Dwarkesh Patel Podcast - Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028
What this covers
The conversation between Dylan Patel and Dwarkesh Patel is an economics tour of AI compute: who pays for it, how the supply chain reacts, and what the spending does to financial markets. Dylan's starting point is that compute's marginal price is set by what a buyer can earn per megawatt. Someone with a GB300 rack serving open-weight models can earn back $10-15 million per megawatt, while labs earn far more. He also traces how value capture across the AI stack has moved, from hardware and memory suppliers a year ago to labs now, with Jane Street and Meta capturing still more by using tokens. Price rises propagate with lags: Nvidia and memory makers reprice quickly, TSMC slowly, and EUV tool capacity limits how fast anyone can expand.
The macro section is the most speculative. Dylan argues that heavy borrowing could lift the discount rate on equities from 3-5% to 8-10%, hitting long-duration companies like Johnson and Johnson and railways, raising costs for debt-heavy sectors, and pushing high-debt countries such as Pakistan and Nigeria toward default, as in the 1980s. Dwarkesh goes further, flagged as fringe: if AI workers can double the labor force yearly, interest rates approach tens of percent. He cites Damon Binder's input-output analysis and a $100 billion-per-gigawatt revenue estimate.
The last stretch is contested. They discuss China's subsidized chip industry against export controls, and whether lab-backed regulation slows American labs more than open-source Chinese models. They also ask whether withholding models would matter in an RSI regime. Dylan closes on whether AI reverses capitalism's historical edge in decentralizing power.
Patel argues that Anthropic and OpenAI will control most of the world's usable AI compute by 2028 because they can monetize compute far better than anyone else and so outbid them, and that the resulting capex and debt demands could push interest rates up and destabilize non-AI sectors and countries, with regulation the main force that could slow this.
- Frontier labs' revenue per megawatt has risen from roughly $10-15 million to far higher, letting them pay more for compute than other buyers can earn from it
- Total AI capex of trillions per year must be financed largely with debt, which pushes up interest rates and the discount rate applied to all equities
- Regulation, safety-driven model withholding, and political backlash are the main forces that could cap the centralization trend
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Even when prices and profits signal demand, the semiconductor supply chain responds slowly: capacity for EUV tools and their mirror optics is limited, so the market takes years to 'whip' into expansion, with Carl Zeiss planning around 100 EUV tools per year only by the end of the decade.
“The supply chain doesn't react immediately. In fact, you go talk to someone at Carl Zeiss... We need to make 100 EUV tools by the end of the decade.”
Countries with high debt, low tax revenue and short-duration debt, such as Pakistan and Nigeria, are likely to default in a higher interest rate regime, and this could repeat the 1980s pattern in which roughly 40 countries, mostly in Latin America, defaulted after Volcker raised real rates to about 8%.
“Those countries, like Pakistan or Nigeria, I think are just going to be very fucked in this new interest-rate regime.”
If interest rates rise, the discount rate on future cash flows rises, so long-duration stable companies such as Johnson and Johnson or railways lose much of their value even if the S&P 500 index as a whole holds up, and Patel says the relevant discount rate may move from 3-5% to 8-10%.
“Why would I pay this much for Johnson & Johnson? ... Why the fuck would I invest that much if my discount rate isn't 3% or 5%? It's now 8% or 10%.”
Because chips deployed this year, such as GB300, TPUv7 and Trainium3, deliver 3-5x more performance per watt than prior generations, each new gigawatt adds far more effective compute than older ones, so labs' share of usable FLOPs compounds beyond their share of watts.
“A new watt deployed this year is significantly more efficient than the watts deployed two years ago.”
Patel argues that the regulation labs advocate is slowing American frontier labs much more than it slows open-source Chinese language models, which are not similarly constrained.
“This regulation that they advocate for is actually slowing down the labs a lot more than it slows down the open-source Chinese language models.”
If Patel's projections hold, export controls will have kept China far behind on compute stock by the time automated researchers arrive, which Dwarkesh calls a notable success, though Patel adds that Chinese financial subsidies to semiconductors may close the gap if takeoff is slower.
“If they ship the amount that you're saying, that's a huge difference... I think that's actually a notable success.”
Dwarkesh argues that blocking external deployment of models is counterproductive because external use drives revenue and feedback now, but a delay of six months in a future RSI regime would let labs use models internally for months of progress while the public lags years behind.
“In that six months, they do recursive self-improvement internally.”
The Chinese semiconductor industry receives more subsidies than the rest of the world's semiconductor industries combined, because Chinese financial systems fund chosen industries heavily, which could let China catch up on chips even though American financial systems are more willing to fund risky startups.
“So the Chinese semiconductor industry has significantly more subsidies than the rest of the world's semiconductor industries combined.”
Dwarkesh argues that once AI pushes the economy toward doubling annually, the interest rate should approach the growth rate and rise to tens of percent, making most non-AI countries default and most non-AI stocks worth close to zero because their discounted cash flows collapse.
“Okay. The rate of interest should be pretty close to the growth rate... we'll go into a world, I think in the 2030s, where the rate of interest is tens of percent.”
If AI can double the labor force each year, as Damon Binder's input-output analysis suggests for a fully automated economy, the economy could double every year or grow tens of percent annually, since people cannot double each year but AI workers can.
“in a world where you can also double the labor force every single year, how fast can the economy grow? I think it could double every single year.”
Weighted by quality, the leading US lab's compute in 2028 could exceed all of China's compute in 2029 or 2030, since 50 GW of domestic Chinese chips may be worth only about 20 GW of American chips.
“there is some factor there where that 50 gigawatts is really worth as much as 20 gigawatts from American chips.”
Regulations such as New York banning data centers, Texas moratoriums, and Ohio requiring data centers to pay property taxes across a radius reduce supply and raise costs, and those costs get passed on through the economy.
“These sorts of things are going to decrease supply and increase cost. That's going to get passed on as well.”
Labs will allocate a shrinking share of compute to inference over time and a growing share to training and research, because the marginal internal value of frontier research compute exceeds the value of additional external tokens, which is why Patel believes this is non-consensus.
“The labs are going to allocate less and less compute to inference over time. I think that's very non-consensus.”
The effective AI labor population at frontier labs is growing about 10x per year at a given capability level, so OpenAI could go from roughly 10 million AI workers to 100 million and then a billion, and within a few years a single lab could hold more effective labor than there are people on Earth.
“OpenAI goes from having, say, basically 10 million AI laborers this year to 100 million the next year, to a billion the year after that.”
If a gigawatt could sustain about a million white-collar workers paid around $100,000 each, that would be about $100 billion of revenue per gigawatt, which Dwarkesh calls surprisingly low, and full AGI would make it many hundreds of billions per gigawatt.
“If you have a gigawatt that can sustain a population of, say, roughly a million white-collar workers... That would be $100 billion. That's actually surprisingly low.”
Value capture has moved around the AI stack: a year ago the hardware and memory suppliers captured most margin while model labs lost money, and now model labs have positive margins of around $100 million per megawatt, while Jane Street and Meta capture even more value from using the tokens.
“So the value capture's shifted around a lot, which is very fun for people tracking the market”
As AI-driven borrowing raises the spread between government rates and private rates, debt-heavy industries like consumer packaged goods, telecom, and banks face higher costs, and rising borrowing costs may force the world to build fewer gigawatts than capitalism would otherwise choose.
“This crowding-out effect is the reason it's not YOLO 1 billion gigawatts.”
Most compute is contracted before it is built, so it must be financed, and companies with balance sheets like Meta and SpaceX can build speculative capacity and sell to labs at high margins, creating a power structure different from neoclouds that need a customer and capital before building.
“Meta is effectively hoarding compute. Them and SpaceX are the only plausible #3, because they're hoarding all this compute.”
Price increases spread through the supply chain with lags: when labs pay higher prices for compute, Nvidia, memory makers SK Hynix, Micron and Samsung, and substrate makers raise prices rapidly, while TSMC raises prices slowly, so the supply chain rebalances over time.
“So with the value capture, I think there's a bullwhip effect here. Just because someone has raised prices doesn't mean the entire supply chain rebalances immediately.”
The binding limit on AGI is not how fast research engineers can make progress but how much the rest of the world allows: regulators raising rates, banning data centers and fabs, and restricting model release will reallocate capital and slow the buildout.
“So the limiter on AGI is not how fast the research engineers... can crank the gears. It's actually just how much does the rest of the world let that happen?”
Capitalism has historically worked because it distributes decision-making and power, and super-centralized economies grew more slowly than decentralized capitalist economies, but AI may favor centralized, non-private-ownership structures that grow faster, so AI flips the historical advantage.
“It's why capitalism worked, right? It's decentralized decision-making and decentralized power. And then AI flips all this on its head.”
Economies of scale in AI training, where a model's training cost is amortized across billions of users, and scarce compute that lets the leader charge higher markups, together give more and more advantage to whoever is ahead, and deployment-based learning could extend that lead.
“AI training has huge economies of scale, because any effort you spend on training an AI for a specific skill... gets amortized across billions of sessions or billions of users.”
If investors truly believe AI will reshape the economy, then every stock, including memory names like Micron, SK Hynix and Kioxia, should trade at 2-3x earnings rather than higher multiples, and memory stocks should not rise tenfold again, which implies the equity market should crash.
“then everything should trade at 2 or 3x multiples and the stock market should fucking crash.”
If compute is growing 2-3x per year and recursive self-improvement is underway, delaying deployment by a year forfeits 3-6 years of AI progress, which makes a coordinated slowdown extremely costly.
“If compute is increasing 2 to 3x every single year, they prevent a whole year of AI deployment such that you're a year behind where you would otherwise have been. During RSI, you're getting 3 to 6 years of AI progress in a single year.”
Using a wafer fab model, roughly $6 billion of fab capex (tooling plus cleanroom and shell) yields a gigawatt of compute each year, and at about $100 billion of revenue per gigawatt, that capex generates over a trillion dollars of end AI revenue across its output years even after taking half for middlemen.
“$6 billion of fab CapEx produces a gigawatt every single year. A gigawatt produces right now $100 billion of revenue.”
Anthropic and OpenAI are taking a rapidly growing share of new compute, roughly 30% of compute added this year, 40-50% next year, and by around the end of 2027 more than half of incremental compute worldwide, so they will control most of the world's usable FLOPs by the end of 2028 because they can outbid everyone.
“Anthropic and OpenAI are taking as much as 40% to 50% of compute next year. This centralization doesn't look like it's slowing down or stopping.”
As hyperscalers, labs and their suppliers raise hundreds of billions in debt, households and investors may shift away from mortgages, government debt and credit for homes toward AI and data center debt, and Anthropic is willing to pay around 20% for incremental capacity because renting from SpaceX would cost even more.
“Anthropic's willing to pay 20% rates for the incremental billion dollars to build their capacity.”
The leading Chinese labs have at most 100-200 megawatts of compute, with ByteDance Seed the outlier, yet Chinese models are not far behind publicly, and Patel suggests this gap matters less right now, since a lab's training run needs only about 200 megawatts at a time.
“The leading Chinese labs have 100-200 megawatts total of compute at most, ByteDance Seed being the one outlier”
Governments, especially the US, will likely restrict labs' ability to use their newest models internally as well as externally, citing safety, and Anthropic had reportedly stopped giving Mythos to foreign employees for a time, which Patel treats as a sign of coming restrictions.
“Anthropic had to stop giving Mythos to foreign employees for a bit.”
Regulatory limits on releasing the best models, such as holding Anthropic's Model 2 and OpenAI's Astra back from release, slow lab revenue per megawatt because labs cannot sell their top models, and if this continues competitors catch up and per-megawatt revenue stalls or declines.
“They're clearly not releasing their best models, in which case their revenue per megawatt stalls or can even start to decline again”
China's domestic fabs, including SMIC and CXMT, will start producing millions of chips per year around 2027-2028, adding roughly 5-10 gigawatts of domestically produced compute in 2028 alone, though those chips are worse than Nvidia's or Google's chips in the same year.
“In '28 especially, fabs start to go up from SMIC and CXMT and such, where domestic production is actually reaching many millions of units a year.”
Anthropic's pre-training run for Mythos used under 200 megawatts for about two months, and RL used even less at a single site, so most of a lab's multi-gigawatt fleet goes to research and experimentation rather than a single model's training run, because clusters are hard to coordinate and co-locate.
“when Anthropic trains Mythos, it's sub-200 megawatts. It's sub-200 megawatts for, call it, two months.”
Accounting only for IT hardware understates AI capex: building 100 gigawatts a year requires about $5 trillion of spend on IT alone, and when power plants with 30-year lives and data centers with 15-20-year lives are built ahead of time, the true figure rises to roughly $7-10 trillion per year.
“So the $5 trillion, once you account for future years' growth, is actually going to be more like $7 or $10 trillion of CapEx.”
Of about $11 trillion in total AI-related capex from 2024 to 2029, roughly $6 trillion can be funded from cash flows while about $5 trillion must be raised as new credit, which would push interest rates up across the ecosystem.
“$6 trillion of that is funded with cash, and $5 trillion of that is funded with debt.”
Dwarkesh worries that around 2030 governments may require a six-month delay before labs release new models, during which the labs could run recursive self-improvement internally, leaving the public years behind, and Patel believes such restrictions are likely as governments already limit model release.
“I'm worried about a world where it's 2030 and the government's like, 'We're going to wait six months before you can release your newest model to the public.'”
In a takeoff scenario, labs keep their best model roughly six months ahead of the externally available version because of safety, regulation, and competitive advantage, and that six-month gap becomes a larger differential as progress accelerates.
“In a takeoff scenario, why would Anthropic not have their best model six months ahead of what is externally available?”
Anyone with a GB300 rack and open-weight models can earn back $10-15 million per megawatt of compute by serving tokens on platforms like OpenRouter, which means compute's marginal price is set by how much the buyer can earn, and labs can outbid all others because they earn far more per megawatt.
“anyone can make money off of $10-15 million per megawatt compute today... You'll start generating more revenue than you're paying for the compute.”
Every force in the market, including RSI, economies of scale, and the labs being the most effective users of compute, pushes toward centralization of compute, so Patel finds it hard to construct a framework in which AI does not produce super-concentration of resources.
“Because it seems to me that every force is screeching towards centralization. And that's scary as hell.”
The labs have moved from venture-funded losses to self-funding: serving GPT-4 on Hopper GPUs produced negative gross margin, but serving current models generates revenue that can exceed the roughly $10-15 million per megawatt compute cost by a large multiple, with Anthropic's revenue reaching as high as $50 million per megawatt, so they can reinvest serving profit into training.
“Before, if they served a model —GPT-4being served onNvidiaHopperGPUs— it was generating negativegross marginfor OpenAI. But now... the revenue has gone as high as $50 million per megawatt.”
Anthropic's monthly ARR additions plateaued even as its compute grew each month, so the marginal megawatt is going to R&D rather than inference, and Patel calls this factual and self-evident from what the labs are doing.
“So they are factually increasing their compute towards R&D today. I think this is self-evident if you look enough at what they're doing.”
Patel says the labs' compute budget is roughly 60% training and 40% inference, and that the training share splits into about 50% research, 10% development, with research covering idea generation, architecture tests, data mixes and hyperparameter sweeps.
“So far it's been 60% training, 40% inference. But that training gets broken down further. Actually 50% of the compute is research, 10% of the compute is development, and then 40% is inference.”
Next year's total AI capex is about $2 trillion, while lab revenue is only in the hundreds of billions, so the labs' cash flows cannot yet fund the buildout, and summing wafer fab, data center, accelerator and energy supply chains puts total capex well above $2 trillion.
“the labs are going to generate hundreds of billions of revenue next year. But ultimately, CapEx next year is like $2 trillion.”
Since 2022, US regulation against China has increased American share of new AI compute, so that today about 70% of new AI data center watts are deployed in America and China gets under 10%, partly because some Nvidia chips shipped to China end up in Malaysia and elsewhere.
“So today, 70% of watts are being deployed in America. China is really a very small number. Sub-10% of watts being deployed for data center AI compute is in China.”
The hyperscalers funded most AI buildout until now but no longer generate enough cash to cover capex, so they are spending everything on capex and raising debt, with Meta, Amazon, Google and Microsoft all borrowing, and Google and Meta having previously paid for buildout by halting buybacks.
“They actually spend everything on CapEx. In addition, they raise debt and spend everything on CapEx. You've seen Meta do it, even Amazon, even Google.”
In a world where safety does not constrain releases, the labs could generate $100 million or more per megawatt and pay around $50 million per megawatt for compute, leaving no rational use for compute outside the labs except to sell it to them.
“But in a world where safety doesn't matter, I do believe that's exactly what happens. They can start generating $100 million per megawatt or more”
Current modeling shows credit plus cash-funded infrastructure through 2029 is not enough compute relative to AI model demand growth, so revenue per megawatt has to keep rising, which means labs will pay more and keep outbidding.
“When you take that, this is not enough compute relative to what the demand growth is from the AI models.”
Global AI compute is projected to add about 30 GW this year, 50 GW next year, and 70 GW in 2028, reaching over 200 GW globally by end of 2028, with 90-100 GW added in 2029 and the slope possibly still rising.
“30 this year, 50 next year, 70 in '28. '29 should be on the order of 90-100.”
Most value generated by frontier models goes to end users and downstream firms such as Jane Street, which earns more from trading on the tokens than Anthropic earns in profit from selling them, and Meta gains more from ad optimization and engagement than Anthropic does.
“Jane Street... is generating way, way, way more value out of the tokens they're paying for than Anthropic is generating in terms of profit”
Labs keep their fastest and best models for internal use because the internal value of accelerating research exceeds what external customers pay, so OpenAI's Ultrafast mode is being allocated to internal use as well as external.
“the internal value I'm generating from super-fast AI or the best AI model is way more than what someone external is.”
Labs will rationally keep the most valuable tokens for internal AI research rather than sell them to downstream firms like Jane Street, since internal use can generate hundreds of millions of dollars per megawatt, so power concentrates within the labs.
“They're more valuable inside. So it's the same thing. Why would I let Jane Street make all this money off of these degenerate options traders?”
Dwarkesh argues that AI firms earn very high returns on small investments, which pushes up borrowing costs across the economy, and Patel argues that the spread between government rates and private rates, rather than the policy rate itself, is what transmits the pressure.
“So the rate of return is incredibly high... You turn $1 into $2 or $10 or something at the end of the year. That raises the rate of interest higher.”
Dwarkesh argues the US can manage a debt burden because data centers built domestically can be taxed, while corporate income tax is under 10% of federal revenue, payroll and income taxes make up over 80%, and automation may shrink those, so the US tax base depends on building data centers at home.
“I think the US will be fine at the end of the day. Because if the data centers are built in America, you can fundamentally just tax the data centers.”
Meta currently raises debt at 5-6%, and Patel expects it would happily pay about 8% because compute returns are large, so a 250 basis point rise would push the whole economy's borrowing costs up.
“recently Meta's raised at 5 to 6%. I don't see why they wouldn't pay 8%.”
For the labs to reach 100 gigawatts combined by 2028, they must outbid everyone else and pay $25, $30, or $50 million per megawatt, because at current $10-15M pricing they cannot gobble up 70% of world compute.
“They do have to start paying $25, $30, $50 million a megawatt to really gobble up 70% of the world's compute in 2028”
Labs cannot leverage all of their gigawatts for training because clusters are hard to coordinate and co-locate, multi-site training is hard, and more RL rollouts do not reliably improve models, so as automated research matures the training share of compute should rise.
“It's hard to coordinate all these clusters. It's hard to co-locate all of them. It's hard to do multi-site training.”
US corporate income taxes are under 10% of federal revenue while payroll and income taxes are over 80%, and interest is about 20% of spending; if rates rise 5 percentage points, debt service could exceed 40%, and with $2 trillion of annual borrowing it could reach 60% of tax revenue.
“Suppose the interest rates rise 1%. Over a five-year basis, the fraction of tax revenue that goes towards servicing the debt goes from 20% to 25%.”
SpaceX has sold or leased compute to labs at about $25-40 million per megawatt, pricing high enough that the seller recoups its capex within about a year, showing that the labs' revenue per megawatt already exceeds what compute cost to build.
“I can sell my compute for $25 million a megawatt or $40 million a megawatt to Anthropic and Google... I'll recoup my entire CapEx in a year.”
Patel expects blended revenue per megawatt at Anthropic or OpenAI to be at least $50 million by end of 2027 and possibly $70-80 million or more, since revenue per megawatt depends on having the best model and being allowed to release it.
“I think it could get higher than that, to like $70, $80 million a megawatt, blended across the company, if not higher.”
Using the rule of 70, an economy growing at 3% a year takes 20-something years to double, which is the baseline against which Dwarkesh contrasts the possibility of an economy that doubles every year under full automation.
“If the economy grows at 3% a year, then rule of 70, that's 20-something years.”
A large injection of capital, such as $10 billion given to Carl Zeiss to expand production, could change supply quickly, but this would have to be done across every company in the supply chain, and Patel doubts it will happen in 2026, 2027 or 2028 because the world is capital constrained.
“if you said, 'Carl Zeiss, here's $10 billion. Please fucking just expand production,' that would change things.”
Patel says that across many possible worlds, Anthropic owning the whole world is near-universal, because labs find it more valuable to use compute internally than to sell tokens to downstream firms like Jane Street, so the market structure drives power to concentrate.
“There's 80,000 worlds and in only one of them, Anthropic doesn't own the whole world.”
Even if AI firms remain privately owned, a small number of companies such as Nvidia, Anthropic, OpenAI, and the hyperscalers control most of the roughly 2% of GDP attributable to AI, so concentration is structural regardless of ownership form.
“So it could be private property, but very few companies are involved.”
The government will not stop AI through a conspiracy but through public opinion and politicians, as constituents increasingly hate AI, so real-world constraints on AI deployment will slow progress even if AI eventually arrives.
“Everyone who's elected is going to hate AI. Even the people who are elected already hate AI.”
As automated coding and automated research mature, and as continual learning arrives, the split between research and training compute will become fuzzier or shift toward training, so more compute goes to training the model itself.
“As we get further and further down automated coding and automated researcher, I actually expect the percentage of the compute budget that goes to research versus training to become a lot more fuzzy, or even higher for training.”
Patel predicts most compute will still transact below roughly $20 billion per gigawatt at the end of next year because most compute is financed and contracted before it is built, so the high bids from balance-sheet sellers will not become the market price for everything.
“I think most compute will still continue to transact at sub-$20 billion a gigawatt.”
Meta's roughly $1.5 trillion market capitalization understates its value, given its cash flows, its infrastructure, and the compute it could sell at high prices per watt, and in a fully AI-pilled world all companies should trade at 2-3x earnings.
“I think they're like a $1.5 trillion company. It's like, what? Silly. They're worth way more than that, at least in a logical sense.”
Anthropic started turning a profit in Q2 and OpenAI could turn a profit around Q3, so the labs are increasingly funded by their own revenue rather than new capital injections, though they still take in new capital to accelerate growth.
“Anthropic started turning a profit in Q2. It's believed at some point in Q3, OpenAI could start turning a profit even”
World compute in gigawatts roughly doubles each year while frontier lab compute roughly triples, so the labs' share of the total keeps rising, and if the trend holds the labs' gigawatts go from about 2 at the start of the year to near 6 by year end, about 18 by end of 2027, and about 54 by end of 2028.
“There's this trend where maybe world compute in gigawatts doubles every year, but the compute at the frontier labs triples every single year.”
Dylan jokes that educating people through podcasts could make society smarter and speed up AI progress, so Dwarkesh's show might be accelerating AI development.
“You educate people. Maybe if they're smarter, they're progressing AI faster.”
Host and guest exchange about sponsor mentions and sign-off pleasantries, with no substantive argument.
“This is not an ad. This is not an ad. This is not an ad.”