Epoch AI describes itself as “a data-first research nonprofit investigating the future of artificial intelligence.” It began in 2021 as a volunteer group collecting and analyzing data on AI models and formally incorporated as Epoch AI after the success of its 2022 paper on trends in training compute. Its stated aim is to “improve society’s understanding of the drivers, progress, and impact of artificial intelligence” by building a shared scientific foundation that is neutral to any specific agenda.
The organization’s work is organized around three questions: the drivers of AI progress, such as compute and data bottlenecks; the measurement of AI capabilities over time through independent evaluations; and the economic and societal impacts of the technology. To support this, Epoch maintains widely used datasets covering machine-learning hardware, training-compute estimates, GPU cluster and frontier data-center tracking, model capability benchmarks, and AI investment and adoption data.
Epoch is best known in the wider public for FrontierMath, a benchmark of very hard, original mathematics problems used to probe the limits of frontier models. Its compute and trend estimates are cited heavily by researchers, journalists, and policymakers, and the organization has partnered with Stanford’s AI Index and the UK government’s science and innovation department.
Why business readers should care: Epoch AI is one of the few neutral sources of hard numbers on how fast AI is scaling and how much it costs, making its datasets a common reference for anyone trying to forecast where the technology is headed.