Brussels Puts a Statutory Spine Under "Digital Sovereignty" — and Makes Inference Data a Procurement Test
After three slips, the European Commission today presented its Tech Sovereignty Package, and the headline is that "digital sovereignty" is no longer a slogan — it has a legal definition for the first time. The centrepiece, the Cloud and AI Development Act (CADA), — a proposal that would establish four statutory levels of cloud sovereignty graded on five criteria: control over the service, control over the supply chain, the processing of AI inference data, the location of the infrastructure, and the level of cybersecurity. Member states will be required to run "sovereignty risk assessments" to map how much of their existing infrastructure depends on non-EU providers, then match sub-sectors to the appropriate sovereignty level by use-case sensitivity (Implicator.ai, eu-cloud-ai-act.com). The package ships alongside a revised Chips Act 2.0 carrying crisis powers to override supplier contracts and fines up to €300,000 for withholding capacity data, an open-source software strategy, and an energy-and-AI roadmap — all under Executive Vice-President Henna Virkkunen's tech-sovereignty portfolio (Euronews, Agence Europe).
Editor's note: today's lead is the one Regulation main, against the standing one-per-day cap — justified because the Tech Sovereignty Package publishing today is the defining European event of the week, the regulatory cycle is genuinely dominated by it, and the package's contents (rather than the anticipation of them) are now on the table. The market structure it responds to is stark: the US hyperscalers hold roughly 70% of European cloud infrastructure while European providers have fallen to around 15%, and a parallel cloud provision would bar US platforms from holding sensitive government data (Bird & Bird). The detail worth circling for any regulated buyer is that "the processing of AI inference data" is now one of the five scored criteria. The place a model's prompts and outputs are processed stops being a marketing line and becomes a procurement test — which is precisely the moment running inference on infrastructure you control, the way Stav and the wider European inference cohort do, moves from compliance overhead to a documented buy-side requirement.
MiniMax M3 Pushes the Open-Weight Frontier Into Million-Token Coding — at a Tenth of the Compute
The open-weight cohort got a genuine step forward this week. MiniMax M3, launched 1 June, is the first open-weight model to combine frontier coding, a 1M-token context window, and native multimodality in one release. It scores 59.0% on SWE-Bench Pro — topping the open-weight field and surpassing both GPT-5.5 and Gemini 3.1 Pro on coding while approaching Claude Opus 4.7 — and is built on MiniMax's Sparse Attention (MSA) architecture, which the company says delivers roughly 9x faster prefill and 15x faster decoding at 1M tokens, with per-token compute around one-tenth of the prior generation at that length (MiniMax, MarkTechPost, The Decoder).
Pricing makes the point sharper than the benchmark: M3 listed on OpenRouter at $0.60/$2.40 per million input/output tokens, with a launch promotion near $0.30/$1.20 — a fraction of the proprietary default (SCMP). The weights and a technical report are due within ten days. The timing is pointed: on the same day Brussels names an open-source software strategy a pillar of European sovereignty, the strongest argument for open weights is no longer ideological but economic — an Apache-class coding model within a couple of points of the frontier at a tenth of the hosted cost is exactly the kind of substrate a curated-OSS platform with a router across models, like Stav, is built to deploy on sovereign hardware.
Cognition Raises $1B at $26B as Devin's Run-Rate Jumps 13x in a Year
The coding-agent economy got its clearest valuation signal yet. Cognition, the maker of the autonomous software engineer Devin, has raised more than $1 billion at a $26 billion post-money valuation, co-led by Lux Capital, General Catalyst and 8VC, with Ribbit, Atreides and Founders Fund participating (TechCrunch, Bloomberg). The numbers behind it are the story: Devin's annualised run-rate went from $37 million in May 2025 to $492 million in May 2026 — a roughly 1,230% increase — with customers including Goldman Sachs, Mercedes-Benz and parts of the US government (The Next Web).
The detail that should land with engineering leaders is the dogfooding figure: Cognition says roughly 89% of its own code is now written by Devin (the company's own figure). That is the most concrete data point so far on how far autonomous coding has moved from demo to production, and it reframes the buy-side question for European teams — not whether to adopt coding agents, but on which models and under which jurisdiction those agents run, given that the most capable open-weight option (see above) now sits within reach of the proprietary leaders on exactly the SWE-bench-style tasks these agents are measured on.
Quick Hits
- Trump signs a voluntary early-access AI order — A 2 June executive order asks US labs (Anthropic, OpenAI, Google) to submit frontier models for government testing up to 30 days before release, explicitly barring any mandatory licensing or preclearance — a voluntary counterpoint to the EU's procurement-criteria approach (CNBC, NPR).
- Microsoft ships its own cost-cutting models — At Build, Microsoft unveiled MAI-Code-1-Flash (text-to-code) and put MAI-Thinking-1 into private preview via Foundry, the latest move to lower developer costs and lean less on OpenAI (CNBC).
- NVIDIA's Cosmos 3 is "fully open" — At Computex, Jensen Huang unveiled Cosmos 3, billed as the world's first fully open omnimodel for physical AI, plus the RTX Spark desktop and Alpamayo 2, a 32B open reasoning model for robotaxis (BGR, NVIDIA).
- WWDC is T-5 — Apple's 8 June keynote is expected to centre on a Gemini-assisted Siri overhaul, the AI story the company has owed since 2024 (9to5Mac).
- SpaceX IPO roadshow opens tomorrow — The roadshow starts 4 June, pricing is set for 11 June and the Nasdaq debut for 12 June under SPCX, targeting roughly $75B at a ~$1.75T valuation, with the Anthropic Colossus compute deal a headline AI-revenue line (CoinDesk).