Editor's note. The freshest story of the last few days isn't a model — it's the people who design models. Google DeepMind lost two of its most important researchers in a week, and where they landed is the lead. The second story stays with the Fable 5 saga, now in its eleventh dark day, where the reported terms of the model's eventual return are the news. The third widens the lens to infrastructure: China's plan to engineer its own sovereign AI stack, read alongside Europe's. Regulation appears only in the quick hits, so there are zero regulation mains today.
Google's real loss this week wasn't a benchmark — it was two architects
The competition everyone watches is over models and leaderboard positions. The competition that decides the next five years is over the small number of people who can design the architecture those models are built on, and this week that fight went badly for Google. Within roughly a week DeepMind lost two of its most consequential researchers. Noam Shazeer — co-author of the 2017 "Attention Is All You Need" paper that introduced the Transformer, and a Gemini co-lead whom Google brought back in 2024 via the roughly $2.7bn Character.AI licensing deal — left for OpenAI as its Lead for Architecture Research, with Sam Altman calling the hire "only 10 years in the making." Days later John Jumper, the 2024 chemistry Nobel laureate for AlphaFold, left for Anthropic after nine years; Alphabet shares slipped on the news.
For a European buyer the lesson isn't to bet on whichever lab is now collecting the talent. It's that the people and the roadmaps both sit firmly outside the EU's jurisdiction, and that the field's direction is being set by a handful of researchers moving between three or four US companies. You cannot predict which architecture bet pays off next, and you certainly cannot influence it from a procurement seat in Frankfurt or Paris. The only durable hedge is the same one this brief keeps returning to: build so the model underneath is a swappable component rather than a dependency, because the layer where the real decisions get made is one you will never have a vote in.
Fable 5's return is being priced — in data retention and nationality checks
Eleven days after the 12 June US export-control order, Claude Fable 5 and Mythos 5 are still offline for everyone, per Anthropic's own status and the independent trackers; the "back in coming days" line from mid-June has quietly expired, and the 22 June subscription-credit cutover came and went with the model dark (continued use now nominally requires usage credits, moot while there's nothing to use). What's genuinely new is the shape of the deal being reported for whenever it does return. Multiple outlets describe the terms expected to govern its return. One — nationality-based geo-fencing — remains unconfirmed by any primary source; the mandatory 30-day data retention on all Fable 5 traffic, including third-party platforms, has in fact been a term of the model since its 9 June launch rather than a new restoration condition, with developers also reporting more aggressive fallback to Opus 4.8 on cyber, chemistry and biology prompts. (Status trackers and the "it's back" claims still disagree; treat the retention terms as reported-planned, not independently confirmed live.)
Whether or not it returns this week, those terms rewrite the bargain. A frontier model that logs and keeps everything you send for 30 days with no opt-out — and that may yet gate access by nationality — is a harder product to defend than the launch-day pitch suggested — and a much harder one to square with data-minimisation duties and the CLOUD Act exposure that already disqualifies US-hosted inference for sensitive EU workloads. The episode's lesson keeps hardening in the same direction: the terms of access to a rented frontier model can be rewritten unilaterally, and the rewrite trends toward more logging and more gating, not less. Weights you host yourself don't get re-termed mid-contract.
Two blocs are now engineering sovereignty into the stack itself
China is drafting a roughly $295bn, five-year plan to wire thousands of data centres into a single national computing grid, with a mandate that at least 80% of the underlying technology — AI accelerators included — be sourced domestically from suppliers such as Huawei, effectively designing Nvidia and AMD out. State operators China Mobile and China Telecom would run the grid, with the NDRC drafting and a 2028 target; counting power-grid integration, the total could reach roughly $740bn. The core reporting dates to 9–12 June, refreshed this week — so read it as structure, not a single-day event.
Set beside Europe's 3 June Cloud and AI Development Act, the parallel is striking. Two blocs are now writing supply-chain control directly into their AI infrastructure: China by industrial fiat — a domestic-content quota — and the EU by procurement scorecard, CADA's four "assurance levels" that grade a provider on EU ownership, supply-chain transparency and freedom from third-country interference. The methods are opposite; the premise is identical, that who controls the silicon and the cloud beneath the model is a sovereignty question rather than a price-performance one. For European buyers the usable artefact is the CADA scorecard, which finally gives procurement a shared vocabulary for what "sovereign" actually means — control, supply chain, data treatment, and where the infrastructure physically sits.
Quick Hits
- The EU AI Act Omnibus heads to a 29 June Council vote — after the European Parliament adopted the text on 16 June, the Council is expected to formally adopt it on 29 June, followed by Official Journal publication. Standalone Annex III high-risk obligations move to 2 Dec 2027 and the watermarking grace period to 2 Dec 2026; until OJ publication the original text remains law. A confirmed Council date finally puts an end-point on a thread that has run for weeks.
- Gemini 3.5 Pro is still a preview with seven days left in June — Google's promised general availability (2M-token context, Deep Think) remains limited to Vertex enterprise preview, absent from the public app, AI Studio and the consumer subscription. Prediction markets put a "by 30 June" landing near 50–55%; the top of the closed-model price ladder stays unsettled until it ships.
- Mistral's mega-round is still early-stage — reporting continues to place the ASML-backed lab in talks to raise around €3bn at a ~€20bn valuation, nearly double its Series C mark, to fund owned data centres near Paris and a €1.2bn Swedish build-out. Terms remain unsettled. (Mistral is a direct competitor in European sovereign inference; industry context, not an endorsement.)
- KPMG and Microsoft scale agent governance globally — the two expanded their alliance on 9 June so KPMG can use Microsoft Agent 365 to manage AI agents under its Trusted AI framework while rolling Copilot across its global workforce — another sign the enterprise value is consolidating in the layer that governs and audits agents, not the models they call.