Where the World’s AI Runs: US vs China vs Europe

Frontier AI compute sits inside a handful of American data centers. China’s fast-growing domestic build and Europe’s publicly funded push compete for what is left.

Global AI compute distribution refers to how training and inference hardware for large AI models is physically located across the United States, China, and the European Union, and how that location determines legal jurisdiction, chip access, and power availability. That concentration decides more than bragging rights. It sets where an enterprise can train a large model. Location also sets where the training data may legally sit, and how long a team waits for capacity. Grid limits and chip export rules now belong in the architecture diagram, right beside residency law. This piece maps where the compute physically sits, why it landed there, and what a team deploying this quarter should do about it.

Which Region Holds the Most AI Compute?

Epoch AI’s dataset of AI supercomputers ranks the United States first by tracked computing performance. China ranks second. No single European country registers as a standalone top-tier bloc in the same data.

That ranking explains most of what follows. Compute is concentrated, and the concentration gets harder to reverse each year, because every new tier of cluster costs more than the last.

Enterprises tend to read this as a procurement story. In practice, it works more like an engineering constraint. A workload pinned to one jurisdiction inherits that jurisdiction’s chip supply, power availability, and legal exposure. Therefore the region choice happens before the GPU choice, whether or not the team notices it happening.

The three blocs are also not converging. Each grew from a different funding model, and each hit a different wall. The United States ran on private capital and hit the electrical grid. China ran on domestic substitution and hit an export-control ceiling at the frontier tier. Europe ran on public co-financing and hit the calendar. Knowing which wall applies to your workload matters more than knowing which region is nominally ahead.

Why the United States Leads on Raw Compute

American dominance came from a private capital cycle that outran government research computing years ago.

According to Epoch AI’s data, leading industry systems grew in performance by 2.7x annually between 2019 and 2025. Public sector systems managed only 1.9x annually across the same period. Compounded over six years, that difference is enormous. Industry’s share of aggregate AI supercomputer performance in the dataset climbed from about 40% in 2019 to about 80% in 2025, while the public sector share fell below 20%, per the same Epoch AI figures.

What the underlying research measures

The dataset comes from the paper “Trends in AI Supercomputers,” written by Konstantin Pilz, Robi Rahman, James Sanders, Luke Emberson, and Lennart Heim. It tracks over 500 AI supercomputers built between 2019 and 2025. The authors find that training compute for the most notable AI models has grown by 4.1x per year since 2010.

Sit with that number for a moment. A 4.1x annual rate means the compute behind a leading training run roughly doubles every six months. No university budget cycle moves that fast. As a result, hyperscalers and a small set of AI labs now set the pace, and national labs follow.

Power sets the near-term ceiling

The private-capital advantage carries a physical bill. The International Energy Agency projects that US data centers will consume more electricity for processing data than the entire energy-intensive US manufacturing sector by 2030. That comparison covers aluminum, steel, cement and chemicals production combined.

Interconnection queues and substation capacity now govern how fast new American capacity arrives. Local siting approvals matter just as much. Our field notes on data center energy constraints trace how those timelines decide which sites can host a cluster at all.

China’s Path: Domestic Chips Under Constraint

China reached second place while operating under US export controls designed to slow its access to leading-edge accelerators.

The Center for Security and Emerging Technology examined that dynamic in “Pushing the Limits: Huawei’s AI Chip Tests U.S. Export Controls,” by Jacob Feldgoise and Hanna Dohmen. Their report shows how Huawei’s Ascend 910B chip advances on its predecessor, and where export controls appear to slow that progress. The finding cuts both ways. Huawei is closing ground, yet it does so with friction on every step.

RAND’s August 2025 commentary, “Leashing Chinese AI Needs Smart Chip Controls,” approaches the same question from policy design. It argues that a workable threshold should keep Chinese developers dependent on hardware they cannot yet fully replace domestically. That framing is useful for engineers, not only for policymakers. China holds substantial aggregate capacity. However, it stays constrained specifically at the frontier tier.

How the constraint changes the engineering

Constraints of this kind push labs toward efficiency rather than scale. Chinese teams adopted domestic accelerators years before those parts matched NVIDIA’s top offerings. Consequently, they invest heavily in kernel-level optimization and communication scheduling, and they run larger chip counts to reach a given effective throughput.

Casual regional comparisons miss this. Aggregate cluster count and manufacturing capacity are genuinely strong. Meanwhile, the frontier gap measured against the highest-performance training runs stays wide in Epoch AI’s country-level data. Enterprises weighing multi-region deployment should treat “China compute” and “frontier compute” as two separate questions with two separate answers.

Europe’s Bet: Public Money, Sovereign Control

Europe is funding a small number of large public facilities organized around data sovereignty rather than raw scale.

The European Commission committed to its InvestAI initiative in February 2025. The European Investment Bank confirmed that the Commission pledged €20 billion via InvestAI to establish up to five AI gigafactories across the EU. A later call widened the ambition. The EU’s digital strategy office confirmed a tender for up to seven AI Gigafactories, backed by up to €10 billion in EU and national funding, and expected to unlock at least €20 billion more in private investment.

STL Partners, analyzing the same program, is blunt about the scale gap. It notes that €20 billion is substantial by European public-policy standards but modest next to the AI infrastructure investment US hyperscalers already deploy.

Demand, however, is not the problem. Deep Tech’s coverage reports the initiative received 76 proposals from 16 member states in a June 2025 call, far exceeding initial expectations. That response signals real appetite for sovereign capacity, even where the funded scale trails US private investment by an order of magnitude.

The regulatory layer nobody else carries

The EU AI Act adds a second dimension the other two blocs do not share in the same form. Article 10 requires data governance documentation for high-risk AI systems, with enforcement starting in August 2026, and penalties reaching into the tens of millions of euros.

For infrastructure teams, this converts a legal question into a design question. Lineage tracking and retention boundaries must exist at the storage and orchestration layer, not only in a policy document. Our breakdown of the EU AI Act compliance countdown sets out which of those controls have to be in place before the deadline, apart from the model-governance rules most coverage focuses on.

The Three-Region Comparison

Sources: Epoch AI, CSET, RAND, European Investment Bank, European Commission digital strategy office, STL Partners.

Named Systems: What the Frontier Gap Looks Like in Practice

Named systems make the gap concrete in a way ranking tables cannot.

As of May 2025, Epoch AI notes that Lawrence Livermore’s El Capitan, the largest known public AI supercomputer, achieves less than a quarter of the computational performance of xAI’s Colossus, the largest known industry cluster. One comparison captures the public-private divide better than any aggregate share. Government labs still build serious machines. Private clusters, meanwhile, now operate at a different order of magnitude.

China’s frontier systems do not appear at the top of Epoch AI’s public rankings in the same way. Export controls cap the highest-end accelerators available to Chinese buyers, which explains part of the absence. Absence from a leaderboard still leaves plenty of working capacity in place. It means the largest Chinese clusters draw from a narrower slice of accelerator generations. Teams in that environment compensate with chip count and software tuning rather than per-chip performance.

Europe’s timing problem

Europe has no named system competing with El Capitan or Colossus on the same leaderboard today. Its gigafactory tenders remain in the proposal and site-selection phase as of the initiative’s 2025 launch.

That timing gap deserves attention. A European enterprise weighing sovereign compute right now may be choosing infrastructure that has not finished construction. Elsewhere, comparable infrastructure has been running production workloads for a year or more. A site still in tender cannot run a training job next quarter. Therefore procurement teams should treat announced capacity and available capacity as separate line items.

What This Means for Data Residency Decisions

Choosing a region means choosing a risk profile, and the three profiles differ sharply.

US-based compute offers the deepest hardware selection and the fastest path to new silicon. Yet it carries growing exposure to power-constrained siting, as data centers compete with heavy manufacturing for grid capacity. China-based compute offers scale beneath a lower frontier-chip ceiling, which limits any workload that needs the newest accelerator generation. EU-based compute offers the clearest legal footing for data that must stay under EU jurisdiction. That footing comes with a smaller and younger buildout still working through its first tenders.

Sovereign AI mandates are pushing more governments toward the European model: public capacity paired with private operators who move quickly inside a fixed jurisdiction. Our guide to sovereign AI infrastructure shows where that pattern has already spread past Europe and what it changes for operators.

Geography now sits alongside the hardware question rather than behind it. A workload that must stay in-region for legal reasons has answered half its infrastructure question before anyone opens a spec sheet.

How Enterprises Actually Use This Map

Few enterprises pick one region and stay there. Instead, teams split workloads by function and route each piece to the region that suits it.

Model pretraining demands the largest contiguous clusters and the newest accelerator generations. Therefore it still points toward US capacity. Fine-tuning and inference carry tighter latency budgets and stricter data-handling rules tied to end users. Those workloads usually belong closer to the users, which increasingly means EU-based capacity for European customers.

Cloud teams have split workloads this way for years. AI sharpens the tradeoff, because the compute itself is scarce and the regulatory stakes are higher. Our comparison of training versus inference infrastructure sets out the memory, interconnect, and latency requirements that pull the two workload types apart. A single-region plan usually breaks at the first inference deployment outside the home jurisdiction.

The export-control wrinkle

Companies operating across US and Chinese jurisdictions face a specific problem. Running inference for Chinese users on Chinese soil means operating under China’s chip-availability ceiling, whatever the parent company can access elsewhere.

That ceiling shapes model choice as much as infrastructure choice. A model trained on frontier-tier US compute may not run efficiently on the accelerator generation available domestically. As a result, enterprises building for multiple jurisdictions increasingly design for the lowest common hardware denominator across their footprint, then optimize upward where local capacity permits.

What to Watch Over the Next 18 Months

Two open bets will settle much of this map. One is whether Europe’s gigafactories close real ground at frontier scale once the first sites commission. The other is whether China’s substitution curve bends fast enough to matter at the top tier. Neither outcome needs to resolve before a team commits this quarter. Match the workload to the constraint you can live with, then stand up the second region before the first one becomes a dependency.

Reserve capacity at portal.axecompute.com, or contact us at info@axecompute.com to plan a multi-region footprint.

About Axe Compute

Axe Compute Inc. (NASDAQ: AGPU) is a neocloud AI infrastructure platform built on a fundamental premise: AI innovation should not be constrained by hardware choice or inventory limitations. Axe Compute gives enterprises and AI innovators choice across hardware, geography, and deployment speed through two delivery models: Axe Compute Access, providing the latest GPU compute options in as fast as 48 hours across numerous global locations, and Axe Compute Build, enabling enterprises to access large-scale dedicated AI factories, all backed by enterprise-grade SLAs and support.

Frequently Asked Questions

Which country has the most AI compute today?

The United States holds the largest share of frontier AI compute, owning about 75% of total computing power among AI supercomputers tracked by Epoch AI. China follows in second place at roughly 15%.

Why does China lag in frontier AI compute despite its manufacturing scale?

US export controls limit China’s access to the highest-performance AI chips, which caps the frontier tier specifically rather than overall compute capacity. Chinese labs compensate with larger chip counts and domestic accelerators, but the top-tier performance gap remains wide.

What is the EU’s InvestAI initiative?

InvestAI is a European Commission program that commits public funding, alongside expected private investment, to build a small number of large AI gigafactories across the EU. It is designed to expand sovereign compute capacity rather than match US hyperscaler scale.

Does data residency law affect where AI workloads should run?

Yes. The EU AI Act requires data governance documentation for high-risk AI systems, with enforcement starting in August 2026, which affects infrastructure choices for any workload processing EU data.

Should an enterprise pick one region for all AI workloads?

Most enterprises split workloads by function instead. Training often favors US capacity for scale and chip access, while inference and fine-tuning often move closer to end users for latency and data residency reasons.

What is limiting further AI compute growth in the United States?

Grid power and data center siting are becoming the binding constraint, according to the International Energy Agency, which projects US data centers will consume more electricity than the country’s entire energy-intensive manufacturing sector by 2030.

Is Europe’s AI compute buildout operational yet?

As of the InvestAI program’s 2025 launch, most EU gigafactory sites remain in the proposal and site-selection phase, meaning much of the announced capacity has not yet been built.

About Axe Compute

Sources

• Konstantin Pilz, Robi Rahman, James Sanders, Luke Emberson, Lennart Heim, “Trends in AI Supercomputers”, Epoch AI / arXiv
• Konstantin Pilz, Robi Rahman, James Sanders, Luke Emberson, Lennart Heim, “Trends in AI Supercomputers”, arXiv
• Konstantin Pilz, Robi Rahman, James Sanders, Luke Emberson, Lennart Heim, “Private-sector companies own a dominant share of GPU clusters”, Epoch AI
• Jacob Feldgoise, Hanna Dohmen, “Pushing the Limits: Huawei’s AI Chip Tests U.S. Export Controls”, Center for Security and Emerging Technology (Georgetown)
• RAND Corporation, “Leashing Chinese AI Needs Smart Chip Controls”, RAND
• STL Partners, “The EU’s AI Gigafactory Initiative: What it means for digital infrastructure?”, STL Partners
• Deep Tech, “The €37 Billion Race: EU AI Gigafactories Europe and the Sovereignty Bet”, deeptech.build
• European Investment Bank, “EIB Group and European Commission join forces to finance AI gigafactories”, EIB
• European Commission, “EU launches AI Gigafactories call to boost Europe’s computing capacity”, Digital Strategy, European Commission
• International Energy Agency, “AI is set to drive surging electricity demand from data centres”, IEA