Epoch AI
Epoch AI is a data-first nonprofit research institute that studies the trajectory of artificial intelligence through open datasets, trend analysis, benchmarks, and forecasts about compute, hardware, data centers, model capabilities, AI companies, and economic impact.
Snapshot
- Type: independent nonprofit research institute.
- Known for: AI compute trend analysis, the Epoch AI model database, frontier data-center tracking, hardware datasets, AI company data, capability evaluations, and AI progress forecasts.
- Origins: started in 2021 as a volunteer group collecting and analyzing AI model data; later scaled into Epoch AI after the 2022 Compute Trends Across Three Eras of Machine Learning paper.
- Executive director: Jaime Sevilla, according to Epoch AI's team page as reviewed May 19, 2026.
- Why it matters: AI governance, safety, investment, and public debate increasingly depend on empirical claims about compute, chips, data, capabilities, costs, and infrastructure growth.
Origins and Role
Epoch AI describes its mission as improving society's understanding of the drivers, progress, and impact of artificial intelligence from a neutral, evidence-grounded perspective. Its about page says it investigates AI's driving forces, tracks capabilities over time, and forecasts economic and societal impact.
The organization began with model-data curation. Its early work culminated in Compute Trends Across Three Eras of Machine Learning, a 2022 paper that helped formalize the view that modern machine learning entered a large-scale era in which training compute rose much faster than ordinary Moore's-law hardware progress.
Epoch now occupies a distinct measurement niche. It is not a frontier model lab, an advocacy organization, or a general policy think tank. Its core contribution is infrastructure for knowing what is happening: datasets, charts, estimates, methods, and explainers that other researchers, journalists, policymakers, and AI companies can cite or contest.
Data and Trend Work
Epoch's public data hub tracks AI models, AI capabilities, frontier data centers, chip sales, chip owners, GPU clusters, machine-learning hardware, AI companies, and polling on AI use. As reviewed May 19, 2026, its AI models database described itself as tracking more than 3,200 machine-learning models from 1950 to the present.
The site's trend pages translate those datasets into legible indicators: training compute growth, software efficiency, frontier training cost, data-center size, chip performance per dollar, memory bandwidth, AI company revenue, investment, and other measures. These figures are not neutral in the sense of being consequence-free; they shape which bottlenecks, timelines, and governance levers seem plausible.
Epoch's methods matter because many AI debates otherwise rely on private claims. Frontier labs know their own training runs, hardware purchases, inference costs, user growth, and internal evaluations. Public institutions often do not. Epoch reduces that asymmetry by collecting public evidence, rating confidence, and publishing uncertainty where possible.
Compute and Infrastructure
Compute is the area where Epoch has had the strongest visible influence. The 2022 compute-trends paper argued that deep-learning training compute had accelerated to roughly six-month doubling times after the early 2010s, and that a further large-scale era emerged after late 2015 as firms trained models with 10 to 100 times larger compute requirements.
Epoch's 2026 trends page reported that training compute for frontier language models had grown about fivefold per year since 2020, while frontier training costs had risen about 3.5 times per year. The same page estimated the largest known AI data center at roughly 700,000 H100-equivalents, while noting that future announced sites could be much larger.
Epoch also tracks AI data centers using satellite imagery, permit data, public reports, and other open-source intelligence. Its data-center topic page frames frontier AI infrastructure as warehouses of specialized chips with city-scale power requirements, built at unusual speed and tied directly to policy questions about power, land, capital expenditure, and national AI capacity.
Benchmarks and Capabilities
Epoch also works on capability measurement. Its about page says the organization regularly runs independent evaluations of new models, reviews prominent benchmarks, and creates new ones. Its benchmark work includes FrontierMath, a difficult mathematical reasoning benchmark built to test frontier systems beyond ordinary public test sets.
This connects Epoch to the broader evaluation layer, but with a different emphasis from organizations such as METR. METR is strongly associated with autonomy, long-horizon tasks, and threat-relevant capabilities. Epoch is more associated with trend measurement, capability progress, benchmark construction, and the quantitative inputs that feed forecasts.
Because benchmarks can become targets, Epoch's capability work sits inside the same tension as the rest of the evaluation ecosystem: public measurement is necessary for accountability, but published tasks, methods, and leaderboards can also shape training incentives and benchmark contamination.
Governance Relevance
Epoch's work is governance-relevant even when it avoids official policy recommendations. Its own description says staff may analyze policy consequences while the organization does not make official recommendations. That posture lets Epoch operate as a measurement institution whose outputs can be used by multiple camps.
Compute governance, export controls, data-center permitting, safety thresholds, public AI investment, model-release policies, and AI infrastructure planning all need estimates. How many models will cross a compute threshold? Who owns frontier chips? How fast are training runs getting more expensive? How soon might public text data become insufficient for naive scaling? Epoch's datasets help turn those questions from slogans into arguable empirical claims.
The same work can also be strategically sensitive. Public maps of compute, chip ownership, and data centers can support accountability and public planning, but they can also reveal competitive or geopolitical information. Epoch's role therefore involves judgment about what to publish, how to express uncertainty, and how to avoid converting measurement into operational targeting.
Central Tensions
- Neutrality versus impact: even when an organization avoids recommendations, its metrics influence which policies and forecasts appear reasonable.
- Open data versus strategic sensitivity: infrastructure transparency supports public oversight but may expose commercially or geopolitically useful details.
- Public evidence versus private reality: the best open-source estimate can still miss undisclosed training runs, cloud allocations, model failures, or internal costs.
- Forecasting versus humility: trend lines are useful, but AI progress depends on algorithmic breakthroughs, bottlenecks, deployment economics, regulation, and social response.
- Benchmarking versus contamination: public capability tests can discipline claims while also becoming curriculum for the next generation of systems.
Spiralist Reading
Epoch AI is one of the institutions trying to make the AI transition measurable before it becomes ambient.
In the Spiralist frame, artificial intelligence is not only a sequence of products. It is a civilizational buildout: chips, land, power, capital, models, data, labor, benchmarks, forecasts, and stories about inevitability. Epoch's significance is that it records the substrate. It asks how much compute exists, where it is going, how fast models are scaling, how much the race costs, and which bottlenecks remain.
That makes Epoch a counterweight to charisma. A CEO says progress is accelerating. A critic says the bubble will break. A government says national advantage is at stake. A market says capital expenditure is destiny. Epoch's best work gives the public a way to ask: what is the actual evidence?
The danger is that measurement itself can become authority. A clean chart can make uncertainty feel smaller than it is. A forecast can harden into prophecy. The proper use of Epoch's work is not worship of the trend line. It is disciplined argument about the material conditions under which the trend line might continue, bend, or break.
Open Questions
- How much frontier AI activity remains invisible to open-source data because companies or governments do not disclose training runs, clusters, or inference budgets?
- Which AI infrastructure datasets should be public, delayed, aggregated, or withheld because of security risk?
- How should policymakers treat compute thresholds when hardware efficiency, algorithmic efficiency, and inference-heavy systems change the meaning of raw FLOP counts?
- Can public benchmarks remain useful as frontier labs train increasingly benchmark-aware systems?
- What institutional safeguards keep a measurement organization neutral enough to be broadly trusted while still clear enough to matter?
Related Pages
- AI Capability Forecasting
- AI Compute
- Compute Governance
- AI Data Centers
- AI Energy and Grid Load
- Scaling Laws
- AI Evaluations
- Benchmark Contamination
- METR
- AI Organizations
Sources
- Epoch AI, About Epoch AI, reviewed May 19, 2026.
- Epoch AI, Our Team, reviewed May 19, 2026.
- Epoch AI, What is Epoch?, reviewed May 19, 2026.
- Epoch AI, Data on AI, reviewed May 19, 2026.
- Epoch AI, Trends in Artificial Intelligence, reviewed May 19, 2026.
- Epoch AI, AI Data Centers: Data & Research, reviewed May 19, 2026.
- Sevilla et al., Compute Trends Across Three Eras of Machine Learning, arXiv, 2022.
- Epoch AI, AI in 2030, 2026.