Editor's note. This week the brief has tracked who captures the value, who controls the inputs, and who authorizes the frontier. Today the arc reaches its logical end: who owns it. The lead is OpenAI's proposal to hand Washington a 5% stake — the state moving from regulator to shareholder. The counterpoint comes from Lisbon, where Portugal just shipped the other kind of sovereign AI: a €7 million model whose weights you can actually download. Between them sits Together AI, whose $800M round shows sovereign capital buying into the open-model serving layer. No fresh EU-regulation story today; the 2 August clock ticks on without news.
OpenAI offers Washington a piece of itself
OpenAI has proposed giving the US government a 5% equity stake — roughly $42.6 billion at its $852 billion valuation — the Financial Times reported on Thursday. Sam Altman's pitch goes further than one company: he reportedly wants every leading US AI lab — Anthropic, Google, Meta — ceding similar stakes into an Alaska-style public fund. The timing is not subtle: the proposal surfaced days after Washington delayed GPT-5.6's general release, and President Trump has already called government ownership of AI companies "a beautiful thing." The FT characterises the talks as early-stage and conceptual; implementing any deal might require an act of Congress.
Follow the week's sequence: the government gates GPT-5.6, drafts voluntary release standards deciding who may access frontier models "at home and abroad" — and now the leading lab proposes making that same government its shareholder. Whatever the fund structure, the conflict is structural: the entity that writes the safety rules would hold a $42.6 billion interest in the rule-taker's commercial success. For a European buyer the calculus is simpler still. A vendor whose regulator is also its investor is entangled with a foreign state at every layer — access, terms, and now equity. The models that don't renegotiate when a cap table changes are the ones whose weights sit under your own jurisdiction.
Portugal ships the other kind of sovereignty — for €7 million
While the frontier negotiates with Washington, Portugal launched Amália on 1 July at Instituto Superior Técnico in Lisbon: the first large language model built for European Portuguese, developed by a university consortium for €7 million in public money. Built on the open EuroLLM-9B base and extended with Portuguese data, image understanding and a larger context window, the model, its dataset and its code are published under Apache 2.0 on Hugging Face — weights, downloadable, today.
A 9B national model will not trouble any leaderboard, and that is not the point. Twenty-four hours ago this column covered China's biggest "open" release of the year, MIT-licensed with weights "coming soon." Amália is the inverse: modest scale, real checkpoint. It is also the first delivery in what is shaping up as Europe's open-model summer — Amália now, OpenEuroLLM's first models due by 31 July, the 400B-parameter EUROPA build behind them. The lesson for European institutions is the price tag: deployable sovereignty in your own language and jurisdiction cost one government €7 million, not a sovereign wealth fund.
Sovereign money buys the open-model serving layer
Together AI — the platform serving open-weight models from Llama to DeepSeek behind a single API — raised $800 million at an $8.3 billion valuation on 1 July, led by Saudi Aramco's venture arm with NVIDIA, Vista and General Catalyst participating. The company's annual bookings have passed $1.15 billion as enterprises hedge their closed-model dependency — demand that reportedly spiked during Fable 5's 18-day outage, when hosted open weights became the emergency fallback. The valuation has 2.5×'d since February 2025, and the plan is to scale compute roughly 50× over five years.
The signal worth keeping: enterprises' need for a second, open supply line is now a billion-dollar business — and the largest new stake in it belongs to a state oil company's venture fund. That is diversification, not sovereignty. An open model served from someone else's infrastructure still lives in someone else's jurisdiction, under someone else's investors; the Fable outage that drove Together's growth made exactly that distinction. The full hedge remains the one no funding round changes: weights you can pull down and serve on infrastructure you control.
Quick Hits
- Geneva becomes AI governance capital next week — the UN and ITU launched the AI for Good Global Commission on 1 July, co-chaired by Marc Benioff and Rwanda's Paul Kagame, with Jensen Huang, Andy Jassy, Brad Smith and Anthropic's Jack Clark among members; its first meeting on 8 July runs alongside the inaugural UN Global Dialogue on AI Governance (6–7 July) — the first multilateral forum where the Fable-style unilateral export control will be on the table.
- A CVSS for jailbreaks — out of the Fable 5 restoration, Anthropic is co-developing a jailbreak severity framework with Amazon, Microsoft and Google scoring exploits on capability gain, scope, ease of weaponization and discoverability — a shared standard that could replace ad-hoc emergency bans; watch the 1 August EO benchmarking deadline.
- California buys Claude at half price — Governor Newsom signed a first-of-its-kind agreement (29 June) giving all state agencies, cities and counties Claude at a 50% discount via the SITeS portal — the largest US government AI deployment yet, signed while federal policy toward the same vendor was still in flux.