From 2 August 2026 the European Commission acquires direct enforcement powers against providers of general-purpose artificial intelligence models, including the authority to impose administrative fines of up to thirty-five million euro or seven percent of global annual turnover, whichever is higher. The date has been on the AI Act timeline since the regulation entered into force in August 2024, but the operational implications are only now sharpening. For model providers, AI Office case-handlers, and the national authorities that will sit alongside them, the next ten weeks are the window in which compliance posture is being decided rather than performed.
The architecture of the August enforcement window has three distinct components. First, the substantive obligations on GPAI providers, which became applicable in August 2025, are now backed by the Commission’s power to act. These cover transparency duties, including the obligation to publish a sufficiently detailed summary of training content, the maintenance of copyright-compliance policies, and, for models classified as presenting systemic risk, the implementation of additional model-evaluation, adversarial-testing and incident-reporting requirements. Second, the Commission has retained discretion over whether to open formal proceedings, with the AI Office expected to prioritise systemic-risk cases. Third, the voluntary General-Purpose AI Code of Practice, published in July 2025, sits in the middle of this enforcement structure as a presumption-of-conformity mechanism.
The Code’s positioning is the part of the regime that has changed the most since the original political agreement. A voluntary code is a familiar instrument in EU regulation, typically used to operationalise high-level obligations in technical terms. What is novel about the GPAI Code is its function as a procedural anchor for enforcement. Providers that sign and follow the Code benefit from a regulatory assumption that they comply with the corresponding statutory obligations, which the AI Office may rebut only on the basis of contrary evidence. Providers that decline to sign do not face additional substantive obligations but lose the procedural cushion. Their compliance is assessed directly against the text of the regulation, which carries more interpretive uncertainty in transparency provisions and in the boundary between general-purpose models and high-risk systems.
For the providers of the largest foundation models, the calculus around Code adherence has been the dominant compliance topic of the last six months. Major providers have publicly committed to the safety-and-security chapter, which is the area most heavily scrutinised by the AI Office. Adherence to the transparency chapter, which contains the training-content summary template, has been more measured. The template requires categories of data sources and explanations of measures taken to respect text and data mining reservations under the Digital Single Market directive. Several providers have argued that the template, in its current form, exposes commercially sensitive information without producing proportionate compliance gains.
For smaller providers and open-weight model communities, the August window introduces a different problem. The threshold for systemic-risk classification, set by reference to compute used in training, places only a small number of models above the line. Models below the threshold are subject to a lighter regime, but the precise contours of that regime are being filled in through guidance rather than primary text. Open-weight providers, in particular, have argued that the obligations relating to downstream-fine-tuning chains are not yet workable in a context where derivative models multiply quickly and provenance tracking is technically incomplete.
The fines structure itself merits careful reading. The seven-percent ceiling tracks the General Data Protection Regulation in headline severity but does not replicate its tendency, in practice, to produce settlements at the lower end of the range. The AI Office has signalled that it expects to use its enforcement discretion in a calibrated way, with first-year actions focused on declaratory clarification rather than maximum-penalty cases. Whether that signal holds in practice will depend on whether a high-visibility incident, such as a safety failure in a deployed model or a copyright dispute that escalates into a regulatory referral, lands inside the first twelve months.
Two further factors will shape the enforcement landscape after August. The first is the interaction with national market-surveillance authorities for high-risk AI systems, whose own enforcement powers come online in the same window. The boundary between a general-purpose model and a high-risk system integrated by a downstream deployer is the most frequently contested compliance question in pre-market briefings. The second is the AI Office’s resourcing trajectory. Pre-launch reporting indicated staffing levels below the projection set in the original implementation roadmap, and case-handling capacity will determine whether the August powers translate into visible decisions or remain latent.
August 2026 marks the point at which the AI Act stops being a roadmap and becomes a live enforcement regime. The Code of Practice is the instrument through which most providers will manage that transition. Whether it remains a procedural cushion or becomes a substantive standard will be decided by the first cycle of cases.
