Source: SpaceDaily

On 2 August 2026 the EU’s AI rulebook moves from phased deadlines into active enforcement: the European Commission can supervise makers of the most powerful general-purpose models and impose fines for non‑compliance. The legislation had already activated prohibitions in February 2025 and duties for general‑purpose model makers from August 2025; the August 2026 date ends a one‑year grace period and triggers broader supervisory powers. General‑purpose AI is split into a lighter tier that applies to all providers (requiring publication of a training‑data summary using an AI Office template released in July 2025, updated at least every six months or sooner if material changes occur) and a heavier tier targeting systems judged to carry “systemic risk.”

The heavier tier applies to models presumed systemic if their training used more than 10^25 floating‑point operations, a deliberately high, blunt capability proxy said to capture only a handful of the most capable global models. Providers of such models face obligations to stress‑test and attempt to break their systems, assess and mitigate risks to health, safety, security, rights and society, report serious incidents to the AI Office, and maintain security. Non‑compliance can attract fines up to 3% of global annual turnover or €15 million (whichever is higher). Commentators cited in the piece (including Joel Christoph writing in Lawfare) describe the AI Office gaining broad authorities — from document requests to source‑code access and the ability to demand remedial action including market removal. Two central implementation gaps are highlighted: unresolved detail around what counts as a “sufficiently detailed” training summary (WilmerHale noted the template omits how to measure scraped content) and limited enforcement capacity, with the AI Office employing just over 125 staff for many Act responsibilities and recommendations to scale GPAI supervision to at least 160 staff by 2030. Recruitment difficulties for frontier‑AI talent are reported, and the article concludes that the legal tools are in place but whether they change how the biggest models are built depends on people and funding — “the tools are on the table. The question is whether anyone picks them up,” in Christoph’s words.