Why two AI labs could soon control most of the world’s compute

Over the next few years, the center of gravity of the global economy is likely to shift toward a handful of AI labs. As AI models get better and the value of compute explodes, more of the world’s capital, energy, and talent will be redirected into building and operating data centers for a tiny number of frontier systems.

This isn’t just about bigger GPUs or faster models. It’s about a world where two labs could end up controlling most of the usable compute on Earth—and with it, a growing share of the world’s effective workforce.

AI labs are becoming the engine of global investment

In the last year, a surprising share of US GDP growth has come from AI infrastructure: chips, data centers, and power for training and running large models. That trend is accelerating fast.

Global tech and AI-related capital expenditure (CapEx) is already above $1 trillion per year. By around 2028, projections suggest it could exceed $2 trillion annually, with AI labs taking an ever-larger slice of that pie.

Two companies in particular—OpenAI and Anthropic—are at the center of this shift. They’ve gone from burning venture capital on unprofitable models to running highly profitable AI services, especially on their latest generations like GPT-5.6 and Anthropic’s newest Claude-class models.

From losing money to minting cash per megawatt

To understand how this works, it helps to think in terms of megawatts and gigawatts of compute. A megawatt here roughly corresponds to the power needed to run a rack of high-end AI accelerators (like Nvidia GB300s or Google TPUv7s) at full tilt.

Today, building AI compute capacity typically costs around $10–15 million per megawatt, once you include servers, networking, and supporting infrastructure. A year or two ago, serving frontier models on this hardware often generated negative gross margins for the labs.

That has flipped. For the newest models, Anthropic is reportedly generating up to $50 million in annual revenue per megawatt of inference capacity. In other words, spending $10–15 million on compute can now bring in several times that in yearly revenue. The excess can then be plowed back into training even larger and more capable models.

This dynamic—turning every dollar of compute into multiple dollars of revenue—is what’s driving the massive build-out of AI infrastructure.

How two labs are swallowing the world’s new compute

The most dramatic shift is in who gets the incremental compute that comes online each year. At the start of 2024, OpenAI had around 2 gigawatts of compute capacity, and Anthropic slightly less. By the end of the year, both are expected to be above 5 gigawatts—a roughly 3–4x increase.

When you look at new capacity being added, the numbers are even more striking:

• In 2024, roughly 30% of all new compute deployed globally is expected to end up serving OpenAI and Anthropic.

• Based on signed and planned deals, that share could rise to 40–50% in 2025.

Because the total amount of compute is growing so quickly, the “incremental” capacity being added each year will soon dominate the installed base. If half of that incremental capacity is effectively controlled by two labs, they rapidly approach a position where they own or direct most of the usable AI compute in the world.

Why newer compute is worth more than old compute

Not all gigawatts are equal. Each new generation of AI chips delivers far more performance per watt than the previous one. Modern accelerators like Nvidia GB300, Google TPUv7, or Amazon Trainium3 can be 3–5x more efficient than their predecessors.

That means a watt deployed in 2028 could be several times more powerful than a watt deployed in 2024. If frontier labs are capturing 40–50% of the new, most efficient compute each year, they don’t just get a big share of the total power—they get a disproportionate share of the best power.

On current trends, by the late 2020s, OpenAI and Anthropic together could control most of the world’s effective AI FLOPs, even if they don’t literally own most of the physical servers.

Who actually builds all this compute?

The labs don’t build everything themselves. Much of the capacity is financed and constructed by cloud providers and other large players, then leased back to the labs.

Key actors include:

• Hyperscalers like Microsoft, Google, Amazon, and Meta, who fund and operate huge data centers and often run the labs’ models as managed services.

• SpaceX, which is emerging as a major new compute builder, constructing large GPU clusters and leasing them to labs at premium prices.

• Specialized data center operators and infrastructure investors, who build facilities and power capacity and then sign long-term contracts with AI tenants.

At the same time, labs are starting to vertically integrate. OpenAI is working on its own chips; Anthropic is buying TPUs from Google and deploying them with partners like Fluidstack. Over time, this could give them more direct control over their compute destiny.

Why compute prices are likely to soar

Right now, it’s still possible for many players to make money at $10–15 million per megawatt of compute. You can buy a rack of GB300s, download strong open-source models, wrap them with efficient inference frameworks like vLLM or SGLang, and sell access via marketplaces like OpenRouter.

But if frontier labs can earn $50 million or more per megawatt, they can afford to pay far higher prices than everyone else. To capture 70% or more of the world’s compute by 2028, they may need to bid up prices to $25, $30, or even $50 million per megawatt.

As those prices rise, the entire supply chain will try to capture more of the value: chipmakers, memory vendors, substrate manufacturers, data center builders, and power producers. There’s a “bullwhip effect” where higher end-user willingness to pay slowly ripples back upstream, raising costs everywhere.

The physical bottlenecks: fabs, mirrors, and power

Even if the economics are wildly attractive, the physical world moves slowly. Building more compute requires:

• More advanced semiconductor fabs with extreme ultraviolet (EUV) lithography tools.

• More high-bandwidth memory (HBM) production.

• More data centers, cooling systems, and transmission lines.

• More power plants and turbines to feed those data centers.

Just one gigawatt of cutting-edge AI compute can require tens of thousands of advanced wafers and huge amounts of DRAM capacity every year. Building enough EUV tools, mirrors, and supporting equipment is a multi-year effort. Even if you handed key suppliers like ASML or Carl Zeiss an extra $10 billion tomorrow, they couldn’t instantly double output.

So while capitalism will push hard to expand supply, there’s an unavoidable lag between rising demand and the physical build-out of fabs and infrastructure.

Regulation is already slowing the frontier labs

One surprising twist is that regulation and self-imposed safety policies are currently slowing down the frontier labs more than they slow down open-source or foreign competitors.

Examples include:

• Delays or partial releases of top models (e.g., internal systems like Astra or unreleased Mythos checkpoints).

• Temporary pauses in training runs due to regulatory or safety reviews.

• Restrictions on who can access the most capable internal models, including foreign employees.

If labs are prevented from deploying their best models externally—or even internally—their revenue per megawatt grows more slowly. That, in turn, limits how much they can outbid others for compute and slows the centralization of power.

This dynamic also connects to broader worries about control and alignment. For a deeper dive on why advanced AI may be hard to control, see this discussion of superintelligent AI and control problems.

Training vs inference: why labs may starve their own customers

Most people assume that as AI demand grows, most compute will go toward inference—serving users, running agents, and powering apps. But there’s a strong argument that labs will increasingly prioritize training and research instead.

Today, a rough breakdown for a big lab might look like:

• ~50% of compute on research experiments (trying new architectures, data mixes, and techniques).

• ~10% on the big development runs that create new frontier models.

• ~40% on inference (serving users and customers).

As models get more profitable, labs face a choice. They can:

• Allocate more compute to inference and harvest huge short-term profits, or

• Divert more compute into training and research to push toward AGI and even higher long-term returns.

Given the stakes, it’s likely that leading labs will gradually reduce the share of compute devoted to inference, even as each megawatt of inference becomes more lucrative. The logic is simple: if using that compute internally for AI R&D creates more future value than selling tokens to customers, the rational move is to keep more of it in-house.

How big could global AI compute get?

On current projections, the world could be adding something like:

• ~30 gigawatts of new AI compute in 2024

• ~50 gigawatts in 2025

• ~70 gigawatts in 2026–2028

That would put total global AI compute north of 200 gigawatts by 2028. If the trend continues, annual additions could reach 90–100 gigawatts or more by the end of the decade.

At current cost levels, building 100 gigawatts of cutting-edge compute in a year implies around $5 trillion in IT CapEx alone. Once you include data centers and power plants—assets that must be built years in advance—the total annual CapEx could approach $7–10 trillion by 2030, or close to 10% of global GDP.

It’s not obvious that capital markets, politics, and local communities will tolerate that level of reallocation toward data centers and fabs, especially as they start to crowd out other investments.

China’s delayed but powerful AI ramp-up

Export controls and financial differences have put China at a significant disadvantage in AI compute. In 2022, China accounted for roughly 30–35% of new global data center compute. Today, it’s under 10% for AI-specific deployments, while the US and close allies account for about 70%.

By 2028, China may still have only around 30 gigawatts of AI compute, much of it based on less advanced domestic chips. In effective performance terms, that could be equivalent to far less than 30 gigawatts of Western hardware.

However, China is extremely good at scaling manufacturing once it commits. As domestic fabs like SMIC and memory makers like CXMT ramp up, and as more smuggled or gray-market chips find their way into Chinese data centers, a sharp hockey-stick in Chinese AI capacity is plausible around 2028–2029.

In a slower-takeoff world where AI progress stretches out over more years, China has more time to catch up. In a fast-takeoff world, frontier Western labs might reach very advanced capabilities before China can match their compute stock.

The coming interest rate shock

All this AI CapEx has macro consequences. When you can reliably turn $1 of investment into $10 or $100 of AI-driven revenue, the natural pressure is to borrow and build as fast as possible. That pushes up the demand for credit across the entire economy.

If hyperscalers, labs, and infrastructure players collectively need to raise trillions of dollars in debt between now and 2029, interest rates on corporate borrowing will rise. Even a 200–300 basis point increase (for example, from 5–6% to 8%) would have huge knock-on effects:

• Non-AI companies that rely on cheap debt—telecoms, utilities, consumer goods, banks—face much higher financing costs.

• Developing countries with large dollar debts could be pushed into default, echoing past debt crises.

• Higher discount rates crush the present value of long-dated cash flows, slashing equity valuations for traditional “steady” businesses.

Some economists have compared this to a potential “second Volcker shock,” where rising real interest rates triggered a wave of sovereign defaults. The difference this time is that the driver isn’t inflation-fighting policy—it’s the extraordinary returns available from AI infrastructure.

When AI becomes most of the world’s workforce

Right now, AI systems are far from being full replacements for white-collar workers. But if we assume that:

• Compute at the frontier grows 3–5x per year, and

• The compute needed to reach a given capability level falls 3x per year,

then the effective AI “population” at a given capability level grows roughly 10x per year.

Once models can do the work of a competent remote worker, software engineer, or researcher, this becomes staggering. A single lab could go from the equivalent of 10 million AI workers one year, to 100 million the next, to a billion shortly after—even before you hit full-blown recursive self-improvement.

On current trajectories, it’s plausible that by the end of the decade, a single frontier lab could have more effective AI labor than the entire human population, at least for certain categories of cognitive work.

This is where concerns about control and centralization become existential. If most of the world’s “minds” are concentrated inside a few labs, any misalignment or misuse of those systems affects almost everything. For more on how this kind of runaway capability can slip out of human control, see this overview of recent AI breakthroughs and control challenges.

Centralization vs decentralization in an AI-first economy

Classical capitalism works well partly because it decentralizes decision-making. Many firms compete, no single actor controls everything, and markets allocate resources through prices.

AI upends this logic. Training frontier models has huge economies of scale: once you’ve paid the cost to train a powerful model, you can amortize that cost across billions of users and tasks. If you’re slightly ahead in capabilities and can better monetize scarce compute, you can outbid everyone else and pull even further ahead.

There are also feedback loops: models that are deployed more widely can learn more from real-world data, and better models can help design and train even better successors. All of these forces push toward centralization.

That leaves us with a stark choice:

• Allow a small number of private labs to centralize most of the world’s effective labor and intelligence, and hope they remain aligned and benevolent, or

• Use regulation and public control to slow and shape AI deployment, at the cost of delaying or diffusing some of the economic gains.

In practice, we’re likely to see a messy mix of both: intense centralization of compute and capabilities inside a few labs, combined with growing political pressure to slow them down, restrict deployment, and redistribute value.

What this means for the next decade

Putting it all together, the next five to ten years could look like this:

• Global AI CapEx climbs into the trillions per year, absorbing a rising share of world investment.

• OpenAI and Anthropic capture 40–50% of new frontier compute by the mid-2020s and potentially control most of the world’s usable AI FLOPs by the late 2020s.

• Compute prices rise sharply as labs outbid other users, and the entire supply chain re-prices around AI demand.

• Interest rates drift higher as trillions in new debt are issued to fund data centers, fabs, and power plants.

• Regulation and politics increasingly constrain how fast labs can deploy and use their most advanced models.

• The effective AI workforce inside a few labs grows exponentially, eventually rivaling or surpassing the human workforce in many domains.

Whether this story ends in a broadly shared prosperity, a highly centralized AI oligopoly, or something stranger will depend on choices we make now—about regulation, taxation, openness, and how we want AI to be integrated into the global economy.

What’s clear is that AI is no longer just a technology story. It’s rapidly becoming the main driver of where capital flows, how fast the world grows, and who holds real power in the 21st century.