10^25 FLOP threshold checker

Article 51(2) of the EU AI Act presumes that a general-purpose AI model has 'systemic risk' if the cumulative compute used to train it exceeds 10^25 floating-point operations. That's 10 followed by 24 zeroes. In practical terms, GPT-4 is estimated at roughly 2×10^25 FLOPs (above the threshold), Llama 3 70B is around 6×10^24 FLOPs (below). The threshold targets the frontier of model scale: roughly the top 5-10 foundation models in the world at any time.

Article 55 obligations attach: (1) model evaluations against state-of-the-art methods, (2) adversarial testing including red-teaming, (3) serious-incident reporting to the AI Office under Article 55(1)(c) on the same 2/10/15 day clock as Article 73, (4) cybersecurity protections appropriate to the model. Plus an obligation to keep documentation that can be requested by the Commission. The voluntary GPAI Code of Practice operationalises all of these. Full Commission enforcement starts 2 August 2026.

Yes. Article 51(1)(a) allows the Commission to designate any GPAI model as having systemic risk based on capabilities, even if it falls below the compute threshold. This is the safety valve for models that don't train at frontier scale but achieve frontier capabilities (e.g. via algorithmic improvements, fine-tuning, or specialised architecture). In practice, designation is reserved for cases where capability tests demonstrate frontier-level performance.

Article 51(2) talks about cumulative training compute. The Commission's draft guidance treats pre-training as the primary anchor, with substantial fine-tuning runs added in. Lightweight RLHF or LoRA fine-tuning on top of an existing model doesn't typically push the cumulative compute over 10^25 unless the base model was already near threshold. Provider documentation should account for the full pre-training run and any material continued-training or fine-tuning rounds.

Not directly. Article 51 + Article 55 are GPAI-provider obligations. Deployers don't have to count FLOPs. What deployers DO need to know: GPAI providers must hand them downstream documentation under Article 53(1)(b), and if a provider's model has systemic risk, that documentation cascade includes the systemic-risk evaluation results. So as a deployer of a GPT-class model, you should expect more detailed safety + evaluation documentation than for a sub-threshold model.