A Model Went Dark for 18 Days. AI Sovereignty Is Now an Operations Problem
On 12 June, a frontier AI model that European businesses had built into live workflows went offline – not from an outage, but from a government order. The US Commerce Department applied export controls to Anthropic’s two most capable models, Fable 5 and Mythos 5, barring access for foreign nationals. Because Anthropic could not verify user nationality in real time, it switched both off for everyone, worldwide. Companies in finance, healthcare, and critical infrastructure lost tools embedded in production with no notice and no transition period.
The controls were lifted on 30 June and access was restored on 1 July. Eighteen days, then back to normal. If you run production and you are tempted to file this under “resolved,” don’t. The outage is over; the exposure it revealed is not.
What actually happened
This was the first time export controls – the rules that govern shipping weapons components and advanced chips abroad – were applied to a running commercial AI service rather than to hardware or code. The legal mechanism treated giving a foreign national access to the model as an export. The practical effect was that a tool hundreds of millions of people were using could be switched off by a decision taken in another country, with no contractual protection and no notice. It has since been resolved, and the provider agreed to closer government review of future releases. The precedent stands regardless.
The overreaction and the useful reaction
Europe’s political response was loud. One French presidential candidate compared the shutdown to a blockade of a strategic shipping strait. Brussels had, days earlier, unveiled a Tech Sovereignty Package, and later that month funded a consortium to build a homegrown open-source frontier model covering all EU languages. That model does not exist yet – the consortium won the right to build it. Useful for 2028; irrelevant to your next quarter.
The more useful reaction came from industry. Siemens’ digital-industries chief made the point to Reuters that sovereignty is not the same as self-sufficiency, and that the real requirement is flexibility, not cutting yourself off. The response from European firms was not to drop US models but to stop depending on any single one: diversify providers, keep alternatives ready, and know which workloads can move.
What this means if you are building AI into operations
Provider concentration is now an operational risk, in the same category as single-sourcing a critical part. You would not run a line on one supplier with no second source and no inventory. Wiring one proprietary, remotely hosted model into a production-critical workflow is the software version of exactly that.
Three practical moves follow. First, design for swappability: your automation layer should let you route to a different model without rebuilding the workflow, so no single provider is load-bearing. Second, sort your AI uses by sensitivity and tolerance for downtime. The ones touching regulated or sensitive data, or that cannot stop, are candidates for an EU-hosted model or an open-weight model you run on your own infrastructure – options that exist today, even if a European frontier model does not. Third, map the dependency and write down the fallback. “It is cheap and best-in-class right now” is a procurement note, not a continuity plan.
The models came back in under three weeks, so the panic reaction – rip out every US model and wait for a sovereign one – is as wrong as the complacency that preceded it. The mature response is architectural, not political: know what you depend on, keep a second option wired in, and put the workloads you cannot afford to lose somewhere you control. June was a cheap warning. The next one might not come with a fast reversal.
“Content generated using AI”