Editor's note: A quiet Sunday before WWDC, but the week's substantive news clusters around one theme — who controls compute. We lead on the burst of national compute alliances (Enterprise/sovereignty), the story closest to a European buyer's decision; carry a Models main on the frontier-release queue now stacking up for mid-June; and one Regulation main, the full reading of the White House's AI executive order now that the legal analysis has landed (it deepens the 3 June quick hit rather than repeating it). Open Source had no fresh drop today — MiniMax M3's weights remain pending — so it is handled in the quick hits.
Compute Becomes Statecraft: Washington, Tokyo and Ottawa Move in Step
In the space of a few days, compute access moved to the top of national policy agendas on both sides of the Pacific. The United States and Japan launched a $1 billion AI and science partnership — $500M each over five years — making Japan the first international partner in the Department of Energy's Genesis Mission, pairing twelve DOE national laboratories with twelve Japanese institutions and granting access to DOE high-performance systems and Japan's Fugaku supercomputer. It lands on older plumbing: the UK and Canada signed their AI-compute memorandum of understanding back in January 2024, an early marker of the same shift.pm.gc.ca/en/news/news-releases/2026/06/04/prime-minister-carney-launches-ai-all-canadas-new-national-artificial) a national strategy explicitly framed around "reinforcing control of our sovereignty," targeting a lift in AI adoption from roughly 12% to 60% by 2034.
The common thread is that compute is no longer treated as a commodity to be rented but as infrastructure to be governed. Canada is now being described as bidding to lead the "middle powers" on AI sovereignty — building bilateral compute ties (reporting describes a series of bilateral AI arrangements since early 2025, though no official list has been published) rather than depending on a single hyperscaler bloc. For European buyers this is the same logic the EU's Cloud and AI Development Act encodes into procurement, now playing out as foreign policy: access, cost, data location and jurisdiction are becoming the terms on which nations — and the regulated firms inside them — decide where their inference runs. The question Stav keeps returning to, where is this governed, is now being answered at the level of treaties, not just contracts.
The Frontier Release Queue Stacks Up for Mid-June
A cluster of frontier model releases is converging on the next two weeks, and the common theme is price, not raw capability. Google's Sundar Pichai confirmed at I/O that Gemini 3.5 Pro would ship "next month," and as of this weekend it has not yet dropped; OpenAI benchmark leaks point to a GPT-5.6 positioned as an efficiency upgrade — roughly 20–30% fewer tokens for equivalent output — rather than a capability leap; and developer communities are anticipating a Claude Sonnet 4.8, on the strength of a source-map string that correctly predicted Opus 4.7's April release. None of these is officially dated, and the Sonnet release in particular should be treated as a high-confidence rumour, not a roadmap commitment.
The reason it matters for buyers is the economics. If GPT-5.6 and a new Sonnet both land in the same fortnight at reduced token pricing — and MiniMax M3's open weights arrive around 11 June as the cohort's reproducibility test — the cost of running production agentic workloads could reset materially in a single cycle. That is the pattern tracked all spring: capability has largely plateaued at the top while price keeps falling, which is precisely why, for a European regulated buyer, the deciding variables are increasingly cost and jurisdiction rather than which model tops a leaderboard.
Washington's Voluntary Cyber Order, Fully Read
The White House's executive order "Promoting Advanced Artificial Intelligence Innovation and Security," signed 2 June, drew its first wave of legal analysis this week — and the substance goes well beyond the early-access provision flagged in passing on 3 June. The order directs Treasury to stand up, within 30 days, an AI cybersecurity clearinghouse that coordinates vulnerability scanning, validation and patch distribution in voluntary collaboration with industry and critical-infrastructure operators. It also tasks agencies with a voluntary framework for secure deployment of frontier models — including the up-to-30-day pre-release government access window — while expressly barring any mandatory licensing, pre-clearance or permitting requirement.
The clearinghouse is the part worth watching, because it speaks directly to a problem Europe has been living through. Anthropic's Project Glasswing surfaced more than 10,000 critical vulnerabilities in a month, yet only a fraction were patched upstream — discovery far outran remediation. A federal coordination body for validating and distributing fixes is one answer; the EU's answer, in its Cloud and AI Development Act, is to make sovereignty and inference-data handling scored procurement criteria. Two markets, two instruments: the US reaching for voluntary, disclosure-based coordination that avoids binding obligations, the EU reaching for binding procurement rules. For a regulated operator the divergence is the planning input — a US framework that asks for cooperation is not a substitute for the contractual and jurisdictional guarantees an EU supervisor will expect.
Quick Hits
- WWDC opens Monday. Apple's 8 June keynote is expected to centre on a rebuilt, system-wide Siri running on a custom ~1.2-trillion-parameter Google Gemini model licensed for around $1B/year, with personal-context access and a standalone Siri app. The EU data-processing-location question — whether the heavy lifting runs on Apple's own private servers or Google's — remains the open buyer-side detail.
- Open-weight cost counter holds; MiniMax M3 weights still pending. The open weights for MiniMax M3 (59% SWE-Bench Pro, 1M-token context) are due around 11 June and remain the cohort's next reproducibility test. Alibaba's Qwen 3.7 Max continues to draw developer attention at roughly half the input and a quarter of the output cost of comparable frontier models.
- A coding tool worth $60B. Cursor reportedly signed an agreement giving SpaceX the right to acquire it for $60 billion — a number that, true or not, marks how central agentic coding tooling has become to the enterprise stack, and how aggressively the largest players are moving to own it.
- SpaceX IPO prices 11 June. The roadshow opened 4 June; pricing is set for 11 June with Nasdaq trading under SPCX from 12 June, targeting up to $75B. The AI line under scrutiny remains Anthropic's reported ~$15B/year compute commitment — a single-vendor dependency the prospectus discussion will have to address.
- The trust paradox, quantified. The Stanford HAI 2026 AI Index, still generating analysis this week, finds usage and distrust rising together: 88% of organisations now report using AI even as trust in AI companies keeps falling.