62
▲ 77% vs 2024
models over 10²³ FLOP, 2025
20202025
What the chart shows
Training compute of 10²³ FLOP was the frontier in 2020, when GPT-3 used about 3×10²³. By 2025 it was routine: Epoch AI counts 104 models released that year above it, 62 from Chinese organisations, 31 from US ones and 11 from elsewhere.
Chinese developers release many mid-sized models, while the very largest runs are mostly American (see the related chart). 2026 so far: 12 from China and 10 from the US.
Key numbers
| Measure | Value |
|---|---|
| United States, 2025 | 31 |
| China, 2025 | 62 |
| Rest of the world, 2025 | 11 |
| All countries, 2025 | 104 |
| United States, 2023 | 21 |
| China, 2023 | 26 |
How to read it
- Epoch AI’s database of large-scale models: every model with a training-compute estimate above 10²³ FLOP, notable or not. Models whose compute Epoch could not estimate are left out.
- A model from organisations in several countries counts once for each group.
- Full calendar years only.
Source: Epoch AI, 2026: Data on AI models, read 3 November 2026. Chart and text by AI Adoption.
Numbers checked against the published data on 3 November 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 3 November 2026.
Download the data (CSV) · Cite this chart
Discussion