Brussels Starts the Clock
Since Saturday, the EU can fine frontier-model providers 3% of global turnover. The enforcement era of AI regulation has begun — and the first thing it punishes is silence.
On August 2, the European Commission's AI Office gained the power to investigate general-purpose AI providers and fine them up to €15 million or 3% of worldwide annual turnover, whichever is higher [1]. For a frontier lab with hyperscaler backing, 3% of turnover is not a compliance line item. It is a strategy question.
The design is worth understanding, because it will be copied. Brussels did not start by regulating what models may do. It started by regulating what their makers must reveal: training-data summaries, capability documentation, incident reporting, model access for evaluators. The first enforceable offense is not building something dangerous — it is declining to show your work.
Who feels it first
Every lab that shipped a frontier model in the last twelve months is in scope for the new-model track: the July wave alone included releases from Anthropic, OpenAI, Google, xAI, DeepSeek and Alibaba [3]. High-risk system obligations, by contrast, were postponed to late 2027 and 2028 by the AI Omnibus [1] — so the near-term pressure lands on model providers, not on the businesses deploying them. Expect the labs' EU legal teams, not their research teams, to set release calendars for the next year.
The second-order effect is the interesting one for markets. Compliance at this depth is a fixed cost, and fixed costs favor incumbents. A regime designed to check the largest labs may, in practice, price mid-tier European competitors out of the frontier — an outcome Brussels explicitly says it wants to avoid, and has not yet shown it can.