Everyone Went Looking for the AI Bear Case. Nobody Found One.

Gavin Baker runs Atreides Management, has been in Nvidia for the better part of 25 years, and just went on Invest Like the Best with Patrick O’Shaughnessy to explain why the 40 to 60 percent AI drawdown that just ran through July doesn’t line up with anything actually happening on the ground. His framing at the top of the conversation is the whole piece in one line: “I want to be scared, you know. I don’t want to feel like a lunatic watching these stocks get more cheaper thinking the expected forward returns are going up.” He went looking for a reason to be scared. He came back empty.

That’s a methodology, and it’s the right one. Baker’s actual question, put to O’Shaughnessy and to everyone he talked to for a month straight, was blunt: “Have you heard a single negative quantitative metric about AI? A single instance of deceleration?” The answer, from him and from every operator he spoke with, was nothing, no soft number, no slowdown. “In fact, every metric is accelerating.” When the guy who’s been holding Nvidia since before most of Wall Street took AI seriously spends a month actively hunting for bad news and comes back with none, that’s worth more than a hundred hot-take threads about a bubble.

The first thing worth sitting with is that GPU prices are doing the opposite of what everyone modeled. Nobody in 2024 or 2025, bull or bear, thought the price to rent a GPU would still be going vertical in 2026. The consensus bull case was a slow, gentle decline as supply caught up. Instead: “we’re up, you know, depending on the starting point, 50 to 60 percent in six or seven months.” Baker isn’t citing an index for this. He’s citing a company he talked to that morning, renting a cluster of Blackwells for roughly $2 per GPU hour and bracing to pay just under $4 when the contract rolls off in seven months. Same cluster. Same chips. No differences. That’s a market where demand for compute is outrunning supply badly enough that a doubling in price over half a year barely makes anyone blink.

The second thing is the one Wall Street keeps getting backwards: hyperscaler operating cash flow is accelerating, not stalling. Microsoft, Meta, and Amazon’s reported operating cash flow growth went from 28 percent to 32 percent this quarter, and once you strip out an unusually large batch of one-time legal charges (mostly EU fines), the real number moved from 28 to 35. That is, in Baker’s words, “a material acceleration at this scale,” and it happened before any of these companies have even lit up the next generation of chips, which arrive at a meaningful pricing premium once current contracts reprice. The market spent July treating Meta renting out excess compute as a sign of a capex retreat. It was the opposite: a company discovering it could sell trading-optimized clusters into the spot market at a massive premium to its own contracted rates, then presumably use that as proof of return before raising more capital. Nothing in the actual capex telemetry moved. If anything it got more aggressive.

The third thing is the open source panic, and this is the argument I think more people need to actually sit with, because it’s counterintuitive and it’s correct. When GLM 5.2 and Kimi K3 landed, the market read it as bearish for the AI infrastructure trade, cheaper open models eating into frontier-model demand. Baker’s answer to that is the cleanest piece of reasoning in the whole conversation: “the reality is a token is a token, and you need the exact same amount of compute to make a token… It takes the same amount of flops, the same amount of memory, the same amount of watts.” Open source models taking share doesn’t shrink compute demand. It shrinks the margin that used to sit inside the frontier labs and pushes those margin dollars down into the infrastructure layer instead, the same GPUs, the same power, the same clouds, just fewer dollars captured by OpenAI and Anthropic and more captured by everyone building underneath them. That’s a margin story moving, not a demand story turning negative, and it’s a far less scary thing to be caught inside of.

The fourth thing, and the one honest scare in the whole conversation worth taking seriously, is credit. Real yields went up. Spreads widened. A Meta bond priced worse than you’d expect a Meta bond to price. If this buildout genuinely needed debt to get built, that would be a real problem, the kind that ends capital cycles badly, and Baker doesn’t wave it away. But his answer is the most interesting math in the piece: model out the gigawatts hyperscalers are guiding to, assume they monetize at a discount to current Blackwell pricing rather than at the aggressive case, and you get roughly two trillion dollars of operating cash flow available to fund the buildout, enough to take something like 700 billion dollars of assumed credit demand off the table entirely. As the installed base of compute reprices from old contracted rates up toward the current spot market, the whole picture gets funded out of cash flow the same companies are already generating, not out of debt markets that might not be there when it counts.

And then there’s the game theory nobody’s pricing in at all: long-term supply agreements. Baker’s read: there are effectively four players who matter at scale in AI silicon, and breaking a long-term supply agreement to chase a cheaper spot price isn’t a pricing decision anymore, it’s existential. Break the agreement, and the moment leverage shifts back toward the chip and memory suppliers in the next cycle, you’re the company that gets deprioritized on allocation. Nobody’s going to blow up years of relationship for a short-term discount when market share for the next decade is effectively being determined right now by who locked in supply and who didn’t. That’s the behavior of an industry that thinks the shortage is structural, not one bracing for a slowdown.

We’d add one more data point, and it’s timely: we wrote about Palantir’s Q2 print less than 24 hours after Baker said all this. Revenue up 93 percent year over year, U.S. commercial revenue up 149 percent for the fourth straight quarter, the stock ripping nearly 30 percent in a single session, and an expanded Nvidia partnership sitting quietly underneath the number as one of the reasons U.S. commercial growth accelerated instead of slowing. That’s Baker’s thesis playing out in real time, from a company that reported the day after this conversation happened. Alex Karp’s own line on it, “for the first time people believe us,” is the corporate version of what Baker is saying about the whole sector: the fundamentals stopped being a story you had to take on faith and started being a number you could just read off the page.

Baker’s honest about where he thinks the real risk actually lives: “I think regulatory has to be the biggest risk.” Not a GPU glut, a credit crunch, or open source eating the frontier labs’ lunch, but bad policy built on bad information, like the water-usage myth he mentions that was off by four orders of magnitude and still gets repeated. That’s a real risk and we’re not going to pretend otherwise. But it’s a different category of risk than “the fundamentals are rolling over,” and conflating the two is exactly how you end up down 40 to 60 percent in a straight line on a month where, as Baker put it, “I haven’t been able to find one that is like a quantitative metric” pointing the other way.

We agree with him, and we’d go further. A market that sells off 40 to 60 percent in a straight line on a month with zero negative fundamental data is pricing in a narrative that outran the numbers, not a slowdown.