Global · AI compute race
The AI race: US vs China vs Europe across the stack, and where Ireland sits
The US leads on compute, chips and frontier models: about three-quarters of GPU-cluster compute (Epoch AI, May 2025), 45% of world data-centre electricity (IEA, 2024) and 19 of 25 models above 10^25 FLOP (Epoch). China has 15% of compute, 5 such models and leads routed usage (59% of OpenRouter tokens). Europe has 15% of data-centre electricity, about 8% of compute and none. Ireland's data centres use 23% of its electricity (CSO, 2025).
Updated 12 October 2026 · data build 2026-10-10 · how we know · corrections
The US leads compute, chips, models and money; China leads usage; Europe trails both
Rows are layers of the AI stack; every cell links to its source and date. Europe = EU-27 plus UK, Norway, Switzerland and Iceland unless the source defines it otherwise. Reported
– means no comparable figure we could source. Ireland is part of Europe's totals.
Download CSV ↓Source: IEA Energy and AI · Epoch AI, GPU clusters, AI data centres, chip owners, models (CC BY 4.0) · Stanford AI Index · CSO · SEC EDGAR · as of 2026-10-11 · Compiled by StackPath
Shares side by side: Europe has power demand, not compute
Europe's share of data-centre electricity is far larger than its share of AI compute, chips or model usage. Measured
Data-centre electricity, 2024 (IEA) · share of 415 TWh
GPU-cluster compute (Epoch data, our sum) · 460 clusters operating to Jul 2025
Large AI data-centre power tracked (Epoch) · 93 sites; the rest are elsewhere (mainly Malaysia)
AI-chip capacity by designer's home (Epoch) · cumulative to Dec 2025
Tokens routed on OpenRouter, 30 days · the rest is undisclosed, other or unclassified
Source: IEA · Epoch AI, data (CC BY 4.0) · OpenRouter (CC BY 4.0) · as of 2026-10-11 · Compiled by StackPath
Notable AI models per year: the US far ahead, Europe in single digits
Models Epoch AI counts as notable, by the bloc of the developer (a joint model counts for each bloc). Measured
Source: Epoch AI, notable models (CC BY 4.0) · as of 2026-10-12 · Stanford AI Index 2025 counted 40 US, 15 Chinese and 3 European notable models in 2024 · Compiled by StackPath
Europe's large AI sites in Epoch's tracker: 243 MW across 7
Sites Epoch tracks by satellite and permits, current power. Not a census. Measured
| Site | Country | Current MW | Owner |
|---|---|---|---|
| Google Waltham Cross | United Kingdom | 88 | |
| Nebius Mantsala | Finland | 75 | Nebius |
| CoreWeave Norway | Norway | 42 | CoreWeave |
| Start Campus Sines Data Campus | Portugal | 33 | Nscale |
| Nscale Keflavik | Iceland | 5 | Nscale |
| EcoDataCenter 2 | Sweden | 0 | Mistral AI |
| Microsoft Narvik Norway | Norway | 0 | Nscale |
None in Ireland. For comparison, the US total in the same tracker is 11,851 MW.
Download CSV ↓Source: Epoch AI, AI data centres (CC BY 4.0) · as of 2026-10-11 · Compiled by StackPath
Our reading. The race is not close on the layers that decide who trains frontier models: compute, chips and lab funding. Europe's strengths in this table are demand-side (data-centre load and spend through US clouds) and public money (gigafactories), which arrive after the private US build-out. The counter-argument, that sovereignty spending is protectionism with little payoff, is tested on the Europe scorecard; this page only measures position.
Epoch's GPU-cluster data covers clusters operating to mid-2025 and an estimated 10–20% of world GPU-cluster performance; Chinese entries are partly anonymised, so China's share is understated in our sum. Epoch's own published shares (US about 75%, China 15%) are shown alongside. Epoch's AI data-centre database tracks the largest AI sites (mostly in the US) by satellite and permits; it is not a census, and a site missing from it is not proof there is none. Chip-owner figures are Epoch model estimates (median) of cumulative AI-chip capacity in H100-equivalents; owners outside the named hyperscalers and China are pooled as 'Other', so Europe can't be split out. IEA figures are 2024 estimates and 2030 base-case projections; Europe is the IEA's own regional definition. Hyperscaler capex for Chinese companies is not in this dataset; see hyperscaler capex.
Questions people ask
- Who is winning the AI race, the US or China?
- On the numbers here, the US: about three-quarters of GPU-cluster compute against China's 15% (Epoch AI, May 2025), 19 of the models above 10^25 FLOP in Epoch's data against 5 for China, and $285.9bn of private AI investment in 2025 against $12.4bn (AI Index 2026). Chinese developers lead on tokens routed through OpenRouter (58.8% in the last 30 days).
- Where does Europe stand in the AI race?
- Behind on compute, chips and models. Europe uses 15% of world data-centre electricity (IEA, 2024) but holds about 8% of tracked GPU-cluster compute, designs none of the AI chips in Epoch's capacity data, and has 0 models above 10^25 FLOP with published estimates.
- How much AI compute does Europe have?
- About 8.3% of the GPU-cluster compute in Epoch AI's data (clusters operating to mid-2025), led by Germany, Norway, France and Switzerland. Epoch's AI data-centre tracker lists 243 MW of large AI sites in Europe against 11,851 MW in the US.
- How much electricity do data centres use in the US, China and Europe?
- The IEA estimates data centres used about 415 TWh worldwide in 2024, 1.5% of world electricity: 45% in the US, 25% in China and 15% in Europe. It projects about 945 TWh by 2030.
- How does Ireland fit into the AI race?
- Ireland hosts many cloud data centres, which used 7.7 TWh or 23.2% of its metered electricity in 2025 (CSO). It has no GPU clusters or large AI sites in Epoch's trackers and no frontier lab HQ, and has committed €10m to a French-led AI gigafactory bid.