llms.txt
Compare with other sites' llms.txt files →# Epoch AI
> Epoch AI is an independent nonprofit research institute that tracks and forecasts AI development through empirical, data-driven methods. Founded in 2021, Epoch maintains the most comprehensive public databases on AI training compute, model capabilities, hardware, data centers, and chip markets. Our research is cited by the Stanford AI Index, the UK government, and major policy institutions. All data is freely available under CC BY 4.0.
## Datasets
- [AI models](https://epoch.ai/data/ai-models): Database of 3,200+ ML systems from 1950 to present, tracking training compute (FLOP), parameters, dataset size, cost, and hardware. [CSV download](https://epoch.ai/data/all_ai_models.csv). [Documentation](https://epoch.ai/data/ai-models-documentation).
- [Benchmarks](https://epoch.ai/benchmarks): Performance data for frontier AI models across 40+ evaluations, including MMLU, GPQA, MATH, and coding benchmarks. [ZIP download](https://epoch.ai/data/benchmark_data.zip).
- [Epoch Capabilities Index (ECI)](https://epoch.ai/eci): A single aggregated capability score across multiple benchmarks.
- [FrontierMath](https://epoch.ai/frontiermath): A benchmark of extremely challenging mathematics problems for evaluating frontier model reasoning.
- [AI data centers](https://epoch.ai/data/ai-data-centers): AI data centers tracked via satellite imagery and permit data, with power capacity, location, and construction timelines. [CSV download](https://epoch.ai/data/data_centers/data_centers.csv). [ZIP download](https://epoch.ai/data/data_centers/data_centers.zip).
- [GPU clusters](https://epoch.ai/data/gpu-clusters): 500+ large-scale compute clusters used for AI training and deployment. [CSV download](https://epoch.ai/data/gpu_clusters.csv).
- [AI chip sales](https://epoch.ai/data/ai-chip-sales): Semiconductor shipment volumes and revenue across major vendors (NVIDIA, AMD, Intel, etc.). [Chip types CSV](https://epoch.ai/data/ai_chip_sales_chip_types.csv). [Organizations CSV](https://epoch.ai/data/ai_chip_sales_organizations.csv). [Timelines by chip CSV](https://epoch.ai/data/ai_chip_sales_timelines_by_chip.csv). [ZIP download](https://epoch.ai/data/ai_chip_sales.zip).
- [AI chip owners](https://epoch.ai/data/ai-chip-owners): Distribution of global AI chip capacity across major companies. [Cumulative by designer CSV](https://epoch.ai/data/ai_chip_owners_cumulative_by_designer.csv). [Quarters by chip type CSV](https://epoch.ai/data/ai_chip_owners_quarters_by_chip_type.csv). [Cumulative by chip type CSV](https://epoch.ai/data/ai_chip_owners_cumulative_by_chip_type.csv). [ZIP download](https://epoch.ai/data/ai_chip_owners.zip).
- [AI chip users](https://epoch.ai/data/ai-chip-users): Estimates of how much compute the leading AI developers (OpenAI, Anthropic, Google DeepMind, Meta Superintelligence Labs, SpaceXAI) use, in H100-equivalents with 90% credible intervals. [Year-end by lab CSV](https://epoch.ai/data/ai_chip_users_year_end_by_lab.csv). [Model intermediates CSV](https://epoch.ai/data/ai_chip_users_intermediates_by_lab.csv). [ZIP download](https://epoch.ai/data/ai_chip_users.zip).
- [AI chip components](https://epoch.ai/data/ai-chip-components): Estimates of the advanced logic wafer capacity, CoWoS packaging, and HBM memory consumed by leading AI chip designers. [ZIP download](https://epoch.ai/data/ai_chip_components.zip).
- [AI chip imports](https://epoch.ai/data/ai-chip-imports): Where AI chips go, estimated from national customs records: quarterly and cumulative trade value, chip units, and H100-equivalents per destination country since 2022. [By country CSV](https://epoch.ai/data/ai_chip_trade/ai_chip_trade_by_country.csv). [By chip designer and chip type CSV](https://epoch.ai/data/ai_chip_trade/ai_chip_trade_by_country_breakdown.csv). [ZIP download](https://epoch.ai/data/ai_chip_trade/ai_chip_trade.zip).
- [AI companies](https://epoch.ai/data/ai-companies): Revenue, funding, staffing, and compute capacity for frontier AI organizations. [CSV download](https://epoch.ai/data/ai_companies.csv). [ZIP download](https://epoch.ai/data/ai_companies.zip).
- [Machine learning hardware](https://epoch.ai/data/machine-learning-hardware): Catalog of 170+ AI accelerators (GPUs, TPUs, custom ASICs) with specs and pricing. [CSV download](https://epoch.ai/data/ml_hardware.csv). [ZIP download](https://epoch.ai/data/ml_hardware.zip).
- [Biology AI models](https://epoch.ai/data/biology-ai-models-documentation): Machine learning models trained on biological data, with training compute and model details. [CSV download](https://epoch.ai/data/biology_ai_models.csv).
- [Polling on AI usage](https://epoch.ai/data/polling): Survey data on AI adoption patterns across demographics and organizations. [CSV download](https://epoch.ai/data/polling/polling_on_ai_usage_oct_2026.csv). [July 2026 wave CSV](https://epoch.ai/data/polling/polling_on_ai_usage_jul_2026.csv). [March 2026 wave CSV](https://epoch.ai/data/polling/polling_on_ai_usage_mar_2026.csv).
- [ChatGPT usage](https://epoch.ai/data/chatgpt-usage): Our database of aggregate metadata from recorded ChatGPT histories of 5,000 US YouGov panelists from December 2022 to December 2025. [Excel workbook](https://epoch.ai/data/chat_stats/chat_stats.xlsx). [Monthly CSV](https://epoch.ai/data/chat_stats/chat_stats_monthly.csv). [Weekly CSV](https://epoch.ai/data/chat_stats/chat_stats_weekly.csv). [Monthly demographic subgroups CSV](https://epoch.ai/data/chat_stats/chat_stats_monthly_subgroups.csv). [Documentation](https://epoch.ai/data/chatgpt-usage-documentation).
- [Cyber vulnerabilities](https://epoch.ai/data/cve): Trends in software and hardware vulnerability (CVE) counts and severity since 2020, broken down by reporting organization. [Data source](https://github.com/CVEProject/cvelistV5).
- [Frontier lab hiring](https://epoch.ai/data/frontier-lab-hiring): Daily archive of Anthropic's and OpenAI's job postings, with open roles by category over time and every posting's lifetime and advert text. [ZIP download](https://epoch.ai/data/frontier_lab_hiring.zip).
- [AI use in math research](https://epoch.ai/data/arxiv): Explorer tracking AI use acknowledgments in math research papers submitted to arXiv. [CSV download](https://epoch.ai/data/math_acknowledgment_trends_monthly.csv.gz).
- [AI use in research](https://epoch.ai/data/ai-research-use): Explorer tracking large language model use acknowledgments in physics, statistics, and math preprints on arXiv and biology preprints on bioRxiv. [Monthly trends CSV](https://epoch.ai/data/ai_research_use_monthly.csv). [Use cases CSV](https://epoch.ai/data/ai_research_use_cases_monthly.csv). [Biology AI tools CSV](https://epoch.ai/data/ai_research_tools_monthly.csv). [Documentation](https://epoch.ai/data/ai-research-use-documentation).
## Research
- [Papers & reports](https://epoch.ai/blog): Peer-reviewed and working papers on AI compute trends, scaling laws, hardware economics, and capability forecasting.
- [Data Insights](https://epoch.ai/data-insights): Short-form analyses and visualizations of AI trends.
- [Gradient Updates](https://epoch.ai/gradient-updates): Newsletter with expert commentary on important questions in AI.
- [Trends in AI](https://epoch.ai/trends): Key numbers on the growth and impact of artificial intelligence.
## Key research findings
- Training compute for frontier language models has grown at roughly 5x per year since 2020 (90% CI: 4x-6x). The largest known training run to date is Grok 4, at approximately 5x10^26 FLOP.
- The cost of LLM inference at a fixed performance milestone has fallen by between 9x and 900x per year, depending on the milestone.
- The largest known AI data center (Anthropic-Amazon New Carlisle) has computing power equivalent to roughly 700,000 NVIDIA H100 chips, with about 1.1 GW of power capacity and $35 billion in capital costs.
- As of Q4 2025, five hyperscalers (Amazon, Google, Meta, Microsoft, and Oracle) collectively hold around 71% of global AI compute, up from 63% in Q1 2024.
## About
- [About Epoch AI](https://epoch.ai/about): Mission, history, and approach.
- [Team](https://epoch.ai/about/team): Researchers, engineers, and operations staff.
- [Consultations](https://epoch.ai/about/consultations): Commission custom research or expert briefings.
- [Transparency](https://epoch.ai/about/transparency): Funders, collaborations, and independence principles.
- [Press](https://epoch.ai/about/press): Resources for journalists.
- [Contact](https://epoch.ai/contact): General inquiries and data feedback (data@epoch.ai).
## Optional
- [Podcast: Epoch After Hours](https://epoch.ai/epoch-after-hours)
- [Careers](https://epoch.ai/about/careers)
- [Privacy policy](https://epoch.ai/privacy)
- [Donate](https://epoch.ai/about/donate)