What is FLOP – the number of floating-point operations (FLOP) of cumulative compute used during a model’s training?
What is the limit as per the current EU AI act – 1025 FLOP
What does it mean if a model crosses this limit – If a general-purpose AI model is trained using more than 1025 FLOP, it is automatically presumed to have “high-impact capabilities” and is classified as “AI model with systemic risk” under the Act.
This would mean that these models will be subject to the heightened compliance and transparency obligations.
Estimated FLOP usage for some of the publicly available models (some of the numbers are estimated based on the scaling laws)
| Model | Estimated Training FLOP Usage |
| Grok 4 | ~6 × 10²⁷ FLOP |
| O3 | 8 × 10²⁵ – 4 × 10²⁶ FLOP |
| Claude Opus 4 | 1.5 × 10²⁶ FLOP |
FLOP limit for systemic risk under the EU AI Act applies from 2 August 2025. Let’s see if these are revised given the pace of change, and most of the newer models exceed these, some by a factor of two.
Thinking Space 