Dylan Patel – Two labs will soon control most of the world's workforce

1h 16m 375K views 25 Aug 2026

Is it worth your time?

You get a detailed breakdown of how frontier AI labs acquire physical infrastructure, fund massive capital expenditures through debt, and concentrate future workforce capabilities.

This conversation examines how the explosive growth of AI infrastructure is reshaping global energy markets, financial systems, and the future distribution of labor. It explores the economic mechanisms that allow a couple of private labs to absorb the world's compute and capital.

Frontier labs consuming global compute. Patel expects that by 2028, OpenAI and Anthropic will control most of the usable flops on Earth because they generate up to $50 million in revenue per megawatt and can outbid the rest of the market. He claims that their CapEx will exceed $2 trillion by 2028, with these two labs absorbing nearly half of all incremental global compute.

CapEx and interest rates. Patel warns that the massive borrowing required to fund trillions of dollars in data centers will drive up global interest rates and potentially trigger a sovereign debt crisis for developing nations. He argues that traditional industries will struggle to compete with AI labs willing to pay high interest rates for capital because their returns on compute are so immense.

Concentration of future AI labor. Patel contends that as frontier models scale, almost all future labor will be concentrated within a tiny number of AI companies. He calculates that the effective population of AI laborers at these labs increases tenfold every year, leading to extreme resource centralization unless governments intervene or slow progress down.

Two labs will soon control most of the world’s compute

The world economy is increasingly a function of lab economics and the compute market. By the end of last year, most American GDP growth was just AI infrastructure. This year, about a third of the compute coming online is for labs like OpenAI and Anthropic. CapEx is a little over a trillion dollars this year and will exceed $2 trillion by 2028. The labs are evolving from companies spending tens of billions a year to hundreds of billions, forecasting trillions by the end of the decade.

Anthropic started turning a profit in Q2, and OpenAI could do the same in Q3 with the rise of Codex and 5.6. Over the last year and a half, their margins have skyrocketed. While the base cost of compute is around $10, $13, or $15 million per megawatt, the economics have shifted. Previously, serving GPT-4 on Nvidia Hopper GPUs generated negative gross margin for OpenAI. Now, when OpenAI serves GPT-5.6 or Anthropic serves Opus 5 or Fable 5, revenue generation has passed well beyond that incremental $10-15 million per megawatt; in Anthropic's case, revenue has reached $50 million per megawatt. This enables them to spend 10 bucks on inference capacity, generate 50 bucks of revenue, and spend all that profit on training.

At the beginning of this year, OpenAI started at 2 gigawatts and Anthropic at less than 2; by the end of the year, they are both above 5. This represents about 30% of the compute added this year, and they will take 40% to 50% of compute next year. A big new entrant, SpaceX, is building a ton of compute and will likely lease much of it to Anthropic and OpenAI because they have the marginal capability to pay the highest price. Additionally, OpenAI is building its own chips, and Anthropic is purchasing TPUs from Google and deploying them with Fluidstack. By the end of next year, half of the incremental compute will go to these two labs.

Within a year and a half or two years, most of the world's compute will be owned by or serving the demand of these two labs. While world compute in gigawatts may double every year, compute at frontier labs triples. Following this trend, it goes from 2 at the start of this year to close to 6 at the end, 18 by the end of 2027, and 54 by the end of 2028. If incremental compute adds 30 gigawatts this year, 50 next year, and roughly 70 the year after, an interesting phenomenon occurs: a new watt deployed this year is significantly more efficient than those from two years ago. I am deploying GB300s, TPUv7s, and Trainium3s, which offer 3-5x more performance per watt than prior-generation chips.

If Anthropic and OpenAI take 45% of compute next year, they will have taken half of the world's incremental new compute by December '27. Because that new compute is at a higher performance than everything before it, if this trend continues—and I see nothing stopping it—they will control most of the usable flops in the world by the end of 2028.

$6 billion in fab capex enables $1t+ of end revenue

I am confused why you think we only add 80 gigawatts in 2028 if the value of compute increases so much. When I interviewed you a few months ago, you said that to make a gigawatt of Vera Rubins, you need 55,000 N3 wafers, 6K N5 wafers, and 170K DRAM wafers.

I had an LLM run your wafer fab equipment model to figure out the annual tooling costs to produce a gigawatt of compute, which it estimated at $3-4 billion. Including cleanrooms and shell, $6 billion of fab CapEx produces a gigawatt every year. Since a gigawatt currently produces $100 billion of revenue, $6 billion in CapEx generates a gigawatt that produces $100 billion annually. Over five years, the first gigawatt generates five years of profits, the second four, and so on. Thus, $6 billion of fab CapEx generates over a trillion dollars of end AI revenue.

There is a 100x discrepancy between fab CapEx and end revenue, meaning you can turn $1 into $100. Yet the emergency is so great that Anthropic and OpenAI feel they could make a trillion dollars right now if they weren't bottlenecked by the mirrors in ASML machines. You have seen people perform funny arbitrages, buying and reselling turbines because their value increases as the bottleneck for data centers.

If anyone had $400 million and could convince ASML to sell them an EUV tool, they should buy one and then sell it for north of a billion dollars.

At Carl Zeiss, they now say they need to make enough mirrors for 100 EUV tools a year by the end of the decade, whereas during our episode earlier this year, they didn't think they needed that many. Currently, 100 ASML tools for 2030 is roughly the right number. But if you gave Carl Zeiss $10 billion to expand production, and did this with every company in the supply chain—perhaps via an AGI-pilled private equity firm maximizing production—that would change things.

Even if the top labs combined generate a trillion dollars in revenue next year, they aren't yet at the point where their cash flows can fund this. Total CapEx next year will be well north of $2 trillion, including $200 billion for the wafer fabrication equipment supply chain, and even more for the data center, accelerator, and energy supply chains.

Compute prices will rise if the labs outbid everyone

They will never get to that point because you want to keep your CapEx higher than your returns. Those gigawatts drive many-fold more throughput and performance by 2028 because hardware has improved; flops per watt have increased and hardware is better at working with AI workloads. This may be difficult given that by 2028 they’ve taken 70-80% of incremental compute. When Amazon serves Bedrock Anthropic models, that counts as Anthropic compute in our worldview because it is effectively revenue for Anthropic, despite revenue shares and credits.

If they reach 100 gigawatts combined by 2028, they will have done disruptive things to the market. Anyone can make money off $10-15 million per megawatt compute today: get a GB300 rack, download Kimi weights and vLLM or SGLang, and set it up with help from Codex and Fable. This has already caused compute pricing to inflect up. To reach 100 gigawatts by 2028, you must believe labs can outpay for compute, as anyone can make money at $10 to $15. If recursive self-improvement occurs—where labs are uplifted by internal models they aren't releasing—this is even more likely. They must start paying $25, $30, or $50 million a megawatt to gobble up 70% of the world’s compute to reach that aggressive 100 gigawatt goal.

The regulation they advocate for slows the labs more than it slows open-source Chinese language models. OpenAI not releasing Astra or stopping training for two weeks, and Anthropic not releasing Model 2 (believed to be the next version of Mythos), shows they aren't releasing their best models. Consequently, their revenue per megawatt stalls or declines as other models become competitive. Their ability to buy incremental compute at a higher price diminishes, potentially preventing them from reaching 100 gigawatts. However, they could generate $100 million per megawatt or more and pay $50 million a megawatt.

As an intuition pump: what if AI models were as good as fully automated software engineers? White-collar workers earn six figures or more. If a gigawatt can sustain roughly a million white-collar workers, that would be $100 billion. With full AGI, it would be many hundreds of billions of dollars per gigawatt.

Which layer will capture most of the surplus?

Most of the value these models generate does not go to OpenAI and Anthropic. Jane Street, with their exclusive contract for GPT-5.6 Ultrafast mode and as one of Anthropic’s biggest customers, generates far more value from tokens than Anthropic generates in profit because they make money off the market. Similarly, Meta, rumored to be as much as 10% of Anthropic’s business, makes more money using these models to optimize ad algorithms and increase engagement time than Anthropic does.

If a million new software engineers appeared, their cost would fall. If the market reaches equilibrium, the price of compute should be close to what Anthropic and OpenAI can generate from it, with a small markup. Right now, it is weird that there is a 4x or more difference between what compute sells for and how much money Anthropic makes from it. If Anthropic’s ability to monetize a gigawatt doubles or triples as revenue per gigawatt increases, it would be weird if that gap continued to grow. Anthropic can take something that cost them $10 and turn it into $100 just by having some weights.

The end user generates the most value, which is why they pay so much. The app layer has generated very little value so far. The model layer, which had negative gross margins a year ago, now has massive positive gross margins and is on a path to generating $100 million per megawatt. A year ago, the hardware supply chain captured all the gross margin while everyone else lost money. OpenAI, Anthropic, and other startups plowed in VC money, and hyperscalers built infrastructure without knowing if there was a payoff. Because tokens were sold for less than the infra cost, the model layer created negative value, and all value was captured at the chip and fab. Initially in 2023, memory providers made no money off HBM or AI memory despite delivering humongous value. Now, TSMC captures far less value than the memory guys. This shifting value capture is exciting for market participants like Jane Street.

I think most compute will continue to transact at sub-$20 billion a gigawatt. If Meta, Microsoft, Amazon, and SpaceX build compute without a customer, they control the situation. Most compute is contracted before it is built, which is what Elon took advantage of. He knew Anthropic was making $60-plus billion per gigawatt, whereas random clouds build compute only after taking a customer's commitment to the credit markets to raise capital.

Meta and SpaceX are the only plausible #3 because they are hoarding compute, using their balance sheets and capabilities to build without an end customer. With their balance sheets, they can access the credit market. Meta and SpaceX now have the optionality to decide if internal use cases make more money or if they should sell to Anthropic or OpenAI at crazy margins. We have entered a regime where SpaceX and Meta can build compute and rent it out for $25, $50, or more, rather than $13.

Will datacenter regulation slow down AI?

I don’t see why it wouldn’t be $50-plus million a megawatt by the end of ’27. It could get higher, perhaps $70 or $80 million a megawatt blended across the company. If I’m Anthropic, incremental compute is worth it; maybe I spend $40 million a megawatt on SpaceX compute. If I’m SpaceX, I look to the supply chain and wonder why Jensen doesn't raise his prices, especially since Elon says they’re exclusive to Nvidia. Then SK Hynix, Micron, and Samsung might raise their prices too.

There is a bullwhip effect with value capture: TSMC raises prices slowly, but memory and substrate companies raise them quickly. Elon wouldn’t have sold if it was $15, but he’s selling because it’s $25+.

I’m surprised you think revenue per gigawatt doesn’t increase way more than 100 per gigawatt by the end of next year. My problem is that the best model in the world was trained in February. OpenAI says they’re not training models for two weeks, and Astra isn’t widely deployed internally. You’re talking about a GPT-4o to Mythos 2-size leap by the end of 2027. We can’t use it to optimize inference performance, but that’s the only way I see it being only $100 million per megawatt by the end of next year. As long as the model gets better, the value generated improves.

Regulation is increasingly about New York banning data centers, Texas holding moratoriums, or Ohio trying to require paying everyone’s property tax in a certain radius. These things will decrease supply and increase cost. In a takeoff scenario, Anthropic would likely have their best model six months ahead of what is externally available. If progress accelerates, that six-month difference is an even bigger differential. As these companies go public and become accountable to investors, they may have close to 20 gigawatts by the end of next year.

Labs are shifting compute from inference to R&D

If 10% of compute is 2 gigawatts, and a lab wants to increase training compute from 60% to 70%, investors might argue they are saying no to $200 billion of revenue, assuming they can generate $100 billion per gigawatt. While most people believe most compute will go to inference, labs will actually allocate less and less to it over time, favoring forward passes for training.

If they generate $30-40 million per megawatt today and allocate 40% to inference, but that figure rises to $60-70 million, Anthropic and OpenAI—including their boards—won't just do dividends and share buybacks; they will build AGI because it is more profitable. They will ratchet up the percentage of compute dedicated to training, even as each increment becomes more profit-generating if dedicated to inference. I would allocate it internally because the value generated from the best AI model is far greater than what an external user provides. While they may not explicitly calculate the incremental progress of turning $100 million per megawatt toward AI research, it makes more sense to dedicate more compute internally.

The only reason to have such large inference compute is to grow the training fleet, which I think they have been doing over the last three months. Every month, Anthropic has added more compute than the prior month; despite some noise from deals like SpaceX, the amount of compute is a curve up. In January, they added less compute than in December, yet their revenue skyrocketed. Since they aren't adding $25 billion of ARR every month, the marginal megawatt is going to R&D at a higher percentage than to inference. They are factually increasing their compute towards R&D today.

China gets less than 10% of new compute, but its labs need less

If you add up the numbers, it would be over 200 gigawatts of world compute by the end of 2028: 30 this year, 50 next year, and 70 in ’28, with ’29 on the order of 90-100. I think the slope can continue upwards, though it is hard to predict more than four years out. We don't know if we are in an RSI regime or when the world economy will grow at 10% a year, because 100+ gigawatts a year implies absurd GDP growth. If RSI kicks off in the West before China has a large amount of compute, we may live in a different world.

In 2022, the US added 45-50% of the world’s compute and China added 30-35%. Since then, big regulations against China and a dramatic increase in America mean 70% of watts are now deployed in America, while China has sub-10% of incremental new compute. China will likely have 30 gigawatts of AI compute or less by 2028, but I expect a big uplift then. In 2026, they still mostly rely on smuggled chips, chips TSMC made for companies that ended up being Huawei, or HBM from Samsung. By ’27 and especially ’28, fabs from SMIC and CXMT will go up, with domestic production reaching millions of units a year. They will incrementally add 5-10 gigawatts of domestically produced chips in 2028 alone.

Those chips will be worse than what Nvidia, Google, or OpenAI will have in 2028. Much depends on whether the US passes the MATCH Act, if tools remain export-controlled, and how fast China builds new domestic equipment. Ultimately, China will hockey stick because they are great at scaling manufacturing quickly. I imagine they will extract more foreign chip purchases or close the gap in what the US allows Nvidia to sell them. It is reasonable that China could do 50 gigs in 2029, but if most are domestic, those 50 gigawatts might only be worth 20 gigawatts of American chips. This projects a world where the leading lab in 2028 has more compute than all of China in ’29 or ’30, if gigawatts are weighted by quality.

Governments and politicians are starting to act, whereas China will not slow down AI; in fact, they will only accelerate it. When I interviewed Jensen about export controls, as a libertarian, I wasn't sure what I thought. I steelmanned the opposite view to hash out ideas, suggesting that cooperating with China might be better since they control so much of the supply chain and robotics needs. However, I didn't realize the compute situation was as dire as you're saying. If the shipping amounts are what you claim, the export controls make a huge difference. By the time we have automated coders and researchers, China will be far behind on compute stock.

Some of this is due to export controls, but some is the financial systems. American systems are more willing to YOLO into startups, but once Chinese systems choose an industry, they subsidize it far more. The Chinese semiconductor industry has significantly more subsidies than the rest of the world's combined. If takeoff takes longer than you imply, China will ultimately catch up drastically on semiconductors and compute.

Chinese companies are not that far behind in AI models relative to their compute. Leading Chinese labs have 100-200 megawatts total, with ByteDance Seed as the outlier; Kimi is not running a gigawatt. Meanwhile, Anthropic will be more than 5 gigawatts by the end of the year. Right now, this difference doesn't matter much. In a lab's compute budget, 50% is research, 10% is development, and 40% is inference. Research involves generating ideas and testing architectures, data mixes, hyperparameters, and attention techniques. When Anthropic trains Mythos, it's sub-200 megawatts for about two months, and RL is even less. While they had multiple gigawatts, most went to research because you cannot always leverage all your compute onto training.

As we move toward automated coding and research, I expect the compute budget for training to increase or the line between research and training to blur. If you reach 100 gigawatts a year, that would be $5 trillion of CapEx annually, not including the need to build power plants well in advance.

Data centers are 15- to 20-year assets. Once you account for future growth, the $5 trillion estimate becomes $7 or $10 trillion of CapEx. When people discuss AI CapEx, they cite $40 or $50 billion, but that only covers critical IT: servers, networking, fiber, transceivers, and optical communications. It ignores the data centers and power plants being built ahead of time. If I'm building 100 gigawatts this year and 150 next year, the buildings for that 150 gigawatts must be built this year. If I'm building 200 gigawatts the year after, I have to buy the turbines now.

Incremental CapEx could plausibly reach $10 trillion by the end of 2030, nearly a tenth of the world economy and a quarter to a third of the US economy. To enable 100 gigawatts by 2030, or even 2028, we are looking at $3 or $4 trillion of CapEx: over $2.5 trillion for IT and another $1 to $2 trillion for data centers, energy, and the downstream supply chain like semiconductors.

It was simple for Google and Meta to stop buybacks and buy computer infrastructure. But by 2028, hyperscalers and their supply chains will be raising hundreds of billions of dollars in debt. Who pays for this? Semiconductor companies like Nvidia and Broadcom and memory companies are funding some CapEx. Traditional infrastructure investors are gathering capital for data centers instead of bridges. Finally, people are realizing they should buy hyperscaler, data center, or Anthropic debt instead of homes, home credit, or government debt. Anthropic is willing to pay 20% rates for an incremental billion dollars to build capacity because their revenue will be huge, and 20% is better than renting from SpaceX for $50 billion a gigawatt. This creates immense contention.

Will AI cause a sovereign debt crisis?

We have been debating whether AI will cause a sovereign debt crisis. It is a huge problem for everyone who cannot turn a little money into a lot of money. The rate of return is incredibly high; if you build a data center and rent it to Anthropic or OpenAI for 10x the depreciated cost, you turn $1 into $2 or $10 by the end of the year. This raises the rate of interest for the entire economy.

As people borrow more, they compete against the lending the government, other companies, or mortgage buyers would have done, making it more expensive for everyone else to borrow. I think the US will be fine because if data centers are built in America, you can fundamentally tax them. However, corporate income is currently less than 10% of federal revenues, while 80%-plus is payroll and income taxes, which will shrink as automation increases. Currently, 20% of tax revenue spending goes toward servicing the debt. Since much of the debt is short duration and rolls over every five years, a 1% rise in interest rates would increase that fraction from 20% to 25%. If it rises 5 percentage points, it goes north of 40%. Accounting for the $2 trillion the government borrows every year, that could exceed 60% of tax revenue just for interest payments.

While the US tax base will increase if we let data centers be built here, other countries are absolutely fucked. Countries like Pakistan or Nigeria, which have high debt, little tax revenue, and debt that is serviced frequently, will suffer in this new interest-rate regime. This crowding-out effect is why it is not YOLO 1 billion gigawatts. Impoverished countries will default, as will debt-heavy industries like consumer packaged goods companies—the ones making things you see at Trader Joe’s—telecoms, and banks. If the spread between the federal rate and what the market charges rises because Amazon wants to raise $100 billion in debt, you face a challenging problem of where the cash comes from. Some is funded by rising cash flows, but the logical move is to invest far more than your cash flows so that future returns are amazing.

Pushing down on the delta are regulations against data centers and AI, consumer and politician anger, and AI labs withholding latest models for safety. Rising interest rates influence all of these, bending the curve from pure capitalism toward a complex system where fewer gigawatts are built than should be.

Our modeling shows about $11 trillion of CapEx from 2024 to 2029. Even funding as much as possible with cash flows, you still need north of $5 trillion in credit. Specifically, $6 trillion is funded with cash and $5 trillion with debt. Raising $5 trillion of debt across the ecosystem makes interest rates go up.

AI revenue will increase, but not forever without hitting constraints. Labs will have incentives and cash flow, though they aren't always the ones building the compute. There is an upper limit on how fast their revenue can grow versus the value they deliver, so they can't cash-flow fund everything. The optimal scenario is using as much credit as possible because you want to build more than cash flows allow. However, $5 trillion of credit and $6 trillion of cash-funded investments through ’29 is not enough compute relative to demand growth from AI models.

One answer is that revenue per megawatt keeps going up, and labs keep inference allocations large, accumulating profit across the S&P 500 as everyone pays to reduce costs. Recently, Meta raised at 5 to 6%. I don't see why they wouldn't happily pay 8% because the return from the compute they build is humongous. The flip side is that a 250 bps increase over today's 5% to 6% makes everyone else in the economy pay 250 bps more. Banks will scream because if their credit spread blows up, their debt reprices faster than their assets, and they ultimately lose tons of money.

If interest rates rise, the discount rate increases, meaning the discounted cash flows of all equities crater. Even if the S&P 500 is fine, individual stocks—especially the stable, Berkshire-type stocks that pay good cash flows for 30 years—will crater in value. Why would I invest that much if my discount rate is 8% or 10% instead of 3% or 5%?

For developing countries, my friend and economist Basil Halperin suggests we'll see a second Volcker shock. In the ’80s, Fed Chair Paul Volcker raised real interest rates to about 8% to fight inflation, causing some 40 countries, mostly in Latin America, to default. I think that will probably happen again.

This all happens before singularity. Interest rates rise 2-3%, and eventually, the world economy will likely double every single year. Researcher Damon Binder has done great work on this. In a fully automated economy, if you are no longer bottlenecked by the fact that you can't double the human population every year, how fast can the economy grow? I think it could double every year, or at least grow by tens of percent.

Since the rate of interest should be similar to the growth rate, we'll enter a world in the 2030s where interest rates are tens, or perhaps hundreds, of percent. Every country not involved in AI production defaults. Every non-AI stock becomes worth basically zero because discounted cash flows are worth nothing. If the federal government can't figure out how to tax AI, servicing the debt will exceed current tax revenue, and you'll have other effects, like being unable to get a mortgage.

We'd enter a totally different growth regime where the opportunity cost of capital increases a ton. The economy is essentially saying the cost of government borrowing to pay pensions is extremely high because that money could instead build a robot factory that builds a robot factory that builds a robot factory.

As interest rates rise, equity markets get pummeled, including AI companies. Some people wonder why Micron, Hynix, or Kioxia trade at 2 or 3 times earnings; if you're truly AI-pilled, everything in the economy should trade at 2 or 3 times earnings. If you're not, then they're over-earning. This is why I think memory stocks won't 10x again. If there is enough demand for memory to cause such a drastic economic change, then everything should trade at 2 or 3x multiples and the stock market should fucking crash.

Meta is a $1.5 trillion company, but logically they are worth way more. Look at their cash flows, the infrastructure they are hoarding, and the compute they can sell for crazy amounts of dollars per watt, either as tokens because their lab works, or to Anthropic and OpenAI.

Ultimately, you have to reallocate all capital to AGI by pricing everyone else out. The limiter on AGI is not how fast research engineers, like our roommate Sholto, can crank the gears, but how much the rest of the world lets that happen. They will regulate, increase interest rates, and forbid data centers and fabs. As company equity values tank, businesses will struggle to pay for AI, forcing Anthropic and OpenAI to design their own chips and build their own infrastructure and data centers in the next couple of years.

I hope a slow takeoff is possible because of the economy and regulatory world. Governments will soon say, "Don't release your models," or "You can't even use your models internally that much." I am most worried about a singularity, which external deployment actually helps prevent. Right now, deployment leads to more revenue because models are incapable of RSI.

I worry about 2030, where the government makes you wait six months before releasing a new model to the public. In those six months, they do recursive self-improvement internally, while the rest of us are stuck with models years behind. Even if the whole world conspires to slow AI by a year, and compute increases 2 to 3x annually, RSI provides 3 to 6 years of progress in a single year.

Governments will not just limit compute, but also the lab's ability to use models internally; for example, Anthropic had to stop giving Mythos to foreign employees for a bit. The US government will not want Anthropic to use Mythos 4 internally. There will be real-world constraints on the progress, development, and deployment of AI. Even though it will happen eventually, we could tear ourselves apart before we get there.

Will the world's future workforce belong to a few companies?

I find it crazy how much of the world’s future labor supply ends up in very few companies and how fast that supply grows. If frontier compute in FLOP terms grows 4-5x a year and the compute required for a level of capabilities decreases 3x a year, the effective AI population size at frontier labs increases 10x year over year. If this trend continues, OpenAI could go from 10 million AI laborers this year to 100 million next year, and a billion the year after. Even if compute scaling slows, it won't take many years before each company individually has more labor equivalence than there are people on Earth; I think that's plausible by the end of this decade.

We don’t think enough about moving into a regime where most "people," in terms of work output, are concentrated within two labs consuming more and more of the world’s compute. If these AIs are misaligned, most of the world is misaligned because most of the world’s minds are there. Even if they aren't, very few companies have immense influence or control. If you believe in RSI and that labs are the most effective users of compute, the only result is centralization of compute.

The effective population of the frontier is currently increasing 10x year over year for a given level of capabilities. If you reach the level of a very competent remote worker, software engineer, or researcher, that population is increasing 10x, 100x, or 1,000x a year at the current rate of growth.

AI training has huge economies of scale because effort spent training a specific skill or knowledge is amortized across billions of sessions or users. Additionally, if you are slightly ahead in the AI race while compute is in shortage, you can charge a higher markup by better economizing this scarce resource. If models learn from deployment, the most widely deployed model gets the most real-world data. Whether it is user deployment and continual learning, economies of scale in training, or RSI where the best AI model helps create the next best model, everything points to centralization.

One of the big intellectual projects I will spend time thinking about is: what is a vision of a decentralized, broadly empowered future after AGI that takes these economies of scale seriously?

The alternative is that the government controls it. I don’t trust the government, and I don’t trust Dario or Sam. Ex ante, it’s hard to see how we avoid choosing one source of centralization. Super-centralized capitalist economies actually grew slower than super-decentralized ones, but AI flips this on its head. Private ownership may not be the most efficient economy and therefore grows slower than a centralized AI economy.

AI is maybe 2% of the economy right now—$1 trillion divided by 30. A large share is happening from very few companies: Nvidia, Anthropic, OpenAI, and the hyperscalers. Unless AI progress slows or governments regulate the fuck out of it, we are headed for a world where either we have super concentration of resources and pray one company gets everything right, or governments and people slow everything down to create a balance of power. Even as we go towards AGI, ASI, and RSI, everything leads to someone capturing more resources. It is hard to find a framework in which AI doesn’t lead to super concentration.

One positive is that today Anthropic does not capture most of the value. They may have gone from $20 million to $100 million per megawatt, but they still pay $13 million for much of the compute they buy. Jane Street is capturing $300 million or $500 million per megawatt, or Dwarkesh is capturing some amount from his podcast and learning about credit.

However, the logic of reallocating inference to AI R&D is that returns to labor inside AI labs are much higher than outside. In 80,000 worlds, in only one does Anthropic not own the whole world. Why would Anthropic allocate compute to Jane Street making money off degenerate options traders if Anthropic can generate hundreds of millions of dollars per megawatt by using that compute internally? That’s what’s happening.

Mentioned

People

Basil Halperin Paul Volcker Damon Binder Gavin Baker

Companies

Anthropic OpenAI SpaceX Google Fluidstack ASML Carl Zeiss Jane Street TSMC Meta Amazon Microsoft Nvidia Micron Samsung SK Hynix SMIC CXMT ByteDance Broadcom Berkshire Hathaway Johnson & Johnson Kioxia