Anthropic restored global access to Claude Fable 5 and Mythos 5 on July 1, ending the 19-day blackout that followed the US Commerce Department's June 12 export controls. European enterprises spent three weeks discovering how much of their AI stack sat on a switch Washington could flip without warning.
The models are back, but the lesson isn't going anywhere. Mythos 5 remains restricted to a handful of vetted institutions; European banks are still locked out. The blackout is the reason Mistral's sovereignty pitch and Germany's new frontier-lab programme both landed with more force this week than they would have a month ago. (VentureBeat)
Research & Labs
Germany starts funding its bid for three sovereign frontier AI labs
SPRIND's €125 million Next Frontier AI challenge moved from selection to cash this week. Ten teams begin 24 months of funding, compute access and company-building support, chosen from a jury that finalised its picks in late June.
The goal is blunt: build three labs by 2028 capable of raising billion-euro scale-up rounds and competing with US and Chinese frontier developers. It's a DARPA-style bet that direct government funding, not just favourable regulation, is what's been missing from Europe's AI strategy. (SPRIND)
Mistral teases a new open-weight model, early access in July
CEO Arthur Mensch confirmed in a July 4 interview that Mistral has "a very exciting model to come this summer," open-weight, with early access opening this month. He gave no parameter count or benchmark detail.
The timing matters more than the specifics. Mensch has spent the past month arguing that open weights are insurance against exactly the kind of access cutoff Anthropic customers just lived through, and a July release would let Mistral make that case with a product rather than a quote. (TechCrunch)
Deployment & Applications
Europe's banks are the AI winners so far, but not uniformly
NatWest booked more than £100 million in additional cost savings in Q1 2026 from AI-driven productivity gains. UniCredit chief executive Andrea Orcel is pushing the opposite lesson from the same trend, warning publicly against "indiscriminate" AI rollout across the bank.
Both banks are scaling AI. The split is over how: NatWest is citing hard savings numbers; Orcel is arguing that benefits only show up with targeted deployment and measurable returns, not blanket adoption. It's the clearest sign yet that European financial services has moved past the pilot stage and into arguing over method. (GAM)
Governance & Safety
The AI Act Omnibus still isn't law, and the clock is tight
The Council gave political sign-off on June 29, but the amending regulation still hasn't been published in the Official Journal. Until it is, the original AI Act deadlines remain legally binding, not the delayed ones everyone has been planning around.
Publication has to happen by July 30 for the new timeline to take effect on August 2. A two-week gap between "agreed" and "law" is normally a formality. This time it's the difference between an August 2 deadline and a December 2027 one. (DLA Piper)
MedTech Europe puts a number on the compliance overlap: €3.3 billion a year
This week's clarification confirmed that AI-enabled medical devices stay under both the AI Act and the existing MDR/IVDR device rules, rather than a single sector-specific regime. MedTech Europe calls the dual layer "an unnecessary layer of complexity" and estimates the added cost at €3.3 billion annually across the industry.
Device makers now have a number to put in front of regulators instead of a general complaint. Whether that changes anything before the guidelines are finalised is the open question. (Healthcare Digital)
Compute & Infrastructure
Quantum Systems raises $1.2 billion, the largest private defence-tech round in European history
The Munich drone maker closed its Series D on July 2 at an $8 billion valuation, co-led by Blackstone, Airbus Defence and Space, and Advent International. Its Vector reconnaissance drone has flown more than 19,000 missions in Ukraine.
Airbus isn't just an investor here. The two companies simultaneously deepened a strategic partnership that pairs a defence prime with the kind of autonomous-systems startup it might otherwise have tried to outcompete. That's a different model for building sovereign capability than the public compute programmes Europe has leaned on so far. (Silicon Canals)
One to watch
The AI Office's real enforcement powers switch on August 2
Once the GPAI rules take full effect, the AI Office can demand technical documentation and training-data summaries from model providers (Article 91), commission independent evaluations with model access (Article 92), and order mitigations or a model's withdrawal from the market (Article 93).
Until now, the AI Office has mostly been building the compliance apparatus. August 2 is when it gets teeth. Watch whether it uses Article 92 against a general-purpose model in its first few months; that's the test of whether the enforcement machinery is real or theoretical. (EU AI Act Explorer)