Fable 5 Went Dark for 20 Days. Kimi K3 Can’t Be Switched Off.

On June 12, 2026, a US export-control directive switched off the best coding model in the world. Not for American companies. For everyone. My terminal in Stockholm returned the same error as a terminal in San Francisco, except nobody in Washington had thought about my terminal at all. The full account of those 20 days is in Fable 5 Is Back* (For Now). This post is about what they felt like from Europe, and what a European operator should have built before they happened.

Here is the version I lived. Much of European B2B sells into worlds where data sovereignty is not a slide in the security deck but load-bearing: contracts specify where data lives, who can touch it, and under which country's law. Teams in those markets spend months answering exactly those questions about their AI stack. Then a directive from a government they do not vote for and cannot appeal to turned off a core dependency overnight.

The texture of those 20 days is worth recording. Workflows tuned to one model ran worse on its substitutes. Evals calibrated against Fable's outputs stopped meaning anything. Nothing broke loudly; everything got slower and slightly worse at once, which is how a dependency failure presents when the dependency is intelligence itself.

That is the test most European AI architectures have never run. Sovereignty is not a flag on a data center, and it is not an EU-region endpoint on an American cloud, which is a residency arrangement wearing a sovereignty costume. It is three concrete capabilities: the weights inside your network, so the thing that does the reasoning is a possession rather than a subscription; egress control, so you decide what leaves the building, which is the promise your contracts actually encode; and continuity, the ability to keep operating on the day a foreign directive lands. I wrote about the mechanics in the on-prem piece. The blackout turned that argument from insurance into evidence.

The market has already priced this

You do not have to take my word that export-control risk is real, because vendors now put its absence in their marketing. Ten days into the blackout, Sakana AI launched Fugu, and the pitch on the launch page says it plainly: the Fugu models stand shoulder to shoulder with Fable 5 "while delivering frontier capability without the risk of export controls." The launch tweet used the same phrase. When a lab puts export-control immunity next to its benchmark table, the risk is a line item buyers ask about, and sellers answer.

One detail on that page deserves a European pause. Fugu, marketed as immune to American export controls, is not available in the EU or EEA while it works toward GDPR compliance. The one jurisdiction that turned dependency risk into a purchasing criterion is also the one the alternatives route around. We are very good at regulating what we can buy. We are not yet good at building what we need.

The uncomfortable symmetry

The obvious European answer to "the American model got switched off" is "run an open-weight model instead." And the strongest open models right now are Chinese: Kimi K3 from Moonshot, GLM from Z.ai, DeepSeek, Qwen from Alibaba. Which is why the Financial Times spent July asking, in so many words, "What is the risk of using Chinese open AI models like Kimi K3?"

It is not a silly question, and I will not dismiss it. Swap a rented American frontier for a downloaded Chinese one and you have traded one geopolitical dependency for another. Provenance matters: know who trained your model and on what. Evaluation matters: test it on your work, not a leaderboard. Hosting matters: a Chinese model behind a Chinese API is exactly as revocable as Fable was.

But the symmetry breaks in the one place that counts. A weights file on your own hardware cannot be remotely revoked, whatever flag it was trained under. Beijing cannot reach into your rack any more than Washington can. June proved that an API is a permission and a weights file is a possession, and only one of those survives a change of policy in someone else's capital.

Do not expect the tension to resolve itself either. The two governments are now openly circling the question:

Read that as an operator, not a diplomat. The weights already published cannot be unpublished. But the pipeline of future releases sits inside a dispute between two governments, neither of which is yours.

Europe's empty column

In early August a post went around that needed no argument, just a list. USA has ChatGPT, Grok, Claude, Gemini, Llama, Copilot. China has DeepSeek, Qwen, Ernie, GLM, Kimi, MiniMax. Europe has?

The honest answer is Mistral, and the honest footnote is that Mistral signed the open-weights letter in July but is not at the open frontier. That is not nothing. It is also not a sovereignty strategy for a continent whose procurement contracts assume one. The American column is closed models you rent. The Chinese column is open models you can hold. Europe's column is a purchasing department deciding which of the other two to depend on.

Could the answer be to wait for controls on the Chinese models too? Aaron Levie made the argument for why that fails, and it applies to Brussels as much as Washington:

Gatekeeping your own frontier models does not stop capable open models from existing. It just decides whose stack the world builds on. The weights are coming regardless. The only question Europe controls is whether it has the muscle to run them.

What a European operator does today

Not in the next EU framework programme. Today, inside one company, three moves:

1. Make the architecture model-agnostic. Your context, your skills, your workflows, and your evals are the layer you own; the model underneath must be swappable. This is Nadella's switching test from Part 1: if you cannot switch out the generalist model without losing the expertise built into your learning system, you have built on sand, and June proved the sand shifts.

2. Stand up an open-weight fallback and drill it quarterly. Kimi K3 or GLM on hardware you control, behind your own egress rules. Once a quarter, replay last month's real workload through it and write down the gap. The gap is real, especially on judgment and taste, but a measured gap is a plan and an unmeasured one is a prayer. On June 12 the companies with a tested fallback had a bad day. Everyone else had a bad month.

3. Keep buying the frontier where judgment demands it. This is not autarky. For high-stakes work, rent the best reasoning money can rent. Sovereignty is not refusing the API. It is being able to lose it and keep shipping.

The temptation is to file the blackout as a weird three weeks. From Stockholm it looked like the first live demonstration of who actually holds the off switch on your company. The Americans hold theirs. The Chinese cannot reach a file on your rack. And Europe, which wrote data sovereignty into its contracts before anyone else, is the last to build what the clause requires.

Own your off switch.