One day. Four structural files. European AI's regulatory backbone lands.

The European Commission yesterday released draft guidelines on the classification of high-risk AI systems and opened a targeted public consultation running until 23 June. The text is intended to support providers, deployers and other relevant actors in determining whether an AI system falls within the high-risk category under the AI Act. The Commission noted that worked examples in the guidance “strive to cover all areas and use cases, but they are not to be considered as exhaustive and may be updated over time.”

The three-phase guide is structured around general principles for classifying high-risk AI, followed by two dedicated sections on the two high-risk categories under Article 6. Article 6(1) and Annex I cover AI systems used in the context of product safety. Article 6(2) and Annex III cover high-risk systems deployed across eight specified areas, including biometrics, education, employment, law enforcement, migration and border control. The Commission has not committed to a final publication date and has not said whether subsequent drafts will be put to further consultation.

The guidelines were originally due 2 February ahead of the AI Act's high-risk provisions taking effect on 2 August. A spokesperson told Euractiv in February that the draft was “currently subject to a revised timeline” to incorporate stakeholder feedback. The delay played a key role in the expedited path of the Digital Omnibus on AI, finalised politically on 7 May, which now pushes stand-alone high-risk system obligations to 2 December 2027 and product-embedded ones to 2 August 2028. More than 110 EU-based businesses, organised as the AI Champions group, had lobbied for the two-year pause on enforcement.

Why it matters. Until yesterday, every European company building or deploying AI had to guess at the operative boundary of “high-risk.” The text just landed. It is the most consequential single document the AI Office has produced this year and the basis on which European AI compliance teams will spend the next two years organising. Thirty-five days of consultation is not a long window. The companies that intervene meaningfully now will shape the final boundary.

The European Parliament yesterday voted through the new EU Foreign Investment Screening Regulation, which replaces the 2019 framework with mandatory screening across all 27 member states. The regulation, provisionally agreed with the Council in December 2025, names artificial intelligence, quantum technologies and semiconductors in Annex I among the strategic sectors requiring screening. Aerospace, energy, transport, digital infrastructure, financial-system entities and electoral infrastructure are also covered. A separate, narrower carve-out covers research and development in general-purpose AI models suitable for defence or space applications, and GPAI models with systemic risk.

The most substantive change is the extension of screening to intra-EU investments by companies ultimately owned by third-country investors. This closes the structuring loophole that allowed investors to anchor in member states with lighter controls. Member State authorities retain final say on individual transactions but are now required to give “due consideration” to comments from other member states and the Commission, with mandatory information-sharing through a strengthened cooperation mechanism. Formal Council adoption is pending; the regulation will fully apply 18 months after publication, likely end-2027. Germany and Spain have signalled intent to implement earlier.

Raphael Glucksmann, the rapporteur who led the file, framed the vote as a structural turning point. “We are closing a chapter of European naivety,” he said in a statement carried by Reuters, with the work continuing through the proposed Industrial Accelerator Act. Phase I review is set at 45 calendar days, non-extendable. Each Member State gets a call-in right of between 15 months and five years for investments outside the mandatory list.

Why it matters. Mistral, Helsing, Aleph Alpha, Lovable, Synthesia and the rest of the European AI cap-table just had their ownership question put on a permanent regulatory footing. Every future fundraise by every European AI company in a covered sector now sits inside this regime. The screening floor, not ceiling, lets ambitious member states go further. This is the structural answer to last week's Mensch testimony: who can own what is now defined.

NextEra Energy unveiled an all-stock acquisition of Dominion Energy valued at approximately $67 billion on Monday 18 May, with European coverage carrying through Tuesday 19 May. The transaction creates the largest regulated electric utility in the world by market capitalisation: NextEra shareholders will hold 74.5% of the combined entity, Dominion shareholders 25.5%. The combined company will have a market capitalisation of around $249 billion and an enterprise value of approximately $420 billion. The deal is expected to close in mid-to-late 2027.

The strategic logic is explicit. Dominion is the utility responsible for powering northern Virginia, the world’s most concentrated data-centre market: more than 450 data centres from over fifty customers, including every major hyperscaler. NextEra is the largest renewable energy developer in the United States. The combined construction backlog is 130 gigawatts — enough, NextEra CEO John Ketchum said on the analyst call, to power 100 million homes out of the roughly 150 million in the country. “Electricity demand is rising faster than it has in decades,” Ketchum said in the joint statement. “We are bringing NextEra Energy and Dominion Energy together because scale matters more than ever.”

European Business Magazine framed the deal for European readers as a direct read-across to the Draghi report’s argument that scale is no longer optional in industries where AI-era investment requirements exceed what fragmented players can deliver. If American utilities are consolidating into $420 billion entities to serve hyperscalers, the question European policymakers face is whether the continent’s fragmented utility sector can compete for the same data-centre investment — or whether the build-out will simply go where the power is reliable, abundant and cheap. The four hyperscalers spent $413 billion combined on capital expenditure in 2025, projected to reach $700 billion in 2026 (Motley Fool).

Why it matters. Stories 01 and 02 define the regulatory and ownership floor for European AI. Story 04 defines the political urgency. This story defines the physical constraint. The 27 May Tech Sovereignty package includes a roadmap for AI in energy — whether the package treats data-centre electricity supply as a continental scale problem will, in retrospect, be the first test of whether Europe is reading the same map the American utilities just acted on.

Renew Europe, the centrist Liberal group in the European Parliament, has formally requested and secured a plenary debate this week on what it terms “the need for an urgent EU response to Mythos and other advanced AI systems to safeguard European cybersecurity.” The session is scheduled for this afternoon in Strasbourg, framing the political case the Commission's 27 May Tech Sovereignty package will have to answer. It is the first time Anthropic's Mythos model has appeared on the Parliament's formal agenda by name.

The Renew demand list, published with the plenary priorities yesterday, has three operative items: full implementation of NIS2 across all member states, a dedicated EU AI cybersecurity strategy, and a structural reduction in dependence on non-European cloud, AI and semiconductors. The framing positions the 27 May Tech Sovereignty package — the Cloud and AI Development Act, Chips Act 2, open-source strategy and AI-in-energy roadmap — as the substantive response. The political logic is that if the package lands without those three items addressed, Parliament will say so on the record.

The debate sits inside a wider European Parliament week that has already passed the FDI Screening Regulation (Story 02), heard the Order of Merit ceremony for Merkel, Wałęsa and Zelenskyy, and will host Question Time with High Representative Kaja Kallas. The Anthropic-EU access negotiation, on which Commission spokesperson Thomas Regnier said discussions remained at a “different stage” than OpenAI's, continues in parallel. OpenAI's GPT-5.5-Cyber rolled out to EU partners on 11 May. Mythos remains unavailable.

Why it matters. Last week's Mensch verbatim testimony has become this week's parliamentary debate. The political vocabulary of European cyber sovereignty has moved, in fourteen days, from a CEO's testimony to a formal plenary item. The Commission's 27 May text now has a fixed political audience waiting to score it.

Founded in Munich in March 2021 by Torsten Reil, Gundbert Scherf and Niklas Köhler, Helsing builds AI software for military applications across air, land, sea and sub-surface domains. The company pledges to sell only to democratic governments. Its software-first architecture pairs the Altra battlefield management platform with mission-specific systems including Cirra (electronic warfare for fighter aircraft), Lura AI (swarm management for autonomous gliders) and Centaur (the multi-domain AI agent powering the CA-1 Europa unmanned combat aircraft unveiled near Munich in September 2025, scheduled for first flight in 2027).

Funding: €102.5M Series A (2021, Prima Materia / Daniel Ek), €209M Series B (2023, General Catalyst), €450M Series C (2024, General Catalyst with Saab, Accel, Lightspeed), €600M Series D (June 2025, Prima Materia) at €12B valuation. Reported in advanced talks for $1.2B at $18B led by Dragoneer with Lightspeed (Financial Times, 11 May). Subsidiaries in Estonia, France, UK. Total capital to date: €1.37 billion.

Government traction. Helsing produces the HX-2 12-kilogram loitering munition with ~100km range and GPS-denied operation; the Bundestag budget committee in February 2026 approved an initial €269M contract under a framework capped at €1B per producer with parliamentary approval required for spending above €2B combined with rival Stark Defence. Several hundred HX-2 units per month delivered to Ukraine as of January 2026, with demand from over six Ukrainian army units. Strategic partnership with HENSOLDT (announced 12 February 2026) equips the CA-1 Europa with radar, optronics, electronic warfare and the MDOcore data-fusion suite. Spanish robotics company Keybotic acquired January 2026. Stockholm presence opened 20 April 2026.

Editorial rationale. Yesterday's FDI Screening Regulation (Story 02) names “general-purpose AI models suitable for defence” explicitly. Helsing is the European company that pairing is written about. The 9.0 reflects both the operational footprint — Bundeswehr contracts, Ukraine deployment, NATO eastern-flank deterrence, partnerships with Saab and Airbus — and the structural fact that Helsing is the only European defence AI company at this scale. The honest deductions: Ukraine combat performance has been disputed (Politico reported HX-2 hit five of fourteen attempts in the Donbas; Helsing cited Russian electronic warfare), and US-led capital now exceeds the original European-investor framing. The Glucksmann vote is, in part, about that.