AI Safety Tracker

Search

every page and section, and 303 records

Pages 58

  • About/about/An index of AI capability, adoption, governance and safety. Scope, sourcing, limits and construction.
  • AI adoption/adoption/Who is using AI: enterprises by country, size and industry, set against the share of the population using it.
  • AI safety and alignment/alignment/The measurable state of AI alignment: solved capability tests, developers' safety promises and release disclosures, safety research volume, and incidents.
  • AI in Australia/australia/Australia's AI law, the share of Australians using AI tools, its governance readiness score, AI datacentres, GPU clusters and notable models, each figure cited.
  • AI training compute and cost/capability/Training compute, disclosed cost, benchmark saturation, release decisions and who builds frontier models. Every figure carries its source.
  • Charts/charts/Every chart and map on the site on one page: training compute, datacentre locations and electricity, token volumes, safety research and law, each cited.
  • AI in China/china/AI law in China, use of AI tools, governance readiness, AI datacentres, GPU clusters, notable models and frontier developers, each figure cited.
  • Corrections/corrections/How to report an error on this site, how reports are handled, and the list of corrections made.
  • AI by country/countries/AI law, use of AI tools, governance readiness, AI datacentres and notable models for Australia, the United States, China, the European Union and the United Kingdom.
  • AI datacentres/datacentres/AI datacentres on a map that pans and zooms: every site Epoch AI tracks, Australian sites from developers' statements, and AI's share of datacentre power.
  • Amazon/developers/amazon/Amazon: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • Anthropic/developers/anthropic/Anthropic: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • DeepSeek/developers/deepseek/DeepSeek: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • Google DeepMind/developers/google-deepmind/Google DeepMind: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • Frontier AI developers compared/developers/The eight frontier AI developers side by side: models, largest training run, safety framework, cards at release, datacentres and token volumes.
  • Meta/developers/meta/Meta: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • Microsoft/developers/microsoft/Microsoft: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • OpenAI/developers/openai/OpenAI: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • xAI/developers/xai/xAI: notable models, training compute, safety framework, release disclosures, datacentres, energy and usage, each from its source.
  • Environmental impact of AI/environment/What frontier AI developers publish about training energy, emissions and power sources, set against the size of their training runs, and what they do not.
  • AI in the European Union/european-union/AI law and the compute threshold in the European Union, AI use by people and enterprises in each member state, datacentres and models, each cited.
  • What are the environmental impacts of AI?/explainers/environmental-impact-of-ai/The electricity, carbon emissions and water behind AI, in plain words, from what operators, developers, the IEA and LBNL publish, and what none of them does.
  • How fast is AI improving?/explainers/how-fast-is-ai-improving/How fast AI is improving, in plain words: test scores and how quickly tests are solved, the computing power behind training, and how far open models trail.
  • How much energy does AI use?/explainers/how-much-energy-does-ai-use/How much energy AI uses, in plain words: datacentre electricity, the power of the largest AI datacentres, a training run and a prompt, and what is unpublished.
  • Frequently asked questions/faq/Short answers to common questions about AI's environmental footprint, energy per prompt, datacentre electricity, AI law and release safety testing, each drawn from a cited record.
  • Embeddable figures/figures/The eight key figures as SVG images for other pages, each printing its sources, licences, date and address, with the HTML that places it.
  • AI glossary/glossary/Definitions of the AI capability, governance and safety terms used on this site, each linked to the page that applies it.
  • The numbers behind AI/A cited index of AI capability, governance and safety. Current figures on model progress, adoption, law and incidents, with the source and retrieval date on every number.
  • Exposure against governance readiness/map/Where AI use has run ahead of the machinery to govern it, and where it has not. Two maps, one scatter, and the table behind both.
  • Governance Readiness Index/methodology/governance-readiness/The scoring rubric behind the governance readiness figure: five dimensions, what each measures, and what the score does not mean.
  • Methodology/methodology/How the data is fetched, processed, checked and published, and the judgement calls behind it.
  • AI Alignment news/news/alignment/Getting a model to do what was intended: interpretability, reward modelling, oversight, deception and specification. Mostly preprints, which are not peer reviewed.
  • AI news/news/general/About AI and nothing narrower. Items land here when no category rule matched, which is common for funding announcements and ministerial visits.
  • AI Governance news/news/governance/The machinery around the law: standards, frameworks, audits, assurance, procurement, institutes and who supervises whom.
  • AI news: policy, research and incidents/news/Government publications, research preprints and reported incidents on AI, sorted into categories by a rule you can read.
  • AI Policy news/news/policy/Law being made or applied. Bills, statutes, consultations, executive orders, enforcement and court decisions. Departmental communications are political statements about policy, not neutral descriptions of it.
  • AI Progress news/news/progress/Capability and how it is built. Benchmarks, training, compute, scaling, releases and what models can newly do.
  • AI Safety news/news/safety/Risk, harm, misuse, evaluation and things that have already gone wrong. Includes reported incidents, which are reported rather than sampled: an absence of incidents in a country usually means an absence of reporting.
  • AI law and policy in Australia/policy/australia/Every instrument relevant to AI in Australia, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in Brazil/policy/brazil/Every instrument relevant to AI in Brazil, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in Canada/policy/canada/Every instrument relevant to AI in Canada, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in China/policy/china/Every instrument relevant to AI in China, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in European Union/policy/european-union/Every instrument relevant to AI in European Union, tagged by how binding it is, with a link to each primary source.
  • AI law and policy/policy/An index of the law that actually applies to AI across eleven jurisdictions, tagged by how binding each instrument is, with a link to every primary source.
  • AI law and policy in India/policy/india/Every instrument relevant to AI in India, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in Japan/policy/japan/Every instrument relevant to AI in Japan, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in Singapore/policy/singapore/Every instrument relevant to AI in Singapore, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in South Korea/policy/south-korea/Every instrument relevant to AI in South Korea, tagged by how binding it is, with a link to each primary source.
  • AI policy timeline/policy/timeline/Every indexed AI law and policy instrument by year of adoption, newest first, with how binding each one is.
  • AI law and policy in United Kingdom/policy/united-kingdom/Every instrument relevant to AI in United Kingdom, tagged by how binding it is, with a link to each primary source.
  • AI law and policy in United States/policy/united-states/Every instrument relevant to AI in United States, tagged by how binding it is, with a link to each primary source.
  • Privacy/privacy/What this site collects, which is almost nothing, and what the host logs on its behalf.
  • AI progress on benchmarks/progress/What AI models can do, grouped by skill and by difficulty: benchmark scores, how long each test stayed useful, and where the current ones are headed.
  • Sources/sources/Every source behind every figure on this site, with its licence, its update cadence and its known limitations.
  • Terms/terms/Terms of use: no advice, no warranty, licensing of material published here and of material reproduced from third parties.
  • AI in the United Kingdom/united-kingdom/AI law in the United Kingdom, use of AI tools, governance readiness, AI datacentres, GPU clusters, notable models and frontier developers, each cited.
  • AI in the United States/united-states/AI law and compute thresholds in the United States, use of AI tools, governance readiness, AI datacentres, notable models and frontier developers, each cited.
  • AI usage: tokens and users/usage/How many tokens Google, ByteDance, OpenAI and Microsoft say they process, how many people use each product, and what the use is for.

Sections 235

  • People against their employersAI adoptionShare of the working-age population using AI tools, against the share of enterprises doing the same. 25 of 27 economies have the population ahead.
  • By number of people employedAI adoptionShare of enterprises using at least one AI technology, by size band.
  • By what the firm doesAI adoptionShare of enterprises using at least one AI technology, by economic activity.
  • What they use it forAI adoptionThe seven technologies the survey asks about individually, as a share of all enterprises.
  • By countryAI adoptionEnterprises with ten or more people employed, using at least one AI technology, 2025.
  • How close is AGI?AI safety and alignmentThere is no agreed definition and therefore no agreed distance. These are the measured quantities that bear on the question, each with its source.
  • What developers have promisedAI safety and alignmentPublished frontier safety frameworks, and what each commits to. Self-written and self-assessed, with no common scale between them.
  • What developers published at each releaseAI safety and alignmentEvery release Epoch AI lists from these developers since January 2025, against the card published for it: when it appeared, whether it reports dangerous-capability results, which government institutes it names as testers before release, and what it says about the developer's own thresholds.
  • How much research goes whereAI safety and alignmentWorks indexed by OpenAlex whose title or abstract matches each search phrase, by publication year. Logarithmic axis; every series spans several orders of magnitude.
  • Reported incidentsAI safety and alignmentThe most recent entries in the AI Incident Database. Headline and short extract only, linking to the original report.
  • Law and policyAI in Australia28 instruments that apply to AI in Australia, grouped by how binding each is. 11 are binding law, and one of those names AI or deepfakes in its title.
  • Use of AI toolsAI in AustraliaShare of the working-age population using AI tools, by reporting period, and the economies ranked either side of Australia in 2026-Q2. Microsoft's estimate from its own telemetry, scaled to each population.
  • Governance readinessAI in Australia6 of 10 points on a five-dimension rubric recording how much machinery exists to govern AI, not how strict it is. 4 of 31 nationally coded jurisdictions score higher and 4 score the same.
  • AI datacentres and computeAI in Australia13 AI datacentre sites recorded from developers', tenants' and planning authorities' statements, by stated power: 2 operating, 2 under construction, 2 approved and 7 proposed.
  • Notable modelsAI in Australia6 models in Epoch AI's notable models list with an organisation in Australia among their developers, 4 of them from academic groups. The largest training compute recorded for any of them is 8.5e22 FLOP, for StarCoder.
  • Not coveredAI in Australia
  • Training compute over timeAI training compute and costEvery notable model with a reported or estimated training-compute figure, on a logarithmic scale. Each step of the axis is ten times the one below it.
  • What the training runs costAI training compute and costDisclosed or estimated cost of the final training run, in 2023 dollars, for the 178 models where a figure exists. Also logarithmic.
  • How models shipAI training compute and costNotable models by publication year and how they were made available: open weights, API access, or never released.
  • How far behind the open-weight frontier isAI training compute and costLargest training run to date over all models, against the same over models released with open weights. Both are running maxima, so neither line can fall.
  • Who builds themAI training compute and costNotable models published since 2023, by organisation and by the country of the organisation that built them.
  • The compute thresholds written into lawAI training compute and costEvery training-compute threshold set by a statute, an executive action or an official policy document, the day each took its final form, and how many models with a published compute figure sat above it that day and sit above it now.
  • Training compute over timeChartsEvery notable model with a reported or estimated training-compute figure, on a logarithmic scale. Each step of the axis is ten times the one below it.
  • How far behind open weights areChartsLargest training run to date, all models against open-weight models. 8 months between the overall frontier reaching 3.8e25 FLOP and an open-weight model doing so.
  • Compute thresholds in lawCharts4 of 6 thresholds were written after a model had already crossed them. Each mark is the day an instrument took its final form.
  • Where AI datacentres areChartsEvery site in Epoch AI's AI Data Centers hub, sized by IT power. The ring is the power planned at the site's last projected stage, the disc the power operating now. Dashed rings are sites the hub does not hold, sized by the capacity their developers state.
  • Datacentre electricity to 2035ChartsElectricity used by all datacentres worldwide: the International Energy Agency's estimate for 2024 and 2025, and its Base Case for 2030 and 2035. The 2035 range spans its four cases. The AI portion is marked where the agency states one.
  • Datacentre power by countryChartsAI datacentre capacity by country in Epoch AI's hub, and every datacentre's electricity by region in the IEA's estimate for 2024. The two measure different things.
  • Training runs against published emissionsChartsEvery run since 2020 above 1.0e21 FLOP with a compute estimate, by developer, logarithmic scale. 1 of the 24 models at or above 1e25 FLOP has a published training-emissions figure.
  • Safety research against capability researchChartsWorks indexed by OpenAlex per year by research area, logarithmic scale. In 2025, 14 works on safety, interpretability, evaluation or catastrophic risk for every 100 on frontier models; 330 in 2020.
  • Releases against their system cardsChartsEvery release Epoch AI lists from the frontier developers since January 2025. 48 of 69 shipped with a system card on the day, 4 had one later and 17 have none.
  • Tokens processed, as each company states itChartsEvery aggregate token count a company has stated, converted to tokens per month, logarithmic scale. The largest is ByteDance, Doubao models at 5.5 quadrillion tokens a month in June 2026.
  • Law that applies to AI, by jurisdictionCharts86 instruments indexed in 11 jurisdictions. 48 are binding law, and 10 of those name AI in their own title.
  • Law and policyAI in China7 instruments that apply to AI in China, grouped by how binding each is. 7 are binding law, and 2 of those name AI or deepfakes in their titles.
  • Use of AI toolsAI in ChinaShare of the working-age population using AI tools, by reporting period, and the economies ranked either side of China in 2026-Q2. Microsoft's estimate from its own telemetry, scaled to each population.
  • Governance readinessAI in China9 of 10 points on a five-dimension rubric recording how much machinery exists to govern AI, not how strict it is. 1 of 31 nationally coded jurisdictions score higher and 0 score the same.
  • AI datacentres and computeAI in China5 AI datacentre sites in Epoch AI's hub in China, by the IT power planned at the last stage of each site's timeline, with the power operating now.
  • Notable modelsAI in China190 models in Epoch AI's notable models list with an organisation in China among their developers, 60 of them from academic groups and 65 published since January 2025. The largest training compute recorded for any of them is 2.5e25 FLOP, for Doubao-pro.
  • Frontier developersAI in ChinaThe developers in the Frontier Safety Framework Index with notable models Epoch AI credits to an organisation in China. DeepSeek is based there. A developer is based where most of its credited models were built.
  • Not coveredAI in China
  • Where they areAI datacentresEvery site in Epoch AI's AI Data Centers hub, sized by IT power. The ring is the power planned at the site's last projected stage, the disc the power operating now. Dashed rings are sites the hub does not hold, sized by the capacity their developers state.
  • AI against every datacentreAI datacentresAI's share of all datacentres by number, electricity and water, as each publisher states it. A dash means the publisher gives no figure.
  • Every site in the hubAI datacentresEpoch AI's sites, largest planned first, with owner, users, status and IT power.
  • Sites added from developers' statementsAI datacentresAI sites in Australia that Epoch AI's hub does not hold, largest stated capacity first and buildings hosting an AI tenant last, with status, power supply and cooling as each developer states them.
  • What this page cannot showAI datacentresQuestions about each site that no field in the hub answers.
  • ModelsAmazonEvery model Epoch AI judges notable from Amazon, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresAmazonThe frontier safety framework Amazon has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsAmazonHow many of Amazon's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresAmazonSites in Epoch AI's data-centre hub that Amazon owns or is named as using, largest planned IT power first. Those it owns plan 4.0 GW of IT power between them.
  • Energy and emissionsAmazonWhat Amazon has published about the energy behind its models, and what permits and press reporting add.
  • UsageAmazonThe token volumes and user counts Amazon has stated, each in its own scope.
  • ModelsAnthropicEvery model Epoch AI judges notable from Anthropic, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresAnthropicThe frontier safety framework Anthropic has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsAnthropicHow many of Anthropic's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresAnthropicSites in Epoch AI's data-centre hub that Anthropic owns or is named as using, largest planned IT power first.
  • Energy and emissionsAnthropicWhat Anthropic has published about the energy behind its models, and what permits and press reporting add.
  • UsageAnthropicThe token volumes and user counts Anthropic has stated, each in its own scope.
  • ModelsDeepSeekEvery model Epoch AI judges notable from DeepSeek, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresDeepSeekThe frontier safety framework DeepSeek has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsDeepSeekHow many of DeepSeek's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresDeepSeekSites in Epoch AI's data-centre hub that DeepSeek owns or is named as using, largest planned IT power first.
  • Energy and emissionsDeepSeekWhat DeepSeek has published about the energy behind its models, and what permits and press reporting add.
  • UsageDeepSeekThe token volumes and user counts DeepSeek has stated, each in its own scope.
  • ModelsGoogle DeepMindEvery model Epoch AI judges notable from Google DeepMind, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresGoogle DeepMindThe frontier safety framework Google DeepMind has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsGoogle DeepMindHow many of Google DeepMind's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresGoogle DeepMindSites in Epoch AI's data-centre hub that Google DeepMind owns or is named as using, largest planned IT power first. Those it owns plan 5.9 GW of IT power between them.
  • Energy and emissionsGoogle DeepMindWhat Google DeepMind has published about the energy behind its models, and what permits and press reporting add.
  • UsageGoogle DeepMindThe token volumes and user counts Google DeepMind has stated, each in its own scope.
  • ModelsMetaEvery model Epoch AI judges notable from Meta, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresMetaThe frontier safety framework Meta has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsMetaHow many of Meta's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresMetaSites in Epoch AI's data-centre hub that Meta owns or is named as using, largest planned IT power first. Those it owns plan 5.6 GW of IT power between them.
  • Energy and emissionsMetaWhat Meta has published about the energy behind its models, and what permits and press reporting add.
  • UsageMetaThe token volumes and user counts Meta has stated, each in its own scope.
  • ModelsMicrosoftEvery model Epoch AI judges notable from Microsoft, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresMicrosoftThe frontier safety framework Microsoft has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsMicrosoftHow many of Microsoft's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresMicrosoftSites in Epoch AI's data-centre hub that Microsoft owns or is named as using, largest planned IT power first. Those it owns plan 4.1 GW of IT power between them.
  • Energy and emissionsMicrosoftWhat Microsoft has published about the energy behind its models, and what permits and press reporting add.
  • UsageMicrosoftThe token volumes and user counts Microsoft has stated, each in its own scope.
  • ModelsOpenAIEvery model Epoch AI judges notable from OpenAI, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresOpenAIThe frontier safety framework OpenAI has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsOpenAIHow many of OpenAI's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresOpenAISites in Epoch AI's data-centre hub that OpenAI owns or is named as using, largest planned IT power first.
  • Energy and emissionsOpenAIWhat OpenAI has published about the energy behind its models, and what permits and press reporting add.
  • UsageOpenAIThe token volumes and user counts OpenAI has stated, each in its own scope.
  • ModelsxAIEvery model Epoch AI judges notable from xAI, by year, with the largest training run each year where Epoch has a compute estimate.
  • Safety framework and release disclosuresxAIThe frontier safety framework xAI has published, and what it published at each release since January 2025.
  • Its models against the compute thresholdsxAIHow many of xAI's models with a compute estimate sit at or above each training-compute threshold written into an instrument, and the first to cross it.
  • DatacentresxAISites in Epoch AI's data-centre hub that xAI owns or is named as using, largest planned IT power first. Those it owns plan 1.9 GW of IT power between them.
  • Energy and emissionsxAIWhat xAI has published about the energy behind its models, and what permits and press reporting add.
  • UsagexAIThe token volumes and user counts xAI has stated, each in its own scope.
  • What we knowEnvironmental impact of AIWhat each developer has published about the energy behind its models, and what permits and press reporting add. 12 of 28 entries record nothing published.
  • How big the training runs areEnvironmental impact of AIEvery run since 2020 above 1.0e21 FLOP with a compute estimate, by developer, on a logarithmic scale. The largest per developer is named.
  • Training footprints that have been publishedEnvironmental impact of AIEvery published figure for the energy or emissions of training a model, with who measured it and on what basis. Emissions per 1e24 FLOP are computed here where both figures exist.
  • Energy per promptEnvironmental impact of AIEvery per-prompt figure a developer has published. Each measures a different statistic.
  • Renewable claims against the gridEnvironmental impact of AIWhat the operators of frontier data centres report for their own electricity. Annual matching counts renewable energy bought over a year; hourly carbon-free energy counts what was clean in the hour and on the grid where it was used.
  • Datacentre electricity to 2035Environmental impact of AIElectricity used by all datacentres worldwide: the International Energy Agency's estimate for 2024 and 2025, and its Base Case for 2030 and 2035. The 2035 range spans its four cases. The AI portion is marked where the agency states one.
  • Datacentre power by countryEnvironmental impact of AIAI datacentre capacity by country in Epoch AI's hub, and every datacentre's electricity by region in the IEA's estimate for 2024. The two measure different things.
  • What this page cannot showEnvironmental impact of AIQuestions a reader brings that no published source answers.
  • Law and policyAI in the European Union8 instruments that apply to AI in the European Union, grouped by how binding each is. 6 are binding law, and 2 of those name AI or deepfakes in their titles.
  • Use of AI toolsAI in the European UnionShare of the working-age population using AI tools in each member state Microsoft covers, 2026-Q2, with each state's rank among 147 economies. Microsoft's estimate from its own telemetry, scaled to each population.
  • Use of AI by enterprisesAI in the European UnionShare of enterprises with ten or more people employed using at least one AI technology, 2025: 19.9% across the EU27, and by member state, by size and by technology. Eurostat's survey of enterprises.
  • Governance readinessAI in the European Union10 of 10 points on a five-dimension rubric recording how much machinery exists to govern AI, not how strict it is. The EU is scored once, and the score applies to each of its 27 member states. 0 of 31 nationally coded jurisdictions score higher and 1 scores the same.
  • Compute thresholdsAI in the European UnionThe training-compute threshold in EU law, the day it took its final form, and the models with a published compute figure above it that day and now. Models with a developer in the European Union are counted separately.
  • AI datacentres and computeAI in the European Union3 AI datacentre sites in Epoch AI's hub in the European Union, by the IT power planned at the last stage of each site's timeline, with the power operating now.
  • Notable modelsAI in the European Union106 models in Epoch AI's notable models list with an organisation in the European Union among their developers, 58 of them from academic groups and 1 published since January 2025. The largest training compute recorded for any of them is 7.7e23 FLOP, for Mixtral 8x7B.
  • Official publicationsAI in the European Union17 publications on AI from EU government feeds in the news feed's 120-day window, the latest 6 below, from 1 publisher in all. Headline and a short extract, each linking to the publisher.
  • Not coveredAI in the European Union
  • How much electricity do datacentres use, and how much of it is AI?What are the environmental impacts of AI?
  • How much carbon does that electricity emit?What are the environmental impacts of AI?
  • How much energy does it take to train an AI model?What are the environmental impacts of AI?
  • How much energy does one AI prompt use?What are the environmental impacts of AI?
  • How much water does AI use?What are the environmental impacts of AI?
  • What is not published?What are the environmental impacts of AI?
  • How is progress in AI measured?How fast is AI improving?
  • How quickly are AI benchmarks being solved?How fast is AI improving?
  • What can AI models do now?How fast is AI improving?
  • How hard are the problems AI can solve?How fast is AI improving?
  • How fast is the computing power behind AI growing?How fast is AI improving?
  • How far behind are openly released AI models?How fast is AI improving?
  • How many new AI models are released?How fast is AI improving?
  • What these measures cannot showHow fast is AI improving?
  • What is the difference between power and energy?How much energy does AI use?
  • How much power do AI datacentres draw?How much energy does AI use?
  • How much electricity do datacentres use?How much energy does AI use?
  • How much energy does it take to train an AI model?How much energy does AI use?
  • How much energy does one AI prompt use?How much energy does AI use?
  • How much AI datacentre capacity is planned in Australia?How much energy does AI use?
  • What is not known?How much energy does AI use?
  • Which AI model or service has the smallest environmental footprint?Frequently asked questions
  • How much energy does one AI prompt use?Frequently asked questions
  • How much electricity do datacentres use?Frequently asked questions
  • How much law applies to AI?Frequently asked questions
  • Do AI developers publish safety testing when a model is released?Frequently asked questions
  • Does this site make forecasts?Frequently asked questions
  • Where does the data come from, and may it be reused?Frequently asked questions
  • How are errors reported and corrected?Frequently asked questions
  • 7 of 18 tests are finishedThe numbers behind AI
  • The numbers written into lawThe numbers behind AI
  • 14.2 GW operating, 38.2 GW plannedThe numbers behind AI
  • 1 of 24 frontier runs has a published footprintThe numbers behind AI
  • 14 safety papers per hundredThe numbers behind AI
  • 5.5 quadrillion tokens a monthThe numbers behind AI
  • Who is actually using itThe numbers behind AI
  • 48 of 86 are bindingThe numbers behind AI
  • Official publications and preprintsThe numbers behind AI
  • Every jurisdictionExposure against governance readinessOrdered by gap, widest use-ahead first. Country maps open an enlarged view with policy coverage.
  • How readiness is scoredExposure against governance readinessCoded here from primary instruments. The rubric and the instruments behind each score are published in the methodology.
  • Every score, and what it was coded fromGovernance Readiness IndexNationally coded jurisdictions. The 27 EU member states share the bloc score and are listed on the map page.
  • StatutesAI law and policy in AustraliaAct of Parliament. Binding on everyone it applies to.
  • Rules made under an ActAI law and policy in AustraliaRules, regulations or codes made under an Act. Binding.
  • Binding on government onlyAI law and policy in AustraliaBinds government agencies only, not the wider market.
  • Regulator guidanceAI law and policy in AustraliaHow a regulator states it will apply existing law. Creates no obligations of its own.
  • Voluntary standardsAI law and policy in AustraliaGuidance with no legal force. Compliance is a choice.
  • Proposed, not lawAI law and policy in AustraliaProposed, consulted on, or before Parliament. Creates no duties yet.
  • Reviews and inquiriesAI law and policy in AustraliaAn inquiry or review. Evidence and recommendations, not obligations.
  • About this indexAI law and policy in Australia
  • StatutesAI law and policy in BrazilAct of Parliament. Binding on everyone it applies to.
  • Proposed, not lawAI law and policy in BrazilProposed, consulted on, or before Parliament. Creates no duties yet.
  • About this indexAI law and policy in Brazil
  • StatutesAI law and policy in CanadaAct of Parliament. Binding on everyone it applies to.
  • Binding on government onlyAI law and policy in CanadaBinds government agencies only, not the wider market.
  • Voluntary standardsAI law and policy in CanadaGuidance with no legal force. Compliance is a choice.
  • Proposed, not lawAI law and policy in CanadaProposed, consulted on, or before Parliament. Creates no duties yet.
  • Reviews and inquiriesAI law and policy in CanadaAn inquiry or review. Evidence and recommendations, not obligations.
  • About this indexAI law and policy in Canada
  • StatutesAI law and policy in ChinaAct of Parliament. Binding on everyone it applies to.
  • Rules made under an ActAI law and policy in ChinaRules, regulations or codes made under an Act. Binding.
  • About this indexAI law and policy in China
  • StatutesAI law and policy in European UnionAct of Parliament. Binding on everyone it applies to.
  • Binding on government onlyAI law and policy in European UnionBinds government agencies only, not the wider market.
  • Voluntary standardsAI law and policy in European UnionGuidance with no legal force. Compliance is a choice.
  • About this indexAI law and policy in European Union
  • What each jurisdiction hasAI law and policyIndexed instruments by how binding each one is.
  • StatutesAI law and policy in IndiaAct of Parliament. Binding on everyone it applies to.
  • Rules made under an ActAI law and policy in IndiaRules, regulations or codes made under an Act. Binding.
  • Voluntary standardsAI law and policy in IndiaGuidance with no legal force. Compliance is a choice.
  • About this indexAI law and policy in India
  • StatutesAI law and policy in JapanAct of Parliament. Binding on everyone it applies to.
  • Voluntary standardsAI law and policy in JapanGuidance with no legal force. Compliance is a choice.
  • About this indexAI law and policy in Japan
  • StatutesAI law and policy in SingaporeAct of Parliament. Binding on everyone it applies to.
  • Regulator guidanceAI law and policy in SingaporeHow a regulator states it will apply existing law. Creates no obligations of its own.
  • Voluntary standardsAI law and policy in SingaporeGuidance with no legal force. Compliance is a choice.
  • About this indexAI law and policy in Singapore
  • StatutesAI law and policy in South KoreaAct of Parliament. Binding on everyone it applies to.
  • Rules made under an ActAI law and policy in South KoreaRules, regulations or codes made under an Act. Binding.
  • Binding on government onlyAI law and policy in South KoreaBinds government agencies only, not the wider market.
  • About this indexAI law and policy in South Korea
  • Instruments by yearAI policy timelineYear of adoption for each indexed instrument. 15 of 86 were adopted before 2015 and do not mention AI.
  • Every instrument, newest firstAI policy timeline
  • StatutesAI law and policy in United KingdomAct of Parliament. Binding on everyone it applies to.
  • Binding on government onlyAI law and policy in United KingdomBinds government agencies only, not the wider market.
  • Regulator guidanceAI law and policy in United KingdomHow a regulator states it will apply existing law. Creates no obligations of its own.
  • Voluntary standardsAI law and policy in United KingdomGuidance with no legal force. Compliance is a choice.
  • About this indexAI law and policy in United Kingdom
  • StatutesAI law and policy in United StatesAct of Parliament. Binding on everyone it applies to.
  • Rules made under an ActAI law and policy in United StatesRules, regulations or codes made under an Act. Binding.
  • Binding on government onlyAI law and policy in United StatesBinds government agencies only, not the wider market.
  • Voluntary standardsAI law and policy in United StatesGuidance with no legal force. Compliance is a choice.
  • Reviews and inquiriesAI law and policy in United StatesAn inquiry or review. Evidence and recommendations, not obligations.
  • About this indexAI law and policy in United States
  • Capability by categoryAI progress on benchmarksBest score any evaluated model has reached in each area, with the benchmark it came from. Categories are ours; the scores are Epoch's.
  • By who else could do itAI progress on benchmarksThe same benchmarks ordered by the difficulty of the questions rather than the subject. Difficulty tiers are our own reading of what each test contains.
  • If the current rate heldAI progress on benchmarksEach open benchmark's frontier score over the last 24 months, extended in a straight line to 90.0%. An extrapolation of a fitted line, not a forecast.
  • The frontier on every benchmarkAI progress on benchmarksRunning best score by month. The dashed line is 90.0%, the saturation point, past which a test stops separating the leading models.
  • Most recent notable modelsAI progress on benchmarksThe last 12 entries added to Epoch's model database, newest first, to 2026-10-12.
  • Notable models released per quarterAI progress on benchmarksModels meeting Epoch's notability criteria, by publication quarter since 2019.
  • Registered so that using them fails the buildSourcesNeither licence permits republication. Both are registered with a redistribution status that stops the pipeline writing their data and the site rendering it.
  • What is oursSources8 datasets on this site are compiled here from primary documents rather than obtained from a publisher. All are licensed MIT.
  • DownloadsSourcesThe 16 datasets behind those indexes, as JSON, and as CSV for the 13 whose records are flat. Each JSON file carries its licences and a citation.
  • Pipeline statusSourcesThe daily run of 28 October 2026 at 13:08, UTC. Fetched data is confirmed on every run that succeeds; our own indexes are dated by their last review. Every module ran.
  • Law and policyAI in the United Kingdom7 instruments that apply to AI in the United Kingdom, grouped by how binding each is. 4 are binding law, and none of those names AI or deepfakes in its title.
  • Use of AI toolsAI in the United KingdomShare of the working-age population using AI tools, by reporting period, and the economies ranked either side of the United Kingdom in 2026-Q2. Microsoft's estimate from its own telemetry, scaled to each population.
  • Governance readinessAI in the United Kingdom6 of 10 points on a five-dimension rubric recording how much machinery exists to govern AI, not how strict it is. 4 of 31 nationally coded jurisdictions score higher and 4 score the same.
  • AI datacentres and computeAI in the United Kingdom1 AI datacentre site in Epoch AI's hub in the United Kingdom, by the IT power planned at the last stage of each site's timeline, with the power operating now.
  • Notable modelsAI in the United Kingdom89 models in Epoch AI's notable models list with an organisation in the United Kingdom among their developers, 38 of them from academic groups and 1 published since January 2025. The largest training compute recorded for any of them is 6.3e23 FLOP, for Gopher (280B).
  • Frontier developersAI in the United KingdomThe developers in the Frontier Safety Framework Index with notable models Epoch AI credits to an organisation in the United Kingdom. None is based in the United Kingdom. A developer is based where most of its credited models were built.
  • Official publicationsAI in the United Kingdom12 publications on AI from UK government feeds in the news feed's 120-day window, the latest 6 below, from 1 publisher in all. Headline and a short extract, each linking to the publisher.
  • Not coveredAI in the United Kingdom
  • Law and policyAI in the United States10 instruments that apply to AI in the United States, grouped by how binding each is. 6 are binding law, and 2 of those name AI or deepfakes in their titles.
  • Use of AI toolsAI in the United StatesShare of the working-age population using AI tools, by reporting period, and the economies ranked either side of the United States in 2026-Q2. Microsoft's estimate from its own telemetry, scaled to each population.
  • Governance readinessAI in the United States6 of 10 points on a five-dimension rubric recording how much machinery exists to govern AI, not how strict it is. 4 of 31 nationally coded jurisdictions score higher and 4 score the same.
  • Compute thresholdsAI in the United StatesThe 4 training-compute thresholds in US law and policy, the day each took its final form, and the models with a published compute figure above each that day and now. Models with a developer in the United States are counted separately.
  • AI datacentres and computeAI in the United States77 AI datacentre sites in Epoch AI's hub in the United States, by the IT power planned at the last stage of each site's timeline, with the power operating now.
  • Notable modelsAI in the United States700 models in Epoch AI's notable models list with an organisation in the United States among their developers, 189 of them from academic groups and 117 published since January 2025. The largest training compute recorded for any of them is 9.4e26 FLOP, for GPT-6 Astra.
  • Frontier developersAI in the United StatesThe developers in the Frontier Safety Framework Index with notable models Epoch AI credits to an organisation in the United States. 7 are based there. A developer is based where most of its credited models were built.
  • Official publicationsAI in the United States18 publications on AI from US government feeds in the news feed's 120-day window, the latest 6 below, from 6 publishers in all. Headline and a short extract, each linking to the publisher.
  • Not coveredAI in the United States
  • Tokens processed, as each company states itAI usage: tokens and usersEvery aggregate token count a company has published or stated, converted to tokens per month. Logarithmic: each step of the axis is ten times the one below it.
  • Who uses themAI usage: tokens and usersDisclosed user counts per product, largest first. Weekly and monthly counts are different measures and are labelled.
  • What the tokens are forAI usage: tokens and usersShares of use by topic or task, as each study states them. Each study counts a different unit, and none counts tokens.
  • What this page cannot showAI usage: tokens and usersQuestions a reader brings that no published source answers.

Economies 173

  • Afghanistan6.3% of the working-age population using AI tools, 2026-Q2.
  • Albania19.2% of the working-age population using AI tools, 2026-Q2.
  • Algeria13.8% of the working-age population using AI tools, 2026-Q2.
  • Angola11.4% of the working-age population using AI tools, 2026-Q2.
  • Argentina23.4% of the working-age population using AI tools, 2026-Q2.
  • Armenia7.7% of the working-age population using AI tools, 2026-Q2.
  • Australia40.7% of the working-age population using AI tools, 2026-Q2.
  • Austria35.4% of the working-age population using AI tools, 2026-Q2.
  • Azerbaijan18.7% of the working-age population using AI tools, 2026-Q2.
  • BahamasPolicy instruments for Bahamas are not yet indexed. This does not mean the country has no AI policy.
  • Bangladesh8.0% of the working-age population using AI tools, 2026-Q2.
  • Belarus10.3% of the working-age population using AI tools, 2026-Q2.
  • Belgium40.3% of the working-age population using AI tools, 2026-Q2.
  • BelizePolicy instruments for Belize are not yet indexed. This does not mean the country has no AI policy.
  • Benin10.4% of the working-age population using AI tools, 2026-Q2.
  • BhutanPolicy instruments for Bhutan are not yet indexed. This does not mean the country has no AI policy.
  • Bolivia13.3% of the working-age population using AI tools, 2026-Q2.
  • Bosnia and Herz.23.2% of the working-age population using AI tools, 2026-Q2.
  • Botswana15.0% of the working-age population using AI tools, 2026-Q2.
  • Brazil20.2% of the working-age population using AI tools, 2026-Q2.
  • BruneiPolicy instruments for Brunei are not yet indexed. This does not mean the country has no AI policy.
  • Bulgaria30.7% of the working-age population using AI tools, 2026-Q2.
  • Burkina Faso10.4% of the working-age population using AI tools, 2026-Q2.
  • Burundi8.0% of the working-age population using AI tools, 2026-Q2.
  • Cambodia6.1% of the working-age population using AI tools, 2026-Q2.
  • Cameroon9.1% of the working-age population using AI tools, 2026-Q2.
  • Canada38.5% of the working-age population using AI tools, 2026-Q2.
  • Central African Rep.9.1% of the working-age population using AI tools, 2026-Q2.
  • Chad9.1% of the working-age population using AI tools, 2026-Q2.
  • Chile24.0% of the working-age population using AI tools, 2026-Q2.
  • China17.5% of the working-age population using AI tools, 2026-Q2.
  • Colombia25.9% of the working-age population using AI tools, 2026-Q2.
  • Congo9.1% of the working-age population using AI tools, 2026-Q2.
  • Costa Rica29.7% of the working-age population using AI tools, 2026-Q2.
  • Côte d'Ivoire13.7% of the working-age population using AI tools, 2026-Q2.
  • Croatia27.2% of the working-age population using AI tools, 2026-Q2.
  • Cuba6.8% of the working-age population using AI tools, 2026-Q2.
  • Cyprus8 instruments indexed. EU-level coverage; national measures are not separately indexed.
  • Czechia31.2% of the working-age population using AI tools, 2026-Q2.
  • Dem. Rep. Congo9.1% of the working-age population using AI tools, 2026-Q2.
  • Denmark32.4% of the working-age population using AI tools, 2026-Q2.
  • DjiboutiPolicy instruments for Djibouti are not yet indexed. This does not mean the country has no AI policy.
  • Dominican Rep.25.8% of the working-age population using AI tools, 2026-Q2.
  • Ecuador20.3% of the working-age population using AI tools, 2026-Q2.
  • Egypt15.5% of the working-age population using AI tools, 2026-Q2.
  • El Salvador20.1% of the working-age population using AI tools, 2026-Q2.
  • Eq. GuineaPolicy instruments for Eq. Guinea are not yet indexed. This does not mean the country has no AI policy.
  • Eritrea8.0% of the working-age population using AI tools, 2026-Q2.
  • Estonia8 instruments indexed. EU-level coverage; national measures are not separately indexed.
  • eSwatiniPolicy instruments for eSwatini are not yet indexed. This does not mean the country has no AI policy.
  • Ethiopia8.0% of the working-age population using AI tools, 2026-Q2.
  • Falkland Is.Policy instruments for Falkland Is. are not yet indexed. This does not mean the country has no AI policy.
  • FijiPolicy instruments for Fiji are not yet indexed. This does not mean the country has no AI policy.
  • Finland30.6% of the working-age population using AI tools, 2026-Q2.
  • Fr. S. Antarctic LandsPolicy instruments for Fr. S. Antarctic Lands are not yet indexed. This does not mean the country has no AI policy.
  • France49.6% of the working-age population using AI tools, 2026-Q2.
  • Gabon15.7% of the working-age population using AI tools, 2026-Q2.
  • Gambia11.5% of the working-age population using AI tools, 2026-Q2.
  • Georgia21.6% of the working-age population using AI tools, 2026-Q2.
  • Germany32.3% of the working-age population using AI tools, 2026-Q2.
  • Ghana10.4% of the working-age population using AI tools, 2026-Q2.
  • Greece21.5% of the working-age population using AI tools, 2026-Q2.
  • GreenlandPolicy instruments for Greenland are not yet indexed. This does not mean the country has no AI policy.
  • Guatemala17.3% of the working-age population using AI tools, 2026-Q2.
  • Guinea10.4% of the working-age population using AI tools, 2026-Q2.
  • Guinea-Bissau10.4% of the working-age population using AI tools, 2026-Q2.
  • Guyana11.3% of the working-age population using AI tools, 2026-Q2.
  • Haiti8.9% of the working-age population using AI tools, 2026-Q2.
  • Honduras14.5% of the working-age population using AI tools, 2026-Q2.
  • Hungary33.3% of the working-age population using AI tools, 2026-Q2.
  • IcelandPolicy instruments for Iceland are not yet indexed. This does not mean the country has no AI policy.
  • India18.5% of the working-age population using AI tools, 2026-Q2.
  • Indonesia14.8% of the working-age population using AI tools, 2026-Q2.
  • Iran12.8% of the working-age population using AI tools, 2026-Q2.
  • Iraq13.0% of the working-age population using AI tools, 2026-Q2.
  • Ireland49.9% of the working-age population using AI tools, 2026-Q2.
  • Israel38.7% of the working-age population using AI tools, 2026-Q2.
  • Italy31.4% of the working-age population using AI tools, 2026-Q2.
  • Jamaica25.0% of the working-age population using AI tools, 2026-Q2.
  • Japan24.7% of the working-age population using AI tools, 2026-Q2.
  • Jordan31.2% of the working-age population using AI tools, 2026-Q2.
  • Kazakhstan16.9% of the working-age population using AI tools, 2026-Q2.
  • Kenya8.9% of the working-age population using AI tools, 2026-Q2.
  • Kuwait22.6% of the working-age population using AI tools, 2026-Q2.
  • Kyrgyzstan10.0% of the working-age population using AI tools, 2026-Q2.
  • Laos8.4% of the working-age population using AI tools, 2026-Q2.
  • Latvia8 instruments indexed. EU-level coverage; national measures are not separately indexed.
  • Lebanon27.7% of the working-age population using AI tools, 2026-Q2.
  • Lesotho9.9% of the working-age population using AI tools, 2026-Q2.
  • Liberia10.4% of the working-age population using AI tools, 2026-Q2.
  • Libya15.5% of the working-age population using AI tools, 2026-Q2.
  • Lithuania25.1% of the working-age population using AI tools, 2026-Q2.
  • Luxembourg8 instruments indexed. EU-level coverage; national measures are not separately indexed.
  • MacedoniaPolicy instruments for Macedonia are not yet indexed. This does not mean the country has no AI policy.
  • Madagascar11.4% of the working-age population using AI tools, 2026-Q2.
  • Malawi11.4% of the working-age population using AI tools, 2026-Q2.
  • Malaysia22.6% of the working-age population using AI tools, 2026-Q2.
  • Mali10.4% of the working-age population using AI tools, 2026-Q2.
  • Mauritania10.4% of the working-age population using AI tools, 2026-Q2.
  • Mexico21.4% of the working-age population using AI tools, 2026-Q2.
  • Moldova19.1% of the working-age population using AI tools, 2026-Q2.
  • Mongolia17.9% of the working-age population using AI tools, 2026-Q2.
  • MontenegroPolicy instruments for Montenegro are not yet indexed. This does not mean the country has no AI policy.
  • Morocco12.0% of the working-age population using AI tools, 2026-Q2.
  • Mozambique11.4% of the working-age population using AI tools, 2026-Q2.
  • Myanmar10.4% of the working-age population using AI tools, 2026-Q2.
  • Namibia15.4% of the working-age population using AI tools, 2026-Q2.
  • Nepal14.6% of the working-age population using AI tools, 2026-Q2.
  • Netherlands43.5% of the working-age population using AI tools, 2026-Q2.
  • New CaledoniaPolicy instruments for New Caledonia are not yet indexed. This does not mean the country has no AI policy.
  • New Zealand44.1% of the working-age population using AI tools, 2026-Q2.
  • Nicaragua12.5% of the working-age population using AI tools, 2026-Q2.
  • Niger10.4% of the working-age population using AI tools, 2026-Q2.
  • Nigeria10.4% of the working-age population using AI tools, 2026-Q2.
  • North KoreaPolicy instruments for North Korea are not yet indexed. This does not mean the country has no AI policy.
  • Norway49.4% of the working-age population using AI tools, 2026-Q2.
  • Oman27.6% of the working-age population using AI tools, 2026-Q2.
  • Pakistan11.9% of the working-age population using AI tools, 2026-Q2.
  • PalestinePolicy instruments for Palestine are not yet indexed. This does not mean the country has no AI policy.
  • Panama24.2% of the working-age population using AI tools, 2026-Q2.
  • Papua New Guinea7.6% of the working-age population using AI tools, 2026-Q2.
  • Paraguay12.6% of the working-age population using AI tools, 2026-Q2.
  • Peru17.3% of the working-age population using AI tools, 2026-Q2.
  • Philippines20.7% of the working-age population using AI tools, 2026-Q2.
  • Poland32.3% of the working-age population using AI tools, 2026-Q2.
  • Portugal27.4% of the working-age population using AI tools, 2026-Q2.
  • Puerto RicoPolicy instruments for Puerto Rico are not yet indexed. This does not mean the country has no AI policy.
  • Qatar43.5% of the working-age population using AI tools, 2026-Q2.
  • Romania18.0% of the working-age population using AI tools, 2026-Q2.
  • Russia9.9% of the working-age population using AI tools, 2026-Q2.
  • Rwanda7.6% of the working-age population using AI tools, 2026-Q2.
  • S. Sudan8.0% of the working-age population using AI tools, 2026-Q2.
  • Saudi Arabia31.4% of the working-age population using AI tools, 2026-Q2.
  • Senegal14.4% of the working-age population using AI tools, 2026-Q2.
  • Serbia25.3% of the working-age population using AI tools, 2026-Q2.
  • Sierra Leone10.4% of the working-age population using AI tools, 2026-Q2.
  • Slovakia27.1% of the working-age population using AI tools, 2026-Q2.
  • Slovenia30.1% of the working-age population using AI tools, 2026-Q2.
  • Solomon Is.Policy instruments for Solomon Is. are not yet indexed. This does not mean the country has no AI policy.
  • Somalia8.0% of the working-age population using AI tools, 2026-Q2.
  • South Africa23.9% of the working-age population using AI tools, 2026-Q2.
  • South Korea40.6% of the working-age population using AI tools, 2026-Q2.
  • Spain45.1% of the working-age population using AI tools, 2026-Q2.
  • Sri Lanka7.6% of the working-age population using AI tools, 2026-Q2.
  • Sudan8.0% of the working-age population using AI tools, 2026-Q2.
  • Suriname11.3% of the working-age population using AI tools, 2026-Q2.
  • Sweden37.3% of the working-age population using AI tools, 2026-Q2.
  • Switzerland39.1% of the working-age population using AI tools, 2026-Q2.
  • Syria7.4% of the working-age population using AI tools, 2026-Q2.
  • Taiwan33.8% of the working-age population using AI tools, 2026-Q2.
  • Tajikistan6.3% of the working-age population using AI tools, 2026-Q2.
  • Tanzania8.0% of the working-age population using AI tools, 2026-Q2.
  • Thailand13.4% of the working-age population using AI tools, 2026-Q2.
  • Timor-LestePolicy instruments for Timor-Leste are not yet indexed. This does not mean the country has no AI policy.
  • Togo10.4% of the working-age population using AI tools, 2026-Q2.
  • Trinidad and TobagoPolicy instruments for Trinidad and Tobago are not yet indexed. This does not mean the country has no AI policy.
  • Tunisia13.9% of the working-age population using AI tools, 2026-Q2.
  • Turkey18.9% of the working-age population using AI tools, 2026-Q2.
  • Turkmenistan6.3% of the working-age population using AI tools, 2026-Q2.
  • Uganda8.0% of the working-age population using AI tools, 2026-Q2.
  • Ukraine9.5% of the working-age population using AI tools, 2026-Q2.
  • United Arab Emirates73.3% of the working-age population using AI tools, 2026-Q2.
  • United Kingdom43.7% of the working-age population using AI tools, 2026-Q2.
  • United States33.0% of the working-age population using AI tools, 2026-Q2.
  • Uruguay25.7% of the working-age population using AI tools, 2026-Q2.
  • Uzbekistan7.6% of the working-age population using AI tools, 2026-Q2.
  • VanuatuPolicy instruments for Vanuatu are not yet indexed. This does not mean the country has no AI policy.
  • Venezuela11.3% of the working-age population using AI tools, 2026-Q2.
  • Vietnam27.9% of the working-age population using AI tools, 2026-Q2.
  • W. SaharaPolicy instruments for W. Sahara are not yet indexed. This does not mean the country has no AI policy.
  • YemenPolicy instruments for Yemen are not yet indexed. This does not mean the country has no AI policy.
  • Zambia13.3% of the working-age population using AI tools, 2026-Q2.
  • Zimbabwe8.9% of the working-age population using AI tools, 2026-Q2.

Glossary 30

  • AGI Artificial general intelligence. There is no agreed definition and no agreed test, so this site does not report a distance to it. The alignment page lists the measured quantities that bear on the question.
  • API access A release in which a model can be queried over a network but its weights are not distributed. One of three release types recorded for notable models.
  • Benchmark saturation The point at which the best recorded score on a test stops separating leading models. This site uses 90 per cent of the maximum score as the threshold and reports the month it was first passed.
  • Binding instrument A law, regulation or other instrument that creates legal obligations. Distinguished on this site from guidance, standards and voluntary commitments, which are recorded separately.
  • Compute threshold A quantity of training compute written into a statute, executive action or official policy document, above which additional obligations apply. Recorded here with the instrument that sets it and the number of models above it.
  • Dangerous-capability evaluation A test of whether a model gives meaningful help with chemical, biological, cyber or autonomous harm. Counted here only where a card reports results for that model.
  • Facility power The whole electrical capacity of a datacentre: IT power plus cooling, lighting and power conversion.
  • FLOP Floating-point operation. Training compute is reported as the total number of operations used in the final training run, on a logarithmic scale, so each axis step is ten times the one below it.
  • Frontier model A model at or near the highest recorded training compute at its date of publication. The term has no standard definition; where this site plots a frontier it means the running maximum over all models published to that date.
  • Frontier safety framework A document in which a developer states the capability levels at which it will take specified action. The frameworks are self-written and self-assessed, and share no common scale.
  • Governance readiness A score compiled here from primary instruments, recording the machinery a jurisdiction has for governing AI. It records presence, not quality or enforcement.
  • H100 equivalent A unit of compute capacity: the number of Nvidia H100 chips with the same total performance as the chips a site actually holds. Written H100e.
  • Hourly carbon-free energy The share of electricity that was carbon-free in the same hour and on the same grid in which it was used. Google is the only data-centre operator that publishes it.
  • IT power The electrical power available to a datacentre's computing equipment, before cooling, power conversion and other overheads. Recorded per site as the power operating now and the power planned at the last stage of the site's timeline.
  • Location-based emissions Emissions from electricity counted at the average carbon intensity of the grid a data centre drew from, whatever contracts the operator holds.
  • Market-based emissions Emissions from electricity after subtracting renewable energy the operator bought by contract. A company matching all its annual use with renewable purchases can report close to zero while its data centres draw fossil power at some hours.
  • Notable model A model meeting Epoch AI's notability criteria, which is the population used for every model count on this site.
  • Open weights A release in which the trained parameters are distributed and can be run by a third party. It does not imply an open-source licence, and the licence attached to a given release may restrict use.
  • Pre-release testing Testing of a model by a party other than its developer before public release. Counted here only where a card names a government institute that tested a pre-release version of that model.
  • Provenance marker The marker beside a figure on this site. It opens the source, publisher, licence, update cadence, retrieval date and stated limitations for that figure.
  • Renewable matching Buying as much renewable electricity over a year as a company uses, through contracts or certificates. Says nothing about whether the power used in a given hour, on a given grid, was clean.
  • Reported incident An entry in the AI Incident Database, which records publicly reported harms involving AI systems. Entries are reports, not adjudicated findings, and the database is not a complete record.
  • Running maximum The highest value recorded up to each date. It rises or holds level and never falls, so a line fitted through one cannot slope downward.
  • Soft law Guidance, standards, codes of practice and voluntary commitments that create no legal obligation. Indexed here alongside binding instruments and tagged by force.
  • Sui generis database right An EU right, separate from copyright, protecting substantial investment in obtaining or presenting a database. It is the reason several third-party policy trackers cannot be bulk-extracted for use here.
  • System card A document a developer publishes for a model, describing its evaluations, safeguards and limitations. Also called a model card or safety report. Dated here by when it first appeared.
  • Systemic risk threshold The training-compute level in Article 51(2) of the EU AI Act at which a general-purpose model is presumed to carry systemic risk. It is a rebuttable presumption rather than a fixed boundary.
  • Token The unit a language model reads and writes: a word, part of a word, or a piece of an image or audio. Each company splits text differently, so a token count from one company is not the same unit as another's.
  • Tokens per month A disclosed token volume converted to a month by this site, so that per-minute, per-day and per-year figures can be drawn on one axis. A quarterly or annual total becomes a monthly average and is marked as one.
  • Working-age population The denominator used for the population adoption figures, as defined by the publisher of that dataset. Adoption shares on this site are shares of that population, not of total population.

Laws and policies 86

Model releases 79

Datacentres 108