Epoch AI is a research institute dedicated to investigating the key trends and issues that will shape the trajectory and governance of artificial intelligence, with the mission "to investigate the trajectory of AI for the good of society."
Purpose: The organization provides independent research and reliable data on AI development trends, helping researchers, policymakers, and the public understand model capabilities, compute, training costs, Scaling Laws, and the long-term direction of AI development.
Main capabilities: Epoch AI tracks and publishes several key trend indicators, including: model capabilities (since the introduction of reasoning models, the ECI frontier has improved by 14 points per year), growth in the stock of compute (total AI chip compute is growing at 3.4x per year, doubling every 6.8 months), training compute (since 2020, the training compute of frontier language models has grown at 5x per year, doubling every 5.2 months), data center construction (gigawatt-scale data centers can be built in about 2 years; the largest known AI data center has computing power equivalent to 1.1 million NVIDIA H100 chips), and chip price-performance (since 2023, AI chip performance per dollar has grown by 49% per year, doubling every 1.7 years). In addition, the organization maintains open databases covering more than 3,200 machine learning models (from 1950 to the present), model performance on key benchmarks, large AI data centers (using satellite and permitting data to show AI compute, power usage, and construction timelines), chip shipments and revenue, and more.
How to use: Users can access these databases and visualizations through its website (some interactive visualizations require JavaScript to be enabled). The organization also provides the latest research updates and weekly AI news commentary, and accepts custom research and consulting services. For deeper understanding, you can contact its team.
Applicable scenarios: Suitable for AI researchers, policymakers, industry analysts, journalists, and organizations and individuals interested in long-term AI trends, helping with data-driven analysis and decision-making.
Limitations: Some data coverage varies by manufacturer; for example, data for Nvidia and Google starts in 2022, while other vendors start in 2024; some estimates are inferred based on revenue data, financial disclosures, and analyst reports. Interactive visualizations require JavaScript to be enabled. The specific content, fees, and prices of custom research and consulting services are not provided in the materials.
