Top AI Investor Gavin Baker: July Sell-Off Misaligned with AI Infrastructure Fundamentals

Organized & Compiled: DeepChain TechFlow

Guest: Gavin Baker, Founder and Chief Investment Officer of Atreides Management

Host: Patrick O’Shaughnessy, Host of Invest Like The Best and CEO of Positive Sum

Podcast source: Invest Like The Best

The AI Selloff Doesn’t Match the Data | Top AI Investor Explains

Broadcast date: August 4, 2026

Disclosure: This guest, Gavin Baker , who founded Atreides Management, manages approximately $5 billion in assets and holds significant positions in NVIDIA (for over 20 years), Astera Labs (approximately 10% of the fund), Micron, Cerebras, and other AI infrastructure stocks discussed in this episode. His views are directly aligned with his portfolio holdings.

However, he is indeed one of America’s top technology investors, having served as Managing Director at Fidelity Investments and managed technology funds exceeding $10 billion. As an early and long-term investor in NVIDIA, Tesla, and SpaceX, Gavin brings over two decades of experience investing in semiconductors and cutting-edge technologies. In the AI wave, he is renowned in Silicon Valley and Wall Street for his deep insights into chip supply chains, compute bottlenecks, and the commercialization pathways of AI.

Gavin Baker is the founder of Atreides Management and has maintained a heavy position in NVIDIA for the past 20 years, making him one of the most steadfast bulls in AI infrastructure. This episode was recorded after the July plunge in AI stocks, during which he returned from Silicon Valley to "stress-test" all of his assumptions. His core conclusion: the July sell-off was completely disconnected from fundamental data—all quantitative metrics were accelerating, yet the stock price was in free fall.

Baker described July as "compressing 2022 into a single month." A large number of AI stocks dropped 40% to 60% from their highs, yet GPU spot prices rose instead of falling (up 50%-60% over six months); the operating cash flow of hyperscale cloud providers accelerated from 28% to 35% (after adjusting for one-time items); and token issuance continued to grow. He believes the market has made three misjudgments: misinterpreting Meta’s leasing of compute capacity as oversupply (when it’s actually mimicking SpaceX’s strategy of selling compute at premium prices), misreading the surge in open-source models as declining demand (when it merely reflects profit shifting from frontier model layers to infrastructure layers), and misinterpreting widening CDS spreads as a sign of credit crisis (when it simply reflects banks hedging their commitments). The only thing that truly concerns him is the rise in real yields and tightening credit markets, but his calculations show that if existing installed compute capacity were repriced at current spot prices, hyperscale cloud providers could largely fund AI infrastructure investments through operating cash flow alone, with minimal reliance on debt.

1. Regarding the divergence between selling pressure and fundamentals

I spent two months in Silicon Valley and didn’t hear a single negative quantitative metric—GPU availability, GPU rental prices, DRAM spot prices, token growth—all indicators are accelerating.

NVIDIA is currently at its lowest forward P/E ratio in the past decade. The market is 100% convinced that AI companies are severely overvalued in terms of earnings.

OpenAI is accelerating, Anthropic is experiencing strong growth, and open-source models are surging dramatically. However, data on Anthropic and OpenAI is not visible in public markets, creating information asymmetry.

2. On open-source models and computational requirements

A token is a token. Whether generated by a cutting-edge model or an open-source model, it requires the same amount of compute, memory, and electricity to produce.

Open-source models take away profit margins from frontier models, not compute demand. They turn tokens with 90% gross margins into tokens with 30% gross margins, while consuming even more GPU hours at the underlying level.

Jensen is the world’s largest supporter of open source. If this were bad for his business, would he really make it a signature issue?

3. Regarding Claude's impact on the market

Claude is essentially the Walter Cronkite of the stock market. Everyone feeds news into Claude and trades based on Claude’s interpretations. Claude is smart, but it’s not always right.

Someone posted a chart of Japanese capacitor stocks saying, "We've gone through the entire capacitor cycle in six weeks." The fundamentals haven't caught up yet, but the stock price has already surged and then crashed.

4. Game Theory on the Memory Market

Suppose in 2027 or 2028 you want to terminate an LTA to get a lower price, but if leverage returns to memory manufacturers over the next two to three years, you’ll be out of the game—you could destroy your entire business.

What NVIDIA is doing is essentially providing GPU buyers with a credit wrapper and a revenue share above the base price, allowing them to rapidly build a massive cloud business through royalties.

5. Regarding SpaceX and orbital calculations

Over the past three years, SpaceX has accessed more computing power at lower prices than anyone else. Now there are reports they’re moving to 8 gigawatts of computing power—I never bet against Elon, but that is indeed an astonishing number.

Orbital computing is becoming more real every day. Benchmark invested in StarCloud, an orbital computing company without SpaceX’s internal launch costs. If Benchmark thought this wasn’t viable, they wouldn’t have invested.

6. Regarding Regulatory and Public Relations Crises

Data centers are the best thing I've ever seen for blue-collar wages. But the Democratic Party, which should represent blue-collar workers, is pushing their jobs overseas.

Someone overstated the water consumption of data centers by ten thousand times in an academic book. The author has admitted the error multiple times, but it’s just like the story of spinach’s iron content: by the time the lie has traveled around the world, the truth hasn’t even finished putting on its pants.

Patrick O’Shaughnessy: What happened this month?

Gavin Baker: I would describe July as "compressing 2022 into one month." There were certainly some negative fundamental factors to address, but overall, the fundamental balance has improved significantly. A large number of AI stocks dropped 40% to 60% from their highs, in a straight-line decline. I ask you, during your summer in Silicon Valley, did you hear even one negative quantitative metric about AI?

Patrick O’Shaughnessy: A signal of slowdown?

No. In fact, every metric is accelerating—whether it's GPU availability, GPU rental prices, this month's DRAM spot prices, or token growth, all are accelerating.

I believe a large part of the issue is that the public markets lack visibility into Anthropic and OpenAI. Meanwhile, there are open-source inference clouds like Fireworks, Baseten, Modal, and Together commercializing inference. Open-source models are accelerating rapidly due to GLM 5.2 and Kimi K3, with Nemotron also making steady progress. OpenAI is accelerating, Anthropic is growing strongly, and it’s almost certain they are generating substantial free cash flow.

Everyone has seen that chart: semiconductor cash flow is rising, while the free cash flow of hyperscale cloud providers is falling. But you missed the private companies. OpenAI and Anthropic’s cash flows aren’t on that chart. I think the chart is missing something even more important: in 2024 and 2025, even the most bullish among you believe GPU rental prices will decline gradually. The bears think they’ll crash. I don’t think anyone in 2024 or 2025 anticipated that old GPU prices would still be surging in 2026.

Patrick O’Shaughnessy: How will the financing environment look over the next six months? How much credit is needed for this build?

This involves the classic capital cycle problem. If there is a supply-demand imbalance, things can unravel very quickly, as happened with the dot-com bubble. If you believe that hyperscale cloud providers can fund this build-out through operating cash flow, then credit tightening isn’t as concerning.

My calculation is as follows: The consensus assumes that hyperscale cloud providers monetize their compute at the rate of Ampere (two generations behind), using Blackwell and Rubin capacity. Hyperscale cloud operators generate operating cash flow of $1.3 to $1.4 trillion. If they monetize at a rate higher than Ampere but lower than Blackwell, the figure would be closer to $2 trillion, reducing credit demand by approximately $700 billion. Additionally, as installed compute is repriced and operating cash flow continues to accelerate, credit metrics will improve, making financing easier.

Signals from the credit market are certainly hard to ignore. Meta issued debt last week, and the pricing was worse than you might think. CDS spreads for all companies are widening. Real yields are rising. These are all facts. If we need debt to finance this build-out, then this is indeed a significant negative signal. But if hash rate is repriced at current spot prices, we may not need much debt at all.

Patrick O’Shaughnessy: What specific data did you hear in Silicon Valley?

This morning, I spoke with a company that rented a cluster of several thousand Blackwells at around mid-$2 per GPU hour. Seven months later, they hope to renew the same cluster with the same number of B200s for less than $4. That’s a 50% to 60% increase in seven months.

Another reasoning cloud company publicly stated on a podcast that they plan to pay 100% more for Blackwell when the contract expires, indicating that all hyperscale cloud providers are underreporting revenue.

There were several sell catalysts in July. First, Meta announced it would lease out computing power, leading the market to believe this indicated excess capacity and plans to cut capital expenditures. This is completely untrue. Meta observed that SpaceX possesses a large amount of installed computing power and is selling optimization-optimized clusters on the market at prices far above contract rates. Meta saw an opportunity: demonstrate high IRR on a small portion of capacity first, then raise equity capital, and subsequently increase capital expenditures. Meta’s capital expenditure telemetry data has shown no change whatsoever; if anything, it has become even more aggressive. Shortly after, they released Muse 1.1, the best model in a long time.

Then came Kimi K3, sparking renewed market panic over open source. Meanwhile, Silicon Data’s token index flattened. But this is due to the rising share of open-source tokens, whose weighting in the index leads to structural flattening. A token is a token—whether it comes from a cutting-edge model or an open-source model, the compute, memory, and power required to produce it are the same. Open source takes away the profit margins of cutting-edge models, not their compute demand.

Patrick O’Shaughnessy: How do you think the market is digesting these messages?

There’s something I’ve been thinking about. Mike Mauboussin has a theory that the collapse of diversity leads to bubbles and crashes. Right now, in the public markets, whether retail or institutional investors, every piece of news is fed into Claude, sometimes Claude Code or Claude Agent. Claude is probabilistic, but the way people interpret news is becoming remarkably similar.

Claude is essentially the Walter Cronkite of the stock market. Everyone believes what it says and trades accordingly. Claude is smart, but it isn’t always right. The stock market is fundamentally a probabilistic Bayesian interpretation of the future. You see a news headline, it’s fed into Claude, Claude interprets it in some way, and then a large group of people trade based on that interpretation.

An anonymous semiconductor industry account called TBU posted a chart of Japanese capacitor stocks, saying, "We’ve gone through the entire capacitor cycle in six weeks." The stock prices doubled, tripled, quadrupled, then crashed. The fundamentals haven’t even caught up yet—you’ve compressed a cycle that normally takes three years into just six weeks.

Patrick O’Shaughnessy: If you had to identify a scenario that would truly scare you, what would it be?

Operating cash flow is no longer accelerating, and this will be the most critical negative signal. This largely depends on the overall performance of Anthropic, OpenAI, Grok, Cursor, xAI, and open-source projects.

If GPU spot prices experience a sustained sharp decline, that’s also scary. But have you ever heard anyone say they have too many GPUs? Not a single person. In fact, it sounds more like a black market.

From a technical standpoint, I think the most interesting potential risk is continual learning and sample-efficient learning. If these are solved, it could mean a temporary gap in training demand—you train a model on 300 trillion tokens, then reduce it to just 10 trillion tokens and deploy it into the world to learn sample-efficiently, which isn’t great for training demand. But the share of training in semiconductor demand is approaching a small, yet non-zero, number; inference is the dominant factor. SSI says they’re releasing a model in August, and a new wave of labs is focusing on this direction. This is good for the world, but it’s hard to say whether it’s positive or negative for infrastructure demand.

Patrick O’Shaughnessy: Everyone is talking about LTA—could you elaborate on that?

We need to shift from "exploiting short-term numbers" to "exchanging durability for long-term agreements (LTA)." Customers prepay a sum, with both a price floor and ceiling. This is similar to the "labor hoarding" discussed years ago when companies were reluctant to lay off employees.

Let’s think about the game theory behind tearing up an LTA. Four companies are truly significant at scale: Amazon’s Trainium, Google’s TPU, AMD, and NVIDIA—which is larger than the other three combined. Suppose in 2027 or 2028, you’re considering tearing up an LTA to get a lower price. But your market share over the next few years will largely be determined by supply allocation and your pre-orders. If you tear up the LTA and the leverage shifts back to memory vendors, you’re out.

What happens if Google pulls out of LTA? It likely signals oversupply, falling prices, and natural capacity contraction. Then, this cyclical industry will shift from oversupply to undersupply. At that point, how do you think memory manufacturers will allocate Google’s volume? You might destroy your entire business and brand.

It wasn't like this before. Apple was the largest buyer and could call the shots because no one could match their scale. But this time is different—you have at least four buyers competing, plus a host of startups. You broke the LTA, and memory manufacturers can say, "Fine, you violated the price agreement, so we'll break the volume agreement and give our volume to your competitors."

Patrick O’Shaughnessy: What do you think of NVIDIA’s current strategy?

NVIDIA has introduced a very clever new business model, which I would describe as a "credit packaging" arrangement combined with a revenue share above a floor price. This could enable them to rapidly build a massive cloud business through royalties. It’s also another way to mitigate cash flow mismatches.

This is not traditional supplier financing. They are not lending money to GPU buyers. Someone else is lending to the buyers, while NVIDIA only provides equity investment and credit enhancement. NVIDIA includes in all equity investment agreements that "this money cannot be used to purchase NVIDIA chips," but money is fungible.

If I were the CEO of SK Hynix, I would do exactly what NVIDIA does: go to GPU buyers and say, I’m participating in NVIDIA’s credit packaging, I have some cash to front, and I want a share of ongoing revenue. This is an extension of the LTA logic: you’re trading short-term upside for durability, and now you can also earn a royalty on ongoing revenue.

NVIDIA's competitive advantages are stronger than ever. If you need to finance chips, nothing is easier to fund than NVIDIA GPUs. Do you need land and power? They’re exceptionally good at securing both. Their revenue per gigawatt is increasing, and their competitive position is growing stronger.

Patrick O’Shaughnessy: What’s the worst-case scenario for the AI field? Regulation?

Regulation is undoubtedly the biggest risk—this is obvious. One of the reasons I stepped out this week was to get a reality check. I didn’t want to be the person watching the stock get cheaper while expected returns rise, yet feeling like I’m crazy. Fundamentals have actually improved significantly from June to July, but I still believe regulation remains the most significant, unavoidable risk.

New York has enacted a moratorium on data centers. We live in a political world that is post-fact and post-logic. The AI industry has done a very poor job with public relations. The narrative in Washington and among many ordinary Americans is: data centers raise your electricity bills, deplete your water supply, and take away your jobs.

But the reality is exactly the opposite. The agreements signed by today’s data center developers aren’t just about buying new cars for police and fire departments—they’re building hospitals, schools, new police stations, and fire stations for the community, while also lowering residents’ electricity bills. Due to behind-the-meter agreements, electricity rates for nearby residents typically decrease after a data center is established. And this work is ongoing, not one-time: you need plumbers, electricians, and HVAC technicians to maintain and upgrade these facilities. Data centers are the best thing I’ve ever seen for blue-collar wages.

The issue of water consumption is even more absurd. An academic book overestimated the water usage of data centers by a factor of ten thousand—not one or two orders of magnitude, but four. The author has repeatedly admitted the error, and the claim has been thoroughly debunked. But this is exactly like the story of spinach and iron: an academic book misplaced a decimal point by two places, and everyone came to believe spinach was the iron-rich food, a belief that still persists 80 years later. While the lie has traveled halfway around the world, the truth hasn’t even finished putting on its pants.

At the ASCO (American Society of Clinical Oncology) annual meeting, this year’s atmosphere was “the most scientific breakthroughs ever seen at a single conference,” with a large portion related to AI. If you have a child, parent, or loved one who is ill, AI is substantially increasing their chances of recovery. These stories need to be told. The industry needs to run ads during NFL games, college football, and the World Series to let people know what data centers are truly doing.

People in Silicon Valley think all of this is obvious, so they assume everyone else knows it too. They can’t understand that for most Americans, this is a completely different—or even opposing—perspective. Even those in deeply red, growth-oriented states are saying: You haven’t told your story well; if you had, we could help retell it, but you’re the experts.

Patrick O’Shaughnessy: What do you think about China having DUV machines?

Both interpretations could be correct. For example, DUV is like a propeller plane, while EUV is a jet turbine. China didn’t have it before, but it’s now said to have it—that’s a paradigm shift. Even though that jet engine is 25 years behind, it’s still significant and shouldn’t be ignored. However, the market may be overreacting. If this truly impacts ASML’s orders, it might not happen for another five years, and in the meantime, the market will likely forget and remember it several times over.

This is indeed very important for China. They are very smart and hardworking, treating it as a critical national priority. But you can't accelerate the learning curve—you have to genuinely go through those learning cycles.

Patrick O’Shaughnessy: What is the impact of open-source models on the industry?

Open-source models are closing in on the cutting edge, and inference clouds like Fireworks make custom models incredibly easy to deploy—this is a golden opportunity for the software industry. AI-native companies can use open-source models to perform their own RL fine-tuning, reducing reliance on cutting-edge models from 100% to just 30%-60% of token consumption, while using their own models for the rest. You’re no longer just a "ChatGPT wrapper"—you now have proprietary data and real defensibility.

Cheap tokens may make the most advanced tokens even more valuable. If you can run a 120 IQ open-source model affordably, isn’t a 160 IQ cutting-edge model that can orchestrate them even more valuable?

Another overlooked aspect is the SRAM accelerator. These chips are not constrained by HBM DRAM production bottlenecks and are typically manufactured using older processes, avoiding competition for capacity with the latest GPUs. When breaking down inference, it consists of two stages: prefill and decode. Decode further splits into attention and feed-forward network. The ideal scenario is to run prefill on chips without HBM, attention on high-performance chips with HBM, and feed-forward on SRAM-based chips. No matter how you adjust the ratios of computation, HBM DRAM, and SRAM, workloads are always changing—splitting them into these three parts significantly improves the overall ROI of AI.

Patrick O’Shaughnessy: Are there any "dark horse" players currently out of the spotlight that could become major players on the level of Game of Thrones?

Lee Buu might be a dark horse. Lynn from Fireworks is also an absolute powerhouse. And our friend Scott Wu from Cognition is also worth watching.

Patrick O’Shaughnessy: What about SpaceX? Does the market understand this company?

SpaceX's fundamentals have improved since its IPO. Grok 4.5 and the acquisition of Cursor have clearly accelerated progress. Over the past three years, they have accessed more computing power at lower prices than anyone else. They entered at the peak of spot prices, but the market has fully absorbed the massive amount of computing power they’ve deployed, and the freight train hasn’t slowed down at all.

A Substack writer cited public reports claiming SpaceX is aiming for 8 gigawatts of computing power. I never bet against Elon, but this is truly an astonishing feat. They’re currently generating revenue at a rate of about $50 billion per gigawatt, with consensus estimates for next year at $73 billion. Even without accounting for Starlink V3, direct-to-phone connectivity, Grok 4.5, or Cursor, if they truly connect even close to 8 gigawatts of computing power, it would far exceed market expectations.

The market’s current interpretation isn’t like this. There are hedge funds in New York shorting the market, believing that spot hash rate prices will drop 90% and that the massive amount of computing power connected to SpaceX won’t generate that much revenue. Maybe. But Elon’s company has been doing incredible things for years. In his words, “We’re good at making the impossible late.”

Orbital calculations are becoming more realistic every day. Benchmark invested in StarCloud, an orbital calculation company without SpaceX’s internal launch cost advantages. If Benchmark’s team thought this wasn’t viable, they wouldn’t have invested. So either I’m crazy, or Elon is crazy, or Benchmark is crazy, or SpaceX’s engineers are crazy. The odds of that are slim.