TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Podcast Notes | Semiconductor and Tech Investment Guru Gavin Baker Goes All-in on AI Infrastructure: July Sell-off Completely Divergent from Fundamentals, NVIDIA at Decade-Low Forward P/E Ratio

Podcast Notes | Semiconductor and Tech Investment Guru Gavin Baker Goes All-in on AI Infrastructure: July Sell-off Completely Divergent from Fundamentals, NVIDIA at Decade-Low Forward P/E Ratio

2026.08.06
Share
"I spent two months in Silicon Valley and didn't hear a single negative quantitative metric."

Compiled & Translated: TechFlow

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

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

Source: Invest Like The Best

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

Air Date: August 4, 2026

Disclosure: The guest for this episode, Gavin Baker, founded Atreides Management, which manages approximately $5 billion in assets with heavy positions in NVIDIA (over 20 years), Astera Labs (~10% of the fund), Micron, Cerebras, and other AI infrastructure names discussed in this transcript. His views are directly tied to his portfolio holdings.

That said, he is undeniably one of America's top tech investors, having served as Managing Director at Fidelity where he managed a tech fund exceeding $100 billion. As an early and long-term investor in NVIDIA, Tesla, and SpaceX, Gavin brings over two decades of experience in semiconductor and frontier tech investing. Throughout the AI boom, he has earned a reputation in Silicon Valley and Wall Street for his deep insights into chip supply chains, computing bottlenecks, and the commercialization path of AI.

Key Takeaways

Gavin Baker is the founder of Atreides Management and has maintained a heavy long position in NVIDIA for the past 20 years, making him one of the most steadfast bulls in the AI infrastructure space. This episode was recorded after the July AI stock selloff, during which he returned specifically from Silicon Valley to "stress test" all his assumptions. His core conclusion: July's selloff completely diverged from fundamental data. All quantitative metrics are accelerating, yet stock prices are in freefall.

Baker described July as "compressing 2022 into a single month." A massive swath of AI stocks fell 40% to 60% from their highs, yet GPU spot prices rose instead of falling (up 50%-60% over six months), super-scale cloud vendors' operating cash flow accelerated from 28% to 35% (adjusting for one-time items), and token production continued to grow. He argues the market made three critical misjudgments: misreading Meta's decision to rent out compute as capacity oversupply (it's actually mimicking SpaceX by selling compute at premium prices), misreading the explosion of open-source models as demand contraction (profits are simply shifting from the frontier model layer to the infrastructure layer), and misreading widening CDS spreads as a credit crisis signal (it's merely banks hedging their commitments). The only thing that genuinely keeps him up at night is rising real yields and tightening credit markets. However, his calculations show that if installed compute is repriced at current spot values, super-scale cloud vendors could essentially fund their AI infrastructure investments through operating cash flow alone, without needing to rely heavily on debt.

Key Highlights

1. On the divergence between selloff and fundamentals

“I’ve spent two months in Silicon Valley and haven’t heard a single negative quantitative metric regarding AI. GPU availability, GPU rental prices, DRAM spot prices, token growth—all indicators are accelerating.”

“NVIDIA currently trades at its lowest forward P/E in a decade. The market is 100% pricing in severe earnings overestimation for AI companies.”

“OpenAI is accelerating, Anthropic is growing strongly, and open-source models are exploding. But public markets have no visibility into Anthropic’s and OpenAI’s actual data, creating information asymmetry.”

2. On open-source models and compute demand

“A token is a token. Whether generated by a frontier model or an open-source model, it requires the exact same amount of compute, memory, and power 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, but the underlying GPU-hour consumption actually increases.”

“Jensen is the world’s biggest open-source supporter. If this were bad for his business, would he really make it a personal signature issue?”

3. On Claude's impact on the market

“Claude is basically the Walter Cronkite of the stock market. Everyone feeds news into Claude and then trades based on its interpretation. Claude is smart, but it isn't always right.”

“Someone posted a chart of Japanese capacitor stocks saying, ‘We just completed an entire capacitor cycle in six weeks.’ Fundamentals hadn't even caught up, but prices had already skyrocketed and crashed.”

4. On game theory in the memory market

“Imagine trying to break an LTA (Long-Term Agreement) in 2027 or 2028 to secure lower prices. But if leverage shifts back to memory vendors over the next few years, you’re locked out. You could torpedo your entire operation.”

“What NVIDIA is doing is essentially offering GPU buyers a credit wrapper plus an above-floor revenue share. This effectively allows them to rapidly build a massive cloud business through royalties.”

5. On SpaceX and orbital compute

“SpaceX has onboarded more compute at lower costs faster than anyone else over the past three years. There are reports they’re aiming for 8 gigawatts of compute. I never bet against Elon, but that is a staggering number.”

“Orbital compute is becoming a reality day by day. Benchmark invested in StarCloud, an orbital compute company without SpaceX’s internal launch cost advantage. If Benchmark felt it wasn't viable, they wouldn't have invested.”

6. On regulation and PR crises

“Data centers are the best thing for blue-collar wages I’ve ever seen in my life. Yet Democrats, who are supposed to represent blue-collar workers, are actively pushing their jobs away.”

“Someone cited an academic book that overestimated data center water consumption by ten thousand times. The author has admitted the error multiple times, but it’s just like the spinach iron content story: by the time the truth gets its pants on, the lie has already traveled halfway around the world.”

Chapter 1: “Compressing 2022 Into One Month”

Patrick O’Shaughnessy: What happened this month?

Gavin Baker: I’d describe July as “compressing 2022 into a single month.” There are definitely some headwinds in the fundamentals worth discussing, but overall, the fundamental balance is significantly improving. A huge block of AI stocks dropped 40% to 60% from their peaks in a straight line. Tell me, while you’ve been spending the summer in Silicon Valley, have you heard a single negative quantitative metric about AI?

Patrick O’Shaughnessy: A single deceleration signal?

Not one. In fact, every metric is accelerating. GPU availability, GPU rental pricing, this month’s DRAM spot prices, token growth—all are accelerating.

I think a major part of the problem is that public markets lack visibility into Anthropic and OpenAI. Then there are these open-source inference clouds—Fireworks, Baseten, Modal, Together—commercializing inference. Open-source models are accelerating rapidly due to GLM 5.2 and Kimi K3, while Nemotron continues its steady progress. OpenAI is accelerating, Anthropic is growing robustly, and it’s almost certain they’re generating significant free cash flow.

Everyone has seen that chart showing semiconductor cash flows trending up while super-scale cloud vendors' free cash flows trend down. But it misses those private companies. OpenAI’s and Anthropic’s cash flows aren’t on that chart. I also think the chart omits a crucial point: in 2024 and 2025, even the staunchest bulls expected GPU rental prices to slowly decline. Bears expected a crash. I don’t think anyone in 2024 or 2025 foresaw that old GPU prices would continue climbing in 2026.

Chapter 2: Who Will Pay for AI Buildout?

Patrick O’Shaughnessy: What will the funding environment look like over the next six months? How much credit does this buildout require?

This touches on classic capital cycle dynamics. If supply and demand go out of whack, things can unravel incredibly fast, like the internet bubble. If you believe super-scale cloud vendors can self-fund this buildout via operating cash flow, then tight credit isn't so scary.

Here’s my calculation: The consensus assumes super-scale vendors monetize Blackwell and Rubin compute at Ampere’s pace (two generations behind). With $1.3 to $1.4 trillion in super-scale operating cash flow, if they monetize at a rate lower than current Blackwell but higher than Ampere, that closer to $2 trillion. That slashes roughly $700 billion in credit demand. Moreover, as installed compute reprices and operating cash flow continues to accelerate, credit metrics will improve and financing will become easier.

Credit market signals are indeed hard to ignore. Meta priced bonds worse last week than expected. CDS spreads are widening across the board. Real yields are rising. These are facts. If we need debt to fund this buildout, it’s a major bearish signal. But if compute reprices at current spot levels, we might not need much debt at all.

Chapter 3: GPU Spot Prices Rise Instead of Fall

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

I spoke with a company this morning that leased several thousand Blackwell chips at around mid-$2 per GPU hour. Seven months later, for the same cluster and scale of B200s, they expect to renew at under $4. That’s a 50% to 60% increase in seven months.

Another inference cloud publicly stated on a podcast that they plan to pay 100% more for Blackwell when contracts expire. This implies all super-scale vendors are underreporting revenue.

There were several catalysts for July's selloff. First, Meta announced it would rent out compute. The market interpreted this as capacity oversupply and planned CAPEX cuts. It’s completely backwards. Meta saw that SpaceX sits on a massive pool of installed compute and is optimally marketing compute-optimized clusters at prices far above contract rates. Meta spotted an opportunity: demonstrate high IRR on a small slice of capacity, raise equity, then ramp CAPEX. Meta’s CAPEX telemetry shows absolutely no change; if anything, it’s becoming more aggressive. Shortly after, they released Muse 1.1, arguably the best model in a long while.

Then Kimi K3 launched, triggering open-source panic on the market. Meanwhile, Silicon Data’s token index flattened. But that’s because the share of open-source tokens is rising, and the way these tokens are weighted in the index creates a structural flatline. A token is a token. Regardless of whether it comes from a frontier or open-source model, producing it requires the same compute, memory, and power. Open-source takes frontier model profit margins, not compute demand.

Chapter 4: Claude Is the Stock Market’s Walter Cronkite

Patrick O’Shaughnessy: How do you think the market digested this news?

One thing I’ve been thinking about. Mike Mauboussin has a theory that the collapse of diversity leads to bubbles and crashes. Right now, in public market investing, whether retail or institutional, every piece of news is fed into Claude, sometimes via Claude Code or Claude Agent. Claude is probabilistic, but people’s interpretations of news are surprisingly homogenous.

Claude is essentially the Walter Cronkite of the stock market. Everyone trusts what it says, then trades accordingly. Claude is brilliant, but it’s not always right. The stock market is fundamentally a probabilistic, Bayesian interpretation of the future. You see a headline, it gets fed into Claude, Claude interprets it in a certain way, and then a massive crowd trades on that basis.

There’s an anonymous semiconductor circle handle called TBU who posted a chart of Japanese capacitor stocks claiming, “We just ran an entire capacitor cycle in six weeks.” Stocks doubled, tripled, quadrupled, then crashed. Fundamentals hadn't even materialized, and you've already compressed a normal three-year cycle into six weeks.

Chapter 5: What Would Break the Bull Case?

Patrick O’Shaughnessy: If forced to identify a scenario that would genuinely frighten you, what is it?

Operating cash flow ceasing to accelerate—that would be the core bearish signal. It largely hinges on how Anthropic, OpenAI, Grok, Cursor, xAI, and the broader open-source ecosystem perform.

A sustained, sharp drop in GPU spot prices would also be terrifying. But have you heard anyone say they have too many GPUs? Not a soul. In fact, it sounds more like a black market.

Technically speaking, I find the most interesting potential risk lies in continuous learning and sample-efficient learning. If those breakthroughs are realized, it could temporarily fracture training demand. Moving from training models on 300 trillion tokens to training on just 10 trillion tokens and then deploying them to learn efficiently from real-world samples would be tough on training demand. But training’s share of semiconductor demand is already approaching a small but non-zero number. Inference is the bulk. SSI says they’re releasing models in August, and a wave of new labs are focusing on this direction. It’s great for the world, but it’s hard to say whether it’s net positive or negative for infrastructure demand.

Chapter 6: Game Theory in the Memory Supply Chain

Patrick O’Shaughnessy: Everyone is talking about LTAs. Can you elaborate?

We need to shift from “chasing short-term numbers” to “trading durability for Long-Term Agreements (LTAs).” Customers prepay a sum with a price floor and ceiling. It’s akin to the “labor hoarding” narrative from a few years ago when companies hesitated to lay off workers.

Think about the game theory of breaking an LTA. Four companies truly matter at scale: Amazon’s Trainium, Google’s TPU, AMD, and NVIDIA, which dwarfs all three combined. Suppose in 2027 or 2028, you want to break an LTA to get a better price. But over the following years, market share is dictated by supply allocation and your pre-purchase volume. If you break the LTA and leverage shifts back to the memory vendor, you’re shut out.

What happens if Google breaks an LTA? It likely triggers oversupply, declining prices, and natural capacity contraction. Then this cyclical industry swings from oversupply to shortage. At that point, how do you think memory vendors will allocate inventory to Google? You could torpedo your entire business and brand.

This wasn’t the case before. Apple was the largest buyer, calling the shots because no one could match their volume. But this time is different. You have at least four competing buyers, plus a swarm of startups. Break an LTA, and the memory vendor can say, “Fine, you broke the price pact, so we’ll break the volume pact and give your volume to your competitors.”

Chapter 7: NVIDIA’s New Playbook

Patrick O’Shaughnessy: How do you view NVIDIA’s current strategy?

NVIDIA has rolled out a brilliantly clever new business model, which I’d describe as a “credit wrapper” plus an above-floor revenue share. This could allow them to rapidly build a massive cloud business through royalties. It’s another way to alleviate cash flow mismatch.

This isn’t traditional supplier financing. They aren’t lending money to GPU buyers. Others are lending to buyers, and NVIDIA is just taking equity stakes and providing credit enhancement. NVIDIA explicitly writes into all equity deals, “This capital cannot be used to buy NVIDIA chips,” but money is fungible, after all.

If I were CEO of SK Hynix, I’d do exactly what NVIDIA is doing. Go to GPU buyers, say, “I’m participating in NVIDIA’s credit wrapper, I have some upfront capital ready, and I want a recurring revenue share.” It’s the logical extension of an LTA: trading short-term upside for durability, now with a royalty stream on top.

NVIDIA’s competitive moat is wider than ever. If you need financing for chips, nothing is easier to finance than NVIDIA GPUs. Need land and power? They excel at chess-like positioning. Their revenue per gigawatt is rising, and their competitive standing is strengthening.

Chapter 8: Regulation Is the Biggest Risk; AI’s Greatest Enemy Is Itself

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

Regulation is undoubtedly the biggest risk, which is painfully obvious. It’s partly why I stepped out this week wanting to be “spooked.” I don’t want to be the guy watching stocks get cheaper, expected returns rise, but still feeling like a fool. July’s fundamentals actually improved significantly compared to June’s, yet I still believe regulation remains the most substantial, non-negligible risk.

New York has enacted a data center moratorium. We live in a “post-fact, post-logic” political world. The AI sector has performed terribly at public relations. The narrative in Washington and among everyday Americans is: data centers will spike your electricity bills, drain your water, and steal your jobs.

But reality is completely the opposite. Current data center developers aren’t just buying new police cars and fire trucks for communities. They’re building hospitals, schools, new precincts, and fire stations, while lowering residents' electricity rates. Thanks to behind-the-meter agreements, data center sit-ins typically lead to lower local electricity prices. And these jobs are recurring, not one-off. You need plumbers, electricians, and HVAC technicians to maintain and upgrade these facilities. Data centers are the best thing for blue-collar wages I’ve ever seen in my life.

The water consumption issue is even more absurd. An academic book overestimated data center water usage by ten thousand times—not one or two orders of magnitude, but four. The author has conceded the mistake multiple times, getting it thoroughly debunked. But it’s identical to the spinach iron content saga: an academic book misplaced a decimal by two places, everyone believed spinach was the highest-iron food, and 80 years later, people still do. By the time the truth gets its pants on, the lie has already traveled halfway around the world.

At this year’s ASCO (American Society of Clinical Oncology) annual meeting, the prevailing mood was “the greatest number of scientific breakthroughs in a single conference ever,” a large portion of which ties directly to AI. If you have a sick child, parent, or loved one, AI is materially increasing their chances of recovery. These stories need to be told. The industry needs to run ads during the NFL season, college football, and the World Series to explain what data centers actually do.

People in Silicon Valley assume it’s all obvious, so they default to assuming others know it too. They fail to grasp that for most Americans, it’s a completely different, even opposing, reality. Even staunchly Republican states pro-growth are saying: “You haven’t told your story well. If you did, we’d happily amplify it, but you’re the experts.”

Chapter 9: China’s DUV, Open-Source Ecosystem, and SpaceX Orbital Compute

Patrick O’Shaughnessy: How do you view China possessing DUV machines?

Both readings could be valid. Picture it this way: DUV is a propeller plane, EUV is a jet turbine. China didn’t have it before, and reportedly now does. It’s a phase change. Even though that engine is 25 years behind, it’s significant and shouldn't be ignored. But the market may be overreacting. If this actually impacts ASML orders, it’s likely five years down the line, with the market forgetting and remembering repeatedly in between.

For China, it is indeed critically important. They’re incredibly smart and hardworking, treating it as a national-level mission. But you can’t speedrun the “learning curve.” You have to actually walk through those learning cycles.

Patrick O’Shaughnessy: Impact of open-source models on the industry?

Open-source models closing the gap to frontier, combined with inference clouds like Fireworks making custom models easy to deploy, is a godsend for software. AI-native companies can use open-source models for their own RL fine-tuning, shifting from relying on 1-3 frontier models to frontier models accounting for only 30%-60% of token consumption, with the rest run on their own models. You’re no longer a “ChatGPT wrapper”; you possess proprietary data and genuine defensibility.

Cheaper tokens could actually make the most frontier tokens more valuable. If you have a 120 IQ open-source model running cheaply, doesn’t a 160 IQ frontier model capable of commanding it become even more valuable?

Another overlooked element is SRAM accelerators. These chips aren't constrained by HBM DRAM capacity bottlenecks, usually manufactured on older nodes, and don’t compete with the latest GPUs for capacity. If you break down inference, it has two phases: prefill and decode. Decode splits further into attention and feed-forward networks. The ideal setup: run prefill on chips without HBM, run attention on high-performance HBM-equipped chips, and run feed-forward on SRAM chips. Regardless of how you tune compute, HBM DRAM, and SRAM ratios, workloads constantly shift. Splitting it into these three components drastically boosts AI’s overall ROI.

Patrick O’Shaughnessy: Any dark horses not yet on everyone’s radar that could become Game of Thrones-level players?

Lee Buu could be one. Lynn from Fireworks is an absolute killer. Our friend Scott Wu at Cognition is also highly noteworthy.

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

SpaceX’s fundamentals have improved since its IPO. Grok 4.5, the Cursor acquisition—Cursor is clearly accelerating. Over the past three years, they’ve onboarded more compute at lower costs faster than anyone else. They entered precisely at peak spot prices, but the market has fully absorbed their massive compute deployment, and the freight train hasn't slowed down at all.

A Substack writer quoted public reports claiming SpaceX aims for 8 gigawatts of compute. I never bet against Elon, but it’s genuinely a staggering feat. They’re currently monetizing at roughly $50 billion per gigawatt, with consensus estimating $73 billion for next year. Even ignoring Starlink V3, direct-to-cell, Grok 4.5, and Cursor, if they actually onboard anywhere near 8 gigawatts, it vastly exceeds market expectations.

The market isn’t reading it that way. There’s a NYC hedge fund shorting it, believing compute spot prices will drop 90%, and SpaceX’s massive compute deployment won’t generate that much revenue. Maybe. But Elon’s company has consistently done incredible things for years. As he puts it, “We excel at making the impossible arrive late.”

Orbital compute is becoming reality daily. Benchmark invested in StarCloud, an orbital compute company lacking SpaceX’s internal launch cost edge. If Benchmark thought it was unviable, they wouldn’t have invested. So either I’m crazy, Elon is crazy, Benchmark is crazy, or SpaceX’s engineers are crazy. That probability is low.