OpenAI and Anthropic's Race to Control Global

Dylan Patel, founder of semiconductor research firm SemiAnalysis, has a habit of making the near future sound both inevitable and slightly terrifying. In a recent conversation with Dwarkesh Patel on the Dwarkesh Podcast, he mapped out a scenario in which OpenAI and Anthropic—two private companies—come to control the majority of the world's usable computing power within the next two to three years. Not a majority of AI compute. A majority of all compute.

The argument is worth sitting with carefully, because it is not primarily about technology. It is about economics, debt markets, and the political economy of infrastructure—territory where the implications extend well past the AI industry's usual concerns.

The Arithmetic of Concentration

Patel's case starts with a simple ratio. At the beginning of this year, OpenAI and Anthropic each commanded roughly two gigawatts of compute capacity. By the end of this year, he estimates both will exceed five gigawatts each. Incremental new compute flowing to the two labs has gone from roughly a third of global additions to somewhere between 40 and 50 percent next year—numbers he says are already contracted and inked.

The mechanism driving this is not market dominance in the traditional sense. It is monetization efficiency. Where a year ago both labs were generating negative gross margins on inference—serving GPT-4 on Nvidia Hopper GPUs cost more than it earned—the economics have inverted sharply. Patel estimates Anthropic's revenue per megawatt of compute has reached roughly $50 million, against a base infrastructure cost of $10–15 million per megawatt. That spread is what lets the labs outbid everyone else for scarce capacity.

"Anyone can make money off of $10 to $15 million per megawatt compute today," Patel notes. "But to get to that 100 gigawatts in 2028, you have to believe that the labs can outpay for compute."

SpaceX's compute business is the clearest illustration of the dynamic. When SpaceX had excess capacity, it found a ready buyer in Anthropic and other frontier labs—because those labs could justify paying $25–40 million per megawatt when competitors could not. That deal, which secured hundreds of thousands of GPUs for Anthropic's use, was not an anomaly. It was price discovery in a market where the buyer with the highest-revenue model wins every auction.

The longer-term projection, if current trends hold: by late 2028, the combined compute of OpenAI and Anthropic—potentially 100 gigawatts—would represent the majority of the world's usable FLOPs, weighted by hardware generation. The newer chips are three to five times more efficient per watt than older generations, so taking 50 percent of incremental new compute actually understates the effective share of processing power.

The Inference-to-Training Pivot

Patel makes a second argument that is less discussed but structurally important: as revenue per megawatt rises, the rational move for labs is to reduce the share of compute dedicated to external inference and redirect it toward internal research and training. He believes this is already happening.

The logic is counterintuitive but coherent. If a gigawatt can generate $100 million in inference revenue, but applying that same gigawatt to training produces a better model that generates $200 million next year, the discounted-cash-flow calculation favors training. For organizations whose explicit mission is building AGI—and whose boards reflect that—the pressure to monetize current capacity competes directly against the pressure to build the next model faster.

This creates a tension that will become acute if either lab goes public. As Dwarkesh Patel puts it: "You're basically saying no to $200 billion of revenue in order to increase your training compute. As investors are like, 'What the f***?'"

Patel's answer is that lab leadership and boards will choose training anyway, and that public market investors will have to live with it—or won't get the chance to object because the labs' governance structures are designed to limit that kind of shareholder pressure. Whether that holds post-IPO is an open question. The monetization pressures building toward IPO are real, and inference revenue is the most legible number on a prospectus.

Regulation as Both Brake and Accelerant

The regulation picture is genuinely complicated, and Patel does not pretend otherwise. Safety-motivated restrictions—labs declining to release their most capable models externally, training pauses, restricted internal deployment—slow the external revenue curve. If Anthropic's best model is withheld from the market, competitors running publicly available weights narrow the gap, and Anthropic's ability to charge a premium per megawatt weakens. A weaker revenue-per-megawatt story means less ability to outbid for compute. The centralization thesis depends, in part, on the labs being allowed to deploy their best work.

But the regulatory interventions at the infrastructure level cut the opposite way. New York has moved to restrict new data center construction. Texas has considered moratoriums. Ohio has contemplated requiring data center operators to fund local property tax relief. These supply constraints don't slow AI development—they raise its cost and that cost gets passed through. Compute becomes more expensive, token prices rise, and the labs with the deepest pockets and highest revenue per unit maintain their purchasing advantage over everyone else.

The irony Patel identifies is precise: "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." The labs lobbied for safety-oriented restrictions and are now discovering that those restrictions cut their own revenue while doing little to constrain competitors who operate outside the regulatory perimeter.

China's compute situation reinforces this asymmetry. Patel estimates China currently receives sub-10 percent of globally deployed AI compute watts, down from 30–35 percent in 2022, largely due to US export controls on advanced semiconductors. Domestic Chinese production from fabs like SMIC is ramping but remains years behind on performance per watt. A leading Chinese lab like Kimi operates at perhaps 100–200 megawatts total—a fraction of Anthropic's current capacity. The export controls appear, by Patel's numbers, to have created a meaningful and widening gap at precisely the moment when the frontier labs are shifting toward compute-intensive recursive self-improvement research.

The Debt Cascade

The more speculative—but not easily dismissed—portion of the conversation concerns what happens to global credit markets when AI infrastructure investment reaches the scales Patel projects. His modeling at SemiAnalysis puts total AI-related capital expenditure from 2024 through 2029 at roughly $11 trillion across the full stack: servers, networking, data centers, power generation, and semiconductor fabrication. Of that, he estimates roughly $5 trillion will need to be debt-financed, because lab revenue—even growing rapidly—cannot fund the entire build-out from cash flows.

Hyperscalers are already borrowing heavily to fund capex. The incremental demand for credit, competing against government debt, mortgages, and corporate bonds across every other sector, puts upward pressure on market interest rates independent of central bank policy. Patel envisions a scenario where Meta and similar companies find themselves paying meaningfully higher borrowing costs than they do today—and that spread ripples through the entire credit market.

An economist Patel cites draws a parallel to the early 1980s Volcker shock, when sharply rising US interest rates triggered debt crises across roughly 40 countries, mostly in Latin America. The mechanism Patel describes is structurally similar: a spike in the opportunity cost of capital, driven not by anti-inflation policy but by the extraordinary returns available in AI infrastructure, reprices debt globally and leaves high-debt, low-revenue countries exposed.

"Every country that is not involved in the production of AI defaults," Patel says, pushing the argument toward its logical extreme. He frames this not as a fringe scenario but as the natural consequence of a sustained, massive shift in where capital earns its highest return.

The US is relatively insulated—it hosts most of the infrastructure and could tax the resulting income. Countries with high debt loads, short-duration obligations, and no position in the AI supply chain are not. This is a policy question that no government appears to be actively modeling at the scale Patel describes, which is itself notable.

The Labor Concentration Problem and What Law Has to Offer

The conversation's closing thread is the one that maps most directly to questions of power and accountability. Dwarkesh Patel observes that if effective AI capability at the frontier is growing 10x per year—a function of both raw compute and improving hardware efficiency—then the "effective AI population" within OpenAI or Anthropic could plausibly exceed the human population of Earth before the end of the decade. Not as a metaphor. As an accounting exercise.

Patel's response is direct: "Every force is screeching towards centralization. And that's scary as hell."

The structural forces he enumerates—economies of scale in training, the monetization premium that flows to whoever has the best model, continual learning advantages from wider deployment, and eventually recursive self-improvement—each independently favor the lab already ahead. Together, they describe something that existing competition law was not designed to address.

This is where the essential facilities doctrine becomes relevant, and where regulators will eventually have to reckon with whether it applies. The doctrine, developed under US antitrust law and referenced in EU competition cases, holds that where a resource is genuinely indispensable to competition and cannot be practically replicated, the controller of that resource can be compelled to provide access on non-discriminatory terms. The canonical examples are railroad networks and telecommunications infrastructure. The question now is whether frontier AI compute—and the training runs that only a handful of organizations can fund—meets that threshold.

My read: the structural conditions Patel describes are closer to the essential facilities paradigm than to conventional monopoly analysis. The barrier is not market share in a product; it is the capital stack required to stay at the frontier at all. If that stack becomes accessible only to two private entities, the Sherman Act's existing tools are probably insufficient, and the EU AI Act's tiered obligations—which attach to general-purpose AI models above certain compute thresholds—represent the more tractable regulatory instrument, however imperfect.

Neither framework was designed for a world where the entity with the most compute also employs the largest effective workforce on Earth. Getting the legal architecture right before that world arrives seems like the more pressing problem than anything currently on the legislative calendar.

Samira Barnes covers technology policy and regulation for Buzzrag.