25
AI models trained with more than 10²⁵ FLOP, the EU AI Act's systemic-risk threshold (18 US, 5 China)
What the chart shows
The EU AI Act presumes that a general-purpose model trained with more than 10²⁵ FLOP poses systemic risk, which brings extra duties. Epoch AI counts 25 models above that line so far: 18 from US organisations, 5 from Chinese ones and 2 from elsewhere.
The first was Gemini 1.0 Ultra (2023); 7 have been released in 2026 so far. Many of these figures are Epoch’s estimates, because developers rarely publish their training compute.
Key numbers
| Measure | Value |
|---|---|
| United States | 18 |
| China | 5 |
| Rest of the world | 2 |
| Released in 2026 so far | 7 |
| First above it: Gemini 1.0 Ultra (2023) | 5.0×10^25 |
How to read it
- Models with an Epoch AI training-compute estimate above 10²⁵ FLOP, from its large-scale models database; a model from organisations in several countries counts once for each group.
- The EU AI Act threshold is in Article 51; the Commission can change it.
- The current year is incomplete.
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.
Updated 3 Nov 2026: Updated from Epoch AI's data read 2026-11-03: the key number went from 24 to 25. Every chart change →
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 3 November 2026.
Updated 3 Nov 2026: Updated from Epoch AI's data read 2026-11-03: the key number went from 24 to 25. Every chart change →
Download the data (CSV) · Cite this chart
Discussion