EU legislators handed the AI industry a significant reprieve last week. On May 7, the Council and European Parliament reached a provisional deal on the AI Act Digital Omnibus, pushing the compliance deadline for most high-risk AI systems back 16 months, from August 2026 to December 2027.
This isn't deregulation. The fines framework stays intact (up to €35 million or 7% of global turnover), and a new prohibition on AI-generated non-consensual intimate imagery takes effect December 2026.
What changed is timing, and for companies running healthcare triage tools, hiring systems, or credit-scoring models, that breathing room is real. Formal adoption is expected by July. (Council of the EU)
Research & Labs
SAP bets €1 billion on Freiburg tabular AI.
SAP announced on May 4 it will acquire Prior Labs, a Freiburg-based startup whose TabPFN model, published in Nature and downloaded more than 3 million times, holds state-of-the-art on structured data benchmarks across hundreds of independent studies. SAP commits more than €1 billion over four years to scale it into what it calls a "globally leading frontier AI lab."
Prior Labs stays independent, with a research team originally recruited from Google, Apple, and Jane Street. This is notable: an enterprise software company making a credible research bet rather than wrapping a third-party API and calling it AI. (SAP)
Ineffable Intelligence raises Europe's largest ever seed.
David Silver, UCL professor and former lead of DeepMind's reinforcement learning team, closed a $1.1 billion seed round at a $5.1 billion valuation in late April. Sequoia and Lightspeed co-led, with Nvidia, Google, and the UK Sovereign AI Fund participating.
Silver's thesis is that RL self-play, learning from experience without human-labelled data, unlocks qualitatively different AI capability. The UK Sovereign AI Fund's co-investment signals a deliberate effort to anchor frontier research in London rather than watch it migrate to San Francisco. (TechCrunch)
Deployment & Applications
WHO maps clinical AI across all 27 EU member states.
An April 20 WHO/Europe report, the first of its kind, found that all 27 member states recognise improved patient care as a primary driver of AI development, with nearly three quarters already deploying AI-assisted diagnostic tools. 63% use chatbots for patient engagement.
The report is careful not to conflate deployment with effectiveness: governance frameworks lag adoption in most countries, and clinical training for staff remains patchy. The recommendation is to establish centres of excellence for structured evaluation before scaling; a meaningful brake on the current tendency to treat pilots as proof. (WHO/Europe)
Cohere and Aleph Alpha merge to pitch sovereign AI.
Announced April 24, Cohere is absorbing Germany's Aleph Alpha in a deal valued at roughly $20 billion, anchored by a $600 million commitment from German retail group Schwarz. Both Canadian and German Digital Ministers attended the Berlin announcement; that is a visible signal this is a government-backed consolidation play, not purely a market transaction.
The explicit pitch: enterprise and public-sector data stays in European infrastructure, outside US cloud stacks. Aleph Alpha's relationships in German public administration and defence give Cohere the regulated-sector foothold it needs to compete for the contracts it couldn't credibly chase before. (CNBC)
Governance & Safety
What the EU AI Act Omnibus actually changed.
Beyond the headline deadline shift, the May 7 agreement contains specifics worth tracking. High-risk systems embedded in regulated products (medical devices, lifts, radio equipment under Annex I) get an additional one-year extension to August 2028.
Watermarking requirements on generative AI outputs are deferred four months to December 2026. The nudifier prohibition (AI-generated non-consensual intimate imagery) is entirely new.
Industry associations had pushed for substantive scope reductions; what they received was timing relief. The enforcement architecture, liability exposure, and penalty structure remain unchanged. (Latham & Watkins)
Healthcare AI faces a structural compliance fork.
Smaller European AI developers in the medical space now face two parallel conformity pathways: CE certification under the Medical Devices Regulation and an AI Act assessment on top. MedTech Europe calls this "duplicate audits" and flags it as a genuine survival risk for smaller developers, not an administrative inconvenience.
The problem is structural: the MDR and AI Act were designed without coordination, and the Digital Omnibus deal does not fix it. For European health-AI startups, the cost of compliance may prove as decisive as the cost of compute. (MedTech Europe)
Compute & Infrastructure
Mistral's compute buildout takes shape.
Mistral's €722 million debt-financed data centre at Bruyères-le-Châtel, south of Paris, is scheduled to go online in the second half of 2026 with 44MW of initial capacity. A parallel €1.2 billion investment in Sweden adds 23MW by 2027.
Both facilities are Nvidia-powered, targeting inference and training. The timing is pointed: Mistral is building its own compute base at the moment cloud GPU costs put frontier training out of reach for most European labs.
Self-owned infrastructure is a hedge against both cost and political dependency. (Techzine)
AI & Society
AI raises productivity without cutting jobs. So far.
A May 2026 EIB working paper on AI adoption across European firms finds that companies using AI see 4% higher labour productivity on average, with no evidence of employment contraction in the short run; AI-intensive firms are actually 4% more likely to hire. The qualification matters: 40-60% of jobs are expected to undergo significant task changes, entry-level opportunities are tightening in some sectors, and gains cluster around firms that already had digital infrastructure.
The pattern is widening inequality between digitally mature and lagging firms, not mass displacement. (EIB)
One to Watch
EuroHPC AI gigafactories.
The amended EuroHPC regulation now permits AI gigafactory development (large-scale GPU clusters embedded in national AI ecosystems), and the official call for proposals was expected to open in early 2026. These facilities will determine which European research groups and startups can access frontier compute without routing through US hyperscalers.
Selection decisions, expected in late 2026, will shape European AI capability for the decade ahead. The countries that secure allocation, and the access conditions attached to it, are worth watching closely. (EuroHPC)