The AI sector is bloody. Why is Gavin Baker, a well-known investor, bucking the trend and bullish on AI?
Abstract: Gavin Baker said that AI stocks fell 40%-60% in July, which seriously deviated from actual fundamentals - after his field research in Silicon Valley,
No negative quantifiable indicators were found, and GPU rental prices increased by 50%-60%.
He believes that the only real risk is the tightening of the credit market, but if the operating cash flow of hyperscale cloud vendors accelerates as scheduled, the demand for debt financing will be drastically reduced.
Regulation, on the other hand, is listed as the biggest tail risk.
Markets are panicking, but Silicon Valley data tells a different story.Recently, well-known technology investor Gavin Baker revisited the sharp sell-off in the AI and semiconductor sector in July 2026 on the podcast program “Invest Like the Best”.
Baker described this period as “compressing 2022 into one month.”
He also said: After several weeks of field research in Silicon Valley, he was unable to find any negative quantifiable indicators.
The trigger for the sell-off: a series of narratives, not fundamentals
Baker sorted out the trigger chain for the July market decline.
First, Meta announced that it would rent out part of its computing power. The market interpreted this as a sign of “excess computing power and cutting capital expenses,” and the stock price immediately fell.
But Baker thinks this is a misinterpretation — “Meta's capital expenditure plan hasn't changed at all; if anything, they've become more aggressive.
“Meta then released the Llama model, which performed well.Further proof that it did not hit the brakes.Immediately after that, open source models such as Kimi K3 were released.Combined with structural changes in the Silicon Data Token Index, the market is concerned that the open source model will reduce the demand for AI infrastructure.
Baker clearly refutes this: “Generating a token, whether open source or closed source, consumes exactly the same computing power — the same floating-point operations, the same memory,
Same power consumption.Open source grabbed share, but only transferred profits from the cutting-edge model layer to the AI infrastructure layer, and there was no negative impact on computing power demand.”
Then there was the rumor about the DUV lithography machine.It triggered a sharp decline in the semiconductor equipment sector. Baker believes the market is overreacting, but it shouldn't be completely ignored either.
The only real risk: credit markets
Of all the triggers, only Baker took changes in the credit market seriously.
“Real interest rates have risen and credit spreads have widened; these are undeniable facts.” He pointed out that the bonds issued by Meta last week were far less expensive than expected.
CDS interest spreads across the board for major technology companies have widened.The core of the question is: How much debt financing is needed for large-scale AI infrastructure construction?
Baker's judgment is that if the operating cash flow of hyperscale cloud vendors accelerates as planned, the need for debt financing will be drastically reduced.
Citing data, he said that operating cash flow (non-free cash flow) of Microsoft, Meta, and Amazon has accelerated from 28% growth in the previous quarter to 32%.
After excluding one-off items such as EU fines, the actual growth rate reached 35%.” It's a substantial acceleration at this volume.”
He further estimated that if the computing power of hyperscale cloud vendors were to be monetized at the current Blackwell price rather than the lower previous generation price,
Operating cash flow is likely to jump from the current consensus forecast of 1.3 trillion US dollars to 140 million US dollars to about 2 trillion US dollars.
As a result, approximately $700 billion in credit requirements were removed from the market.
“If repriced at current prices, construction over the next few years may be entirely self-financing from operating cash flow.”
GPU price: 50%-60% increase, completely contrary to expectations
Baker shared an example that he thought would best explain the problem.
A well-known startup leased a number of Blackwell clusters 7 months ago for about $2 per GPU per hour.
Expect to pay close to $4 when renewing the contract — an increase of about 50%-60%.
“You would have expected a moderate drop in prices to be beneficial; the result was a vertical rise.”
He also mentioned that an inference cloud company publicly stated on a podcast that it plans to pay a 100% higher Blackwell price after the contract expires.
“This means that all hyperscale cloud vendors are underestimating their profitability.”
Baker said the goal of his research in Silicon Valley was to “try as hard as possible to find negative data points.”
However, apart from the controversy over third-party data about Anthropic's growth rate slowing down,
“I couldn't find any negative quantifiable indicators”.
Game theory in the memory market: no one dares to default on LTA
Baker spent a great deal of time analyzing the strategic significance of long-term supply agreements (LTAs) in the memory market.
He pointed out that memory (HBM DRAM) is currently the most critical bottleneck in AI computing power - “The more memory configured per unit of computing power, the more tokens are produced.
This is the single most important variable to improve the efficiency of token output.”In this context, the cost of defaulting on an LTA is extremely high.
Baker uses a game theory framework analysis: Assuming that in 2027 or 2028, a hyperscale cloud vendor tries to lower prices and default on LTA on the grounds of oversupply in the market,
A memory vendor can completely transfer the production capacity originally allocated to it to a competitor.
“You just lost market share. However, this industry is cyclical; after an oversupply is oversupply, there must be a shortage of supply. What will your next distribution be like?”
He concluded: “In the current environment, defaulting on an LTA could ruin your entire business and market position. This has never been the case before.”
Nvidia's New Business Model: Credit Packaging Plus Revenue Share
Baker believes Nvidia's current valuation is grossly undervalued.
Nvidia is implementing a “credit package plus revenue share” model: providing financing facilities for GPU buyers.At the same time, a revenue share is obtained when the GPU price is higher than a certain reserve price.
“It could actually allow them to quickly set up a large cloud business operating through royalties.”Baker believes that this model has been seriously misunderstood by the market and suggests that Nvidia step up its explanation to the outside world.
His logic is: in the current environment, no asset is easier to finance than Nvidia GPUs;
Nvidia also excels in matching land and electricity resources; in addition to its equity investment in almost all AI laboratories,
Nvidia is actually using tens of billions of dollars of free cash flow to systematically strengthen its competitive moat.
“Nvidia's forward price-earnings ratio is currently at its lowest point in the past two years. 100% of the market believes their profitability is grossly overrated.
Maybe, but I can't find data in Silicon Valley to support this judgment.”
SpaceX: an underrated computing power giant
Baker quoted Substack author Will Funda AI report as saying that SpaceX plans to build 8 gigawatts of computing power within 18 months.
“I would never bet Elon to fail, but it would be a truly incredible feat.”
Baker pointed out that SpaceX's current computing power monetization efficiency is about 50 billion US dollars per gigawatt, while the market's consensus revenue forecast for next year is 73 billion US dollars.
“Forget Starlink V3, forget Grok 4.5 and Cursor, the combined ARR of these businesses alone could soon surpass $10 billion.
And none of this has been fully factored into the valuation.”
He also mentioned that Benchmark has invested in the orbital computing power company Star Cloud, which is collaborating with SpaceX to use Starlink laser technology.” Maybe I'm crazy, maybe Elon is crazy,
Maybe Benchmark is crazy, maybe SpaceX's engineers are crazy too — but that doesn't seem likely.”
Biggest tail risk: regulation, not technology
Baker said, “The New York data center moratorium makes me uneasy.
We live in a post-factual, post-logical political world.” He believes that the AI industry is “doing a very bad job” in PR.
As a result, ordinary Americans have the false impression that “data centers will drive up electricity prices, deplete water resources, and take jobs.”
Baker refuted these narratives one by one: under current data center site selection agreements, electricity prices for surrounding residents usually drop;
An academic book has “four orders of magnitude” of errors in estimating data center water consumption, which the author himself has acknowledged many times;
The blue-collar jobs created by data centers are ongoing, not one-off.
The full interview is below:
AI Selloff and Fundamental Divergence: In-depth Analysis by Top AI Investors
Program: Invest Like the Best (Invest Like the Best) Recorded: August 4, 2026
Program introduction
In this episode of the program, Patrick O'Shaughnessy and Gavin Baker thoroughly analyze the drastic sell-off in the AI and semiconductor markets in July 2026.
Gavin described this period as “compressing the fluctuations of 2022 into one month.”
It also explains why the fear of the open market contrasts sharply with the strong fundamentals observed on the ground in Silicon Valley.
The two discussed in depth the accelerating operating cash flow of hyperscalers (hyperscalers), the sharp rise in GPU spot prices,
Why is the open source AI model instead driving large-scale infrastructure requirements,
As well as the game theory logic behind long-term agreements (LTAs) in the memory market, Nvidia's strategic dominance, and SpaceX's expansive data center layout,
And the regulatory risks facing AI computing power.
timestamp
0:00 Introduction
1:17 AI sell-off and fundamental divergence
10:20 Financing AI infrastructure construction
18:18 GPU prices continue to rise
24:23 Claude stirs up the market
29:01 What factors may break the current logic
36:59 Memory supply war
42:08 Nvidia's new style of play
61:51 Data Center and Governance
71:11 SpaceX and orbital computing
Text records
Gavin (Speaker 1): I wish I could be afraid. Watching these stocks fall and become cheaper, I don't want to be crazy and feel that expected returns are rising.
My main task in Silicon Valley this week was stress testing — finding negative signals and telling me something negative.
But I haven't been able to find any negative data at the level of quantitative indicators; the fundamentals are actually continuing to improve.
In terms of stocks, judging from the price at the time we recorded this episode, Nvidia's forward price-earnings ratio is already the lowest point in the past ten years.
The market believes 100% that their profitability is grossly overrated.
Patrick (Speaker 2): Gavin, it's only been two months since the last time. The pace of model releases is getting faster and faster, and our podcast recording intervals are getting shorter.
We're now basically following the rhythm recording program released by the model. Also, I remember someone pointing out before that our podcast seems to be the market...
Gavin: Part of the top appeared synchronously - no one has been able to say that since this time. What are you thinking right now? This month has been crazy.
I think July is about compressing the trend for the whole of 2022 into one month. There are indeed some negative elements in terms of fundamentals, which we should talk about, but overall,
I think the fundamental balance is improving significantly.A large number of AI-related targets fell 50% to 60% from their high point, and fell unilaterally by 40% to 60% within a month.
I asked you before I left, have you been in Silicon Valley all summer and have you heard any negative quantitative indicators about AI? Did you hear any signs of slowing down?
None at all. In fact, every indicator is speeding up.
Moreover, this is not the kind of blind optimism of people who are passionate about AI - they are supported by data,They all come from different perspectives.
AI sell-off and divergence from fundamentals
No matter where you look at it — GPU availability, GPU rental pricing, this month's DRAM spot price, token growth — everything is actually accelerating.
I think one of the main causes of the problem is that the market lacks transparency about Anthropic and OpenAI.
Also, open source inference clouds in the US that monetize through inference services — such as Fireworks, Modal, and Together — none of these companies are listed companies.
Once you see this data, the picture is quite different. Thanks to GLM 5.2 and Kimi K3, the demand for open source reasoning has greatly accelerated.
Nevatron is also progressing steadily.
We also saw the release of a small, valuable US open source model. OpenAI as a whole is accelerating, and Anthropic continues to grow strongly.
And it's almost certainly already generating significant free cash flow.
Everyone is looking at that picture: semiconductor cash flow is like this, and free cash flow for hyperscale cloud vendors is like that.
But if you leave out these private companies, then the picture will lose some very important information.
More importantly: in 2024 and 2025, even the most optimistic would expect GPU prices to slowly drop.
If you're a pessimist, expect a sharp drop in prices. But I don't think anyone expected it in 2024 or 2025,
By 2026, the price of older GPUs will rise vertically.
So everyone thought at the time to be smart and sign long-term contracts in advance.
To some extent, many emerging cloud service providers (Neo Cloud) have to do the same.
Because they require an off-take agreement (off-take agreement) to finance the GPU.
As a result, there was a huge discount on the contract price of the installed computing power compared to the current spot market.
As contracts expire one after another, computing power is repriced to a higher level, spot discounts will narrow, and computing power prices will continue to rise.
I think as the ROI question is gradually answered, you'll see plenty of signs of acceleration — this quarter has actually begun to show.
Judging from operating cash flow (note that this is operating cash flow, not free cash flow), operating cash flow reported by Microsoft, Meta, and Amazon accelerated year-on-year.
Increased from 28% to 32%.
These hyperscale cloud vendors do now have some unconventional large projects — such as billions of dollars in legal fees (mostly European Union fines)
——The number of one-off projects this quarter was particularly prominent.
If these are excluded, the growth rate jumped from 28% to 35%, which is quite a significant acceleration on this scale.
And this was before the Rubin architecture's computing power came on a large scale — Rubin will bring higher premiums when lit up, and those contracts have yet to be repriced.
It was quite a painful experience.
Replay of events
Let's sort out how this month has evolved.
First, Meta announced that it would rent out computing power to the outside world. The market interpreted this as an extreme shortfall — believing that they had overcapacity and needed to cut capital expenses, which was a disaster.
But in reality, this is not the case at all. Their later earnings report did not cut capital expenses.
The real situation is: they saw SpaceX hold a large amount of installed computing power and sell some of the clusters optimized for transactions to the market at a premium far above the contract price.
At least at the analyst level, they saw an opportunity — there was plenty of speculation that they would go to finance.
Perhaps their idea is to first use a small portion of production capacity to prove that it can generate a strong internal rate of return, and then raise equity capital to speed up the pace of development.
and increase capital expenditure.
In hindsight, that didn't seem like their real intention. However, the market interpreted this in a very negative way and immediately sold it off.
I was pretty sure that wasn't bad at the time. Meta's capital expenditure plan's various leading indicators have not changed at all; instead, they continue to increase.
It didn't take long for them to release the best model in a long time - Llama 1.1, which is indeed a very good model.
It's just overshadowed by the light of Claude 4.5, but it's much better than you expected two years ago. So it's impossible for them to hit the brakes at all.
Kimi was then released, and the market had a huge fear of open source.
At the same time, the Silicon Data Token Index (Silicon Valley Data Token Index) showed some declines and leveling off. These two things are actually related.
The Silicon Data Token Index captures the structure (mix) of tokens; it doesn't see all tokens.
However, due to the advent of GLM 5.2 and Kimi (although penetration takes time), there is a structural shift in this data:
Migrate from more expensive cutting-edge model tokens (the inference is that gross margin may be between 80% and 95%, which is questionable, but in short, extremely high) to open source tokens.
The market didn't know why this was interpreted as bad.
But the reality is: a token is a token, and the computing power required to generate a token is exactly the same — regardless of flops, memory, or power consumption.
Tokens are not completely equivalent, but generally speaking, open source grabbing shares is nothing more than shifting profits away from the cutting-edge model layer.
The elastic effect also stimulated demand for more tokens, thereby driving the demand for more computing power. Profits, on the other hand, have flowed from the cutting-edge model to the AI infrastructure layer — both Anthropic and open source models run on the same underlying cloud infrastructure.
The latter charges the same amount of computing power.
So you're actually just taking profits away from cutting-edge models and letting more profit flow to the AI infrastructure layer. This is one of the catalysts.
Then these things are superimposed and continue to ferment. For example, Jensen Huang is the world's biggest supporter of open source - if open source is bad for his business,
Would he still do that?
Of course he will support it because it's the right thing for the world, but maybe this won't be his most iconic topic.
Incidentally, I think in a world where there are only one or two leading cutting-edge models, and the gross margin is as high as 90%, open source is very important to humans and society.
We need more models to coexist.
Then there was the DUV lithography machine. Since everyone was packed into the same basket, this triggered a large-scale sell-off in the semiconductor equipment sector.
Then I think the real core concern came up: real interest rates are rising — this is reasonable because large-scale investments require financing,
Credit markets are under increasing pressure, although the vast majority of financing still comes from operating cash flow.
As a result, real interest rates rose, and credit spreads widened.
Meta issued a bond last week, which was not well priced, clearly showing that the credit market is under pressure — credit default swaps (CDS) for all companies are expanding.
A very smart private equity capitalist said that this was entirely to be expected; it was just a normal operation for banks to hedge against their own promises.
But at any rate, it doesn't look good on the surface. The widening CDS, widening interest spreads, and rising real interest rates are undeniable facts.
If we need to use debt to finance this round of construction, that would be very worrying indeed.
Financing AI infrastructure
This is why the difference between the spot price and the contract price is so important.
Let me sort out the digital logic: if you look at the operating cash flow of hyperscale cloud vendors corresponding to the installed computing power,
The market's consistent expectations are actually modeled based on the monetization rate of the Ampere architecture (two generations behind, not one generation Hopper).
Calculated in this way, the operating cash flow of hyperscale cloud vendors is about 1.3 to 1.4 trillion US dollars.
However, if a discount was applied to the current Blackwell price, the figure would be closer to $2 trillion, which would reduce credit demand by about $700 billion.
As these installed computing power are repriced, credit metrics will improve, and credit financing will become easier at that time — even if they actually choose to borrow money,
It remains to be seen.
Two months ago, we talked about Blackwell's “air pocket” (air pocket) risk — spending hundreds of billions of dollars to buy Blackwell,
However, this computing power is initially mainly used for training, and training does not produce immediate rewards, which may pose a risk.
We did see this phenomenon in the first quarter. And the reason I'm gradually relieved of this risk is because Anthropic's performance is really impressive.
The market chose to ignore this in April, May, and June, but by July, various factors were compounded.
The market stopped ignoring—just when operating cash flow really began to accelerate. This is a fait accompli.
Microsoft brought in a large amount of new production capacity in June, which was not even reflected in the second-quarter earnings report.
Core issues
At the end of the day, everything depends on whether you believe that Silicon Valley's quantitative demand signal will continue, thereby driving the installed computing power to be repriced to a higher level as the contract expires.
Operating cash flow continues to accelerate and is sufficient to support most or all construction financing. If repricing is based on current interest rates,
Financing needs for the next few years may be entirely met from within operating cash flow.
In 2022, the market has clear concerns — recession, interest rate hikes, inflation. Everyone knows it very well.
The concerns this time — DeepSeek and Liberation Day (“Liberation Day” tariffs) — are also clearly distinguishable, which is somewhat reassuring in a strange sense.
Those concerns — with the exception of credit — feel like they are nonsense on a factual level.
And our analysis convinced me that even if credit was actually needed, the problem would be solved as repricing progresses.
GPU prices continue to rise
I was chatting to a company this morning and they're one of those most popular startups.
They previously leased a Blackwell cluster of a few thousand units, and the price was around $2 per GPU per hour.
Seven months later, they renewed their lease on the exact same cluster — same size, B200, exactly the same — hoping to get the deal for just under $4.
Just heard this news today and it was really crazy. You would have expected only a moderate price drop to be beneficial; as a result, it increased by 50% to 60% within six to seven months.
Stories like this abound. I remember an inference cloud company -- I can't remember the name -- said on a podcast,
They expect the price paid for Blackwell will increase by 100% after the contract expires.
This means that all hyperscale cloud vendors are under-profitable.
My main task in Silicon Valley this week was to do stress tests and try my best to find negative data points.
The main negative sign mentioned is that third-party data shows that Anthropic's growth curve seems to be slightly off track.
This may be true, but OpenAI and open source are accelerating significantly. Taken together, there is still a net acceleration overall.
Open source is like dark matter in the universe to the open market—hard to measure directly, but if you track down what those speculative clouds say,
Whether on a podcast or at a conference, demand is clearly accelerating — and that makes sense, as GLM 5.2 and Kimi K3 brought about a huge leap in capabilities,
And this trend will continue.
Nvidia is also steadily moving Nevatron to the forefront.
This month was indeed full of challenges and setbacks. But every hypothesis has been stress-tested, and the fundamentals are still improving.
Judging from the data from the time we recorded this episode, Nvidia's forward price-earnings ratio was the lowest in the past two years.
The only time semiconductor companies had cheaper valuations were Liberation Day and DeepSeek, and those two were V-bottoms.
This means that 100% of the market believes their profitability is grossly overestimated.
We need to be humble; maybe the market is right. But my mission in Silicon Valley this week was to try as hard as possible to find negative data points.
Generally speaking, coming to Silicon Valley has positive and negative effects, and the overall bias is positive. But this time, apart from Anthropic's third party data,
I couldn't find any negative quantifiable indicators.
And this has also been strongly questioned by Anthropic shareholders — they are desperate to tell you what information they have,
I'm just worried that saying it will affect future IPO quotas.
According to third-party data, Grok and Cursor also underwent quite significant changes in July. The release of Grok 4.5 and Grok Build made July a turning point.
Claude stirs up the market
One observation worth mentioning here is that I received an article by Mike Mauboussin to the effect that the collapse of diversity is the root cause of bubbles and crashes.
Now, the public stock investors I know — whether retail investors or institutions — are immediately entered into Claude or Claude Code.
Sometimes it's Claude intelligence.
Claude is probabilistic, but the way it interprets these messages probably doesn't make much difference.
It's a bit like going back to the days of Walter Cronkite (Walter Cronkite) — there was only one recognized authoritative voice, then the media fragmented and the authoritative voice disappeared;
Now, Claude is in a sense the “Walt Cronkite” of the stock market, and everyone believes what it says.
Claude is really smart, but it's not always right, and its interpretations aren't always right. In the stock market, you are essentially dealing with probabilistic Bayesian judgments about the future.
So what you see is: a message comes out, processed by Claude, comes up with some kind of interpretation, and a large number of market participants trade based on it.
An anonymous semiconductor community member named T.B.U. posted a chart of Japanese capacitor stocks, saying that we experienced a complete capacitor cycle within six weeks.
Those stocks — whether doubling, tripling, or quadrupling — pulled up vertically and then crashed. The actual fundamental changes have yet to occur.
However, a cycle that originally took three years was completed in six weeks.
Patrick: What you feel here, especially what makes me more curious about this matter, is the progress of innovation in all aspects of inference service efficiency and model training.
How will these affect the open market over time? Have you learned any innovative directions that have a long timeline that you're particularly interested or curious about?
Gavin: There's one thing I'm really curious about: it seems like a lot of people think they're very close to solving “continuous learning” and “efficient sample learning” — we've talked about these two topics before.
If these two problems are solved, in a sense, will it result in a temporary discontinuity of demand?
I've heard that some models are trained on around 20 billion tokens, and now these big models are being trained on 300 trillion tokens.
If you could train a model with 10 trillion tokens and then have it learn samples efficiently in the real world, that doesn't sound great for training requirements.
However, training will eventually account for a very small number of semiconductor computing power requirements — not trending towards zero, but it will be very small.
It's probably the most central technical insight of the week, but you don't know if it's a long or short term thing.
SSI said they will release the model in August, and a whole bunch of new laboratories focused on this are also emerging.
If this were to happen, it would be a huge boon for the world. We all want to see this day.
But it's hard to believe this will have a negative impact on AI infrastructure requirements.
Nvidia is deeply tied to all of these startups, so if you force me to answer, what situations would actually make me change my mind and worry?
What factors could break the current logic
The main risk is that operating cash flow will not continue to accelerate. This depends to some extent on the performance of Anthropic, OpenAI, Grok, Cursor, and open source.
Second, if there is a relatively continuous sharp drop in GPU prices, the market will react immediately. If GPUs start to become easy to obtain, that would be a cause for concern.
But have you heard anyone say they have too many GPUs? None at all. It's the complete opposite; it feels like a drug market.
If the user growth of these labs stagnates or even declines, that is a real downside — unless it's just because open source tokens are grabbing a portion of the market while expanding the overall cake.
I do think the future is a pattern of multiple models coexisting, especially for AI-native companies.
They can take an open source model, and deploy it behind a router after supervision, fine-tuning and reinforcement learning customization—the router first sends queries to its own model, and then verifies it by cutting-edge models such as Claude or Grok.
In many scenarios, slightly better results can be obtained at half the cost.
However, when many people hear “halving the cost,” they think this is bad for AI; in fact, it's not at all.
The cost paid by the user is only a function of the token's gross margin. Essentially, it only transfers the token from an expensive cutting-edge model with a gross profit of about 90% to an open source token with a gross profit of about 30%.
However, these tokens consume exactly the same computing power.
Big companies are getting smarter: they set up routers to slow down AI, but that doesn't actually reduce token consumption.
Instead, they may consume more by switching to cheaper open source tokens.
AI-native companies, on the other hand, are fully betting on AI, employing almost no humans, and putting all of their budgets on tokens, with no sign of slowing down.
An idea I can't forget: only 500,000 people around the world use intelligent AI, and about 250,000 of them have a serious shortage of computing power.
There are 7.8 billion people on Earth. What happens when users grow from 500,000 to 1%, to 100 million, to 500 million?
Many people will ask: Where exactly does this operating cash flow come from? Who is the final payer?
Fundamentally, it either comes from faster economic growth brought about by increased productivity, or from labor substitution.
In many AI-native companies, you have indeed seen labor substitutions, but not because of layoffs, but because they don't recruit that many people at all.
The gross profit generated by each employee is increasing dramatically, and compared to previous generations of startups, this figure has risen almost vertically.
Another interesting observation: I heard this morning from a top tech CEO who has founded multiple companies — if you look at the companies that the founders lead and control,
Excluding overrecruitment during the pandemic, there were almost no large-scale layoffs. These people were supposed to be the quickest to embrace AI to improve efficiency.
However, they have hardly carried out large-scale layoffs — this may indicate that they think there are still plenty of opportunities left to humans, plus rapid growth in token consumption.
Judging from Cognition Reup's data on Stripe, the companies that spend the most on AI are growing significantly faster.
The Cognition Index's data is shocking. Although critics say it doesn't fully control industry differences, if you dig deeper,
You'll find that even in blue-collar industries such as plumbers and HVAC contractors, AI is bringing significant improvements.
The memory supply war
I must mention a shift in the market that I think is very important, but I misunderstood before: we are undergoing structural changes, especially in the memory sector.
From pursuing short-term performance beyond expectations, to supply chain agreements that sacrifice short-term upside in exchange for certainty, also known as long-term agreements (LTAs).
Under the LTA framework, customers prepay and lock in a price range.
It made me think about game theory questions. Let's sort out the game logic that breaks LTA.
In this game, there are only four players with real scale influence: Amazon (Trainium), Google (TPU), AMD, and Nvidia, which far surpasses all three others combined.
The reason why memory is so critical is that in a given computing power unit, the more memory, the higher the token output — this is the most critical dimension for improving the output of tokens per unit of computing power.
At the same time, it will also reduce the cost per token, which is why demand has not declined at all as a result.
Now let's say it's 2027 or 2028, and you want to break some LTA and get a lower price.
However, to a large extent, market share over the next few years will be determined by supply chain pre-allocation.
If you break the LTA — assuming it's not a serious oversupply situation — and within the next two to three years, for whatever reason, the bargaining power is back in the hands of the memory manufacturer, your business is over.
Even if Google breaks the LTA when there is an oversupply, it often means a drop in prices and a natural contraction in production capacity.
But what do you think Google's production capacity allocation will look like when the next tight supply cycle hits?
I think in the current environment where memory is the core axis of everything, breaking LTA may ruin your entire business and market position.
This is very different from before. In the past, Apple could do whatever it wanted with memory manufacturers — because it was the overwhelming biggest buyer, had no competitors, and no matter how SK Hynix was treated, Micron would follow suit.
But now it's different. You have at least four big players, and plenty of startups competing.
If you break the LTA, the memory manufacturer can say: OK, if you break the price agreement, we break the quantity agreement and give the share to your competitor.
That's why Nvidia's dominance is so inexplicably underestimated in the current environment. There's nothing easier to finance than a Nvidia GPU.
They are also playing a good game in terms of site selection and electricity acquisition — playing an important role through matching and resource integration.
They also launched a very clever new business model: similar to a “credit package+revenue sharing” structure,
If the GPU price is above a certain bottom line, you can set up a sizable cloud business through royalties very quickly.
Nvidia's new style of play
Nvidia's model, I think, has been seriously misunderstood. Essentially, they're not directly lending money to GPU buyers — they're participating in credit packaging,
Someone else is responsible for the financing. So this isn't supplier financing in the strict sense of the word; it's more like equity investment plus revenue sharing.
They do invest in equity, and almost every time they don't participate in equity investment, it's a mistake in hindsight.
They've invested in almost everything — except long-time memory companies, and they never invested in Anthropic (but have since).
If you have plenty of cash flow and are extremely optimistic about AI, and at the same time, you can also see the progress of all laboratories — keep learning the lab,
Safe Super Intelligence is working with them — Jensen is naturally optimistic about what he sees.
So why not get an increase in equity at the same time, plus a share of revenue? This not only mitigates the mismatch of cash flows among all parties, but also strengthens one's competitive position.
It can also significantly increase revenue per gigawatt.
If I were the CEO of SK Hynix, I would do the exact same thing: find GPU buyers and say I would be willing to participate in Nvidia's credit packaging structure.
Memory companies' business is inherently less stable and predictable, but they may be able to reduce their own risk by prepaying a portion of their cash while seeking a share of ongoing revenue.
I'm sure agencies like Blackstone and Apollo are recommending similar solutions to memory companies.
This is actually a natural extension of LTA logic — by sacrificing part of the upward trend in exchange for sustainability, you can also obtain concessions for recurring revenue.
This is exactly what Nvidia is doing, and I think they should explain this logic more clearly to the market.
From the perspective of game theory, if Anthropic had a positive layout like OpenAI in terms of computing power, it would have been far ahead for a long time.
OpenAI is back at the forefront of competition, and Grok is also in the game. These are all companies with cutting-edge computing power.
Who else is going to relax the throttle after seeing all this? Especially when financing can come entirely from operating cash flow.
Patrick: What do you think of the DUV lithography machine news?
Gavin: I think both statements can be true at the same time. For example: if the DUV machine is a jet turbine engine, then the EUV machine is probably a curved speed engine;
Or DUV is a propeller plane, and EUV is a jet. They didn't have it before, and now they supposedly do. It's a phase change—you go from liquid to solid.
That “jet engine,” or “propeller plane,” may be 25 years behind current technology, but it is still important and should not be ignored.
The market's reaction was indeed grossly excessive. If this matter were to affect ASML orders, it might have to wait until 5 years from now,
And the market will repeatedly worry about it and forget it over and over again during this period. So this may be an overreaction, but it shouldn't be ignored either.
Data Centers and Governance
About other companies
Open source is getting closer to the cutting edge, and companies like Fireworks make it extremely easy to customize—you can train results that are comparable to or even surpass cutting-edge models for specific scenarios.
Furthermore, costs have been drastically reduced, which is a great thing for the entire software industry and a large number of AI-native companies.
Our friend Vijaya said he had never seen so many companies generate $50 million in annual revenue and cash flow in 9 months from inception.
These companies may have previously been derided as “ChatGPT shells,” but now that they have open source, they have accumulated field-specific data.
A real moat was created.
Fireworks has launched a product called Nexus. With just three lines of code, it can ingest your data and fine-tune it through reinforcement learning.
Also, through intelligent router distribution queries - Harvey (before the acquisition), Cursor, Harvey, and Lagora are all vigorously using this set of ideas.
Because if you can transfer 30% to 60% of your token consumption to your reinforcement learning model and only use the cutting-edge model for the rest,
You're no longer just a shell; you have a real moat.
Patrick: The Cursor thing is very interesting -- AI is running fast on our human learning curve: using cutting-edge model planning, assigning tasks to lighter models, increasing efficiency by 15 times.
Gavin: It's really interesting. Perhaps an open source token with a low gross margin — only slightly inferior to the cutting-edge model — would instead make the top cutting-edge tokens more valuable.
If you have a batch of open source models with an intelligence level of 120, and the price is extremely low, wouldn't a model that can coordinate and schedule them and have an intelligence level of 160 would be more valuable?
The last time we talked, I've always been surprised that economic returns are highly focused on cutting-edge models.
This pattern is changing, but open source tokens may instead reinforce the value of the most advanced tokens.
What's most shocking: those open-source-based inference cloud companies are growing almost as fast as early cutting-edge labs, yet they hardly burn money. The Rule of 40 metrics are amazing.
About maximum risk
Regulation is unquestionably the biggest risk, and the most obvious risk.
New York announced a data center moratorium, and we live in a post-factual, post-logical political world.
The AI industry does a terrible job in PR, and the mainstream narrative for ordinary Americans is that data centers will drive up your electricity bill, seize your water resources, and then take your job.
The reality is that according to the agreement signed now, after the data center is installed, the electricity bills of the surrounding residents will actually tend to drop.
Because there are “behind-the-meter deals” (behind-the-meter deals) behind it. Previously, data center developers only needed to buy a few cars for local police and fire stations.
Now they want to build hospitals, schools, police stations, and fire stations, and reduce residents' electricity bills.
Moreover, related jobs are ongoing because data centers require a large number of plumbers, electricians, and HVAC contractors — data centers are, in a sense, the biggest driver of blue-collar wage growth in my lifetime.
However, the Democratic Party, which nominally represents blue-collar workers, is blocking these employment opportunities.
There is a saying about data center water consumption: A scholar made a calculation mistake in the book and overestimated the water consumption by 10,000 times — not an order of magnitude, but four orders of magnitude.
The author himself has acknowledged this mistake many times. This question has long been fully falsified, but rumors are still spreading.
This reminded me of the story of Popeye eating spinach — also because the decimal point was misplaced in an academic book, spinach's iron content was overestimated, and this mistake has been circulating until now.
This industry needs someone to stand up and tell the truth. Maybe a foundation is needed to advertise during the NFL, World Series, college football seasons,
Tell the public what a data center really looks like: if you sign a commitment agreement, your electricity bill will drop, the community will benefit, and a large number of high-paying blue-collar jobs will continue to appear.
It has little impact on water resources and the environment, and it could be built 10 miles away from the outskirts of the city.
At the same time, stories of AI saving lives and curing rare diseases also need to be told.
The atmosphere of this year's ASCO Annual Meeting — “This was the most scientific breakthrough in a single conference in history” — a significant portion of the contribution came from AI.
If you have a sick child, parent, or loved one, AI is actually improving their chances of recovering.
People in the industry think these are all obvious common sense, so they take it for granted that everyone understands them.
But in fact, there is a huge gap between this and what most ordinary Americans think.
The New York ban feels like it's just the beginning. Even in the most business-friendly Crimson states, if you don't take the initiative to tell your story well, no one will tell it for you.
SpaceX and orbital computing
Patrick: What do you think of SpaceX being digested by the open market? Do you think the market really understands this company?
Gavin: I don't understand it. Fundamentals have been improving since the IPO — Grok 4.5, Cursor's acquisition and acceleration,
And they've proven that they can introduce more computing power at a lower cost and faster than anyone else.
When they entered the market, they just hit a high spot price - in a strange sense,
Instead, it's one proof of the strong demand for computing power: they put a lot of computing power into the market all of a sudden, and the market had almost no feeling, like a freight train didn't slow down at all.
One Substack analysis (Funder AI) suggests that SpaceX will try to introduce 8 gigawatts of computing power. I'd never bet Elon lost, but it would be an extremely amazing feat.
Interest rates have risen without falling since they last signed the contract, and their monetization efficiency is around $50 billion per gigawatt. The consensus forecast for next year is $73 billion.
Let's not mention Starlink V3, Starlink Direct Connect, Grok 4.5, and Cursor. Together, these businesses alone can quickly reach $10 billion in annual recurring revenue.
Ignore the core Starlink basic business — if they can actually introduce that much computing power, the consensus estimate of 73 billion is 8 gigawatts times 50 billion per gigawatt,
And apparently they won't all be lit up in early 2027, which seems extremely unlikely to happen to me, and I hardly believe that report.
But it's worth noting: up to now, companies that have been able to introduce more than 500 megawatts of electricity every year,
There are only hyperscale cloud vendors, Coreweave, Weave, Crusoe, and SpaceX, and SpaceX, and SpaceX accomplished this feat at the lowest cost and fastest speed.
Their clusters are really well received by users.
The market's current mainstream interpretation of SpaceX is that the spot price will drop by 90%, and a large amount of computing power will not generate the expected benefits.
Maybe. But I've seen Elon's company make the seemingly impossible possible time and time again.
I think in the current stock price, there are very few expectations from this part of the potential computing power.
It was very difficult to introduce computing power on a large scale, and it was extremely difficult to power up GPUs, but they did, and it didn't feel like it was factored into market expectations or estimates.
What's interesting is that the plot “SpaceX — Data Center Company” seems less and less like a joke now.
I've been at the interstellar base for quite a while, and the orbital calculations feel more realistic every day. Starship landed the other day and it was really cool.
Our Benchmark friend invested in Star Cloud, an orbital computing company, and SpaceX collaborated with it to some extent.
It is said that Starlink laser communication technology will be opened, which is essential for orbital calculations. Benchmark's people are very smart and are not in Elon's ecosystem at all.
But I chose to invest in an orbital computing company at a fairly high valuation without having SpaceX's internal launch cost advantage — this is a good “rational person test” for me.
Maybe I'm crazy, maybe Elon is crazy, maybe Benchmark is crazy, maybe SpaceX engineers are crazy, but these are all crazy things. I think the probability is really low.
Patrick: We're recording this episode at Benchmark's office and sitting at their famous dinner table. Thanks to Benchmark, thank Eric for helping with coordination, and thank all partners.
Gavin: After a year, we'll know who's right and who's wrong. Time will tell. The future is probabilistic, but we're in an exciting time.
Patrick: If we keep on posting rhythm shows with the models, we'll probably see each other again in a few weeks, maybe here at Benchmark.
It's always a pleasure to talk to you.
Twitter: https://twitter.com/BitpushNewsCN
Compare TG Community: https://t.me/BitPushCommunity
Compare TG subscriptions: https://t.me/bitpush
