Epoch AI

Epoch AI is a non-profit research organization founded in April 2022 and headquartered in San Francisco, California, USA. The company consists of a founding team of seven people including Jaime Sevilla, Tamay Besiroglu, Lennart Heim, Pablo Villalobos, Edu Roldán and Marius Hobbhahn. As of the end of 2022, the team size will be 13 people, equivalent to 11 full-time equivalent positions. The company has received $1.96 million in funding.

Epoch AI’s mission is to investigate the future development of artificial intelligence, including analyzing its drivers and predicting its impact on the economy and society. The company emphasizes creating a scientific environment and reviewing AI-related statements. The company's activities cover trend surveys in AI development and deployment such as computing, algorithms, and training data, database construction, predictive model development, and AI system capability assessment.

The company's research areas include the growth rate of AI training computing (increased 4.4 times annually since 2010), computing power trends (doubling every 10 months since 2019), the cost and power consumption of cutting-edge AI model training (doubling annually), and the proliferation of large-scale AI models (as of June 2025, more than 30 models have reached GPT-4 scale, or more than 10^25 FLOPs).

Epoch AI releases the following datasets, tools, and publications:

AI model database: Contains 443 AI model release records from 2010 to 2025, covering language, vision, multi-modal, biology, robotics and other fields. Each record includes model name, domain, training computational load (FLOPs), training power consumption (W), organization, and release date. The database is accessible via the website and is accompanied by documentation describing power consumption estimation methods.

Data insights and visualization tools: Interactive charts are provided, covering NVIDIA chip production, the number of more than 10^25 FLOP models, computing trends after 2010, computing power comparisons of leading technology companies, model training power consumption trends, and training length trends. These tools are generated based on public data and support users to explore AI hardware and model development indicators.

Blog articles: including "How much does it cost to train frontier AI models" (discussing the cost of training frontier AI models), "Training compute of frontier AI models grows by 4-5x per year" (analyzing the training compute of frontier AI models growing by 4-5x per year), and "Trends in machine learning hardware" (overviewing machine learning hardware trends).

The research paper: "Can AI Scaling Continue Through 2030?" examines four constraints such as power, chip manufacturing, data and latency, and predicts that 2e29 FLOP training operations can be achieved by 2030.

Benchmark Project: FrontierMath, a benchmark containing hundreds of unpublished expert-level math problems that take experts hours to days to solve. The company also maintains a database of machine learning trends, providing visibility into the AI ​​landscape.

Epoch AI's output has been cited by a number of organizations, including Our World in Data, the UK Government, the Dutch Government, the New York Times, Time, The Economist and the Financial Times. The company shares updates via Substack and the X (formerly Twitter) account, known as @EpochAIResearch.