- Lead. Starting August 2, 2026, the European Commission’s AI Office can formally demand documentation, restrict AI models from the EU market, and levy fines against providers of general-purpose AI — converting obligations that have been on the books since August 2025 into enforceable rules with real financial consequences.
- Fact. Fines reach 3% of total worldwide annual turnover or €15 million, whichever is higher; four independent violation routes exist, meaning a single non-compliant provider could face multiple concurrent penalties — applied to companies whose global revenue runs into the hundreds of billions of dollars.
- Stake. Any general-purpose AI model placed on the EU market on or after August 2, 2025 faces immediate enforcement scrutiny with no grace period, a category that covers essentially all current flagship language models from OpenAI, Google DeepMind, Anthropic, Meta, and xAI.
What changes on August 2
The EU AI Act’s GPAI provisions — covering documentation requirements, copyright compliance, systemic-risk assessments, and transparency obligations — came into legal force on August 2, 2025. What was missing until now was the Commission’s ability to act on non-compliance. From August 2, 2026, the AI Office gains the full enforcement toolkit: it can request records, conduct technical evaluations of models, order corrective measures, restrict or withdraw a model from EU availability, and issue fines.
The fine structure has four independent legal routes: violating the substantive GPAI rules; ignoring documentation requests; refusing model access for evaluation; and failing to implement ordered corrections. Each can trigger a separate penalty. For a company like Google or OpenAI, 3% of worldwide annual turnover represents a potential exposure of several billion dollars per incident.
Which companies are most exposed
The hardest line in the regulation targets models exceeding 10²⁵ floating-point operations of training compute — a threshold the AI Office treats as presumptive evidence of systemic risk, triggering additional assessment obligations. As of mid-2026, that threshold is believed to capture the largest GPT, Gemini, Claude, and Llama model families, though training-compute disclosures are not standardised and the AI Office has not published a public list.
Non-EU providers must appoint an authorised EU representative by August 2 — a legal requirement that serves as a point of contact for enforcement proceedings. Companies that have not done so are in immediate violation from the first day of enforcement.
The Code of Practice buffer
The AI Office gave companies a path to mitigated enforcement: the voluntary GPAI Code of Practice, to which a number of major AI providers have signed commitments. Signatories have their Code obligations weighed when a fine is calculated and receive good-faith treatment from the Office during any investigation. However, signatures do not block enforcement — a provider can be a Code signatory and still be fined if it has not met its specific commitments.
The EU AI Act’s GPAI enforcement deadline has been on industry compliance calendars for most of 2026. The question now shifts from preparation to execution: which companies have documentation systems, copyright-compliance records, and systemic-risk assessments in place — and which ones are hoping the AI Office’s initial focus will fall on competitors rather than themselves.
What enforcement looks like in practice
The AI Office has not published a formal enforcement priority list, but Commission officials have signalled that monitoring will focus initially on Code signatories’ adherence to their stated commitments. Non-signatories face the full statutory framework from day one. Models already on the market before August 2, 2025 have a two-year compliance runway — until August 2, 2027 — offering some older-generation models a near-term reprieve.
The first formal enforcement actions, if any, are unlikely to produce fines immediately; AI Act proceedings follow an investigation-and-remediation structure that takes months. But the shift from voluntary standards to enforceable law marks a durable change in the operating environment for every AI lab distributing models in the EU — the world’s largest single regulatory market for digital services.