Summary

July was “2022 in a month” — AI names down 40-60% in a straight line — yet Baker’s week of pressure-testing Silicon Valley for “one negative quantitative metric” turned up no clear one: GPU availability, GPU rental pricing, spot DRAM, and token growth are all accelerating. Nvidia now trades at its lowest forward P/E in ten years; “the market 100% thinks they are significantly over-earning… maybe they are.”
The core long thesis is compute repricing: everyone in ‘24-‘25 modeled GPU prices declining, but old-GPU prices are “going vertical in 2026” — one startup rented an identical B200 cluster at mid-$2/GPU-hour and, seven months later, hoped to pay just under $4, while an inference cloud plans to pay 100% more at renewal. As below-market contracts roll off, the hyperscalers’ installed base reprices higher: “essentially all the hyperscalers are under-earning.”
Credit is the only bearish catalyst he calls real — hyperscaler CDS blowing out, a Meta bond pricing poorly, real yields up — but consensus monetizes incoming Blackwell/Reuben gigawatts at Aier-generation rates ($1.3-1.4T of hyperscaler operating cash flow); monetize at merely a discount to current Blackwell and it’s ~$2T, “taking 700 billion of credit demand out.” Failsafe: “if credit’s not there, it just means the flops that are there are going to be even more valuable.”
The open-source freakout was backwards: “a token is a token” — GLM 5.2 and Kimmy K3 shift tokens from ~90%-gross-margin frontier to ~30%-margin open source on the same flops, memory, and watts, moving margin dollars into the infrastructure layer. The tell: Jensen wouldn’t be “the world’s biggest supporter of open source if it was bad for his business.”
Memory LTAs are now franchise-defining: with four buyers at scale and market share set by supply allocations, “you might blow up your entire business and your franchise by breaking an LTA.” Nvidia’s answer — a “credit wrapper with a revenue share if GPU prices are above a floor” plus equity stakes everywhere — is misunderstood, and exactly what he’d copy if he ran likely Hynix.
Market microstructure has changed: everyone feeds every headline into Claude, and “Claude is kind of Walter Cronkite for the stock market” — smart but not always right, compressing a three-year Japanese capacitor cycle into six weeks. His anchor: “the three most important words in investing aren’t margin of safety, but I don’t know.”
Biggest risk is regulatory, not fundamental: New York’s data-center moratorium looks like the first of many while the industry loses a PR war partly built on an admitted 10,000x water-usage error — despite data centers being “the best thing to happen for blue-collar wages in my lifetime.”
SpaceX is the underpriced compute machine: monetizing ~$50B per gigawatt against $73B consensus next-year revenue, with a public report claiming 8GW in 18 months he “almost doesn’t believe,” fundamentals better since IPO (Grok 4.5, Cursor), and “orbital compute feels more real every day.”

Deep dive

1. July was “2022 in a month” — and only one contested negative datapoint emerges

Baker’s label for the month: “2022 in a month,” with AI names down “40 to 60% in a month in a straight line.” His stated mission for the week — “pressure test… tell me something negative” — found no clear negative metric: across GPU availability, GPU rental pricing, the spot price of DRAM, and token growth, “in fact, every metric is accelerating,” and from data, not vibes.
The valuation fact he keeps returning to: Nvidia at its lowest forward P/E of the last ten years, cheaper only at DeepSeek and Liberation Day — both V-bottoms. “The market 100% thinks they are significantly over-earning… and we need to be humble. Maybe they are.”
His place on the spectrum: “essentially everyone out here is more bullish than me.” An essay he read — renting an H100 for a year at ~$250,000, roughly 15x current spot — “wasn’t even in my considered but dismissed as totally unlikely outcomes.” “I look at what’s happening in the stock market and I feel like a foolish optimist” — and in Silicon Valley, he’s the bear.

2. Anatomy of the panic: every catalyst except credit was misread

Meta “renting out compute” traded as excess capacity and coming capex cuts. Baker: “this is not at all what it was” — Meta watched SpaceX sell big training-optimized clusters at “a truly massive premium” to contracted rates and saw an IRR showcase, possibly ahead of an equity raise. Capex telemetry never shifted, capex wasn’t cut, and Muse 1.1 — Meta’s best model in a long time, overshadowed by Grok 4.5 — says the foot stays on the gas.
The Silicon Data token-index dip that read as demand rolling over was mix shift: GLM 5.2 and Kimmy K3 moved tokens from frontier pricing — inference margins of “80, 90, or 95” percent, debatable — toward open source. The market took it as negative; he reads it as margin migrating, not demand shrinking.
China’s DUV machine cratered semicap baskets. His analogy: DUV is a propeller plane to EUV’s jet turbine — a real phase transition even 25 years behind, because chipmaking is learning-by-doing: “you can’t teleport into the future.” Verdict: significant, and the market still probably overreacted. (On the smuggled-EUV rumor he can’t verify: “what a feat of espionage — those things are giant.”)

3. A token is a token — open source takes margin, not compute demand

The center of his rebuttal: “a token is a token, and you need the exact same amount of compute to make a token” — same flops, same memory, same watts. Open source taking share strips margin from the frontier layer and pushes margin dollars, plus demand elasticity, into infrastructure. The tell: would Jensen be “the world’s biggest supporter of open source” if it hurt his business?
Router economics resolve the paradox of falling enterprise AI bills: companies that “20x’d” spend and burned budgets in three months install routers and cut spend — while GPU-hours consumed likely rise — because savings come from swapping ~90%-gross-margin frontier tokens for ~30%-margin open-source ones, “in a lot of cases slightly better outcomes at half the cost.”
“Open source is kind of dark matter to the public markets” — hard to measure, but the American inference clouds monetizing it (Fireworks, likely Baseten, Modal, Together) indicate accelerating demand after the GLM 5.2 / Kimmy K3 capability leap. The lone soft datapoint — third-party data on Anthropic’s curve bending — “may very well be true” but is hotly contested by Anthropic shareholders itching to say otherwise.
He entertains the “Anthropic/OpenAI/Grok maximalist view” — a frontier model hitting RSI distills itself cheaper at every intelligence level, leaving no room for open source — but doubts it: AI natives can RL models on proprietary data (Fireworks’ Nexus is “three lines of code”; Harvey before its acquisition and Cursor leaned in), stop being wrappers, and cheap “120 IQ” open models arguably make the “160 IQ” orchestrator more valuable.

4. Credit is the one real worry — and repricing compute defuses it

The catalyst he can’t wave off: real yields up, spreads wider, CDS “blowing out” for everybody, and Meta’s bond “did not price where you would think a Meta bond would price.” “That would be really, really scary if we needed debt to finance this buildout” — debt-fueled buildouts “demand immediate repayment,” which is what unwound the internet cycle.
His model: consensus effectively monetizes incoming Blackwell/Reuben gigawatts at Aier rates — two generations back — for $1.3-1.4T of hyperscaler operating cash flow. Monetize at merely a discount to current Blackwell and it’s ~$2T, which “takes 700 billion of credit demand out” and improves the very credit ratios spooking the market.
Already printing: Microsoft, Meta, and Amazon operating cash flow accelerated from 28 to 32 this quarter — 28 to 35 stripping an unusual slug of one-timers, mostly EU fines — “a material acceleration at this scale,” before Reuben premiums and contract repricing. And the failsafe: “if credit’s not there, it just means the flops that are there are going to be even more valuable.”

5. Spot vs. contract: the installed base is massively under-earning

The anecdote of the trip: a hot startup rented several thousand B200s at “mid $2 per GPU hour” and, seven months later, hoped to pay just under $4 for an essentially identical cluster — spot up 50-60% when every 2024-25 model, bull or bear, had GPU prices declining. “I don’t think anyone in 24 or 25 thought that the prices of old GPUs would be going vertical in 2026.”
An inference cloud said on a podcast it plans to pay 100% more for Blackwells when its contract expires — “that just means essentially all the hyperscalers are under-earning.” As contracts roll off, the contracted base reprices toward spot even if spot itself declines.
The demand backdrop: maybe 250,000-500,000 people on Earth use agentic AI, amid an acute compute shortage. “What happens when we go from 500,000… to 100 million, to 500 million?” And his favorite check: “Have you heard anyone say they have too many GPUs?” Not one — “sounds like a drug market.”

6. Claude is Walter Cronkite for the stock market

An investor’s (name unclear in the audio) theory that a breakdown in diversity precedes bubbles and crashes, updated: everyone in public equities now feeds every headline into Claude or a Claude agent, and “there’s probably not that much variation in the way it’s interpreting this news… Claude is kind of Walter Cronkite for the stock market” — really smart, “but it’s not always right,” in a game that is a probabilistic Bayesian interpretation of the future.
Exhibit: TBU’s chart of Japanese capacitor stocks — “we’ve had an entire capacitor cycle in 6 weeks,” vertical then whoosh, a would-be three-year cycle compressed before the fundamentals even hit.
A Fidelity friend’s dictum for the era: “just do the dumbest most superficial thing as quickly as possible and just cycle between them” — this month, cutting risk on narratives that “except for credit… are just kind of ridiculous.” Yet the tape keeps falling, and he honors the technician’s warning — “it’s definitionally the bullet you don’t see that gets you” — alongside his own anchor: “the three most important words in investing aren’t margin of safety, but I don’t know.”

7. Memory LTAs and Nvidia’s credit wrapper: trading upside for durability

The transition he admits he got wrong: memory makers swapping short-term price spikes for long-term supply agreements with prepays, floors, and ceilings. Memory is the axis it all revolves around — more memory per flop means more tokens per unit of compute, “the single most important thing you could do,” which is why demand has shown no negative elasticity.
Game theory of breaking an LTA in 2027-28: four buyers matter (Amazon’s Trainium, Google’s TPUs, AMD, and Nvidia, “much bigger than everybody else combined”), and share flows from supply allocations. Break one, and when leverage cycles back to memory, “you’re out of business… you might blow up your entire business and your franchise.” Unlike the Apple era, when the memory makers would always take the biggest buyer back.
Nvidia’s new model: “a credit wrapper with a revenue share if GPU prices are above a floor” — not vendor financing, since a third party lends — plus equity everywhere (“essentially every time they haven’t taken an equity stake in something, it’s been a mistake”). It could build a giant royalty-driven cloud business, lifts revenue per gigawatt, and hardens the moat: rival-chip buyers pay more at TSMC and for HBM DRAM, and “nothing’s more financeable than an Nvidia GPU. Nothing.” Which makes the decade-low multiple “a little hard for me to understand.”
Asked what he’d do running likely Hynix: “the exact same thing Nvidia is doing right now” — put up cash, take a cut of ongoing revenues, the logical extension of the LTA trade. “I’m sure our friends at Blackstone and Apollo are suggesting some variant of this to the memory companies.”

8. Nobody lets off the gas — and the technical wildcards

The lesson every lab absorbed: Anthropic, had it matched OpenAI’s compute aggression, “would have run away with it.” OpenAI is back in the game; Grok is on the Pareto frontier; SpaceX is in via Grok 4.5 and Cursor. Dario’s own dilemma — buy too much compute and go bankrupt, too little and lose — prompts the question: “is anybody going to back off anytime soon, especially if it could be funded out of operating cash flow?”
His biggest technical takeaway: many people feel close to solving continual and sample-efficient learning (SSI says its model comes in August). If you could train something on 10 trillion tokens and let it learn in the world instead of using 300 trillion tokens, training’s share of compute “asymptotes to something not approaching zero but very small” — “awesome for the world,” hard for him to believe it’s negative for infrastructure demand, “but again, trying to be really open-minded.”
Hardware wildcard: disaggregating inference — prefill on chips without HBM DRAM, attention on high-powered HBM chips, and the feed-forward network on SRAM accelerators built on older nodes: “you just can’t beat SRAM for that feed-forward network.” He calls it “really, really positive for the ROI on AI.”
Dark horses at Game of Thrones scale: Lynn at Fireworks (“an absolute killer”), Cognition’s Scott Wu, and one name the audio garbles. The inference clouds themselves are the sleeper: growing almost as fast as the frontier labs did early “but burning very little cash… crazy numbers” on a rule-of-40 basis.

9. Regulation is the biggest risk, and the industry is losing the narrative

“Regulation just has to be the biggest risk… you just can’t ignore New York making a data center moratorium” — which feels like “the first of many,” while “the AI industry has done a terrible job of PR.” The Washington narrative: data centers raise your power bill, take your water, take your job.
His counter: behind-the-meter deals generally lower local electricity prices; pledge-era developers now build hospitals, schools, police and fire stations; and the jobs persist through ongoing maintenance and upgrades. “Data centers are in a lot of ways the best thing to happen for blue-collar wages in my lifetime” — with Democrats, ostensibly blue-collar’s party, opposing them.
The water panic traces to an author’s admitted 10,000x overestimate that still circulates — the Popeye-spinach decimal error, 80 years on. His half-serious fix: a foundation or PAC airing “here’s what a data center does” ads during the Final Four and NFL games, plus telling the AI-cures-diseases story (at ASCO this year, the most scientific breakthroughs seen at a single conference, partly due to AI).

10. SpaceX: the market isn’t pricing the compute machine

Patrick asked whether SpaceX is “the most important new company to be public”; Baker said the fundamentals have improved since the IPO: Grok 4.5, the Cursor acquisition (Cursor “has clearly accelerated meaningfully”), and a three-year record of bringing on compute faster and cheaper than anyone. They dumped vast compute into the market overnight at spot highs “and the freight train didn’t slow down at all” — itself one of the more bullish demand datapoints.
A public report claims SpaceX will bring on 8 gigawatts in 18 months — he “almost doesn’t believe” it, but they monetize around $50B per gigawatt against consensus next-year revenue of $73B, so anywhere near it swamps estimates. Only the hyperscalers, CoreWeave, Crusoe, and SpaceX have brought on more than 500MW in a year. Elon’s phrase: “we specialize in making the impossible late.” The big New York hedge-fund short case needs spot compute down ~90%.
After time at Starbase: “orbital compute feels more real every day.” His sanity check is Benchmark — no Elon-ecosystem ties — funding StarCloud, which SpaceX is kind of partnering with and may let use Starlink laser tech: “maybe I’m crazy and maybe Elon’s crazy and maybe Benchmark is also crazy… that just doesn’t seem that probable.”

Verification Notes

The raw captions do not resolve the name in the diversity theory; the digest leaves that name unnamed.

How the White House Plans to 10x Scientific Productivity | Michael Kratsios | EP #276 Huawei 'Genius Youngster' Enters the Embodied AI Marathon