I had three half-finished one-shots in my terminal. It was Friday evening and I was testing Fable 5, Anthropic's most capable model, for the first time on real work. Not prompting in a chat window, but real workflows. Agents building artifacts, processing messy source material, holding context through long chains. The model was absurdly good. It could take messy input and deliver auditable results in a way that felt like a genuine step forward.
At 5:21 PM the US government sent a letter to Anthropic. By 9 PM the model was gone. Shut down for everyone. No warning, no appeal.
And I sat there, mid-workflow, staring at an empty terminal.
One letter is enough
The official statement was about "foreign nationals." The government claimed that foreign citizens, foreign companies, and foreign governments could not have access to the model. It sounds surgical. It sounds like a targeted export control.
But Anthropic sells globally. They hire globally. They have enterprise customers with global workforces, APIs distributed to hundreds of millions of users. If the rule says no foreign citizen, anywhere, may touch the model, including inside the US, then it is not a targeted restriction. It is a kill switch wrapped in export control language. And everyone involved knows it.
It was likely about a jailbreak path linked to Fable and its sister model Mythos 5. Anthropic argued it was narrow, that it did not justify the scope of intervention. Maybe they were right. But the point is not whether this particular order was justified. The point is that the precedent now exists. Any frontier model can be frozen based on a claim the public cannot review, through a process no one can audit, under a standard no one can apply consistently.
That is not safety governance. That is discretionary power.
And I am a European customer. My business, my clients' businesses, my entire work with AI agents runs on models that live on American servers, governed by American terms, one government letter away from disappearing.
The generator in the garage
Greg Isenberg, who runs one of the bigger AI podcasts, used a metaphor that stuck: local models are the generator in the garage. Most of the time you are happy with the power grid. It is cheaper, simpler, someone else maintains it. But the people who are truly resilient have a generator ready. The hurricane comes, the lights go out, but they can still run.
A couple of years ago, local models were junk. I want to be honest about that. Running a model on your laptop in 2024 was more of a tech experiment than a working tool. But that has changed. Faster than most people expected. Qwen 3.6 from Alibaba performs at levels that surpass models four times their size. Gemma from Google fits in 16 GB of RAM and writes beautifully. DeepSeek solves hard coding problems. Llama has an entire ecosystem of fine-tunings.
A 12-billion parameter model on a Mac with 16 GB of RAM handles an estimated 80 percent of what most people use ChatGPT or Claude for. Not 100 percent. Not frontier quality. But good enough to be private, free, and always on.
What makes local models genuinely interesting is not that they replace cloud models. It is that they give you a layer no one can take away. Your data never leaves the machine. Every query after the hardware is free. And the model on your disk works regardless of whether the company that built it still exists, regardless of whether a government approves, regardless of whether the internet is working.
The European dependency
But here is what I think about most, and what none of the American podcasts discuss.
We in Europe sit in a double dependency. We depend on American cloud models, and we depend on the American government allowing us to use them. One can disappear through a business decision. The other can disappear through a political decision. And we have no control over either.
The Fable 5 incident was formally about safety. But it demonstrates a mechanism. A mechanism where access to intelligence can be throttled with a single letter. If your business runs on one model, one lab, one country's regulatory mood, and one access agreement, you do not have a stable operating plan. You have a dependency.
European alternatives exist. Mistral is building frontier models in Paris. Their models are open, commercially licensed, and competitive. Mistral Large performs in the same class as GPT-4 and Claude Sonnet. Mixtral has become a standard for local installations with its mixture-of-experts architecture. They are European, they operate under European law, and they have no conflict of interest with the US Department of Defense.
Yet almost none of my clients use Mistral. Most have never even heard of them.
That is not a technology problem. It is a brand problem and an ecosystem problem. OpenAI and Anthropic have built their models, but they have also built a narrative. They have built a gravitational field of integrations, documentation, communities, and venture capital-driven visibility that makes everything else feel like a second choice. Mistral does not just need better models. They need European companies to actively choose to diversify their stack.
Where are the Nordic models?
And then there is the question nobody asks. Where are the Nordic models?
We have the expertise. We have the infrastructure. We have data protection legislation that actually works. We have universities producing top researchers in machine learning. But we have no Nordic foundation model.
I would pay for it. Not as a donation to a research project, but as a commercial customer who wants a model trained on Nordic data, optimized for Nordic languages, hosted under Nordic jurisdiction. A model that understands that "semester" in Sweden does not mean "semester" in California. A model that can write contracts in Swedish without sounding like a Google Translate artifact.
The margin does not matter. If a Nordic model costs 20 percent more per token than GPT-5 but gives me the guarantee that it will not disappear with an American government letter, that is cheap insurance.
AI developers in Finland, Norway, Denmark: build that model. Companies like Silo AI (now part of AMD), the Norwegian NorwAI consortium, Sweden's AI Sweden. They have the expertise. What is missing is commercial will and paying customers who say "yes, I want this, and I will pay the margin."
I am saying it now. I will pay the margin.
Own a piece of your stack
The lesson from Fable 5 is not that the cloud is bad and local is good. That would be too simple. Cloud models are smarter. They are more convenient. They have better tool integration. I will continue using them every day.
But the lesson is that you should not build your entire life on something that can vanish with a single letter. Own a piece of your stack. Have the generator in the garage.
In practice, that means three things.
Have a local model ready. Download Ollama or LM Studio. Pull down Qwen 3 or Gemma. Run it on a real task, not as an experiment but as a working tool. Build the instinct for what can run locally and what needs frontier. That instinct will become the most important AI skill you can have in the coming years.
Diversify your providers. Do not use only OpenAI. Not only Anthropic. Test Mistral. Test Google. Build workflows that can switch models without everything falling apart. It is not about finding the perfect model. It is about not having a single point of failure that can take everything down.
Demand European alternatives. Not for ideological reasons, but for business reasons. Every time you choose a provider for an AI service, ask yourself: what happens if this provider is shut down tomorrow? If the answer is "everything stops," you have a problem that no prompt engineering can solve.
Fable 5 will probably come back. Anthropic and the US government have worked together before, and both have commercial reasons to resolve this. But the warning shot has been fired. The next era of AI is not just about model quality. It is about access quality and governance quality.
And for us in Europe, it is about stopping being passengers in someone else's car.