Adoption / 2025-H12025-H22026-Q12026-Q2

The numbers behind AI

A live index of AI capability, governance, impact and safety.

19 datasets 23 sources updated today

Read the figures Charts

Divergence map

Ranked 25 25.9% South Korea, 2025-H1.

Ranked 18 30.7% South Korea, up 4.8 points since 2025-H1.

Ranked 16 37.1% South Korea, up 11.2 points since 2025-H1.

Ranked 12 40.6% South Korea, up 14.7 points since 2025-H1.

Highest 59.4% United Arab Emirates, 2025-H1.

Highest 64.0% United Arab Emirates, up 4.6 points since 2025-H1.

Highest 70.1% United Arab Emirates, up 10.7 points since 2025-H1.

Highest 73.3% United Arab Emirates, up 13.9 points since 2025-H1.

Ranked 11 34.5% Australia, 2025-H1.

Ranked 11 36.9% Australia, up 2.4 points since 2025-H1.

Ranked 11 39.5% Australia, up 5.0 points since 2025-H1.

Ranked 11 40.7% Australia, up 6.2 points since 2025-H1.

2025-H1 2025-H2 2026-Q1 2026-Q2 2026-Q3 not yet published

under 10% 10 to 20 20 to 30 30 to 40 40% and over no data

Find a country

7 / 18 benchmarks solved past 90%

24 models at or above 1e25 FLOP

14 / 100 safety works per hundred on frontier models

147 / 147 economies where use rose since 2025-H1

48 / 86 policy instruments that are binding law

58 jurisdictions scored for readiness

01 / Progress

7 of 18 tests are finished

A benchmark counts as solved once the best score recorded by any evaluated model passes 90%. At that point it has stopped telling models apart.

7 of 18

Scores run to 2026-10. 11 tests still separate the leading models.

FrontierMath-Erdos, the hardest test still open 4%

The best score recorded by any evaluated model.

Explore progress

Data: Epoch AI, AI Benchmarking Hub

02 / Capability

The numbers written into law

Training compute is the one part of AI regulation that is arithmetic rather than judgement. These are the models that cross each threshold.

9.4e26 GPT-6 Astra, the largest disclosed training run 1

1e26 Californian threshold 4

1e25 EU systemic-risk threshold 24

Below Every other model with a reported or estimated figure 516

540 models with a compute figure 3 thresholds in force 2,596 MW documented cluster power

Explore capability

Data: Epoch AI, Notable AI Models. Compute figures are frequently estimates.

03 / Datacentres

14.2 GW operating, 38.2 GW planned

IT power at the 95 AI datacentres Epoch AI tracks, in 12 countries. Planned is Epoch's projection from permits and plans, to 2030.

May 2028 Microsoft Fairwater Wisconsin, chips owned by Microsoft, used by OpenAI (likely) and Microsoft (likely). 369 MW operating now. 2.3 GW

Mar 2028 Anthropic-Amazon New Carlisle, chips owned by Amazon, used by Anthropic. 1 GW operating now. 1.9 GW

Dec 2028 OpenAI Stargate New Mexico, chips owned by Oracle, used by OpenAI. Under construction. 1.8 GW

Jan 2028 Meta Hyperion, chips owned by Meta, used by Meta. Under construction. 1.6 GW

74 of 95 sites operating 62% of operating power where a cloud provider owns the chips 15% of all datacentre electricity on accelerated servers, IEA, 2024

Explore datacentres

Data: Epoch AI, AI Data Centers. Power and compute are Epoch's estimates.

04 / Environment

1 of 24 frontier runs has a published footprint

Training runs at or above 1e25 FLOP, and how many have a published energy or emissions figure. Beyond them, what is known comes from permits, press reports and company-wide sustainability reports.

1 of 24

Google's data centres ran on 65% carbon-free electricity hour by hour in 2025, against 100% matched over the year. No other operator publishes the hourly figure.

12 of 28 developer disclosures record nothing published 9 published training footprints 4 operators' renewable claims

Explore environment

Data: our own index of what developers and operators publish, and Epoch AI compute estimates

05 / Alignment

14 safety papers per hundred

Published work on AI safety counted against published work on frontier models, for 2025, the last complete year. Counted by keyword match rather than by field, so the definition is arguable.

For every hundred works published on frontier models in 2025, 14 were published on safety.

Recent years undercount, because indexing lags publication.

01 7 of 8

Publish a framework

Developers running frontier-scale training that publish a frontier safety framework. The rest publish nothing, and the absence is the finding.

02 4

Commit to stopping

Of those that publish, the number committing to halt deployment or development at a stated capability level rather than to deciding what to do when they reach one.

03 48 of 69

System card on the day

Releases by these developers since January 2025 that shipped with a system card. 17 are named in no card at all.

Explore alignment

Data: OpenAlex, Epoch AI, and our own indexes of safety frameworks and release disclosures

06 / Usage

5.5 quadrillion tokens a month

The largest token volume any company has stated: ByteDance, Doubao models, June 2026. Each figure below counts something different, and 4 tracked developers state none.

Jun 2026 ByteDance, Doubao models 5.5 quadrillion

May 2026 Google, all products 3.2 quadrillion

Jul 2026 Google, API only 964 trillion

Mar 2026 OpenAI, API only 657 trillion

Jun 2025 Microsoft, Foundry, yearly average 42 trillion

No volume stated by Anthropic, DeepSeek, Meta or xAI Tokens per month, converted by this site

Explore usage

Data: our own index of company statements, each linked to where it was made

07 / Adoption

Who is actually using it

Share of each economy's working-age population actively using AI tools in 2026-Q2, ranked, with the change since 2025-H1.

15.4% median economy 73.3% United Arab Emirates, the highest 25 of 27 economies where more people use AI than firms 42% of firms in Denmark, 2025

Economy

Share, 2026-Q2

Change since 2025-H1

United Arab Emirates

73.3%

+13.9

Singapore

64.3%

+5.7

Ireland

49.9%

+8.2

France

49.6%

+8.7

Norway

49.4%

+4.1

Spain

45.1%

+5.4

New Zealand

44.1%

+6.5

United Kingdom

43.7%

+7.3

Netherlands

43.5%

+7.2

Qatar

43.5%

+7.8

Show the remaining 137

Share of working-age population using AI tools by economy in 2026-Q2, ranked highest to lowest, with the change since 2025-H1. Rows 11 to 147.
Australia	40.7%	+6.2
South Korea	40.6%	+14.7
Belgium	40.3%	+6.8
Switzerland	39.1%	+6.7
Israel	38.7%	+4.8
Canada	38.5%	+5.0
Sweden	37.3%	+6.1
Austria	35.4%	+6.3
Taiwan	33.8%	+7.4
Hungary	33.3%	+5.4
United States	33.0%	+6.7
Denmark	32.4%	+5.8
Germany	32.3%	+5.8
Poland	32.3%	+5.9
Italy	31.4%	+5.6
Saudi Arabia	31.4%	+7.7
Czech Republic	31.2%	+5.2
Jordan	31.2%	+5.8
Bulgaria	30.7%	+5.3
Finland	30.6%	+5.0
Slovenia	30.1%	+5.5
Costa Rica	29.7%	+4.6
Vietnam	27.9%	+6.7
Lebanon	27.7%	+2.9
Oman	27.6%	+5.0
Portugal	27.4%	+5.0
Croatia	27.2%	+5.4
Slovakia	27.1%	+5.0
Colombia	25.9%	+5.5
Dominican Republic	25.8%	+3.8
Uruguay	25.7%	+4.8
Serbia	25.3%	+5.6
Lithuania	25.1%	+4.1
Jamaica	25.0%	+2.8
Japan	24.7%	+8.0
Panama	24.2%	+3.9
Chile	24.0%	+4.4
South Africa	23.9%	+4.6
Argentina	23.4%	+5.6
Bosnia And Herzegovina	23.2%	+5.0
Malaysia	22.6%	+4.3
Kuwait	22.6%	+4.9
Georgia	21.6%	+4.3
Greece	21.5%	+3.8
Mexico	21.4%	+4.7
Philippines	20.7%	+3.6
Ecuador	20.3%	+3.3
Brazil	20.2%	+4.6
El Salvador	20.1%	+5.5
Albania	19.2%	+3.4
Moldova	19.1%	+2.5
Türkiye	18.9%	+5.5
Azerbaijan	18.7%	+4.5
India	18.5%	+4.3
Romania	18.0%	+2.7
Mongolia	17.9%	+5.3
China	17.5%	+2.1
Guatemala	17.3%	+3.6
Peru	17.3%	+3.9
Kazakhstan	16.9%	+4.2
Gabon	15.7%	+3.4
Libya	15.5%	+2.8
Egypt	15.5%	+3.0
Namibia	15.4%	+2.4
Botswana	15.0%	+2.2
Indonesia	14.8%	+3.1
Nepal	14.6%	+2.3
Honduras	14.5%	+2.1
Senegal	14.4%	+2.0
Tunisia	13.9%	+1.6
Algeria	13.8%	+2.5
Cote D'Ivoire	13.7%	+2.9
Thailand	13.4%	+4.3
Zambia	13.3%	+1.6
Bolivia	13.3%	+2.4
Iraq	13.0%	+2.7
Iran	12.8%	+3.2
Paraguay	12.6%	+2.5
Nicaragua	12.5%	+2.5
Morocco	12.0%	+1.5
Pakistan	11.9%	+2.2
Gambia	11.5%	+0.9
Angola	11.4%	+2.5
Madagascar	11.4%	+2.5
Malawi	11.4%	+2.5
Mozambique	11.4%	+2.5
French Guiana	11.3%	+3.0
Guyana	11.3%	+3.0
Suriname	11.3%	+3.0
Venezuela	11.3%	+3.0
Benin	10.4%	+1.7
Burkina Faso	10.4%	+1.7
Ghana	10.4%	+1.7
Guinea	10.4%	+1.7
Guinea-Bissau	10.4%	+1.7
Liberia	10.4%	+1.7
Mali	10.4%	+1.7
Mauritania	10.4%	+1.7
Niger	10.4%	+1.7
Nigeria	10.4%	+1.7
Sierra Leone	10.4%	+1.7
Togo	10.4%	+1.7
Myanmar	10.4%	+2.0
Belarus	10.3%	+2.7
Kyrgyzstan	10.0%	+2.4
Lesotho	9.9%	+1.1
Russia	9.9%	+2.3
Ukraine	9.5%	+0.4
Cameroon	9.1%	+2.1
Central African Republic	9.1%	+2.1
Chad	9.1%	+2.1
Congo	9.1%	+2.1
Congo (DRC)	9.1%	+2.1
Kenya	8.9%	+1.1
Haiti	8.9%	+1.8
Zimbabwe	8.9%	+2.0
Laos	8.4%	+2.4
Bangladesh	8.0%	+1.5
Burundi	8.0%	+1.6
Eritrea	8.0%	+1.6
Ethiopia	8.0%	+1.6
Somalia	8.0%	+1.6
South Sudan	8.0%	+1.6
Sudan	8.0%	+1.6
Tanzania	8.0%	+1.6
Uganda	8.0%	+1.6
Armenia	7.7%	+1.5
Papua New Guinea	7.6%	+0.4
Sri Lanka	7.6%	+1.4
Uzbekistan	7.6%	+1.9
Rwanda	7.6%	+1.6
Syria	7.4%	+0.7
Cuba	6.8%	+1.1
Afghanistan	6.3%	+1.2
Tajikistan	6.3%	+1.2
Turkmenistan	6.3%	+1.2
Cambodia	6.1%	+1.5

Explore adoption

Data: Microsoft AI for Good Lab and Eurostat. Microsoft-visible use, scaled, not observed total use.

08 / Policy and governance

48 of 86 are binding

Instruments indexed across 11 jurisdictions, read from the primary source. Binding means a statute or delegated legislation; the rest is guidance, standards and strategy.

48 of 86

58 jurisdictions are separately scored for governance readiness, 31 of them coded individually.

Australia is indexed far more deeply than the rest, because it was indexed first, so the totals are not a comparison between jurisdictions.

Explore policy

Data: our own index of primary legal instruments, alongside the governance readiness rubric

09 / The wire

Official publications and preprints

Government publications, preprints and reported incidents, filed by keyword and linking to the publisher rather than to a copy. 4 jurisdictions in the current feed.

2026-10-28 Research preprint kw-bench: Breaking Frontier Harnesses on Real-World Tasks
2026-10-28 Research preprint Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception
2026-10-28 Research preprint Your Agent Acts Differently When You're Around: How User Presence Changes Agent Behavior in Negotiation
2026-10-28 Research preprint Why and How to Do Science When AI Is Better at It: A Framework Grounded in Human Understanding and Direction
2026-10-28 Research preprint Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind
2026-10-28 Research preprint Where Should the Evaluation Budget Go in LLM-Driven Agent Scaffold Evolution?

Explore the news feed

Feeds: arXiv, the US Federal Register, NIST, GOV.UK, the European Commission and three others