The AI Act has been law in Europe since 2024, but its meaningful effects begin now. The general-purpose AI provisions — transparency mandates, systemic-risk classification thresholds, copyright disclosure, and the obligation to share technical detail with downstream deployers — become enforceable in stages through 2026. The structure is reasonable on paper. The enforcement intensity, particularly around the systemic-risk threshold and copyright disclosure, will decide whether frontier models continue to be developed in Europe or whether the continent becomes an importer.
Key takeaways
- Systemic-risk classification at 10^25 FLOPs covers every leading-edge model.
- Transparency mandates create real but manageable compliance cost.
- Copyright disclosure is the riskiest provision for model developers.
- Code of practice negotiations decide most of the practical impact.
Why the threshold matters
The FLOP threshold defines who falls into the systemic-risk category. It is set low enough that every meaningful model already qualifies and any new frontier model will. That removes the option of escaping the regime by sub-threshold training; it does not remove the option of training elsewhere.
- Systemic-risk obligations. Adversarial testing, model evaluation reporting, incident disclosure.
- Transparency. Watermarking, technical documentation, deployer interface obligations.
- Copyright. Training data disclosure, opt-out compliance demonstration.
Why copyright is the live variable
The copyright disclosure obligation is broad and the EU is the only jurisdiction that requires it at this granularity. Developers face a real choice between disclosure and training elsewhere.
What this does to European-trained models
Mistral, Aleph Alpha and a thin layer of national champions face a structural disadvantage if disclosure pushes copyright costs up. The political will to support them is strong but limited.
Where US developers will land
Train in the US, deploy in Europe under transparency obligations. Compliance is manageable for deployment; it is harder for training.
Compliance cost, simplified
Estimated annual ongoing compliance cost per frontier model developer.
| Obligation | Cost | Hardness |
|---|---|---|
| Systemic-risk testing | $10–30M | Medium |
| Transparency/documentation | $5–15M | Low |
| Copyright disclosure/opt-out | $20–50M+ | High |
| Incident response | $5–10M | Medium |
The Act is workable if enforcement is reasonable and crushing if enforcement is maximalist.
Frequently asked questions
Will the US match these rules?
Not in this form. State-level action and sectoral federal rules are the path; not a horizontal regime.
Does this push training to the US?
At the margin yes; cost differentials remain manageable for well-capitalized players.
What about open-source?
Special provisions soften obligations for open models below the threshold; above it, full obligations apply.
The bottom line
The EU AI Act is the most ambitious horizontal AI regulation in the world. Its impact in 2026 depends entirely on enforcement intensity. Reasonable enforcement makes Europe a competitive deployment hub. Maximalist enforcement makes it an importer of models trained elsewhere.
