AI Isn't Oversupplied — It's Severely Undersupplied
Everyone is worried too many AI data centers are being built. Gavin Baker's conclusion is the opposite: by 2028 the real problem is severe undersupply, and compute prices may rise rather than fall.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
He couldn't find a single metric getting worse
Gavin says he spent the whole summer asking everyone the same question: can you name one quantitative data point in your business that is getting worse, just one. Across July and August he couldn't find a single person. Anthropic may be slowing because of its quiet period, but OpenAI is clearly accelerating, open source is accelerating more, and Grok clearly accelerated after Grokbot. Meanwhile public-market AI names have pulled back sharply over the last two months. His analogy: you can drown in a river that averages two feet deep — at the index level nothing is moving, but individual AI names have drawn down deeply, while the fundamentals are accelerating across the board.
— Gavin BakerAnthropic is an accidental enterprise company
Gavin says Anthropic asks candidates in its culture interview, "how would you feel if the equity went to zero," because it wants mission-aligned people. His comment: we want missionaries, but we also want people to make money, because if the equity goes to zero you can't afford the compute your mission requires. He calls Anthropic an "accidental enterprise company" — the enterprise business is just a byproduct of the mission, whereas OpenAI and SpaceX are more commercial. He also speculates that Anthropic did a revenue-definition rebase at some point, that it is now comparable to OpenAI, and that the next disclosure is quite likely to show reacceleration.
— Gavin BakerRevenue is a variable the labs control themselves
Gavin offers a concrete calculation: suppose a lab has 10 gigawatts of power, 8 gigawatts of it for inference, and at $60 billion of revenue per gigawatt per year that's $480 billion in annualized revenue — less than a year to pay back on a revenue basis. But if there's a major research breakthrough and they decide to switch those 8 gigawatts from inference to training, revenue instantly drops from $480 billion to $120 billion. His judgment is that they will do it. Which means public markets have to get used to one thing: these companies' revenue is largely determined by their own choices about which checkpoint to release, how to price it, and how to allocate compute between training and inference.
— Gavin BakerThe demand side hasn't even started
Gavin says these companies have roughly $80 billion of revenue, behind only about 30 million genuinely heavy paying users, and he himself would "take the under" on that 30 million figure. What a16z sees internally is a power law: the heaviest-using engineers spend 10x, even 100x, the tokens of the median engineer. On the enterprise side, incumbent banks spend about 1%, technology-forward companies spend a high single-digit percentage, and the most AI-native companies spend over 10%. Against 1.5 billion knowledge workers worldwide, his conclusion is that on the demand side we are nowhere near anywhere yet, while the supply side is already severely constrained.
— Gavin BakerGrokbot is another ChatGPT moment
Gavin says token consumption inside a16z rose 100x from March to August. After he got Grokbot Enterprise, with two people using it, token spend could rise another 10 to 20 times in a month. The podcast summarizer, Substack summarizer and X sentiment tracker he built with Claude Code in a few hours each take just 7 to 12 seconds with Grokbot, and come out better. His distinction: Claude Code is reactive knowledge augmentation, while Grokbot directly gives you action recommendations — "here's what I recommend you do today." He is 50, and says that no matter how hard he tries he will never be as natively fluent with AI as a 23-year-old.
— Gavin BakerThe real risk is severe undersupply
Gavin says everyone is worried about oversupply; he is more worried about severe undersupply by 2028. The reason is that all the forecast buildout could be delayed by political factors. If that scenario plays out, you would see the price of acquiring intelligence rise substantially — the opposite direction from what everyone expects. He cites Dorcash's view that token costs could rise 10x. His logic: the reason people are willing to use frontier tokens rather than cheap tokens today is that even at frontier prices there is enormous consumer surplus. If supply runs short, the result could be "compute inequality" — large companies and wealthy people can afford compute — and that would be caused by the data centers' own opponents.
— Gavin BakerOrbital compute is a Starship reusability contest
Gavin says an orbital data center is not a building in space; it's roughly an airplane-sized rack of 72 chips in a sun-synchronous orbit, with radiators permanently in shadow. He runs the numbers: assume $50 billion per gigawatt, of which $35 billion is chips and the remaining $15 billion is power, cooling and labor — and those inflate on Earth, because electricians and materials are both getting more expensive. So the real comparison is launch cost, and Starship reusability can push it below $1 billion, flipping the economics instantly. He concedes that training will always be done on Earth and that latency is a real constraint, but that an ever-larger share of the world's compute will go into orbit.
— Gavin BakerMicrosoft is betting on a portfolio of models, not one frontier
Gavin says Microsoft tried to build a frontier model and failed; Satya said 18 months ago they would have their own competitive model, and now they don't. But the world has become much friendlier to their strategy, because the future is a portfolio of models: no single model is best at every task. He predicts that ten thousand large companies worldwide will take the best open-source model — most likely NVIDIA's — do RL and supervised fine-tuning on their own data, own and control their own intelligence, and put one or two frontier models behind a router for planning. That future is far friendlier to Microsoft, because there are no longer just two dominant frontier models.
— Gavin BakerWithout vertical integration you can't be the low-cost provider
Gavin says becoming the abstraction layer for enterprise intelligence is one of the best positions in business history, but it is very hard to do. He uses American retail as an example: any category in the US is worth $50 billion or more, and you just need a thousand stores across 50 states, with different climates and consumer preferences, inventory that has to be right, prices that have to be right, employees who have to be friendly and not steal, 100% annual turnover, and stores that have to be clean and bright — it sounds simple, and historically the number of people who have pulled it off can be counted on one hand. The abstraction layer is the same. He also says that in the long run, if you don't vertically integrate and don't own your own compute, you cannot become a low-cost provider.
— Gavin BakerAn accelerator with 1% share is worth a hundred billion
Gavin's advice to semiconductor founders: if you're a chip CEO, the only thing you should say is thank you Jensen, how do we work with you. His rule of thumb is that today every 1% of accelerator share is worth roughly $100 billion, so there's no need to take NVIDIA head-on — pick a niche and getting 1% is enough. He points out that Jensen has nine chips, from two kinds of GPU to CPU to Ethernet switch, and that all of his largest customers have products competing with those chips. The key reason is that Jensen's data centers are the easiest to finance: a $50 billion project needs only $15 billion of equity, with the other $35 billion financeable as debt, plus residual value guarantees and revenue sharing.
— Gavin BakerOpen source is good for NVIDIA
Gavin says he cannot accept the idea that Jensen being the biggest advocate of open source is a huge risk to his business. Quite the opposite: open source means tokens generated on NVIDIA GPUs are no longer at 90% gross margin, maybe 40%, so more tokens get consumed and more compute is needed — which in a supply-constrained world is good for him. He says Jensen has locked up 70% to 80% of global supply, plus wafer capacity, DRAM, NAND, lasers, capacitors, everything a rack needs. His incentive is the fragmentation of AI, which aligns perfectly with American interests.
— Gavin BakerIn their own words · checked verbatim
my standard question is, can you tell me one quantitative data point in your business that's getting worse? Just one.
Gavin Baker1:02
we want missionaries, but we also want people to make money. And at the end of the day, you can't afford the compute you want for your mission if you go if the equity goes to zero.
Gavin Baker5:06
this is not an or thing it's an and thing right like this is an and thing
Gavin Baker8:10
every time you've had a real, you know, profound new technology, you know, whether it's the automobile, the TV, the radio, internet, the PC, um, railroads, steel mills, you get a bubble because the markets get really excited and they get ahead of themselves.
Gavin Baker20:24
everybody's worried about over supply. I'm like more worried about massive massively under supply.
Gavin Baker29:30
there's not a physics reason why this can't work.
Gavin Baker36:33
my rule of thumb for accelerators every 1% share today is probably worth a hundred billion
Gavin Baker1:00:47
I just can't take it that people think that Jensen is like the world's biggest advocate for open source and it's somehow the a giant risk to his business.
Gavin Baker1:04:47
Figures
| Growth in token consumption inside a16z | Up 100x from March to August | 16:21 |
| Increase in token spend from Grokbot Enterprise | Could rise 10 to 20 times in a month | 16:21 |
| Number of knowledge workers worldwide | 1.5 billion | 14:19 |
| Nebius compute payback period | 9 to 10 months | 11:14 |
| Orbital data center cost structure | $50 billion per gigawatt, of which $35 billion is chips and $15 billion is power, cooling and labor | 36:33 |
| NVIDIA data center financing structure | A $50 billion project needs $15 billion of equity, with the other $35 billion debt-financed | 1:02:47 |
| Share of global compute supply Jensen has locked up | 70% to 80% | 1:04:47 |
| Kirkland Ellis budget for building its own legal AI | $500 million | 57:45 |
Glossary
- Neocloud
- A cloud provider dedicated to renting out GPU compute, such as Nebius or CoreWeave.
- RVG (residual value guarantee)
- A guaranteed residual value that chipmakers provide on data center assets, reducing risk for the financier.
- ZDR (zero data retention)
- A policy under which enterprises require model providers not to retain their data, affecting whether enterprises are willing to hand their data to frontier labs.
- EV to net PP&E
- Comparing a company's market value with the net value of its compute assets — like a price-to-book ratio for the AI era.
- Jalapeno
- OpenAI's in-house ASIC chip, which Gavin calls the first good in-house chip besides TPU and Trainium.
How to listen
Founders and investors watching AI compute supply and demand, data center investment and the semiconductor ecosystem; anyone who wants to understand why the "AI bubble" narrative may have it backwards.
The opening section on Anthropic's quiet period and IPO game rules can be fast-forwarded.