The European Commission is expected to table the Cloud and AI Development Act (CAIDA) tomorrow, May 27, the centrepiece of its tech sovereignty package. The proposal targets triple EU data centre capacity by 2031 through harmonised fast-track permitting and grid integration support. The market context is unflattering: independent operational benchmarking puts only three EU-headquartered providers in the top GPU cloud tiers, H100 contract prices rose 40% in five months, and all rental capacity is booked through September. Whether CAIDA adopts a risk-based approach or the categorical "effective control" criteria pushed by France and European cloud CEOs will determine whether EU procurement becomes a protected market or accelerates the gap it intends to close.
Research & Labs
Recursive Superintelligence emerges from stealth with $650 million
London and San Francisco-based Recursive Superintelligence came out of stealth on May 13 with $650 million at a $4.65 billion valuation, backed by GV, Greycroft, Nvidia, and AMD. The founding team includes UCL professor and ex-DeepMind scientist Tim Rocktäschel, ex-Meta FAIR director Yuandong Tian, and Vision Transformer co-author Alexey Dosovitskiy.
The company's thesis is that AI systems can autonomously improve their own architecture, training, and research direction in an accelerating loop. A public "Level 1" autonomous training system is targeted for mid-2026; the company has fewer than 30 employees and no shipped product.
London now hosts two frontier labs with this profile: Recursive alongside David Silver's Ineffable Intelligence. Both spun from UCL and DeepMind networks, both valued above $4 billion on seed funding alone.
Mistral acquires Vienna's Emmi AI
On May 19, Mistral completed its second acquisition in three months: Emmi AI, a Vienna-based startup building physics AI models for industrial engineering workflows. Emmi's tools handle computational fluid dynamics, thermal analysis, and materials stress testing across aerospace, automotive, and energy sectors.
The deal adds 30+ engineers to Mistral's science teams and makes Linz an official Mistral office alongside Paris, London, Amsterdam, Munich, San Francisco, and Singapore. The strategic direction is visible: Mistral is positioning itself as the AI infrastructure partner for European industrial enterprise, where data-residency requirements and regulatory complexity make a European-headquartered provider structurally competitive against US hyperscalers.
(Sifted)
Deployment & Applications
EU puts €63 million into operational clinical AI
The Digital Europe Programme opened this month a €63.2 million call for operational AI deployment in clinical settings: image screening and diagnostic tools in actual medical centres, not research grants or procurement consultations. The Cancer Imaging Initiative runs alongside it, assembling 60 million cancer images across member states for AI training and validation by end of 2026.
A standardised evaluation protocol is a funding condition. If it holds, this becomes the cross-European dataset needed to determine which diagnostic AI tools generalise across patient populations, something that national health systems cannot assemble individually.
AI anti-corruption is already running, not planned
Ukraine's Dozorro processed 3.6 million procurement tenders in 2024 using machine learning to flag corruption risk. Albania formalised the step further: AI system Diella now formally participates in public procurement, drafting terms of reference, setting eligibility criteria, and verifying documents.
These are operational systems, not pilots. The EU public buyers community has updated its model AI contractual clauses in 24 languages; Ireland launched a sector-wide AI procurement consultation in February. The distance between cautious EU AI governance language and the operational bets of smaller member states is narrowing faster than the policy documents suggest.
Governance & Safety
High-risk AI classification guidelines arrive 11 weeks before enforcement
On May 19, the Commission published draft guidelines clarifying when an AI system qualifies as "high-risk" under Article 6 of the AI Act, covering all three classification pathways: embedded AI in regulated products (medical devices, machinery), and the eight Annex III use-case categories (employment, credit, law enforcement). Consultation closes June 23; the Annex III enforcement deadline is August 2.
For deployers running hiring tools or credit-scoring models, classification determines the compliance burden. The guidelines are non-binding; authoritative interpretation rests with the CJEU. That structural ambiguity cannot be resolved before the deadline.
GPAI enforcement powers arrive August 2
The AI Office's enforcement powers over general-purpose AI model providers activate on August 2, 2026: documentation requests, capability evaluations, corrective action orders, market restriction, fines. Systemic-risk models (above 10^25 FLOPs) face adversarial testing requirements on top.
How aggressively the Office uses these powers in its first quarter will calibrate expectations for every frontier AI provider operating in Europe. The first enforcement letters, when they arrive, will be read closely.
(EU Artificial Intelligence Act)
Compute & Infrastructure
CAIDA's ambition vs. the actual GPU market
The CAIDA proposal expected tomorrow targets triple EU data centre capacity through fast-track permitting, grid integration support, and capital backing for European providers. The problem is structural: SemiAnalysis ClusterMAX 2.1 data (April 2026) puts only three EU providers (Scaleway, GCORE, Nebius) in the top operational tiers.
On-demand Nvidia GPU rental is sold out through September, and H100 prices rose 40% in five months. No regulatory instrument can address those constraints in this policy cycle.
A procurement rule that redirects workloads to European-only providers without closing the operational gap transfers cost to European users without building capability. The Commission's own JRC has drawn that distinction explicitly; the question is whether the political coalition behind CAIDA has absorbed it.
Stuttgart's HammerHAI arrives this quarter
EuroHPC JU's HammerHAI supercomputer, built by HPE on Nvidia GB200 NVL4 architecture and offering 15 exaflops of peak AI inference performance, is due for delivery to Stuttgart's HLRS this quarter, with operations starting in H2 2026. Access is free for eligible European researchers, startups, and SMEs.
HammerHAI is among the first EuroHPC systems designed specifically for AI inference workloads rather than traditional simulation. That distinction matters: European labs increasingly need inference infrastructure, and publicly funded access to it could reshape who can afford to iterate at scale.
AI & Society
Nobel economist calls for UK-France-Germany AI DARPA
Philippe Aghion, 2025 Economics Nobel laureate, used the Brussels Economic Forum to argue that Europe should abandon 27-country AI coordination and instead build a willing "core", starting with the UK, France, and Germany, with dedicated DARPA-style agencies funding breakthrough AI research. Both Germany and the UK have launched innovation agencies in DARPA's image in recent years; Aghion's argument is that they should act in common rather than separately.
The counterargument is that a core-EU model fragments the single market. But the current model is what the Commission's own documents describe as an "inferior R&I system" compared to US and Chinese peers.
Aghion's provocation lands in that vacuum.
(Sifted)
One to Watch
CAIDA text and the first AI Office enforcement letters
The CAIDA proposal text drops tomorrow, May 27. The critical variable: does it adopt a risk-based framework, or the categorical "effective control" criteria pushed by French and European cloud providers? The answer shapes EU public-sector AI procurement for the decade. Watch simultaneously for the AI Office's first enforcement actions after August 2. Whether documentation requests go to the largest GPAI providers immediately or start smaller will signal whether the Office is calibrating for compliance or deterrence.