The Compute Race

US, China, the Gulf and Europe · Flagship: who holds the compute

AI compute has become a strategic asset that governments race to build, deny and export. The US holds most of the world's AI supercomputer capacity, China plans a roughly fourfold to sixfold increase in intelligent compute by 2030 with a push toward domestic chips, and Washington has borrowed the language of the Manhattan Project for its national AI effort.

The US-China compute race in two measures: Epoch AI estimates the US held about 75% of tracked GPU-cluster performance in May 2025 against about 15% for China; the IEA puts US data centers at about 45% of global data center electricity in 2024 against about 25% for China and 15% for Europe; and the Manhattan Project framing, from the 2024 congressional commission recommendation to the 2025 Genesis Mission WHO HOLDS THE COMPUTE AI supercomputer performance, May 2025 (Epoch AI) US 75% China 15% Data center electricity, 2024 (IEA) US 45% China 25% Europe 15% Other 15% Not comparable: China reports 2,185 FP16 EFLOPS (June 2026), a different unit. THE MANHATTAN PROJECT FRAMING Jun 2024Situational Awareness essay: "The Project" by 2027/28 Nov 2024Congressional commission: fund a Manhattan Project for AGI Mar 2025Superintelligence Strategy argues against it; proposes deterrence Nov 2025Genesis Mission order cites the Manhattan Project
US cluster share
~75%
China share
~15%
China 2030 target
9,800 EFLOPS
Genesis Mission
November 2025
Hub
Society

What it is

Compute as a Strategic Asset

Governments now treat AI compute the way they once treated steel, oil or enriched uranium: something to build at home, keep from rivals and trade for influence. Data centers are where that contest is physically fought.

  • US: America's AI Action Plan of July 2025 set out to win the race through permitting, federal sites for data centers and exports of the full American AI stack. Export controls on advanced chips have swung from bans to licensed sales with a revenue share to the US government.
  • China: East Data West Computing moves compute to cheap western power, state-funded data centers must use domestic chips, and Huawei's Atlas SuperPoDs trade weaker single chips for very large systems. China leads in open-weight model downloads.
  • Gulf: the UAE and Saudi Arabia are investing sovereign capital and cheap power, under US chip-export approvals.
  • Europe: the EU is funding up to seven AI gigafactories; the UK plans AI Growth Zones toward 6 GW by 2030.

The race framing has its critics. The authors of Superintelligence Strategy, including former Google CEO Eric Schmidt, argue a US push for exclusive control invites rivals to sabotage it, and propose deterrence instead. Chinese officials describe US export controls as containment. See The Pace Debate.

Flagship

US and China: Who Holds the Compute

  • Clusters: Epoch AI estimates the US held about 75% of tracked AI supercomputer performance in May 2025 and China about 15%. The dataset covers only known clusters, an estimated 10-20% of the global total.
  • Electricity: the IEA puts US data centers at about 45% of global data center electricity in 2024, China about 25% and Europe about 15%.
  • China's own count: MIIT reported 2,185 EFLOPS of intelligent compute at FP16 in June 2026, up 177% in a year, and targets 9,800 EFLOPS by 2030. The units differ from Epoch's, so the figures cannot be compared directly.
  • Chips: US officials estimated Huawei's 2025 Ascend output at no more than 200,000 chips; Nvidia says its share of China's AI chip market has fallen to near zero as Beijing steers buyers to domestic chips.

Programs

National Programs and Controls

National AI compute programs and controls
ProgramWhatDate
USCC annual reportTop recommendation: a Manhattan Project-like program to race to AGINovember 2024
America's AI Action PlanThree pillars; executive orders on data center permitting and AI exportsJuly 2025
Genesis Mission, EO 14363DOE AI-for-science platform, "comparable in urgency" to the Manhattan Project; partner credits of $2.4B announced October 2026November 2025
Stargate, US sitesMore than 9 GW planned at seven sites; about 0.3 GW operating at Abilene (Epoch AI)April 2026
Export controls on ChinaH20 licensed with 15% revenue share; H200 case-by-case with a 25% tariff; Blackwell and Rubin excluded2025-2026
East Data West ComputingEight national hubs and 10 clusters steering compute to cheap western powerSince 2022
MIIT compute plan to 20309,800 EFLOPS intelligent compute target, up from 1,590 in 2025; domestic chips favoredSeptember 2026
Domestic chip ruleNew state-funded data centers must use Chinese AI chips (Reuters)November 2025
Huawei Atlas 950 SuperPoD8,192 Ascend chips per system; Huawei admits a single-chip gap with NvidiaQ4 2026
Chinese open-weight modelsAbout 41% of 2025 Hugging Face downloads; Qwen has 113,000+ derivativesMarch 2026
Stargate UAE and HumainUS approved up to 35,000 GB300-class chips each for G42 and HumainNovember 2025
EU AI gigafactoriesUp to €10B public funds and at least €20B expected private investment for up to seven gigafactories; call closes 12 November 20262025-2026

Record

Dataset Record

Dataset
US, China, Gulf and EU compute programs, export controls and capacity shares
Sources
USCC, White House, DOE, Nextgov, Epoch AI, IEA, BIS via Crowell and Mayer Brown, MIIT, State Council, Reuters, Financial Times, Huawei, Hugging Face, DatacenterDynamics, EuroHPC, Euronews, CSET, TechCrunch
Last verified
October 2026