Summary

Gavin Baker has spent the summer asking every AI leader “Can you tell me one quantitative data point in your business that’s getting worse? Just one” — and through July and August, nobody could. AI broadly accelerated both months even as some AI names fell into significant drawdowns; the index calm is misleading because “you can drown crossing a river that’s on average two feet deep.” His caveat: Anthropic is in an IPO quiet period.
Both men reject zero-sum framing: this cycle is “not an or thing, it’s an and thing,” where frontier labs, open source, neoclouds, applications, and Nvidia all win. Every LP conversation starts with “How is this all gonna go wrong?”, but the supply-side data shows a roughly nine-to-ten-month payback for Nebius ($50B/gigawatt, 50-60% prepaid by customers), Blackstone/KKR/Apollo financing at low cost, and useful lives extending — true equity payback “might be way inside of a year.”
The demand side is “absolutely nowhere”: the companies’ roughly $180B of revenue rests on maybe sub-10 million heavy users against 1.5 billion knowledge workers. Baker’s fund Atreides grew internal token consumption 100x from March to August; some AI-native companies already spend 10%+ of human compensation on tokens versus ~1% at old-economy firms. Both worry more about undersupply than overbuild through ’28 — Dwarkesh’s scenario of token prices rising 10x is “the opposite direction of where everybody thinks this is gonna go.”
Public markets will have to digest lab revenue as a dial, not a stream: a lab monetizing 8 gigawatts of inference at ~$60B per gigawatt per year could cut revenue from $480B to $120B in this example by reallocating to training — “and I actually think they would do that.” Satya “blinked” on capex and regrets it; Dario chose bankruptcy-avoidance over share, “and OpenAI was aggressive, and now OpenAI is back in the game.”
Baker and George argue that the AI industry must tell its own truth: data centers are “probably the best thing that has ever happened to working-class Americans.” Town tax revenue “10Xs,” Loudoun County pairs America’s highest income with its highest data-center density, and cheap natgas ($2-3 vs ~$20-25 in Europe/Asia) is reindustrializing America — while Baker alleges “an organized CCP-funded campaign… laundered through TikTok” against data centers. His favorite Dario line: “stop talking about curing cancer and actually cure cancer.”
Orbital compute flips on Starship reusability: of $50B per gigawatt, ~$35B is IT either way, while the $15B of terrestrial power/cooling/labor is inflationary — and reusable launch takes the space alternative under $1B. Elon and Jensen have co-designed a Reuben rack targeted for a Q4 ’27 launch; even two quarters late, “that’s 2028,” and per Brad Gershner it’s “happening in plain sight.” Training stays on Earth — latency and speed of light are real.
The endgame is an ensemble of models behind routers, and the “arbiter of intelligence” abstraction layer is the most vied-for position “in the history of business.” Enterprises will RL open-source base models (soon likely Nvidia’s, via Nemotron and the Poolside acquisition) on their own data rather than hand context to frontier labs; Fireworks Nexus is the best instantiation today; Kirkland & Ellis’s $500M in-house build validates the category but understates the difficulty.
On Nvidia, Baker’s rule for semiconductor CEOs is “the only thing you should ever say is ‘Thank you, Jensen’” — while George’s rule of thumb is that every 1% of accelerator share is worth ~$100B, so plug into Jensen’s ecosystem rather than tugging on Superman’s cape. Baker estimates Jensen has locked up roughly 70-80% of the supply chain; his data centers are the most financeable ($15B equity on $50B), and George’s deal hierarchy reads true customer preference — equity investments beat RVGs beat token-priced warrants beat naked warrants.

Deep dive

1. Nobody can name a single worsening data point — while AI stocks sell off

Baker’s standard question all summer: “Can you tell me one quantitative data point in your business that’s getting worse? Just one.” In July and August he found no takers. OpenAI has “clearly accelerated,” open source more so, and Groq saw “a pretty dramatic acceleration” after GroqBot; the hedge he volunteers is that Anthropic is in a quiet period, “so maybe they’ve slowed down a little bit.”
The disconnect he’s trading around: some AI names in “pretty significant drawdowns” over two months while fundamentals broadly accelerate. Little action at the index level, but “you can drown crossing a river that’s on average two feet deep.”

2. “Maybe everyone wins” — and Anthropic’s pre-IPO gamesmanship

Borrowing Eric Fisher’s line from the Patrick O’Shaughnessy podcast — “maybe everyone wins” — Baker lists Anthropic, OpenAI, SpaceX, Meta, Google selling TPUs, open source, neoclouds, and inference clouds. George’s LP version: every conversation starts with “How is this all gonna go wrong?”, and his answer is “this is not an or thing, it’s an and thing” — with Nvidia “at the center of all of it.”
Baker’s hypothesis on Anthropic’s apparent slowdown: they trued up accounting to be comparable to OpenAI on revenue added, tested the waters, and the next disclosure is likely a re-acceleration. Plus checkpoint gamesmanship: Anthropic is “clearly waiting” for OpenAI to release Astra, so Fable 5.1 “magically” ships hours later.
One culture flag: Anthropic interviews now ask “How would you feel if the equity went to zero?” Baker wants missionaries too, but “you can’t afford the compute you want for your mission if the equity goes to zero.” George’s tag: they’re “the accidental enterprise company” — Baker: “the accidental everything.”

3. Lab revenue is a dial public markets haven’t priced

The illustration: a lab with 10 gigawatts of power, 8 on inference at ~$60B per gigawatt per year, is a $480B-revenue business (conservative — “people seem to think Anthropic and OpenAI are both monetizing at a hundred billion dollars a gigawatt today”). A research breakthrough could flip 8 gigawatts to training and revenue would fall to $120B — “and I actually think they would do that.”
The contrast with Meta and Google is that their businesses had no comparable massive cost or infrastructure-to-serve-revenue trade-off; labs face a different dynamic.
The spending-doctrine scoreboard: Satya blinked — “I know I’m good for my eighty billion” at Davos, then slowed down, and “really regrets that.” Dario reasoned publicly that overspending risks bankruptcy, which is worse than losing share, and was conservative — “and OpenAI was aggressive, and now OpenAI is back in the game.” SpaceX was aggressive too.

4. Sub-one-year paybacks, financed by people who underwrite for a living

From Nebius — called “Nebulous” once in the transcript — and CoreWeave disclosures: a gigawatt costs ~$50B, customers prepay 50-60%, leaving $25-30B to recover — a nine-to-ten-month payback, faster on spot. SpaceX is faster still because it brings big clusters on quickly; Baker now prices per megawatt, not per GPU.
His career-level framing: “there haven’t been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub-one-year paybacks.”
On circularity fears: the financiers are Blackstone, KKR, and Apollo at relatively low cost, and one reason is useful lives keep extending while monetization per gigawatt rises — so “the true equity payback might be way inside of a year.”

5. Diffusion is “absolutely nowhere” — and GrokBot is another ChatGPT moment

George’s demand math: the monetization of these companies, roughly $180B “or something in that direction,” rests on maybe 30 million heavy-paying users — Baker takes the under, George concedes “it might be sub ten” million. Inside a16z companies, the top engineers spend 10-100x the median on tokens; some AI-native firms spend 10%+ of human comp, while old-economy companies doing a good job spend about 1%. Against 1.5 billion knowledge workers, “it feels like we’re nowhere on the demand side, and we’re massively supply constrained.”
Baker’s own tape: Atreides’ internal token consumption is up 100x from March through August, and Grok Enterprise with two users looks like another 10-20x in a month. His personal test: things that took hours with Claude Code — podcast, Substack, and X summarizers, a sentiment tracker — “each took seven to twelve seconds with GrokBot. And it’s better.”
The next leg is action-taking: a bot that answers “What are the recommended actions?” from everything the other bots learned. George is “horse racing” GroqBot, Codex, and portfolio company Town on “make me better at my job” — “just wait till everyone does this stuff… it feels like that’s sort of endless token consumption.”

6. Yes, every real technology gets a bubble — but physical constraints are the governor

Baker concedes the pattern: railroads, steel, autos, radio, internet — “you get a bubble because the markets get really excited… that overvaluation leads to an overbuild,” and debt-funded buildouts “demand immediate ROI,” so “you can’t be off on the timing.” Mitigant: a majority of this buildout is still funded from operating cash flow. He also corrects himself on record — he’d claimed the South Sea Bubble was tied to longitude and sailing; “turns out it was not.”
The buildout is straining raw productive capacity — wafers, copper (“everybody in copper, there’s an AI thesis”) — with only several million people driving “a crazy global compute shortage. What happens when that’s five hundred million?” He thinks these constraints slowing things down is “actually good for society.” New headwinds: real rates rising (“it just is what it is”) and regulation — “it’s shocking what’s happening in America.” George: “We’re in a really bad place.”

7. The industry has to tell its own truth: data centers are reindustrializing America

On doomerism, Baker recounts his X exchange with Sholto and Dario: writing one positive and one negative essay isn’t balance when the negative is existential — against Yudkowsky’s “If we build it, everyone will die,” his favorite Dario line is “stop talking about curing cancer and actually cure cancer.”
The untold story: data centers are “probably the best thing that has ever happened to working-class Americans” — college may now be “significantly NPV negative” versus electrician/plumber/HVAC wages; with behind-the-meter power, town tax revenue “doesn’t double. It, like, 10Xs.” Loudoun County — highest-income county, highest data-center density — is the rebuttal to relocation critics: “we’ve done that. And it worked out really well.” The water objection, in George’s words, is “totally debunked”; Baker calls the water use “nothing.”
Baker alleges “an organized CCP-funded campaign, I think, against data centers here in America… laundered through TikTok.” Meanwhile George argues that the Strait of Hormuz closure leaves US natgas at $2-3 versus $20-25 in Europe and Asia — a structural manufacturing-cost advantage compounding the data-center boom: “we are reindustrializing America, and it’s awesome.”
The playbook is Sheryl Sandberg’s: name specific small businesses transformed, such as the Des Moines cake bakery. Every AI company — SpaceX, Anthropic, OpenAI, Nvidia, AMD, Broadcom — should do it: “the truth shall set you free, but only if you tell it.” George’s pushback on the standard message: “we need to stay ahead of China” is “correct but ineffective” — too abstract when voters care about affordability.

8. Undersupply through ’28, possible price spikes, and the compute-inequality trap

Both take the undersupply side: no capacity is available through ’28 and forecast builds will probably slip on politics. Baker: “Everybody’s worried about oversupply. I’m more worried about undersupply.” Consequence: prices could rise for access to intelligence — Dwarkesh posited token costs up ~10x — “the opposite direction of where everybody thinks this is gonna go,” plausible only because frontier tokens carry enormous user surplus today.
Baker’s ironic warning to “data center degrowthers”: the consequence may be “real compute inequality, where big companies and wealthy people can afford compute… and it’s like, well, that happened because of you.” Baker also makes the mass-market point that advertising takes years to build, creating a dangerous gap before low-cost products can be broadly supported.
Open-source myth-busting: open tokens aren’t free — they require roughly the same compute per token as a comparably sized frontier model, with the difference coming from margins charged on top. And the Kimi license stipulates a 30% revenue share (“because it’s open weights, not open source”), while the model is far more token-hungry per task.

9. Orbital compute: “a solved problem” that flips on Starship reusability

Not the Death Star: a rack of 72 chips roughly airplane-sized, solar wings, sun-synchronous orbit, and a radiator always in the rack’s shadow. Baker’s favorite anecdote: a physics-PhD investor friend insisted it was impossible, visited SpaceX, and said “Well, I was wrong.” His meta-point: your hours of thought versus 10,000 SpaceX engineers with hundreds or thousands of hours each — “it’s a solved problem,” simpler than a Starlink satellite.
The math: of $50B per gigawatt, ~$35B is IT either way; the $15B of power, cooling, and labor is inflationary on Earth (electrician compensation, copper, materials). With Starship reusability, launch drops under $1B and “the economics just instantly flip.” Kept caveats: you always train on Earth — “speed of light limitations are a real thing” — and terrestrial data centers are “not going anywhere.”
Timeline: Elon and Jensen have co-designed a Reuben rack to launch in Q4 ’27 — “let’s just say he’s off by two quarters… that’s 2028.” Per Brad Gershner, “nobody’s really paying attention to this, and it’s happening in plain sight.”
Baker’s framing of the Elon companies is “heads you win, tails you win”: first-party AI caught the frontier fast, and any overbuilt capacity earns a sub-six-month compute payback. Baker stacks the TAMs: Starlink mobile is another $800-900B of wireless, roughly $2T with broadband, plus fast-growing AI ARR, the neocloud, and growing X ads — expect a “Starlink, GrokBot, X advertising bundle,” Google-style. Starbase Louisiana: infrastructure for thousands of launches a year, two per pad per day — probably conservative.

10. The ten-year moonshots: asteroid Psyche, Earth “zoned residential,” Optimus on Mars

Baker’s most futuristic call: “asteroid mining is gonna be a very real thing” — Psyche holds more gold, silver, platinum, and every precious metal “than exists in the Earth’s crust”; capture it into stable orbit over an American-owned Pacific atoll, work it with Optimus robots, and “delivery to Earth is free.” He pairs it with Bezos’s line: “Earth is going to be zoned residential” — heavy industry moves to space, answering the pollution objections.
Mars, “at the outside, eight years away”: a fleet of Starships lands, a ramp comes out of the Pez dispenser, “Optimus robots holding American flags” walk down and deploy solar, batteries, and racks of compute — 4K video across Mars, then humans. Bigger than the moon landing, and the century’s frame: “this will be like the age of Elon and Jensen,” who are “fundamentally altering the fabric of human society and civilization.”

11. The ensemble future and the war for the abstraction layer

Microsoft failed at frontier models — Satya said, by Baker’s recollection, that they would have competitive in-house models roughly 18 months earlier and doesn’t have them — but the world got friendlier: the future is “an ensemble of models” on a Pareto curve behind routers. Even GrokBot, in Baker’s understanding, is Gemini 3.7 Flash, Groq 4.6, and some Opus behind a router, with Elon surely pushing to make it all first-party.
The enterprise pattern: take a strong open base model — soon likely an Nvidia one, via Nemotron and the Poolside acquisition — and RL/fine-tune on your own data rather than share your context with a frontier lab, which “may be hazardous for your financial health.” Chip companies can fund open training (“it’s trivial to do a fifty to a hundred billion dollar training run for Jensen”), and Baker wonders if Google’s long game is selling TPUs and letting cash flow decide the winner.
The prize is being “the arbiter of intelligence for global enterprises” — George calls it the most vied-for position “in the history of business,” contested by labs, Microsoft, Databricks, Snowflake, Palantir, inference providers, Fireworks (whose Nexus product Baker calls the best broad instantiation today), Harvey, Legora, Salesforce, and Workday. Kirkland & Ellis’s $500M in-house build is “massive validation of the category” — but it’s not a one-time build. Baker’s retail analogy: run 1,000 clean, well-stocked, well-staffed stores across 50 states and you’re worth $50B — sounds easy, almost nobody in history has done it.
Baker’s point on Cursor: while everyone else was “creating a digital deity,” Cursor “just wanted to make great product” — the most product-focused at the frontier, now part of SpaceX, suiting Elon’s engineering mindset. Coding is uniquely verifiable and perfectly documented; Baker says the broader knowledge-work pie “is gonna be very messy to go get.” Long-run winner is the low-cost provider — hard without vertical integration — hence George’s hyperscaler lens: EV to net PP&E, “kind of an AI version of price to book.”

12. Nvidia: be nice to Michael Jordan

Baker’s rule for semiconductor CEOs: “the only thing you should ever say is, ‘Thank you, Jensen.’” George’s rule of thumb is that every 1% of accelerator share is worth ~$100B, so pick a niche and plug into an ecosystem of nine chips — accelerators, CPUs, Ethernet switches, two DPUs, scale up/out/across/in — rather than going head-on. The Jordan game-tape warning: talk trash in game 50 “and he just looks.” Sometimes Superman “just flies away — that’s what happened to the TPU team.”
Financeability is the moat: a $50B Nvidia data center needs only a $15B equity check, with Blackstone/KKR/Apollo underwriting the rest — and it’s not circularity, because a residual-value guarantee below Jensen’s gross profit is “super NPV positive with very little risk,” plus a revenue share. TPUs are second most financeable at roughly double the equity and higher rates. RVGs also democratize compute against an Anthropic/OpenAI-dominated world — the neocloud playbook again — and Baker asks whether Jensen has locked up 70-80% of the supply chain: fab, DRAM, NAND, lasers, capacitors. Per Dylan at SemiAnalysis, he’s “the central bank of AI.”
Hardware humility from a scarred semis investor: sometimes the chip comes back from the lab, they plug it in, “and it doesn’t work at all” — and you’re back for another billion. (He corrects George on Cerebras: the chips worked; they lacked product-market fit for two generations.) Credit where due: the “Halapeno” ASIC, as spoken, is the first good internal chip he’s seen outside TPU and Trainium — but it’s competitive with one of Jensen’s nine, and Elon partnering rather than building his own was “a very high Elon move”; George says history will judge it a wise decision.
George’s closing analytical tool: in a supply-constrained world you can’t infer preference from sell-outs (even old H100s resell high), so read the deal hierarchy — chip-maker equity investments in customers (Amazon and Google’s TPU/Trainium investments in Anthropic: unable to lose if dollars invested are less than gross profit) beat RVG-financed deals, which beat warrants tied to a fixed price per million tokens (good only “as long as the performance of your chip outruns the performance of your stock”), which beat naked warrants that can be negative NPV. On Nvidia’s deals, George says “there’s a reason that people I consider smart are investing in their deals.”

The OpenAI/Hugging Face attack, clearly explained Vol.232 Industry Watch 44 | Embodied Intelligence + AI4S: What New Species Could Emerge?