DeepMind goes through Washington. Berlin convenes. EEA on AI's footprint.

The Center for AI Standards and Innovation (CAISI) at the US Department of Commerce announced on Tuesday that Google DeepMind, Microsoft, and xAI have signed new agreements granting the US government pre-deployment access to frontier models for national-security evaluation. The agreements expand on previous CAISI partnerships with OpenAI and Anthropic from 2024, which have themselves been renegotiated to reflect the new directives from Commerce Secretary Howard Lutnick and the Trump administration’s AI Action Plan. Under the deals, developers frequently provide CAISI with versions of models that have safety guardrails reduced or removed, allowing the agency to probe for national-security capability and risk before public release. CAISI says it has now completed more than 40 such evaluations.

For Europe, the geographic detail of this arrangement is the entire story. DeepMind is headquartered in London, with its principal research operations at the King’s Cross campus. It is, by any measure, the most consequential European AI lab and one of the most capable frontier labs in the world. Under this agreement, the next-generation Gemini model — built primarily in the United Kingdom — will be evaluated by the US government before any external release. The UK AI Security Institute conducts its own separate evaluations, but no equivalent EU-level pre-deployment access agreement exists for any frontier lab, anywhere.

The pattern this creates is now structural. Anthropic’s Mythos is restricted to American partners and the NSA — European governments demanded access at the Eurogroup on Monday and received no commitment. DeepMind’s Gemini, built in London, now flows first to Washington under formal evaluation. The frontier capability that European institutions need access to in order to defend themselves — and the frontier capability built on European soil — both arrive at European desks via Washington’s decision tree, not Brussels’s. The AI Act regulates how AI is deployed in Europe. It has no mechanism to compel access to frontier models built outside the EU. This week made that gap operationally visible in two consecutive days of policy reality.

The European Environment Agency published two briefings on Tuesday: Navigating Europe’s twin transition — opportunities and challenges of digitalisation in the green transition, and Artificial intelligence and sustainable consumption in Europe (EEA Briefing 09/2026). Both arrive at a moment when AI infrastructure is expanding faster in Europe than at any point in the last decade, and when the carbon, water, and critical-mineral implications of that expansion are still poorly priced into either policy or procurement.

The numbers in the briefings are sharper than the framing. Global AI computing capacity is doubling roughly every seven months, citing Epoch AI 2026. Compute capacity from major AI chip designers has grown by approximately 3.3 times annually since 2022. NVIDIA currently provides over 60% of global AI compute, with Google and Amazon supplying most of the rest. Data centres consumed an estimated 1.5% of global electricity in 2024 and around 3% of Europe’s — but in Ireland the figure already exceeds 20%, and AI-focused capacity is concentrating around the FLAP-D cluster: Frankfurt, London, Amsterdam, Paris, Dublin.

The water dimension is starker. Under a baseline scenario, AI-driven data centres could increase annual water consumption for cooling and electricity generation to approximately 1,068 billion litres by 2028 — an elevenfold increase compared to 2024. AI growth also depends on a narrow set of critical raw materials — gallium, germanium, indium, palladium, tantalum — that have received far less policy attention than the battery materials underpinning the green transition, despite being equally strategic for European autonomy.

The headline conclusion is that AI can deliver real environmental benefits when deliberately aligned with sustainability objectives — better data, smarter logistics, more efficient industrial processes, lower-carbon procurement. But the EEA is explicit that data centres, networks, and devices together are generating an environmental footprint that efficiency gains alone cannot offset. The implication for European policymakers, particularly with the AI Act high-risk deadline approaching August 2 and the gigafactory programme contested, is that “sovereign compute” cannot be debated in isolation from energy, water, and supply-chain constraints. The twin transition is not two separate transitions running in parallel. It is one system with internal trade-offs that current policy frameworks are not pricing.

The Rise of AI Conference 2026, marking ten years of one of Europe’s longest-running and most respected AI gatherings, opened on Tuesday evening with the speaker dinner at Humboldt Carré in Berlin. The full conference programme runs today, Wednesday May 6. Founded by Fabian Westerheide as a Singularity discussion group in 2015, Rise of AI has grown into a curated executive event that, by design, prioritises depth over scale: roughly 300 in-person seats and around 1,400 online viewers, deliberately constrained to keep the conversation substantive.

The two-stage programme — Meta Stage and Applied AI Stage — covers four core themes: applied artificial intelligence, regulation and policy, sustainable and trustworthy AI, and strengthening Europe’s AI ecosystem. Confirmed speakers include Prof. Dr. Jürgen Schmidhuber, director of the AI Initiative at KAUST and a foundational figure in modern deep learning; Prof. Dr. Peter Sarlin, co-founder and CEO of AMD Silo AI; Dr. Irakli Beridze, Head of the UN Centre for AI and Robotics (UNICRI); and Prof. Dr. Feiyu Xu, Professor of Industry AI at the German University of Digital Science.

The convening matters this week for a specific reason. Berlin is the same city where SAP committed €1 billion to Prior Labs on Monday — the most ambitious enterprise AI research investment ever made by a European company. The Rise of AI audience includes the founders, investors, policymakers, and corporate leaders whose collective decisions over the next twelve months will determine whether the SAP commitment is the start of a pattern or an isolated bet. The conference’s framing this year — sovereignty, trust, real-world deployment — reads as a direct response to the gigafactory critique published Monday and the Mythos access crisis that broke the same week.

The Tech, Digital and AI Summit opened in Brussels on Tuesday. The framing in the official programme is unusually direct for an EU institutional convening: “Can Europe deepen digital ties with the U.S. without giving up its push for strategic autonomy and competitiveness?” And: “How can Europe reduce its dependence on American digital infrastructure — from AI models to cloud services — without stifling growth or exposing itself to geopolitical vulnerability?”

The agenda explicitly addresses the Digital Networks Act and EU merger rules, the Digital Markets Act and Digital Fairness Act, AI infrastructure for closing Europe’s innovation gap, and how Brussels and Washington navigate divergent digital regulatory frameworks under transatlantic tariff pressure. The Parliament passed its DMA enforcement resolution in the same week. The Commission is expected to clarify next steps on the Google Android preliminary findings ahead of the May 13 consultation deadline. The Omnibus second trilogue is approximately one week out. The summit’s timing makes the policy questions concrete rather than abstract.

The most consequential subtext is the one not on the agenda. Yesterday’s CAISI announcement gives the US government formal pre-deployment review access to a model built in London. The Eurogroup spent Monday demanding access to a model held in San Francisco. The same week, the EU’s own €20 billion gigafactory programme was publicly criticised in Politico for lacking a viable business case. The Brussels conversation this week is not about whether Europe needs to act. It is about whether the actions Europe is currently taking are the right ones — and what changes when they are not.