Sub-one-year compute paybacks make undersupply the bigger risk than bubble

1. No bear case found — ▶ 1:22

   Baker's standard question to everyone is to name one quantitative data point in their business that is getting worse. Through July and August he found nobody. OpenAI, open source and Grok have all accelerated, and he expects Anthropic's next disclosure after its quiet period to show reacceleration. Meanwhile public AI stocks have "fallen out of bed" over two months. His image: you can drown crossing a river that's on average 2 ft deep. Index levels hide big drawdowns in individual AI names while fundamentals broadly accelerate.

2. Revenue is a dial — ▶ 5:52

   Take 10 gigawatts, with 8 allocated to inference at $60 billion per gigawatt. That is $480 billion a year, a one-year payback on revenue. Shift 8 gigawatts to training after a research breakthrough and revenue falls to $120 billion. Baker thinks labs would make that call. The pushback is that stock volatility would hit morale, recruiting and retention. Either way, revenue is under lab control through which checkpoint they release, where they price on the Pareto curve and how they split compute. Meta and Google never faced a trade-off like that.

3. An "and" thing — ▶ 8:03

   Every LP conversation opens with what is going to crash. Baker rejects the framing: it is an "and" thing, not an "or" thing. Frontier models work, N-1 models work, open source works, applications work, the clouds and the five labs work. Nvidia sits at the center of all of it, and 26 years have taught him not to bet against Jensen Huang. He agrees with Eric Fishery's "maybe everyone wins" view, including Google selling TPUs and neoclouds winning, and calls zero-sum thinking the real error.

4. Labs won't chase free cash flow — ▶ 9:07

   Labs will generate large operating cash flow but not free cash flow anytime soon. It goes into GPUs and XPUs, or into subsidizing first-party products. Belief in scaling laws keeps them focused on compute, not margin. Satya Nadella blinked at Davos with his "good for my $80 billion" line, slowed down and regrets it. Dario Amodei was conservative because bankruptcy is worse than losing shares. OpenAI and SpaceX were aggressive, and OpenAI is back in the game. Returns on that aggression look high regardless of supply-demand mismatches.

5. Paybacks inside a year — ▶ 11:50

   A gigawatt costs $50 billion. Customers prepay 50-60%, leaving $25-30 billion at risk. Nebius and CoreWeave disclosures imply a 9-10 month payback, spot is faster, and SpaceX brings big clusters online faster still. Baker says there have been few chances in his career to deploy tens or hundreds of billions at sub-one-year paybacks. Blackstone, KKR and Apollo finance it at low cost, and useful lives keep extending as models improve. He dismisses the circularity worry. True equity payback may sit "way inside of a year."

6. Demand barely diffused — ▶ 14:34

   The knock on overbuild is demand. Revenue near $80 billion rests on perhaps 30 million heavy paying users, and Baker thinks it may be under 10 million. The highest-spending engineers burn 10-100x the median. Token spend against human compensation is about 1% at good old-economy firms, high single digits at tech-forward ones, and 10%-plus at AI-native companies. With 1.5 billion knowledge workers, diffusion is "nowhere" on a 10-year view. Disappointing real-economy adoption is the risk he concedes, not oversupply.

7. From reactive to autonomous — ▶ 16:41

   Baker's internal token spend rose 100x from March to August. Giving two people Grokbot Enterprise may lift it another 10-20x in a month. Summarizers and sentiment trackers that took hours with Claude Code took 7-12 seconds with Grokbot, and came out better. Engineers went from roughly 20% AI-written code to 90%-plus. Those tools are still reactive, though. The next layer asks the bot for recommended automations, then clicks go. He is horse-racing Grokbot, Codex and Town on that job, and calls the demand it implies endless.

8. Bubbles need constraints — ▶ 20:36

   Automobiles, TV, radio, the internet, the PC, railroads and steel mills all produced overvaluation, then overbuild. Debt-funded builds demand immediate ROI, but most of this one is still funded from operating cash flow. Physical limits bind first. If 10% of the vision comes true, several million users cause a compute shortage, so what happens at 500 million, and how many copper mines? Baker calls those constraints good for society. Rising real rates make sense given the investment. Regulation he calls "a really bad place."

9. Tell the data center story — ▶ 23:56

   Baker says the industry must tell its own truth. Electricians, plumbers and HVAC technicians earn so much that college may be NPV negative. Data centers can lift town tax revenue 10x, Loudoun County, Virginia, pairs the top US income with the densest data centers, and water use is "debunked." US gas at $2-3 against $20-25 abroad is re-industrializing America. "Stay ahead of China" is correct but too abstract. He wants labs to copy Meta's habit of naming real small businesses, and claims an organized CCP-funded, TikTok-laundered campaign opposes data centers.

10. Undersupply through 2028 — ▶ 29:34

    No capacity is available against forecast builds through 2028, and political delays will push those builds out. Baker worries about massive undersupply, not oversupply. That could raise the price of intelligence, which runs against consensus. Dwarkesh floated a 10x token price rise. Users pick frontier tokens because of large consumer surplus. The likely loser is the public: compute inequality caused by degrowthers, since ad-funded mass-market tiers take years to build. Open-source tokens are not free either, needing the same compute, and Kimi's license takes a 30% revenue share.

11. Orbit as swing capacity — ▶ 34:22

    An orbital rack is airplane-sized, about 72 chips, not a Death Star. A sun-synchronous orbit keeps the radiator in the rack's shadow. On Earth, a $50 billion gigawatt is about $35 billion of chips and $15 billion of power, cooling and labor, and that $15 billion is inflationary. Starship reusability takes launch under $1 billion, so economics flip. Training stays on Earth because latency matters. Musk and Huang co-designed a Rubin rack for Q4 2027, and Baker accepts a two-quarter slip. Physics skeptics, including one PhD investor, changed their minds after meeting SpaceX engineers.

12. Heads you win, SpaceX — ▶ 39:42

    Skip orbital compute and SpaceX still works. Starlink mobile addresses another $800-900 billion of wireless revenue, and with broadband the market is close to $2 trillion. First-party AI is growing fast, X ads are growing, and a Starlink, Grokbot and X ads bundle could echo Google's cloud-plus-ads play. If SpaceX overbuilt, scarce supply still gives a sub-six-month compute payback. Starbase Louisiana can support thousands of launches a year, and pads are engineered for two Starship launches a day. Baker calls it "heads you win, tails you win."

13. Asteroids and Mars — ▶ 44:41

    Asteroid mining will be "a very real thing," in Baker's view. Psyche holds more gold, silver and platinum than exist in Earth's crust. With Starship and a lunar base, asteroids could be moved to a stable orbit over an American-owned atoll in the Pacific, worked by Optimus robots with no humans within 50 miles. He cites Jeff Bezos's view that Earth becomes zoned residential and heavy industry moves to space. On Mars, he expects a fleet of Starships within eight years at the outside, carrying Optimus robots, solar panels, batteries and compute racks.

14. Ensemble beats two giants — ▶ 48:15

    Microsoft tried and failed to build a frontier model, yet the future favors it. Companies will fine-tune and RL a strong open model on their own data, which Baker expects to be Nvidia's, and own it. It will sit behind a router with one or two frontier models, the best for planning and cheaper ones for execution. Baker's understanding is that Grokbot runs Gemini 3.7 Flash, Grok 4.6 and some Opus. It is hard to execute, like running a $50 billion retail chain. Kirkland & Ellis spending $500 million shows both the size and difficulty of the prize.

15. Nvidia's financing moat — ▶ 1:00:29

    A $50 billion Nvidia data center needs a $15 billion equity check, and the other $35 billion is financeable. A residual value guarantee is safe while it costs less than the gross profit from chips sold into the site. TPU sites need roughly double the equity and pay higher rates. Nvidia has locked up an estimated 70-80% of supply. Baker's rule: every 1% of accelerator share is worth about $100 billion, so be nice to Jensen and plug into his ecosystem. Dylan calls him "the central bank of AI." Jalapeno rivals only one of his nine chips.

The binding constraint on AI is supply, not demand. Gavin Baker thinks the buildout will be short of capacity through 2028. He says no capacity is available against all the forecast builds, and those builds are now likely to slip because of political resistance. Everyone worries about oversupply. He worries about the opposite. If he is right, the consensus on pricing is backwards as well. Intelligence gets more expensive, not cheaper. Dwarkesh has floated the idea that a token could cost 10x more. Users would pay it because frontier models already hand them a large surplus.

The supply economics are the reason to believe it. Take a gigawatt that costs $50 billion to bring online. Customers put up 50-60% of that in advance, which leaves $25-30 billion at risk. Nebius and CoreWeave disclosures imply a payback of 9-10 months on that basis. Spot pricing pays back faster. SpaceX, which brings big clusters online quickly, does better still. Baker says he has rarely seen a chance to deploy tens or hundreds of billions of dollars at sub-one-year paybacks. The money is cheap too. Blackstone, KKR and Apollo finance these builds at a low cost of capital, and useful lives keep extending as models improve. That is why he waves off the circularity worry. True equity payback, he says, may sit well inside a year.

The obvious objection is that supply economics always look good just before the overbuild. That depends on demand, and demand looks barely started. The lab revenue base, around $80 billion, rests on perhaps 30 million heavy paying users. Baker takes the under and thinks it may be below 10 million. Set that against 1.5 billion knowledge workers. Token spend runs about 1% of human compensation at old-economy firms that are doing well, high single digits at tech-forward ones, and 10% or more at AI-native companies. Inside companies, the heaviest engineers spend 10-100x the median. His own firm's token spend rose 100x between March and August. He expects another 10-20x within a month after giving two people Grokbot Enterprise. The next step is agents that recommend and take actions rather than summarize. Over a decade, he says, we are nowhere on diffusion.

That is why Baker rejects the question he hears in every LP meeting, which is what is going to crash. He thinks it is an and thing, not an or thing. Frontier models work. Models one step behind the frontier work. Open source works. Applications, clouds, the five labs and Nvidia all work. Open source in particular is no threat to the labs' economics. An open-source token takes the same compute as a frontier token on a comparably sized model. Only the margin on top differs. Nvidia is happy to see 40% margins on tokens instead of 90%, because cheaper tokens mean more tokens, and more tokens mean more compute in a supply-constrained world. Open source is also the hedge against a future where only big companies and wealthy people can afford compute.

The harder question is who sits on top. Baker expects enterprises to take a strong open model, run reinforcement learning and fine-tuning on their own data, and keep that context away from the frontier labs. Behind a router, the owned model works alongside one or two frontier models. The most advanced handle planning and cheaper models handle execution. His understanding is that Grokbot already runs Gemini 3.7 Flash, Grok 4.6 and some Opus this way. That future suits Microsoft far better than a two-lab world, even though Microsoft failed to build its own frontier model. But describing it is easy and building it is not. Baker compares it to running a $50 billion retail chain: a thousand stores across 50 states, each stocked right for its region. Kirkland & Ellis is spending $500 million on its own system. That validates the category and shows its cost. Coding and legal are documented and verifiable. Most work for those 1.5 billion people is not. And without owning compute, nobody in this layer is the low-cost provider.

Nvidia is the hinge. Baker's case rests on financing. A $50 billion Nvidia data center needs a $15 billion equity check, and the other $35 billion is financed. A residual value guarantee sits behind it. As long as the guarantee costs less than the gross profit Nvidia earns selling chips into the site, the deal carries little risk for Nvidia. TPU sites need roughly double the equity and pay higher rates on the debt. Nvidia has also locked up an estimated 70-80% of supply, from fab capacity to DRAM, NAND, lasers and capacitors. One analyst, Dylan, calls it "the central bank of AI." The practical advice follows. Every 1% of accelerator share is worth about $100 billion, so a chip startup should pick a niche and plug into Nvidia's ecosystem. Even the best new ASIC Baker has seen, Jalapeno, competes with one of Nvidia's nine chips. In a shortage, customers take any chip, so deal structure is the honest signal of preference. Chip-company equity stakes rank first, then residual guarantees, and plain warrants rank worst.

If Earth cannot supply enough, orbit might. Baker's case is cost arithmetic. An orbital rack is airplane-sized, about 72 chips, and sits in a sun-synchronous orbit so the radiator stays in the rack's shadow. On Earth, a $50 billion gigawatt is about $35 billion of chips and $15 billion of power, cooling and labor. That $15 billion is inflationary, because electricians and copper are getting scarcer. In orbit, it is replaced by launch cost, which falls under $1 billion with Starship reusability. Training stays on Earth because latency matters. But an increasing fraction of the world's compute moves up. Elon Musk and Jensen Huang have co-designed a Rubin rack for launch in Q4 2027, and Baker will accept a two-quarter slip. Even without orbital compute, SpaceX is "heads you win, tails you win." Starlink mobile and broadband address close to $2 trillion, first-party AI is growing, and a sub-six-month compute payback covers any overbuild. Asteroid mining and Mars are further out, but he treats both as likely.

What can break this is politics, not physics. Baker says rising real rates and regulation are the real brakes, and that he is shocked at what is happening in America. His answer is that the industry must tell its own story, and the story is strong. Electricians, plumbers and HVAC technicians earn so much that college may now be NPV negative. A data center can lift a town's tax revenue 10x. Loudoun County, Virginia, has the highest data center density and the highest income in the US. Water use, he says, is debunked. US natural gas costs $2-3 against $20-25 in Europe and Asia, which is re-industrializing the country. The standard "stay ahead of China" argument is correct but too abstract. Meta does better, with Sheryl Sandberg walking through specific small businesses, like a Des Moines bakery that grew to 15 locations. He wants every lab and chip company to do the same. He also asserts, with no evidence offered, an organized CCP-funded campaign against data centers, laundered through TikTok. The cost of losing is compute inequality, with big companies and wealthy people buying scarce compute while everyone else waits for an ad-funded tier that takes years to build.

The strongest counter-case is history, and Baker concedes it. Every profound technology, from railroads and steel mills to the PC and the internet, produced a bubble, then overvaluation, then overbuild. Debt-funded builds need immediate returns and cannot be wrong on timing. His defenses are that most of this build is still paid from operating cash flow, and that physical limits slow it. If 10% of what he describes comes true, several million users already cause a global compute shortage. What happens at 500 million, and how many copper mines does that take? He calls those constraints good for society.

A second weakness is his own revenue arithmetic. A lab with 10 gigawatts, 8 on inference at $60 billion per gigawatt, makes $480 billion a year. Move 8 gigawatts to training after a research breakthrough and revenue falls to $120 billion. Baker thinks a lab would do it. His interlocutor is not so sure, because falling revenue would hit morale, recruiting and the stock. Either way, revenue is under the labs' control, through which checkpoint they release, where they price and how they split compute. The one-year payback is therefore partly a choice. Nadella's blink on capex and Dario Amodei's preference for caution show how contested that choice is. The third weakness is the survey itself. Baker asked people inside the ecosystem for one worsening metric and found none. That is useful. It is also a sample of people living off the boom.

The test comes when OpenAI and Anthropic are public and a training decision shows up in reported revenue. Baker thinks that will help, because investors will finally be able to put the numbers into a model. It will also be the first time the market learns how much of the number is a choice.