In April 2026, Anthropic gave an all-US consortium exclusive access to its most advanced cyber capabilities, AI that can find and exploit software flaws in critical infrastructure on its own. No European institution made the list. The operators who run Europe's critical infrastructure, its national cyber-defence teams, and its defence ministries were all left out, unable to use the technology to protect the systems Europe depends on. Without sovereign capabilities of its own, Europe cannot turn the most advanced AI on the job of defending its own infrastructure. That is exactly the gap the Frontier AI Initiative was set up to close.
Launched in Berlin in November 2025 by France, Germany and the European Commission, the Initiative is a non-profit public-private body, with an executive team, a board and a high-level group of advisors, tasked with giving Europe a sovereign frontier AI capability. Six months in, it is still caught between competing paths: train a European frontier model from scratch, bet on a shift in paradigm, pursue vertical integration, or build out compute. This memo, addressed to the European AI Office and the Initiative's task force, argues for a single, immediately actionable priority: scale and Europeanise a system that already works.
Through its interministerial digital directorate (DINUM), France runs Albert API, the most operationally mature sovereign AI deployment in Europe today, already serving more than 70 public-sector projects. We propose turning it into a European Sovereign Frontier Access Layer, open to public administrations and critical-infrastructure operators across the Union, paired with the capacity to audit and fine-tune models under EU jurisdiction and to run agentic cyber defence. For roughly €1 billion over three years, one to two orders of magnitude less than the cost of training a frontier model from scratch, Europe can reach the minimum level of sovereignty the Initiative has to clear to remain a serious instrument.