The Compute Cartography of AI Power

If you map where the world’s AI training compute actually sits — not where the companies are headquartered, not where the models are accessed, but where the physical servers consume electricity and do arithmetic — you get a picture of power that looks nothing like the political map of the world. The territory of AI is not measured in square kilometers. It is measured in petaflops per second and dollars per kilowatt-hour.

In early 2027, the best available estimates suggest that the United States accounts for roughly 55 percent of global AI training compute capacity. China accounts for perhaps 15 to 20 percent, depending on how you count Ascend clusters and the significant uncertainty around unreported domestic deployments. The European Union accounts for perhaps 8 to 10 percent. Everything else is divided among a long tail of countries that are building AI infrastructure with varying degrees of urgency and coherence.

These are rough numbers. The precision is not available because data center capacity is not uniformly reported, because some significant deployments are not publicly disclosed, and because the definition of “AI training compute” involves contested boundary conditions around chip generations and cluster configurations. But the rough picture is stable: the US dominates, China is a distant second, Europe is a more distant third, and there is an enormous gap between the top tier and the rest.

What makes this map politically significant is that it does not merely describe where compute currently exists. It describes where decisions about what AI gets trained, on what data, for what purposes, with what safety constraints, are made. The companies that control the largest compute clusters — Microsoft Azure, Google Cloud, Amazon Web Services in the US; Alibaba Cloud, Tencent Cloud, ByteDance in China — are not merely infrastructure providers. They are the entities that determine which AI development projects are feasible, how much they cost, and who can access the results.

This is a form of sovereignty that political philosophy has not fully processed. Traditional sovereignty is territorial and legal. Countries control what happens within their borders and can (in principle) regulate or prohibit activities that occur on their soil. AI compute sovereignty is infrastructural and economic. Countries control what happens on their compute clusters, can direct that compute toward state-preferred applications, can restrict access to researchers or companies they deem adversarial, and can establish the training data regimes that shape what values and capabilities the resulting models embody.

The European Union has spent enormous legislative energy on AI regulation — the AI Act, passed in 2024, is the most comprehensive attempt by any jurisdiction to regulate AI deployment — but has relatively little control over the training infrastructure that determines what models exist in the first place. European AI regulation is primarily downstream regulation: rules about how AI can be used after it is built, by whom, with what documentation. The upstream question — who builds the models and on what basis — is answered elsewhere, primarily in Northern Virginia and in the data centers that Alibaba operates in Hangzhou.

The Gulf states have emerged as the most surprising entrants in the compute geography. Saudi Arabia’s Public Investment Fund, through its NEOM project and through the SambaNova partnership and the newly established Saudi Data and AI Authority, has made AI infrastructure investment a central pillar of its post-oil economic strategy. The UAE, through G42 (which signed a contentious partnership with Microsoft in 2024 after considerable US government scrutiny) and through its AI agency led by Omar Al Olama, has positioned Abu Dhabi as a compute hub for the Middle East and Africa.

These investments are not primarily about the Gulf states developing indigenous AI research capacity. They are about becoming nodes in the global AI supply chain — data center locations that serve as neutral ground for companies and countries that want AI compute without full alignment with either US or Chinese ecosystems. The political calculation is explicit: in a world divided between American and Chinese AI infrastructure, a Gulf data center that hosts neutral or multi-aligned workloads has commercial value that pure national capacity calculations underestimate.

The US government has watched these developments with the layered concern of an actor that wants its allies to build AI capacity, wants its ally the UAE to not be too close to Chinese AI companies, and understands that the economic incentives driving Gulf AI investment are not fully aligned with American security preferences. The Microsoft-G42 deal was negotiated over more than a year, with US government review of the terms, specifically because Abu Dhabi’s relationship with Chinese technology companies was a documented concern.

The energy dimension of compute geography is less discussed than it should be. AI training is extraordinarily energy-intensive. A large training run for a frontier model consumes electricity roughly equivalent to the annual residential energy consumption of a small city. The geographic distribution of training compute is therefore also a geographic distribution of energy consumption — and the places where cheap, abundant, reliable electricity exists are the places where AI training can most economically occur.

The US advantage in AI compute is partly the advantage of cheap electricity in specific regions (Virginia, Oregon, Texas), reliable grid infrastructure, and the engineering talent that builds and operates data centers. China’s AI data center build-out has been concentrated in Inner Mongolia and Gansu, which have cheap coal power and are sufficiently distant from population centers to absorb the thermal load of large server facilities. Europe’s AI compute ambition has been constrained by electricity costs that are substantially higher than in the US or China, a fact that the EU’s renewable energy transition has improved but not yet resolved.

Countries with abundant renewable energy and low costs — Iceland, Norway, Canada’s Quebec province — have attracted data center investment but lack the broader technology ecosystem and regulatory environment that makes them attractive for frontier AI development specifically. Compute geography is not simply about electricity. It is about the combination of energy, talent, infrastructure, and regulatory climate that frontier AI development requires.

The most consequential developing story in compute geography is the race to build sovereign AI clusters — compute capacity owned or controlled by national governments rather than private companies. The French Sovereign AI initiative, the UK’s AI Safety Institute compute commitment, India’s IndiaAI mission announced in early 2024, and Japan’s GENIAC program are all versions of the same recognition: that depending on private, foreign-controlled infrastructure for the AI systems that your government, military, and critical services will run on is a strategic vulnerability.

This recognition is correct. The inference is being acted on at national expense by governments that are simultaneously trying to attract private AI investment. There is a tension here that has not been resolved: governments want both the economic activity of private AI investment and the control of sovereign AI infrastructure, and the two objectives pull in different directions. Private AI companies do not want government control over their infrastructure. Governments do not want to pay for infrastructure that benefits private companies without appropriate conditions.

India’s IndiaAI mission is the most ambitious attempt by a non-OECD country to stake a position in this geography. The program envisions compute clusters sufficient to train models with capabilities relevant to Indian languages and Indian development priorities, distributed across multiple data center locations, with governance frameworks that keep the infrastructure under Indian institutional control while allowing access to Indian researchers and companies. Whether the execution matches the vision is an open question. The vision itself is correct: a country with 1.4 billion people and 22 constitutionally recognized languages that wants to be an AI power rather than an AI consumer needs to own some part of the stack.

The compute map of 2027 is a first draft of the AI power map of 2035. The countries and companies that understand this early enough to act on it will have positions that are difficult to displace. The ones that treat compute as an infrastructure commodity that can be purchased on demand will discover, when they need it most, that the price of sovereignty rises sharply when you are trying to buy it in a crisis.