Gavin Baker on why every AI fundamental is accelerating while the stocks collapse
Latent Space · Gavin Baker (Atreides) 1h 18m Recorded: July 2026 selloffDays after the Dalio "AI bubble" bear case, Gavin Baker (a top AI-infrastructure investor) makes the opposite argument: the market has become the bubble — AI stocks are down 40–60% in a month while every underlying fundamental accelerates. His Silicon Valley field research over the summer found zero negative quantitative metrics about AI. Nvidia trades at its lowest forward P/E in a decade. The only real risks he can find: the credit market and regulation — not demand.
This is the essential bull-bear tension for anyone deciding where the AI cycle actually is. Baker's core claim: "A token is a token. All open source taking share does is take margin dollars out of the frontier model layer and drive more margin into the AI infrastructure layer." Demand is the unbroken line.
The context matters — this was recorded during a brutal selloff, not a frothy peak.
Baker frames his whole trip to Silicon Valley as a pressure test: "I want to be scared. I don't want to feel like a lunatic watching these stocks get cheaper thinking the expected forward returns are going up." His conclusion after meeting companies across the ecosystem: the fundamentals got better in July vs June — while the stocks crashed.
"However you cut it — GPU availability, GPU rental pricing, spot price of DRAM, token growth — everything is actually accelerating." The only contested data point: third-party Anthropic token data, which Baker argues is offset by open source (GLM 5.2, Kimi K3) and OpenAI accelerating the comp net.
The market panicked when open-source models (GLM 5.2, Kimi K3) appeared to steal share. Baker's rebuttal: all tokens cost the same compute (same flops, watts, memory). Open source just shifts margin from 90%-gross-margin frontier tokens to ~30% open-source tokens — driving MORE demand for compute, not less. It takes margin from frontier models and funnels it into the infrastructure layer. Jensen Huang being the world's biggest open-source supporter is a tell: "Would he support it if it was bad for his business?"
The contracted base of installed compute trades at a massive discount to current spot prices. As those contracts roll off, compute reprices higher. "The market is modeling a deceleration which I think is unlikely." GPU rental prices rising 50–60% — the opposite of the gentle decline everyone expected — is the cleanest proof of a shortage.
Microsoft, Meta, Amazon operating cash flow accelerated from ~28 to ~32 (or 35 excluding unusual items). "That's a material acceleration at this scale" — and it's before Rubin chips come online at a premium and before contracts reprice.
"Maybe 500,000 people in the world are using agentic AI. There are 7–8 billion people. What happens when we go from 500,000 to 100 million to 500 million?" Plus differential waves: AI natives leaning in (not hiring humans), east-coast companies barely adopted, Europe still regulating. The runway is enormous.
Baker is not blindly bullish — he names the two things that could actually derail the thesis. Notably, neither is "demand."
"The degree to which this buildout is going to require credit... that would be the classic capital cycle. We start to overextend ourselves with debt and that's where things get." Real yields are up, CDS spreads are blowing out, Meta's bond priced below expectations. BUT: Baker models that if the installed compute reprices even at a discount to current Blackwell, hyperscaler operating cash flow hits ~$2 trillion — removing ~$700B of credit demand. The differential between spot and contract pricing is why he thinks credit won't matter.
"Regulation has to be the biggest risk." New York made a data-center moratorium. The political narrative is that data centers raise power prices, take water, take jobs — all of which Baker argues is false (behind-the-meter deals actually lower local power prices and create persistent blue-collar jobs). But the industry has done "a terrible job of PR" and the narrative is spreading.
If solved (models trained on 10T tokens instead of 300T, learning sample-efficiently in the world), it could temporarily disrupt training demand. Baker is open-minded this is the biggest technical unknown — but notes it "would be awesome for the world" and he doubts it's negative for inference/infrastructure demand long-term.
A sustained contraction in GPU prices, or "getting really easy to get GPUs," would be the tell. So far: "I have not met a single person who says they have too many GPUs." If the frontier labs (Anthropic/OpenAI/Grok) plateau or decline without open source growing the pie — that's genuinely negative.
Lay this against the Ray Dalio bubble video. They're looking at the same AI mania and reaching opposite conclusions.
How to reconcile: Dalio is talking about price/valuation mechanics and the credit cycle (a market-level bubble that pops regardless of fundamentals). Baker is talking about real economic fundamentals and demand (a genuine buildout whose value isn't fully priced). Both can be true simultaneously: a great technology can still have a frothy, mean-reverting stock. The difference is degree and timing — and both admit uncertainty.
| Topic | Baker's View | Signal |
|---|---|---|
| Nvidia valuation | Lowest forward P/E in 10 yrs; market thinks it's "significantly over-earning" | Bullish |
| GPU rental prices | Up 50–60% in 6–7 months, against all expectations | Bullish |
| Open source (GLM, Kimi) | Not a negative — shifts margin to infra, drives compute demand | Bullish |
| Hyperscaler OCF | Accelerating ~28→32–35, before Rubin premium & contract repricing | Bullish |
| Credit / CDS / real yields | Real risk, but OCF reprice removes ~$700B of credit need | Watch |
| Regulation (NY moratorium) | Biggest risk; industry losing the PR war | Bearish |
| Continual learning | Wildcard — could disrupt training demand if solved | Unknown |
| LTAs (memory/GPU) | Game theory: breaking one = losing future allocations; nobody will | Bullish |
| xAI / SpaceX | "Data center company"; compute absorbed by market without a blip | Bullish |
| Token spend vs labor | 20–30% of comp in frontier AI natives; $25T knowledge-work pie | Bullish |
A striking observation on market behavior: almost every investor now feeds news into Claude/Claude Code, and there's little variation in how the model interprets it. So the market has become monoculture — "everyone trades on Claude's view." Baker cites a capacitor cycle that ran an entire 3-year cycle in 6 weeks, and Japan's capacitor stocks going vertical then crashing — "the actual fundamentals haven't even hit." This concentration of interpretation is a real fragility — and a contrarian signal. When everyone reasons through the same model, the market can move on narratives that are "factually, except for credit, not true."
Baker's thesis is the strongest fundamentals-first rebuttal to the AI-bubble bear case: every demand signal is accelerating, the compute shortage is real and worsening, and the buildout is increasingly self-funding via operating cash flow. The risks are credit and regulation, not collapsing demand. His humility is the most credible part: "I feel like a foolish optimist in the market, but when I talk to people in this ecosystem, I'm bearish relative to essentially everyone." Whether he's right on timing is unknowable — but the direction of AI infrastructure demand is his strongest card.