On 12 June 2026, at 5:21pm New York time, Anthropic received a US export-control directive: suspend access to Fable 5 and Mythos 5 "by any foreign national, whether inside or outside the United States". Unable to verify nationality in real time, it switched both off for every customer in the world. If you had a feature in production on that model, those are not eighteen days of geopolitics: they are eighteen days of downtime.
Eighteen days, and no SLA that covers them
Anthropic complied while stating it disagreed "that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people". On 30 June the restrictions were lifted and on 1 July Fable 5 returned, with AWS, Google Cloud and Microsoft Foundry re-enabled separately. Mythos 5 came back only for a set of US organisations, and the rest of the Glasswing programme is still waiting.
No contract covers a government order, and here the vendor had no room either. As a European customer you have no visibility into the proceeding and no standing to contest it: this is supply-chain risk, and it is not mitigated by negotiating, it is mitigated by designing.
Brussels is moving, at Brussels speed
On 3 June 2026 the Commission proposed the Cloud and AI Development Act and its four sovereignty levels (SEAL): processing and storage inside the EU (1), demonstrated independence from third countries (2), EU ownership and control down to personnel citizenship (3), no possible external interference (4). Final adoption is expected in late 2027 and the Commission estimates only about 1% of Europe's public services need level 4. For an SME one line matters: CADA would let the Commission extend sovereignty risk assessments to private companies in NIS2-regulated sectors. That question will show up in a tender questionnaire long before it shows up in law.
The defence is architectural
- The model id is configuration, not code: one internal interface, the model read from
AI_MODEL_ID. Switching vendor should cost a deploy, not a refactor. - A named fallback on a different vendor: written in the runbook, in a different jurisdiction, exercised at least once. A fallback never run is not a fallback.
- Tiering by task: DeepSeek V4-flash costs $0.14 per million input tokens (cache miss) and $0.28 output, against $5 and $25 for Claude Opus 5: classification and tagging do not need a frontier model.
- Your own eval suite: twenty real cases from your domain with expected output. Without it you never know whether the backup model is acceptable or merely available.
Open weights are not sovereignty
Kimi K3, released on 27 July, has 2.8 trillion total parameters and 104 billion active per token: self-hostable in Europe in theory, out of reach for an Italian SME in practice. And the licence is a custom "Kimi K3 License", not MIT as several outlets reported.
Verdict
Single-model dependency is a business continuity risk, full stop. You do not need to migrate everything to European vendors, and you do not need regulatory panic: you need every AI feature in production to have an abstract interface, a backup model in a different jurisdiction, and twenty tests that tell you whether it holds up. Half a day per project, against eighteen days of downtime.