
If you have extremely strong free cash flow yourself and are very bullish on AI, why not capture this upside? Jensen Huang can see the progress of almost every lab: OpenAI, Anthropic, xAI, Safe Superintelligence, and various continuous learning labs. What he sees clearly keeps him optimistic.
So NVIDIA can first take equity upside, second take revenue sharing, and third also help fill the funding gap for the entire ecosystem.
NVIDIA is generating hundreds of billions of dollars in free cash flow every year, while many labs currently still have negative free cash flow. Before the operating cash flow of these labs truly accelerates and they can be fully internally financed, there is indeed a funding gap in the ecosystem. By participating in filling this gap, NVIDIA can not only gain extra returns but also increase its revenue per gigawatt, further strengthening its competitive position.
You and I have both invested in chip startups. Suppose a customer wants to use a certain startup's chips, that's great of course, but it will immediately face three problems: the wafer price paid to TSMC is often higher than NVIDIA's; the price of buying HBM DRAM is also higher; and it is almost impossible for these chips to get financing as easily and at the same low cost as NVIDIA GPUs.
If the architecture also relies on HBM DRAM, it will be directly exposed to the risk of high-priced memory, unless it chooses a completely different architecture.
So many things that have happened recently are actually strengthening NVIDIA's position.
This brings us back to game theory. If Anthropic had bought compute as aggressively as OpenAI in the past, it might be leading much further ahead now. Now OpenAI is back at the competitive frontier, and Grok has also entered the frontier. The companies truly on the Pareto frontier all have one thing in common: they have compute.
Patrick: After seeing this round of competition, who else would dare to take their foot off the gas?
Gavin: It's hard.
Dario talked about this about four months ago, and he went very deep. For these labs, buying too much compute means an investment scale large enough to bankrupt the company; but buying too little could mean falling behind directly in the competition.
We have seen this happen with our own eyes. OpenAI just returned to the frontier, and now SpaceX has made a strong entry with Grok 4.5 and Cursor.
From a game theory perspective, in the foreseeable future, I find it hard to see who would voluntarily back down, especially when these capital expenditures can ultimately be more supported by operating cash flow.
Everyone in Silicon Valley Is More Optimistic Than Gavin: Some Even Think H100 Annual Rent Could Rise to $250,000
Patrick: You've met so many people here lately. Have you encountered anyone much more optimistic than you? If so, what do they believe that you don't?
Gavin: Honestly, almost everyone here is more optimistic than me.
I read that article Dwarkesh wrote, and I was just stunned at the time.
Patrick: Was it the judgment about "compute prices rising 3x" or something like that?
Gavin: No, it's more exaggerated, about 15x. His core point is that renting an H100 for a year might be worth $250,000. That's about 15 times the current spot price.
This number is not in my scenario assumptions at all. Let alone what Bayesian probability I would assign to it—I didn't even put it in the tail scenarios of "considered but excluded due to low probability."
This is very Dwarkesh's style. He is very smart and extremely well-connected.
He also pointed out that three variables are rising simultaneously: the profit margin of compute itself is rising, compute volume is rising, and the profit margin of inference is also rising. Multiply the three together, and you can explain why the aggregated growth of various labs and open-source models is so exaggerated.
Of course, the profit margin of open-source models themselves hasn't truly risen. But overall, people here are indeed more optimistic than me.
When I look at the stock market, I feel like a foolish super-bull; but once I talk to people in the labs and the infrastructure ecosystem, I am almost always the more cautious one in the room.
This contrast is very strange.
DUV Lithography: This Is a "Phase Transition," but It Doesn't Mean Catching Up to EUV Overnight
Patrick: How do you view the news about China's DUV? I see market reactions covering almost the entire spectrum. One view is that this is equivalent to ASML's technology from 2001; the other view is that this is the beginning of a completely new story that will change how we understand the global supply of advanced compute.
Gavin: I think both views could be true simultaneously.
To use an analogy, if DUV lithography is equivalent to a propeller plane, then EUV might be equivalent to a jet plane; or you could say DUV is a jet engine and EUV is a warp drive. The metaphor itself isn't important; the point is that there is still a huge generational gap between the two.
But China didn't have this capability before, and now it reportedly does. This in itself is a "phase transition"—like matter changing from liquid to solid. Even if this machine is still 25 years behind, it doesn't mean it's unimportant.
I don't think this should be taken lightly.
On the other hand, the stock market is prone to overreacting to this kind of news. Even if it ultimately really affects ASML's orders, the real impact might be five years from now. Over these five years, the market might forget about it, worry about it again, forget again, and worry many times.
So my judgment is: this is indeed important, but the market may have overreacted in the short term.
If standing from China's perspective, this is obviously a matter of national importance.
But I would never underestimate China's engineering capabilities. They are very smart, very hardworking, and view semiconductor self-reliance as a national strategy.
Patrick: But can they really catch up in one go from a level equivalent to 2001 to 2026, or even 2030?
Gavin: That's the key. Semiconductor manufacturing is highly dependent on "learning by doing." This learning cycle accumulated through hands-on practice cannot be skipped directly by being smart, nor is there a portal that can teleport you to the future. You have to do it round by round, make mistakes round by round, and then improve round by round.
So is it important? Yes. Did the market overreact? Most likely yes.
There is another point that is hard for Americans to truly understand: the decoupling between China and the U.S. is already a process that has started and is self-reinforcing. For better or worse, both sides are now continuously reinforcing each other's actions.
This is unfortunate, but it is the reality. China won't stop, and the U.S. won't stop either.
Open Source Will Not Kill Software Companies; It May Actually Deepen the Moats of AI-Native Companies
Patrick: Apart from those dozen or so most core public companies and a few private labs, how do you view other companies in the U.S.?
Gavin: If you only look at last month, of course you'd feel the market is only about AI. But the month before that wasn't like this—back then, almost every sector except AI was rising in a straight line.
Looking at the software industry itself, I increasingly think that open source gradually approaching the frontier, coupled with companies like Fireworks making it very easy to customize models, is actually a huge gift to the software industry.
In many scenarios, you can use your own proprietary data to tune an open-source model to a level more suitable for specific tasks than general frontier models, and at a significantly lower cost.
This is especially important for a large number of AI-native companies.
Our friend Vichia said about two years ago that he had never seen so many companies achieve tens of millions of dollars in annual revenue so quickly from the very beginning of their founding; some companies can even reach $50 million in annual revenue and generate cash flow within nine months.
In the past, people would dismiss many of these companies as "ChatGPT wrapper apps." At the time, it was indeed hard to judge how much long-term competitiveness they actually had.
But now with open source, the situation is changing. If an application company has accumulated unique data in a certain vertical scenario, it can use that data to train its own models, rather than forever relying on a single frontier model provider.
Fireworks has a product I think is very cool called Nexus. Suppose you use different tools like Claude Code, OpenAI Codex, and Grok Build simultaneously; Fireworks can import your data, use reinforcement learning to train a model that better fits your business, and then route different queries to different models via a router.
The results they get are astonishing.
This could almost become a universal path for many AI-native companies. Companies like Harvey, Cursor, and Legora are all strongly moving in this direction.
In the future, not all tokens will have to be handed over to one or two frontier models. You might only hand 30% to 60% of tokens to frontier models, letting each model do the tasks it is best at, and handing the rest of the requests to your own reinforcement learning-trained models.
At this point, you are no longer just a wrapper app; your moat will be much deeper.
Patrick: There's an example with Cursor that left a deep impression on me. It feels like AI is "speedrunning" the organizational experience humans have already formed: the smartest people are responsible for planning, and then they hand specific tasks to people with less capability.
Gavin: Exactly. Frontier models are responsible for planning, and then they distribute tasks to cheaper, less smart models. In the Cursor case, efficiency seemed to improve by 15 times; I don't remember the exact metric, but the magnitude is astonishing.
This also brings up another interesting counter-argument.
Some staunch supporters of Anthropic, OpenAI, and Grok believe: once frontier models truly reach ASI, the operating cost of the most advanced intelligence will drop rapidly. You can provide a super-intelligent model at a very low price, and then through distillation, meet tasks of different intelligence levels at low cost; in that case, there might be no room for open source to exist.
This view cannot be easily dismissed. After going through this month, I especially want to remain humble. Anything is possible.
But for me, this outcome is still unlikely.
Patrick: Why?
Gavin: First, there are already many AI-native companies that have accumulated a massive amount of domain-specific proprietary data.
Before open-source models truly rose and inference clouds and routers matured, these companies had almost no choice: you just had to accept whatever terms of service the frontier model providers gave.
Now, if they can get off this treadmill and build their own model layer, they can gain a certain degree of independence and also enhance the durability and security of their business.
Moreover, your point just now might hold true in reverse: a large volume of cheap, "120 IQ" equivalent open-source tokens could actually significantly elevate the value of the truly most cutting-edge, "160 IQ" models.
Patrick: Right. If a bunch of 120 IQ models can execute tasks at low cost, wouldn't the 160 IQ model that can orchestrate them and make the highest-level decisions be even more valuable?
Gavin: Yes.
When we last talked, there was something that really surprised me: at that time, the vast majority of economic returns were almost entirely flowing to the frontier model layer. Now, with the rise of inference clouds like Together, Modal, and Baseten, this structure is starting to change.
The business models of these companies are astonishing. They are extremely capital-efficient, and their growth rate almost catches up with the early stages of frontier labs, but they burn much less cash. If you re-evaluate them using some traditional SaaS metrics, like the "Rule of 40," their numbers are very exaggerated.
Patrick: Could the compensation distribution within an organization give us a good analogy? The CEO's compensation might be many times the median employee's; the relationship between frontier tokens and open-source tokens might ultimately be like this.
Gavin: Absolutely possible.
The entire pie is growing rapidly. Frontier models might lose some token share, but still capture a very large portion of the economic value; meanwhile, among the actual number of tokens processed, open-source models might account for the majority.
Going back to the original infrastructure logic: this is actually very beneficial for compute demand. A token is a token. Generating tokens still requires FLOPs, power, space, cooling, and hardware.
Gavin Believes the Biggest Risk for AI Is Not Technology, But Regulation and Public Narratives
Patrick: If you absolutely had to say what the worst thing in the AI field that could truly change long-term judgments is, what would it be? Regulation?
Gavin: I think regulation must be ranked among the biggest risks, and it is the most obvious one.
I came to Silicon Valley this time really wanting to hear something that could truly scare me. I don't want to feel like a lunatic—watching these stocks get cheaper and cheaper while still feeling their future expected returns are getting higher and higher.
Especially since frontline fundamentals were clearly still substantively improving from June to July. So I've been thinking: what exactly am I missing?
After thinking it over, I still believe regulation is the most worrisome risk.
You can't ignore the emergence of data center pause-related policies in New York State. We are now in a very strange political environment where facts and logic don't necessarily automatically win public discussions. And the AI industry has done very poorly in public communication in the past.
I think the industry is now at least starting to realize this problem. It might not be solved yet, but it is realized.
The narrative in the minds of Washington and many ordinary Americans today is: data centers will raise my electricity prices, take away local water, steal jobs, and ultimately AI will steal even more jobs.
Whereas the AI industry itself believes the reality is exactly the opposite.
According to the agreement structures adopted by some large data center projects currently, what I see is exactly the opposite: after a data center lands, the electricity prices for other local users might even drop. The relevant data center commitments pushed by the Trump administration also try to make developers bear more local power and community infrastructure costs.
In the past, when a large project entered a community, it might just add a few cars and some equipment to the police and fire departments. Now, many data center developers are talking about: building hospitals, building schools, building new police and fire departments, while also bearing power infrastructure investments.
Patrick: If a data center that signed such commitments comes to your community, it certainly sounds very different.
Gavin: And the blue-collar jobs brought by data centers are not just one-time construction positions.
Data centers need plumbers, electricians, and HVAC contractors in the long term; they need maintenance, upgrades, and renovations. Technology is constantly changing, and equipment needs continuous updating.
In many ways, this is one of the most beneficial things for blue-collar wages I have seen in my lifetime.
And precisely because of this, the current political narrative feels absurd to me. Some political forces claim on the surface to represent blue-collar workers, yet simultaneously oppose the infrastructure construction that will create these high-paying jobs in the long term.
This reminds me of a saying: a lie can travel halfway around the world while the truth is still putting on its shoes.
There is a particularly typical example about data center water usage. Someone in a book overestimated data center water usage by 10,000 times. Not by a little, not by one or two orders of magnitude, but an extremely exaggerated error. The author later admitted the mistake multiple times, and the claim was refuted, but the original error had already spread.
Patrick: This is a bit like the "Popeye effect." Have you heard the spinach example?
Gavin: No.
Patrick: Popeye eats spinach because for a long time everyone thought spinach had a particularly high iron content. But according to that classic story, the original source was just a decimal point error in academic literature, which then spread all the way. Until today, many people still think spinach has a particularly high iron content.
Gavin: I always thought so too. That's crazy, and that was decades ago.
But data centers are facing a similar problem right now: once an error forms a public narrative, even if it is corrected later, it might continue to exist.
So someone always has to tell the other side of the story. I even feel the entire industry should organize. If no one is willing to do it, maybe some kind of industry organization, foundation, or even a PAC specifically for public communication is needed to run ads during mass events like the NCAA Final Four, NFL, college football, and the World Series, telling ordinary people what data centers actually are.
If there is a data center built according to these commitments in your community, the industry should clearly explain: it might invest in the local power system, bring long-term high-paying blue-collar jobs to the community, and also bring continuous investment to local finances and public facilities.
Many people mistakenly think that once the data center is built, the workers leave. Indeed, there is an employment peak during the construction phase, and subsequently many construction teams will move to the next project, but existing data centers still continuously require maintenance, upkeep, and upgrades.
And data centers can absolutely be built outside the city; they don't necessarily have to occupy the most core urban space.
This story must be told.
We also talked about another more important narrative last time: AI is increasingly and genuinely helping with medical discoveries, saving lives, and treating rare diseases.
I remember the overall atmosphere at ASCO this year was: this was the single conference with the most scientific breakthroughs many attendees had ever seen, and at least a part of it was clearly driven by AI.
If your child is sick, your parents are sick, or a loved one has a serious disease, AI is increasing their probability of getting effective treatment.
The industry must get these stories out.
People in Silicon Valley look at these things every day and feel everything is obvious, so they often don't realize that people outside have no idea. They cannot understand that there might be a huge gap between the facts that are extremely obvious in their eyes and the cognition of most Americans.
This is why I think the AI industry must learn to tell its own story better. New York State might just be the beginning. Even in some deep red states that are very supportive of economic growth, local governments will tell the industry: you haven't told your story clearly, we can't invent this story for you; but if you tell it clearly first, we can help you repeat it.
If you don't proactively tell your real situation, no one will tell it for you.
What Else Is the Market Missing? SRAM Accelerators May Continue to Improve Inference ROI
Patrick: What other important things haven't we talked about?
Gavin: I think most of the current discussions about compute have missed a hardware direction: what will happen when SRAM accelerators unconstrained by HBM DRAM truly enter the system?
These types of accelerators can usually use relatively older process nodes, so they won't compete for the same advanced process capacity as the most advanced GPUs.
When talking about inference, people usually split the process into prefill and decode. But decode can actually be split into two parts further: attention and feed-forward network.
I think a more ideal architecture might be: using a chip not necessarily equipped with HBM to handle prefill; using a high-performance chip with a large amount of HBM to handle attention; and then using an SRAM-type accelerator to handle the feed-forward network.
Whether adding these SRAM accelerators to existing deployed compute clusters or adopting them directly in newly built clusters, the return on investment I see is very attractive.
Especially for the feed-forward network part, SRAM is hard to beat.
If you try to cram everything into a single chip, you have to constantly adjust the ratio between on-chip compute, HBM, and SRAM. But workloads are constantly changing, and different tasks vary greatly; it is almost impossible for you to always hit the optimal ratio.
If the inference process can be split more modularly into three parts, letting different hardware do the most suitable work, I think the overall return on investment for AI infrastructure will significantly improve.
In This "Game of Thrones," What Other Dark Horses Are There?
Patrick: You've always described these companies as a large-scale "Game of Thrones." So is there a dark horse that is not in everyone's center of vision today but could become a major player in the future?
For example, Micron suddenly becoming as important as Anthropic and OpenAI could count as an answer. The players everyone is familiar with today include Microsoft, Amazon, NVIDIA, and SpaceX. Who else?
Gavin: I can think of a few.
Leopold might be a dark horse.
Lin Qiao from Fireworks is also absolutely worth watching; she is a very strong entrepreneur.
And our friend Scott Wu. Cognition is also a bit close to that position, and you know it very well.
These are the few names that come to mind most easily right now.
SpaceX: Has the Market Not Yet Understood It as an AI Infrastructure Company?
Patrick: What about SpaceX? At least from the period since it just entered the public market, how do you feel the market is digesting this company?
Do you think the market truly understands it? It might be the most important company among this batch of newly listed companies.
Gavin: My current feeling is that the market hasn't truly understood it yet.
For me, since its IPO, SpaceX's fundamentals have actually been continuously improving. Grok 4.5 was released, and it completed the acquisition of Cursor, while Cursor's growth has significantly accelerated.
Over the past three years, SpaceX has also proven a very important thing: they can bring large-scale compute online faster and at a lower cost than almost any company.
Now we also know that even if you deduct the huge gap between early contract prices and today's spot prices, their compute construction capability is still very strong.
In a sense, the thing that best proves how strong AI compute demand is is that SpaceX injected a massive amount of new compute into the market in a very short time, but the entire GPU market hardly showed any obvious shock as a result. This compute was completely absorbed by the market, like a freight train, hardly slowing down.
An author who writes for the FundaAI Substack said that SpaceX plans to bring 8GW of compute online.
I never like to bet against Musk not being able to do something, but I must say, 8GW is a truly unbelievable goal.
Since SpaceX signed the last batch of compute contracts, contract prices haven't dropped; they have risen. According to current figures, their monetized revenue per gigawatt of compute is about $50 billion, while the market consensus revenue expectation for next year is about $73 billion.
This doesn't even count Starlink V3 or Starlink Direct to Cell yet. Looking only at Grok 4.5 and Cursor, the combined ARR of these two businesses might soon reach $10 billion. Let's even put aside Starlink's core foundational business for now.
If SpaceX can truly bring online a compute scale close to what external rumors suggest, and the market is still only modeling it with a revenue expectation of around $73 billion, the gap here is very large.
Of course, 8GW cannot be fully powered up right at the beginning of 2027. Frankly, I am also very skeptical of the claim of "building 8GW in 18 months" itself.
But one fact is important: so far, there are very few companies that can truly bring over 500MW of new power capacity online in a year; it's basically the large hyperscale cloud providers, CoreWeave, Crusoe, and SpaceX. And SpaceX has done it at one of the largest scales, fastest speeds, and lowest costs; customers in the market also really like their clusters.
So I think the market still needs to truly see these numbers materialize before it will re-understand this company.
Patrick: Obviously, the market is not looking at SpaceX with this logic right now.
Gavin: Right. A large hedge fund in New York seems to have formed a complete short-selling logic. Their core judgment is roughly: future compute spot prices will drop by 90%, and SpaceX will still put a massive amount of new compute into the market, so the revenue ultimately generated by these assets will be far below bullish expectations.
Maybe they are right. I must always leave room for this possibility.
But I also want to make it clear: over the years, I have seen with my own eyes Musk's companies accomplish some extremely astonishing things.
"Building 8GW in 18 months" is still an exception; the reason I cite it is only because FundaAI has made this claim public, not because I can verify this number myself.
Patrick: Musk has a quote that roughly means: "We specialize in making the impossible arrive late."
Gavin: I haven't heard it before, but that's a great quote and makes a lot of sense.
From my perspective, the current stock price and market performance forecasts for SpaceX hardly fully reflect the scale of compute it might bring online in the future.
Let me emphasize again, my own expectation is far below 8GW. Finding enough power for these GPUs is very difficult, but SpaceX has always been very good at handling these kinds of difficulties in the past.
Patrick: I'm thinking of that very interesting meme: "SpaceX—a data center company?"
Gavin: One hundred percent. It can absolutely be seen this way.
And I've spent a lot of time in Starbase. The matter of orbital computing is becoming more real every day. Seeing Starship land the other day was really cool.
This also reminds me of our friends at Benchmark. They invested in StarCloud. StarCloud is a company doing orbital computing; as far as I know, SpaceX is also collaborating with them, potentially letting them use Starlink's laser communication technology, which is very important for orbital computing.
For me, this is also a very good "sanity check." At least within my circle of knowledge, the investors at Benchmark are very smart, and they are not part of the Elon ecosystem at all.
They are willing to invest in an orbital computing company at a decent valuation, and this company doesn't have the internal launch cost advantage that SpaceX has.
So I ask myself: am I crazy?
Maybe I'm crazy, maybe Elon is crazy, maybe Benchmark is crazy, maybe SpaceX's engineers are also crazy. But to make me believe that all these people are betting in a completely wrong direction at the same time, I don't think the probability is high.
Conclusion: Ultimately, Time Will Prove Who Is Right and Who Is Wrong
Patrick: Should we mention where we are recording today?
Gavin: Sure. We are sitting right next to that very famous dinner table at Benchmark.
Patrick: Then thanks to Benchmark, and special thanks to Eric for helping coordinate this recording.
It should be very interesting to see where these stocks are in a year.
Gavin: Ultimately, time will tell. People might be right, or they might be wrong; the future is inherently probabilistic.
But we are indeed in a very exciting moment right now.
Patrick: If we continue recording shows according to the "model release cycle," we might be meeting again in a few weeks.
Gavin: Yeah, it's really crazy.
Patrick: As always, it was a very enjoyable chat.
Scan the QR code to follow us