Europe · Europe vs the US
Europe vs the US in AI: level on using it, far behind on building it
Europe has about one-eleventh of America's AI compute: 6.6% of world GPU-cluster performance against 74.5% (Epoch AI, May 2025). The gap is smallest where AI is used: 37% of EU firms use generative AI against 36% in the US (EIB, 2025). It is widest where AI is made: 2 notable European models in 2025 against 64, and no European AI-chip designer. Big users pay 1.46× US power prices (Eurostat, EIA, 2025).
Updated 12 October 2026 · data build 2026-10-10 · how we know · corrections
Europe has about 1/11 of America's AI compute
Share of the world's GPU-cluster computing power, May 2025. EU-27 (4.8%) plus Norway (1.8%), as Epoch publishes them; the UK, Switzerland and Iceland sit in Epoch's 'Other'. Reported
Europe uses a third as much data-centre electricity as the US, but holds about 1/11 of its AI computing power. It has firms that use AI at the same rate as American ones, and almost none of the chips, models or capital that produce it. Our own sums of Epoch's cluster file put Europe at 9.8% or 8.3% rather than 6.6%, depending on which countries and clusters count; on any method the ratio is between 1 : 7 and 1 : 11.
Source: Epoch AI, AI supercomputer performance share by country, 2025-06-05 · as of 2026-10-12 · Compiled by StackPath
Layer by layer: Europe trails on 5 of 11 headline measures by more than 4 to 1
One headline measure per layer of the AI stack. If the US = 100, how much does Europe have? Widest gap first. Calculated
| Layer | Europe | US | China | Ireland | Europe ÷ US | Go deeper |
|---|---|---|---|---|---|---|
| Chips cumulative to 2025-12-31 | 0% | 96.4% | 3.6% | No Irish designer in the data | 0 | European AI chip startups · EU Chips Act tracker · AI supply chain |
| Compute May 2025 | 6.6% | 74.5% | 14.1% | 0: no Irish cluster in Epoch's data | 1 : 11 | Europe's AI compute · AI race: US vs China vs Europe · AI compute added each year |
| Data-centre capacity 2024 | 15% | 45% | 25% | 7.7 TWh, 23.2% of metered electricity (CSO, 2025) | 0.33× | Data-centre electricity in Europe · US data-centre construction |
| Power price 2025 (annual); not like-for-like | €0.111 | €0.076 | – | €0.176 | 1.46× | Electricity prices for data centres |
| Grid access Europe Aug–Oct 2026 · US Jun 2026; not like-for-like | about 255 GW (9 countries) | more than 1,000 GW (large loads) | – | about 6 GW of interest; 2 GW contracted | 0.26× | Grid queues in Europe |
| Capex and investment latest 10-K or 10-Q (2025-12-31 to 2026-06-30); not like-for-like | ≤$296bn (all non-US) | $772bn | – | not named in any 10-K since Microsoft's FY2023 ($16.4bn) | 0.38× | Hyperscaler capex · Pledges vs builds in Europe · Who pays for AI |
| Neoclouds latest filing or release, Dec 2025 to Sep 2026; not like-for-like | 251 MW (Nebius 170, Nscale 81) | 2,500 MW (CoreWeave 1,500, Crusoe 1,000) | – | no neocloud site | 1 : 10 | European neoclouds |
| Labs and models 2025 | 2 models | 64 models | 35 models | – | 1 : 32 | AI model rankings · European AI labs · AI lab economics |
| Funding to Oct 2026; not like-for-like | $14.7bn | $359.1bn | $8.7bn | no lab headquartered in Ireland | 1 : 24 | European AI labs · AI lab economics · Who pays for AI |
| Adoption 2025 survey | 37% | 36% | – | – | 1.03× | AI adoption: US vs Europe |
| Policy and public money EU: 18 decisions to Jul 2026 · US: 49 projects to Apr 2026; not like-for-like | €14.8bn approved (18 decisions; €0.9bn in a plant producing today) | €28.1bn ($31.5bn awarded, 49 projects; $13.1bn disbursed by Apr 2026) | – | Intel Leixlip and Analog Devices Limerick: aid amounts not disclosed | 0.53× | EU Chips Act tracker |
– means no comparable figure we could source. Ireland is part of Europe's totals. Full sources and every supporting row are in the table at the foot of the page and the CSV.
Download CSV ↓Source: Epoch AI, AI chip owners (CC BY 4.0), 2026-10-11 · Epoch AI, AI supercomputer performance share by country, 2025-06-05 · IEA, Energy and AI, 2025-04 · Eurostat nrg_pc_205, 2026-10-09 · EIA Electric Power Monthly, table 5.3, 2026-09 · StackPath, data-centre grid queues in Europe, 2026-10-11 · SemiAnalysis, Stop Saying Half of 2026 US Datacenter Capacity Is Canceled (free section), 2026-06-18 · SEC EDGAR 10-K filings, 2026-10-11 · Nscale S-1 (18 Sep 2026), 2026-09-18 · StackPath neocloud dataset (Nebius, CoreWeave, Crusoe filings and releases), 2026-10-11 · Epoch AI, notable AI models (CC BY 4.0), 2026-10-11 · Epoch AI, AI companies (CC BY 4.0), company releases and press, 2026-10-11 · EIB Investment Survey 2025 (press release), 2025-10-14 · European Commission, Chips Act (18 first-of-a-kind decisions), 2026-09-11 · as of 2026-10-12 · 41 rows; 40 computed from StackPath datasets · Compiled by StackPath
Where Europe is ahead or close: 7 measures, most of them about using AI
Rows where Europe's figure is at least three-quarters of America's (for prices: at most a third higher), plus the one chokepoint Europe owns. Read the caveats: several of these flatter Europe. Calculated
Share of TOP500 supercomputer performance (Rmax)
0.79× Europe ÷ US
Europe 29.6% (155 systems) · US 37.5% (161 systems) · China 12.7% · Ireland 2 systems, best #381
2026/06 list. TOP500 counts systems whose owners submit a benchmark, mostly public science machines. Most private AI clusters never submit, so this flatters Europe's public HPC.
Cheapest European large-user power price vs the US industrial average
0.53× Europe ÷ US · not like-for-like
Europe €0.041 (Norway) · US €0.076 · Ireland €0.159
Europe 2025-S2 · US 2025. Norway, France and Sweden paid less than the US industrial average. Europe's dearest, Poland, paid €0.172: the average hides a 4× spread.
Generating capacity added to the grid in a year (all sources, net of closures)
1.88× Europe ÷ US
Europe 74 GW · US 39 GW · China 533 GW · Ireland 2 GW
Europe 2024/2025 · US 2025 · Ireland 2024. Nameplate capacity, mostly solar and wind in Europe, so it overstates firm power for data centres. Europe's grid is growing; its connections for large loads are not keeping pace.
Price to rent the same GPU machine in a European cloud region, vs the cheapest US region
1.3× Europe ÷ US
Europe +30% (median of 42 European region prices at AWS, Azure, CoreWeave) · US cheapest US region = 1 · Ireland +20% (Azure H200, Dublin); AWS lists no B200, B300, H100, H200 in Dublin
list prices, 2026-10-11. CoreWeave prices Europe and North America the same; the premium is at the US hyperscalers.
Firms using generative AI (same survey on both sides)
1.03× Europe ÷ US
Europe 37% · US 36%
2025 survey. The only like-for-like adoption measure we found. Depth differs: 81% of US AI-using firms use it in more than two activities, against 55% in the EU.
Firms using AI, official statistics
0.84× Europe ÷ US · not like-for-like
Europe 20% · US 23.8% · Ireland 19.6% (15 of 28 European countries)
EU 2025 · US 2026-09-07 to 2026-09-20. Not like-for-like. Eurostat leaves out firms with fewer than 10 staff, which use AI less, so it flatters the EU. The US question was reworded in Nov 2025: the old wording last read 10.0%, the new one first read 17.3%.
Makers of EUV lithography machines (the tools that print leading-edge chips)
Europe only no US maker
Europe 1 (ASML, Netherlands) · US none · China none
2026. Europe owns one chokepoint outright: ASML makes the EUV machines every leading-edge chip needs. ASML describes EUV as technology unique to ASML. A company's statement, not an independent market count. ASML guided about 60 low-NA EUV tools for 2026 (Q2 2026 call, unchecked here).
Source: TOP500, June 2026, 2026-06 · Eurostat nrg_pc_205, 2026-10-09 · EIA Electric Power Monthly, table 5.3, 2026-09 · Ember yearly electricity data (CC BY 4.0), 2026-10-11 · AWS price files, Azure Retail Prices API, CoreWeave pricing, 2026-10-11 · EIB Investment Survey 2025 (press release), 2025-10-14 · Eurostat isoc_eb_ai, 2026-06-15 · US Census Bureau, Business Trends and Outlook Survey, 2026-10-08 · ASML, EUV lithography systems, 2026-10-11 · as of 2026-10-12 · Compiled by StackPath
Where Europe is far behind: 13 measures below 1 to 4
Sorted from the widest gap. Chips, compute, models and money. Calculated
| Measure | Europe | US | China | Europe ÷ US | When |
|---|---|---|---|---|---|
| Share of the world's installed AI-chip capacity, by the home of the chip's designer | 0% | 96.4% | 3.6% | 0 | cumulative to 2025-12-31 |
| Models trained with more than 10^25 FLOP (the EU AI Act's systemic-risk line), published estimates | 0 models | 19 models | 5 models | 0 | Epoch pull, Oct 2026 |
| Annualised revenue (run-rate) of the leading AI labs | $1bn (Mistral AI) | $125.24bn (OpenAI, Anthropic, xAI) | $4.4bn (4 labs) | 1 : 125 | latest run-rate, mid to late 2026 |
| Share of tokens routed through OpenRouter, last 30 days | 0.26% | 29.2% | 58.8% | 1 : 112 | 30 days to 2026-10-11 |
| Power of the large AI data centres Epoch tracks, at full planned build-out | 850 MW (1.7%) | 43,410 MW (88.6% of the tracked total) | 1,940 MW | 1 : 51 | Epoch pull, 2026-10-11 |
| Power of the large AI data centres Epoch tracks, operating now | 310 MW | 15,180 MW | 850 MW | 1 : 49 | Epoch pull, 2026-10-11 |
| Private money into AI companies in 2025 (not like-for-like) | over €5.3bn (tech.eu count) | $285.9bn | $12.4bn | 1 : 48 | 2025 |
| Notable AI models released in the year, by developer's home | 2 models | 64 models | 35 models | 1 : 32 | 2025 |
| Compute capacity the labs have contracted (GW), Europe's lab vs OpenAI and Anthropic | 1.4 GW (Mistral AI, 2 deals, $0.83bn) | 36.3 GW (OpenAI 25.5, Anthropic 10.8; $955bn) | – | 1 : 26 | deals to Oct 2026 |
| Total equity raised by AI labs: 10 European labs vs OpenAI, Anthropic and xAI (not like-for-like) | $14.7bn | $359.1bn | $8.7bn | 1 : 24 | to Oct 2026 |
| Share of the world's GPU-cluster computing power | 6.6% | 74.5% | 14.1% | 1 : 11 | May 2025 |
| Live power at the two largest neoclouds headquartered in each region, company-wide (not like-for-like) | 251 MW (Nebius 170, Nscale 81) | 2,500 MW (CoreWeave 1,500, Crusoe 1,000) | – | 1 : 10 | latest filing or release, Dec 2025 to Sep 2026 |
| AI-chip startup funding, January to mid-April 2026 | $0.8bn | $4.7bn | – | 1 : 6 | Jan to mid-Apr 2026 |
Source: Epoch AI, AI chip owners (CC BY 4.0), 2026-10-11 · Epoch AI, AI models (CC BY 4.0), 2026-10-11 · Epoch AI, AI companies (CC BY 4.0), company statements and press, 2026-10-11 · OpenRouter rankings (CC BY 4.0), 2026-10-11 · Epoch AI, data centres and timelines (CC BY 4.0), 2026-10-11 · Stanford HAI, AI Index 2026, 2026 · tech.eu, the AI companies that raised the most in 2025, 2026-05-21 · Epoch AI, notable AI models (CC BY 4.0), 2026-10-11 · StackPath compute-commitments ledger (SEC filings and company releases), 2026-10-11 · Epoch AI, AI supercomputer performance share by country, 2025-06-05 · Nscale S-1 (18 Sep 2026), 2026-09-18 · StackPath neocloud dataset (Nebius, CoreWeave, Crusoe filings and releases), 2026-10-11 · as of 2026-10-12 · Compiled by StackPath
China is ahead of Europe on most of these too: 35 notable models in 2025 against Europe's 2, 58.8% of OpenRouter tokens against 0.26%, and 3.6% of AI-chip capacity by designer against zero.
What the gap is made of: it widens from using AI to making it
The same ratios, ordered from demand (firms using AI) through power and compute to models and chips. Editorial judgement
Source: EIB Investment Survey 2025 (press release), 2025-10-14 · Anthropic Economic Index (Hugging Face), 2026-06-26 · IEA, Energy and AI, 2025-04 · Eurostat nrg_pc_205, 2026-10-09 · EIA Electric Power Monthly, table 5.3, 2026-09 · TOP500, June 2026, 2026-06 · European Commission, Chips Act (18 first-of-a-kind decisions), 2026-09-11 · GAO-26-109121, Semiconductors (Aug 2026), 2026-08 · Epoch AI, AI supercomputer performance share by country, 2025-06-05 · Nscale S-1 (18 Sep 2026), 2026-09-18 · StackPath neocloud dataset (Nebius, CoreWeave, Crusoe filings and releases), 2026-10-11 · Epoch AI, AI companies (CC BY 4.0), company releases and press, 2026-10-11 · as of 2026-10-12 · Compiled by StackPath
Our reading. Europe is a customer of the AI stack, not a producer. Its firms adopt at about the US rate, and it consumes a third as much data-centre power. But the further a layer is from the end user, the wider the gap: 1/11 of the compute, 1/10 of the live neocloud capacity, 1/24 of the lab funding, 1/32 of the notable models, 1/125 of the lab revenue, and none of the AI chips. Three things sit underneath: capital (US labs have raised 24× what ten European labs have), power that costs 1.46× as much and connects slowly, and the absence of a European hyperscaler, so most of the spend that lands in Europe is booked by US companies whose long-lived assets sit 72% in the US (the four that disclose).
Ireland against both: the most data-centre-heavy grid, no AI compute
Every row with an Irish figure. Ireland is counted inside Europe's totals. Measured
Source: Epoch AI, AI supercomputer performance share by country, 2025-06-05 · Epoch AI, data centres and timelines (CC BY 4.0), 2026-10-11 · Ember, Grids for data centres in Europe, 2025-06-19 · US DOE / LBNL, 2024 US Data Center Energy Usage Report, 2024-12-20 · CSO, Data centres metered electricity consumption 2025, 2026-07-07 · Eurostat nrg_pc_205, 2026-10-09 · EIA Electric Power Monthly, table 5.3, 2026-09 · SEMO, T-4 2029/30 auction report, 2026-04-14 · PJM, 2028/29 BRA results, 2026-07-14 · StackPath, data-centre grid queues in Europe, 2026-10-11 · as of 2026-10-12 · Compiled by StackPath
An outside view: the US queue is bigger, and also mostly paper
Paraphrased from the free section of a SemiAnalysis post; their models and paywalled numbers are not used. Editorial judgement
All 41 comparisons, with sources
Headline and supporting rows by layer. 'Fallback' marks a row read from the hand-checked file because its source dataset wasn't published. Reported
Download CSV ↓Source: Eurostat isoc_eb_ai, isoc_ai_iaiu · US Census BTOS · US Census C30 · StackPath Europe and global datasets · as of 2026-10-12 · Compiled by StackPath
Ratios are Europe divided by the US. Where lower is better (power prices, Arena rank), a ratio above 1 means Europe is worse. Rows marked not like-for-like compare different measures, scopes or years; their ratios are indicative only. Europe's scope varies by source (EU-27, Epoch's EU-27 plus Norway, IEA's Europe); each row names it. Values come from other StackPath datasets where they are published; otherwise from the hand-checked research file, flagged as a fallback. Currency: US dollars converted at the ECB rate of 2026-10-09 (1.1206 $/€) for chip aid, and at the ECB 2025 average for 2025 money.
Questions people ask
- How far behind the US is Europe in AI?
- It depends on the layer. Europe has about 1/11 of America's GPU-cluster compute (6.6% vs 74.5% of the world, Epoch AI, May 2025), 2 notable AI models in 2025 against 64 (Epoch), and labs that have raised $14.7bn against $359bn for OpenAI, Anthropic and xAI. But 37% of EU firms use generative AI against 36% in the US (EIB, 2025).
- Does Europe have more AI compute than the US?
- No. Epoch AI put the EU-27 at 4.8% and Norway at 1.8% of world GPU-cluster performance in May 2025, against 74.5% for the US. In Epoch's tracker of large AI data centres, Europe has 310 MW operating against 15,180 MW in the US. Europe does better on the TOP500 list of public supercomputers (29.6% vs 37.5%).
- Is electricity more expensive for data centres in Europe than in the US?
- On average, yes. The EU-27's largest industrial users paid €0.111/kWh in 2025 (Eurostat band IG, excluding VAT and recoverable taxes), 1.46× the US industrial average of €0.076 (EIA). Ireland paid €0.176. Norway, France, Sweden paid less than the US average (latest half-year).
- How much electricity do data centres use in Europe compared with the US?
- The IEA estimates Europe used 15% of the world's data-centre electricity in 2024 and the US 45% (of about 415 TWh). It expects the US to add about 240 TWh by 2030 and Europe more than 45 TWh. Data centres were 4.4% of US electricity in 2023 (DOE/LBNL) and about 3% of Europe's in 2024 (Ember); in Ireland they took 23.2% in 2025 (CSO).
- Do European companies use AI as much as American ones?
- About as many do: 37% of EU firms used generative AI in the EIB's 2025 survey against 36% of US firms. US firms use it more deeply: 81% of AI-using US firms apply it to more than two activities, against 55% in the EU. Official statistics differ in method: 19.95% of EU firms with 10+ staff (Eurostat, 2025) and 23.8% of US employer firms (Census BTOS, Sep 2026).