The European Commission, the Council, and the Parliament agreed Friday to defer the AI Act's Annex III high-risk obligations by sixteen months — from 2 August 2026 to 2 December 2027 — in the Digital Omnibus package first reported on 7 May and confirmed in Brussels this week. [1] The same instrument pushed Article 50(2) watermarking from August to December 2026 and slid national AI regulatory sandboxes to August 2027. [2]

The timing is the story. Friday was Day 4 of Federal Trade Commission enforcement under the United States' TAKE IT DOWN Act, which carries a $53,088 civil penalty per violation and a 48-hour takedown clock. [3] FTC Chair Andrew Ferguson sent letters this week to Meta, Google, X, Microsoft, Apple, Reddit, Snapchat, and TikTok; the agency's first named targets were nudify-image services. [4] The paper's Friday feature on TAKE IT DOWN's first enforcement week reads the European retreat against the American ramp.

Annex III is the spine of the AI Act. It lists the systems Brussels classed as high-risk in 2024: hiring tools, credit scoring, biometric identification in public spaces, critical-infrastructure controllers, exam graders, predictive-policing models. For each, providers were to compile the nine-section Technical Documentation File required under Article 11 before placement on the Union market. [1] That obligation now lands sixteen months later than the original text promised, with formal Council and Parliament endorsement and OJEU publication still to follow. [5]

The Commission's stated reason is simplification. Industry's stated reason, in the spring round of consultations, was that the harmonised standards bodies (CEN-CENELEC JTC 21) had not finished the technical specifications providers needed to demonstrate conformity. [1] Both can be true. What is also true is that no high-risk provider faced a Brussels enforcement penalty in 2026, and now none will until late 2027 at the earliest.

One Brussels addition runs the other way. The Omnibus introduced a new prohibition on "nudifier" applications, the same category the FTC pursued first under TAKE IT DOWN. [3] Two regulators on opposite trajectories converged on the single use case neither could defend leaving alone. The convergence is not coordination; it is the unmistakable shape of the political floor.

For the paper's running ledger on AI and state power — agent pricing, supplier revenue, OAuth incidents, encyclicals, and now deadline arithmetic — the European move is the first instrument in months to record an act of retreat rather than advance. Nvidia booked $75.2 billion in Data Center revenue on Wednesday and lifted next-quarter guidance to $91 billion. [6] Google priced its agentic tier at $200 per month at I/O Day 2. [6] SpaceX's S-1 disclosed Friday that Anthropic is paying it $1.25 billion a month through May 2029 — a contracted run-rate of roughly $45 billion. [7] Against those prints, a sixteen-month deferral of the rules meant to govern the systems built on that compute reads as more than housekeeping.

The deferred items are not symbolic. Annex III covers the AI systems most likely to determine whether a European resident gets a mortgage, a job interview, a school place, or a police stop. The August 2026 date was the load-bearing deadline the Commission cited when it sold the Act in 2024 as the world's first comprehensive AI law. Friday's vote moves the load-bearing deadline behind the next European Parliament's first full year. The composition of that Parliament has not yet been elected.

Brussels will publish the agreed text in the Official Journal in coming weeks. National sandboxes will not be operational until summer 2027. The first Annex III conformity assessments will not be due until winter 2027. Until then, the only AI regulator with an active 48-hour clock on a covered platform sits in Washington, sending letters to the same eight companies the European text was meant to bind. The transatlantic gap is not closing. It is opening on a new schedule.

-- ANNA WEBER, Berlin