Models at or above the EU systemic-risk compute threshold Counted from models with a reported or estimated training-compute figure. Many values are estimates. Notable AI Models Epoch AI Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured.
22
First crossed in 2023. Article 51 sets it at 1e25 FLOP.
Median share of working-age population using AI tools Median across all 147 economies in the dataset. AI Diffusion Report: AI User Share Microsoft AI for Good Lab Derived from anonymised Microsoft telemetry, then adjusted for OS/device market share, internet penetration and a modelled mobile-to-desktop ratio. It therefore measures Microsoft-visible AI use, scaled, not observed total AI use. Systematically less reliable where Microsoft platform share is low, and it cannot see AI embedded in products users do not recognise as AI.
14.8 %
Across 147 economies, 2026-Q1.
Tracked benchmarks scoring above 95 percent A benchmark counts as saturated here when the best recorded score reaches 0.95. AI Benchmarking Hub Epoch AI Mixes Epoch's own evaluation runs with externally reported scores; these are not strictly comparable. Benchmark questions themselves remain the property of their creators and are not redistributed here. External rows retain their original licensing.
6 of 17
A saturated benchmark stops distinguishing between models.
Most recent model with a published compute figure Notable AI Models Epoch AI Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured.
GLM-5.3-Flash
Z.ai (Zhipu AI), 2026-08-20, 3.2e24 FLOP.
Documented power capacity across all reported clusters Power is summed over the subset of clusters with a published capacity figure. GPU Clusters Epoch AI Compiled from public reporting; each row carries a certainty flag ranging from confirmed to speculative. Chinese clusters are systematically less well documented than US ones, so country totals understate opacity as much as they measure capacity.
2,596 MW
466 clusters across 36 countries have a location on record. Power is summed only where it is published.
United Arab Emirates sits at 70.1%, about 12 times Cambodia. The median economy is at 14.8%.
| Economy | Share | Relative magnitude |
|---|---|---|
| United Arab Emirates | 70.1% | |
| Singapore | 63.4% | |
| Norway | 48.6% | |
| Ireland | 48.4% | |
| France | 47.8% | |
| Spain | 44.2% | |
| New Zealand | 43.0% | |
| United Kingdom | 42.2% | |
| Netherlands | 42.1% | |
| Qatar | 41.8% | |
| Australia | 39.5% | |
| Belgium | 39.0% | |
| Israel | 38.1% | |
| Switzerland | 37.8% | |
| Canada | 37.3% | |
| South Korea | 37.1% |
Estimated from anonymised Microsoft telemetry, adjusted for device and internet penetration. It measures Microsoft-visible use scaled by a model, not observed total AI use, and it cannot see AI embedded in products people do not recognise as AI. One signal, not a measurement.
Best score recorded on each benchmark by any evaluated model, as a running maximum. A benchmark at 100 percent has stopped telling you which model is better.
OTIS Mock AIME 2024-2025
100.0 %
Saturated
FrontierMath-Tier-4-2025-07-01-Public
100.0 %
Saturated
FrontierMath-2025-02-28-Public
100.0 %
Saturated
MATH level 5
98.1 %
Saturated
FrontierMath-Tier-4-v2-Private
97.6 %
Saturated
GPQA diamond
95.8 %
Saturated
FrontierMath-Tiers-1-3-v2-Private
93.7 %
Still discriminating
Mystery Game Puzzles
84.0 %
Still discriminating
Mixes evaluation runs performed by Epoch AI with externally reported scores, which are not strictly comparable. The line shows the best available capability at each point in time, not the typical one. AI Benchmarking Hub Epoch AI Mixes Epoch's own evaluation runs with externally reported scores; these are not strictly comparable. Benchmark questions themselves remain the property of their creators and are not redistributed here. External rows retain their original licensing.
The EU AI Act presumes a general-purpose model carries systemic risk once cumulative training compute passes 1e25 FLOP. It is the only compute threshold currently written into binding law anywhere, which is why it is the only one plotted. Of the 537 models here with a compute figure, 22 cross it.
Compute figures are frequently estimates rather than disclosures. Treat borderline cases as indicative, never as a legal finding about any particular model.
China has the most reported clusters. The United States has most of the reported power. That difference measures disclosure as much as capacity.
| Country | Clusters | Documented power | With known capacity | Share of clusters |
|---|---|---|---|---|
| China | 188 | 296.9 MW | 140 of 188 | |
| United States of America | 119 | 1,962.8 MW | 117 of 119 | |
| Japan | 25 | 43.1 MW | 25 of 25 | |
| France | 14 | 33.2 MW | 14 of 14 | |
| South Korea | 12 | 12.7 MW | 10 of 12 | |
| Germany | 11 | 25.2 MW | 11 of 11 | |
| Italy | 9 | 39.9 MW | 9 of 9 | |
| Brazil | 9 | 7.3 MW | 9 of 9 | |
| Russia | 8 | 5.7 MW | 7 of 8 | |
| United Kingdom | 6 | 12 MW | 6 of 6 |
China reports 188 clusters at 296.9 MW documented, against 119 clusters at 1,962.8 MW for the United States. Capacity is published for only a minority of Chinese clusters, so the gap shown here is partly a reporting gap. GPU Clusters Epoch AI Compiled from public reporting; each row carries a certainty flag ranging from confirmed to speculative. Chinese clusters are systematically less well documented than US ones, so country totals understate opacity as much as they measure capacity.
From 3 jurisdictions via official public-domain and open-licence channels. Headline and a short extract only, always linking back to the original.
17 November 2026 The EGG (Brussels) The Apply AI Summit will gather high-level stakeholders - from industrial representatives, policymakers and startups - to accelerate AI implementation in...
European Commission Shaping Europe's digital future European Commission Covers all EU digital policy, not only AI; requires topical filtering. Reuse decision does not extend to third-party material or EU logos.
UKHSA study finds sleep app cough data can detect respiratory illness and signal outbreaks early
GOV.UK GOV.UK news and communications UK Government Keyword-filtered search feed, so relevance varies. Departmental press releases are political communications, not neutral descriptions of policy.
£100 million fund competition will help promising British AI start-ups start in the UK and tackle some of society's biggest challenges
GOV.UK GOV.UK news and communications UK Government Keyword-filtered search feed, so relevance varies. Departmental press releases are political communications, not neutral descriptions of policy.
16 September 2026 The State of the Union (SOTEU) address is an annual speech by the President of the European Commission to the European Parliament, setting the priorities for the Union for the...
European Commission Shaping Europe's digital future European Commission Covers all EU digital policy, not only AI; requires topical filtering. Reuse decision does not extend to third-party material or EU logos.
The Commission has designated ChatGPT as a Very Large Online Search Engine (VLOSE), as well as Reddit and Roblox as Very Large Online Platforms (VLOPs), under the Digital Services Act (DSA). These...
European Commission Shaping Europe's digital future European Commission Covers all EU digital policy, not only AI; requires topical filtering. Reuse decision does not extend to third-party material or EU logos.
20 October 2026 Watch the pitches of the top 10 finalists of the Apply AI Startup Award and learn more about the European AI innovation ecosystem. The Apply AI Startup Award recognises innovative...
European Commission Shaping Europe's digital future European Commission Covers all EU digital policy, not only AI; requires topical filtering. Reuse decision does not extend to third-party material or EU logos.
The NOAA Science Advisory Board (SAB) is the only Federal Advisory Committee with responsibility to advise the Under Secretary of Commerce for Oceans and Atmosphere on long- and short-range...
Commerce Department Federal Register API US Office of the Federal Register Full-text keyword matching on "artificial intelligence" returns documents that merely mention AI in passing. Requires topical filtering before display.
Notice is hereby given, pursuant to the provisions of the rules and regulations of the U.S. Commission on Civil Rights (Commission) and the Federal Advisory Committee Act, that the Maryland...
Civil Rights Commission Federal Register API US Office of the Federal Register Full-text keyword matching on "artificial intelligence" returns documents that merely mention AI in passing. Requires topical filtering before display.
Attribution is structural here, not editorial discipline. It is the one architectural rule the whole project rests on.
Every source is declared once with its licence, attribution, cadence and known limitations.
Each processed figure carries a source id, so attribution is a join rather than a habit.
The pipeline refuses to write data whose licence forbids republication. The site refuses to render it.
Any number opens its own source, licence and retrieval date on click.
Two sources are registered specifically so that using them fails the build: Artificial Analysis, whose free tier forbids redistribution, and the Stanford AI Index, which is licensed CC BY-NC-ND and therefore cannot be re-plotted. Both are linked instead.
No scraping, no paid APIs, no unlicensed reuse. Each of these publishes openly and is credited under its own terms.
Models, benchmarks, clusters, hardware
CC BY 4.0AI diffusion by economy
MITRules, notices, executive actions
Public domainUK government announcements
OGL v3.0EU digital and AI policy
Decision 2011/833/EUUS standards and evaluation work
Public domain