Editor's note. This week's brief has moved through who captures the value, then who controls the inputs. Today it reaches the top of the stack: who authorizes the frontier at all. The lead is the Financial Times report that Washington is about to formalise a framework deciding when advanced models ship and who may access them — at home and abroad. The second story is Microsoft answering Amazon's billion-dollar deployment unit within 48 hours, at 2.5× the size, with European names on the customer list. The third is China's biggest "open" release of the year, which also demonstrates the difference between an open licence and weights you can actually hold. Regulation stays in the quick hits, where the EU AI Act's enforcement clock keeps ticking.
The frontier now ships on Washington's schedule
The US government is in advanced talks with AI companies on voluntary standards for releasing new models, with an announcement possible as soon as next week, the Financial Times reported on Wednesday. The framework would set benchmarks and timelines for advanced models — and, notably, clarify who can access them "in the United States and abroad." It builds on June's executive order directing agencies to test frontier models before release, and it lands at the end of a week in which the machinery was already visibly running: GPT-5.6 remains limited to a small vetted-partner list at the government's request, Google is in talks with Washington ahead of releasing its next advanced coding models, and on Tuesday the Commerce Department lifted the export controls on Anthropic's Fable and Mythos models — less than three weeks after ordering their suspension.
Put those together and something structural has changed: frontier model access is no longer solely a vendor's commercial decision, it is becoming an instrument of state policy — with "voluntary" standards negotiated by companies that are simultaneously preparing IPOs and depending on federal goodwill. For a European buyer the operative word in the FT's summary is abroad. Whether your organisation gets a frontier model, when, and under what conditions is increasingly decided in Washington, by a framework you have no vote in — the same lesson the Fable suspension taught in June, now being written into standing process. The models you can plan on are the ones whose weights and serving infrastructure sit under your own jurisdiction.
Microsoft raises Amazon's deployment bet within 48 hours
Two days after AWS announced its $1B forward-deployed engineering unit, Microsoft launched "Microsoft Frontier Company" — $2.5B and 6,000 employees dedicated to embedding engineers inside enterprise customers to design, deploy and improve AI systems, led by former Microsoft Asia president Rodrigo Kede Lima. Early customers include the London Stock Exchange Group, Unilever, Novo Nordisk and Land O'Lakes, with Accenture among the systems-integration partners. With OpenAI and Anthropic having launched similar ventures in May — Anthropic's backed by Goldman Sachs, Blackstone and Hellman & Friedman at $1.5B — every major US vendor now runs a paid army of engineers whose job is to live inside customer organisations.
The confirmation is the story: when the second hyperscaler commits 2.5× the money within two days, the "deployment last mile is where the margin lives" thesis has become consensus. Note the customer list — LSEG and Unilever are exactly the kind of European enterprises this column writes for, and they are taking embedded engineers from a US vendor into the heart of their AI programmes. The value of that help is real; so is the question it sharpens. An embedded engineer is by design inside your environment, your data and your architectural decisions — which makes "whose infrastructure, whose jurisdiction, whose boundary" a question to settle before the pod arrives, not after.
China's biggest open release comes with an asterisk
Meituan — the Chinese delivery giant — open-sourced LongCat-2.0 on 30 June: a 1.6-trillion-parameter MoE coding model (~48B active per token) with a native 1M-token context, MIT-licensed, and — per Meituan — the first trillion-parameter model trained entirely on Chinese silicon, a 50,000-chip domestic ASIC cluster with no NVIDIA involvement. The credibility comes less from the vendor-reported benchmarks (59.5 on SWE-bench Pro, a one-point edge over GPT-5.5 — though short of Sonnet 5's 63.2) than from production usage: the model ran anonymously on OpenRouter as "Owl Alpha" for two months and became one of the platform's most-used models — billions of tokens a day — before anyone knew whose it was. Promotional API pricing of $0.30 per million input tokens undercuts GPT-5.5 by an order of magnitude.
The asterisk matters as much as the release: at launch the Hugging Face and GitHub pages list the weights as "coming soon" — today LongCat-2.0 is accessible only through hosted APIs. Which makes it a clean illustration of the distinction European buyers should keep drawing: an MIT licence on the announcement is not the same as weights on your own hardware. Until the checkpoint is downloadable, "open" here means cheap hosted access to a model in someone else's jurisdiction — impressive engineering, and this week of all weeks, a reminder that the only supply no framework, export directive or pricing clock can re-term is the one you can pull down and serve yourself.
Quick Hits
- EU AI Act GPAI enforcement is 30 days out — the Commission's enforcement powers for general-purpose models — information requests, model access and penalties up to €15M or 3% of global turnover — activate on 2 August; for providers, the voluntary Code of Practice window is closing.
- Anthropic explores its own silicon — Anthropic is in early talks with Samsung to manufacture a custom AI chip on the 2nm process, per The Information — a week after OpenAI unveiled its Jalapeño inference chip with Broadcom; every frontier lab is now hedging its NVIDIA dependency.
- Zuckerberg tempers the agent narrative — in an internal town hall, Meta's CEO said agentic AI development over the last four months "hasn't really accelerated in the way that we expected" and the restructuring wasn't as "clean" as it could have been — a rare frontier-lab admission that the agent timeline is slipping.