Compute is concentrating into two labs several times over each year, and by 2028 they will control most of the world's compute
Compute is concentrating into Anthropic and OpenAI at a rate of several times a year; by 2028 they will control most of the world's effective compute, may push the price of compute to $50 million per megawatt, and could trigger a sovereign debt crisis.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Two labs are buying up the world's new compute
Global AI capex has already passed $1 trillion this year, and will exceed $2 trillion in 2028. At the start of this year OpenAI and Anthropic each had roughly 2 gigawatts; by year end both will be above 5 gigawatts. Around 30% of this year's added compute went to the two labs, and next year that share rises to 40%-50%. Because each new generation of chips delivers 3-5 times the performance per watt of the last, the compute they are adding is also of higher quality, so by 2028 they will control most of the world's usable FLOP. The logic is simply that they can monetize compute better than anyone else, which lets them outbid everyone for every marginal unit.
— Dylan Patel$600 million of fab capex levers into more than $1 trillion of revenue
Dwarkesh works through the model: $600 million a year of fab capital expenditure sustains the output of one gigawatt of compute, and one gigawatt today generates on the order of $100 billion a year in revenue. Roll that forward over five years and the same $600 million ends up mapping to more than $1 trillion of end AI revenue. Even after every intermediary takes its cut, the leverage is still hundreds of times over. That is why everyone is expanding capacity as fast as they can, and why the supply chain -- EUV lithography above all -- has become the binding constraint. But physical capacity expands slowly, so the world faces years of compute shortage regardless.
— Dwarkesh PatelLabs will bid the price of compute up to $50 million per megawatt
Right now anyone can make money on compute at $10-15 million per megawatt, which is why prices have already started to climb. If the labs are going to absorb 70%-80% of the added compute in 2028 and reach 100 gigawatts between them, they will have to pay $25 million, $30 million, even $50 million per megawatt. Elon Musk has already demonstrated the play: selling SpaceX's compute to Google or Anthropic at $40 billion per gigawatt. As long as the labs' revenue per megawatt grows faster than the price, they have every incentive to keep bidding it up, until there is no margin left for anyone else.
— Dylan PatelMost of the value the model layer creates goes to its customers
Gross margins at the model layer have flipped from negative to positive, but most of the value is not being captured by OpenAI and Anthropic -- it flows to the end users. Jane Street earns far more in trading revenue from GPT-5.6 than it pays in token fees; Meta was at one point rumored to account for around 10% of Anthropic's business, and made back more than that by optimizing ads. The same rotation happens inside the hardware chain: memory makers got nothing out of HBM in 2023, and now earn more than TSMC. Whether the model companies end up renting compute out at a high price the way Elon does, or keep it for internal use, is what will decide the split.
— Dylan PatelCompute will pull back from external inference into internal training
This is the non-consensus view: the labs will devote less and less of their compute to external inference and shift more of it to internal research and training. The reason is that keeping the strongest model in-house produces far more value than selling tokens outside. OpenAI and Anthropic are already doing this: the share of added compute going into R&D is rising, while the best model available to outsiders may stall -- OpenAI stopped training for two weeks, Anthropic does not ship a model that fails safety review. That widens their internal iteration advantage further.
— Dylan PatelChina is tens of times behind on compute but barely behind on models
China's currently deployed data-center AI compute accounts for less than 10% of global additions, and is projected to reach only around 30 gigawatts in 2028, on domestic chips whose performance clearly lags. It may add another 50 gigawatts in 2029, but in quality terms that is equivalent to only 20 gigawatts of American chips. And yet China's leading labs have at most 100-200 megawatts, against Anthropic's more than 5 gigawatts by year end, and the gap in the models is nowhere near that large. Dylan suggests this may mean the current compute gap matters less than it looks -- but if the US keeps tightening export controls while China's domestic fabs ramp in 2027-2028, the catch-up starts.
— Dylan PatelAI's borrowing will crush blue chips and poor countries first
Total AI capex from 2024 to 2029 runs to roughly $11 trillion, of which more than $5 trillion has to be borrowed. Borrowing on that scale pushes interest rates up: Meta pays 5%-6% on debt today and may be willing to pay 8% later. Higher rates shrink the discounted cash flows of non-AI companies sharply, so the prices of long-duration, stable blue chips -- Johnson & Johnson, say -- collapse; countries with heavy debt and thin tax revenue, such as Pakistan and Nigeria, could default en masse. Dylan sees this as a constraint on the expansion; Dwarkesh thinks the US can tax its way through it and that it is everyone else who gets hurt.
— Dylan PatelWithin a few years one lab's labor force exceeds the world's population
Frontier compute grows 4-5 times every 12 months, while the compute needed to reach any given level of capability falls 3 times a year, so the effective AI labor force at a frontier lab swells by about 10 times a year. At that rate OpenAI goes from roughly 10 million AI workers this year to 100 million next year and 1 billion the year after; even if compute growth slows, it takes only a few years before a single lab's labor equivalent exceeds the population of the earth. That means most of the future "working population" may sit inside two companies -- and whether or not they are aligned, the structure of power changes completely.
— Dwarkesh PatelIn their own words · checked verbatim
$6 billion of CapEx at the fab level will have generated over a trillion dollars of end AI revenue.
Dwarkesh Patel10:13
The only reason to have inference compute be so large is so you can grow your training fleet.
Dylan Patel32:26
If it rises 5 percentage points, that would go north of 40%. But if you take into account the fact that the government is borrowing $2 trillion every single year, then that goes from 40% to north of 60%.
Dwarkesh Patel49:31
Well, if you’re really AI-pilled, everything in the economy should trade at 2 or 3 times earnings.
Dylan Patel49:31
We talk often about centralization of power because of nationalization or whatever. But we don't think enough about the fact that we're actually moving very fast into a regime where most “people”, in terms of work output, are concentrated within two labs who are consuming more and more of the world's compute.
Dwarkesh Patel1:07:01
It seems to me that every force is screeching towards centralization. And that's scary as hell.
Dylan Patel1:07:01
Figures
| OpenAI/Anthropic compute scale | roughly 2 gigawatts each at the start of the year, more than 5 gigawatts each by year end | 0:00 |
| Baseline cost per megawatt of compute | about $10-15 million; Anthropic's revenue per megawatt tops out at $50 million | 0:00 |
| The two labs' share of added compute next year | 40%-50% | 7:01 |
| Total AI capex, 2024-2029 | roughly $11 trillion, of which more than $5 trillion is debt-financed | 49:31 |
| Global added AI compute (as framed on the show) | 30 gigawatts this year, 50 next year, 70 in 2028, 90-100 in 2029 | 36:11 |
| China's total AI compute in 2028 | about 30 gigawatts or less | 36:11 |
| China's 2029 additions and their equivalent | may add 50 gigawatts, but roughly equal to 20 gigawatts of American chips | 36:11 |
| Meta's recent cost of debt | 5%-6%, and may be willing to pay 8% later | 49:31 |
Glossary
- FLOP
- The basic unit for measuring AI compute: floating-point operations per second.
- RSI
- Recursive Self-Improvement -- an AI improving its own capabilities, compounding exponentially.
- hyperscaler
- The giant cloud computing companies: Google, Meta, Amazon, Microsoft and the like.
- crowding-out effect
- Borrowing on a vast scale pushes interest rates up and squeezes out other borrowers' demand for credit.
- continual learning
- A model that keeps updating on use after deployment, rather than being trained once and frozen.
How to listen
Investors tracking how AI compute gets allocated, data-center capex, and the path of interest rates; founders making decisions inside the standoff between cloud providers and labs; engineers studying how China's chip industry might catch up.
The sponsor read from 1:19-2:45 is skippable; the rest is dense throughout.