Key Takeaways:
July's AI selloff wiped 40-60% off many AI names in a straight line, yet every quantitative metric Gavin Baker tracks is accelerating — GPU rental prices up 50-60%, hyperscaler operating cash flow growth climbing from 28% to 32%, and Nvidia sitting at its lowest forward price-to-earnings ratio in a decade.
The July 2026 AI complex fell "2022 in a month," Gavin Baker, who runs Atreides Management and whose portfolio spans hyperscalers, GPU clouds and AI-native startups, said on the "Invest Like the Best" podcast, recorded in the first week of August at Benchmark's Menlo Park offices. His stated mission for the week was to find a negative AI data point. "I want to be scared," he said. "I don't want to feel like a lunatic watching these stocks get more cheaper thinking the expected forward returns are going up." He could not find one.
The selloff was narratively liquid but factually hollow, Baker argued. Meta's decision to rent out compute was read as excess capacity and an imminent capex cut, yet its capex telemetry never shifted and it shipped Muse 1.1 days later. Open-source models Kimi K3 and GLM-5.2 reshuffled the measured token mix, but "a token is a token" — the same flops, memory and watts — so open source merely moves margin from 80-95% gross-margin frontier tokens to roughly 30% tokens, with the dollars landing in the infrastructure layer. A reported Chinese DUV breakthrough is real but generations behind. The one legitimate fear — that the buildout needs debt as credit tightens — is the load-bearing argument of the episode.
The repricing math that defuses the credit scare
The master key, Baker said, is that the entire contracted base of installed compute trades at a massive discount to current spot prices, and as those contracts roll off, compute reprices higher even if spot declines. One of Silicon Valley's most sought-after startups rented a cluster of several thousand B200s at "somewhere in the mid-$2 per GPU hour" seven months ago; the identical cluster now rents for "just under $4" — up 50-60%. An inference cloud said on a podcast it plans to pay 100% more for Blackwells when its contract expires.
The flow-through is already visible. Microsoft, Meta and Amazon reported combined operating cash flow (not free cash flow) accelerating from 28% to 32% in Q2 2026 — and to 35% stripping billions in EU fines and legal expenses. Baker's consensus critique: the gigawatts expected online are modeled to monetize at Ampere/A100-era rates, two generations behind. At Ampere rates, hyperscaler operating cash flow runs $1.3-1.4 trillion; at a discount to current Blackwell rates, closer to $2 trillion. The delta removes roughly $700 billion of credit demand, loosening credit markets just as the buildout needs them. "If credit's not there, it just means the flops that are there are going to be even more valuable," he said.
Open source, the model war, and where margin migrates
July was a "transformational month" for models, topped by xAI's Grok 4.5, Grok Build and the Cursor acquisition, with Meta's Muse 1.1, Zhipu's GLM-5.2 and Moonshot's Kimi K3 landing in the same window. The public-market picture omits the private layers: Anthropic, OpenAI and American open-source inference clouds (Fireworks, Baseten, Modal, Together) are growing almost as fast as the frontier labs did early on while burning little cash. Baker's tell on open source is Jensen Huang — "the world's largest supporter of open source" — who would not make it his signature issue if it hurt Nvidia's business.
The lab game theory is the sharpest form of the argument. "Anthropic, if they had been as aggressive on compute as OpenAI had been, they would have run away with it," Baker said, citing Dario's dilemma that buying too much compute risks bankruptcy while buying too little risks losing. OpenAI is "back in the game," Grok is in the game, and after Grok 4.5 plus Cursor, "is anybody going to back off anytime soon, especially if it could be funded out of operating cash flow?"
The demand curve and the sample-efficiency caveat
Baker's rough count: 250,000-500,000 people worldwide use agentic AI today, against 7-8 billion people, at a moment of acute compute shortage. Token spend already runs 20-25% of total compensation at leading AI-native companies in his portfolio, up to 50% at the highest ratio he has heard, against his frame of $5 trillion in knowledge-work spend at stake (20% of $25 trillion). The one caveat that could dent training demand: multiple labs feel "very close to solving continual learning and sample-efficient learning," which would shrink training's share of compute demand to "not approaching zero but very small." SSI's model arrives in August 2026. Baker called it "amazing for the world" but the most interesting open question.
Nvidia's credit wrapper and the memory game
Baker believes Nvidia's new business model is deeply misunderstood: a "credit wrapper with a revenue share if GPU prices are above a floor," where someone else lends the GPU buyer the money and Nvidia takes a cut of ongoing revenue — potentially making it "a really giant cloud business" through royalties. He also flagged the memory long-term supply agreement (LTA) game theory: with only four players at scale in accelerators — Amazon's Trainium, Google's TPU, AMD and Nvidia, "much bigger than everybody else combined" — breaking an LTA to grab lower prices risks losing supply allocations for years. "If you break your LTA and then in the next two or three years for any reason leverage shifts back to the memory guys, you're out of business," he said.
Regulation is the real tail risk
"Regulatory has to be the biggest risk," Baker said, citing New York's data center moratorium in a "post-factual, post-logical political world." He argued the industry has done "a terrible job of PR," with a book author's error overestimating data center water usage by 10,000x — four orders of magnitude — still propagating despite repeated debunking. The untold counter-narrative: behind-the-meter deals mean electricity prices generally fall for surrounding communities, the data center pledge now delivers hospitals and schools, and the jobs are lasting blue-collar work. Baker called data centers "the best thing to happen to blue-collar wages in my lifetime." His prescription: an industry foundation running ads during the Final Four, NFL games and the World Series.
SpaceX and the honest stress test
SpaceX — the most important newly-public company — does not appear to be understood as one, Baker said. A Substack report claimed it is targeting roughly 8 gigawatts of new compute; Baker said he "will never bet against Elon" but finds that scale near-implausible. Still, SpaceX monetizes at roughly $50 billion per gigawatt, consensus revenue for 2027 is $73 billion, and Grok 4.5 plus Cursor alone likely hits $10 billion in annual recurring revenue quickly. He flagged Benchmark's investment in StarCloud, an orbital compute company partnering with SpaceX, as a "rational person check." "Maybe I'm crazy and maybe Elon's crazy and maybe Benchmark's crazy and maybe the SpaceX engineers are all crazy," he said. "That just doesn't seem that probable."
Baker's honest caveat is his own posture: "Essentially everyone out here is more bullish than me, man." He searched for a reason to be scared and could not find one that survived contact with a quantitative metric, while the stocks are cheaper than they have been in a decade. For investors, the tension is between narrative velocity — Claude-as-Walter-Cronkite herding, capacitor cycles compressed into six weeks — and fundamental velocity: spot repricing, operating cash flow acceleration, and the still-unbuilt demand from 500,000 agentic users. Nvidia, trading at its lowest forward P/E in a decade, a multiple previously touched only on "Liberation Day" and after DeepSeek — "both kind of V-bottoms" — is the cleanest expression of the bet.
This article is for informational purposes only and does not constitute investment advice.
