The EU Artificial Intelligence Act is entering full enforcement for general-purpose models. From 2 August 2026, the European Commission — acting through the AI Office — holds exclusive authority to supervise and enforce Chapter V obligations on GPAI providers, per official digital-strategy.ec.europa.eu materials and AI Act service desks.
Powers include requesting documentation, evaluating models, requiring corrective measures and imposing fines. For intentional or negligent GPAI breaches, penalties can reach €15 million or 3% of worldwide annual turnover, whichever is higher. Prohibited practices and other Act tiers contemplate still higher thresholds in some cases.
Governance is hybrid: the AI Office concentrates supervision of the underlying GPAI model, while national authorities remain responsible for most AI systems deployed on their territory. A member state may ask the Commission to act on the foundation model when risk originates there.
Article 50 transparency obligations also apply from 2 August 2026. Systems already on the market before that date generally have until 2 December 2026 to meet machine-readable marking of synthetic content. That same 2 December date brings the strengthened prohibition on AI-generated non-consensual intimate imagery and related illicit material.
For U.S. labs with European users, the contrast with the White House’s “morally binding” Joint Commitment is structural: Brussels sets a single supervisor, deadlines and fines; Washington prioritizes voluntary self-policing. Complying in the EU is not optional because the single market conditions access.
GPAI providers must prepare training inventories, systemic-risk policies where applicable, notification channels and evaluation evidence. The Office can request access to model evaluations; refusing or ignoring information requests is explicitly among sanctionable conducts.
Firms that only integrate third-party models do not fully escape: they remain under national authorities for the deployed system, and can be affected if a GPAI provider receives a corrective order that changes usage or transparency terms down the chain.
The autumn 2026 calendar leaves little room for wait-and-see. Those still treating the AI Act as future text risk findings in internal audits and European customers demanding contract clauses aligned with the AI Office.
Legal analysts stress that effectiveness will depend on the Office’s human and technical resources, cooperation with member states and clarity of compliance guidance. An exclusive authority on paper needs operating capacity so it does not become a bottleneck.
Neural Pulse summarizes the framework from institutional sources: exclusive GPAI supervision from 2 August 2026, fines up to €15M/3%, Article 50 and intimate-image prohibitions toward 2 December, hybrid national-EU model.
Alongside voluntary White House pledges and industry bodies such as the planned SAFA, the EU offers the hardest regulatory counterpoint among large markets. Global providers will need dual-compliance designs, not a single-pole choice.
October 2026 is therefore an enforcement month: GPAI rules are no longer only legislative promise. The question for each lab is whether its evidence dossier would withstand a first formal request from the AI Office.
