Special edition · AI sovereignty
Local AI: What runs in your own house can't be switched off.
On June 9, Fable 5 was the most capable AI model in the world. Three days later, it was gone.
On June 9, 2026, Anthropic released Fable 5, its most capable model to date – available through the major cloud platforms, from Amazon Bedrock to Google Vertex. Companies started building it into agents and workflows.
Three days later, it was over. According to an update on Anthropic's own announcement page, access to Fable 5 – and to its sibling model Mythos 5 – was suspended on June 12 to comply with a US government directive. Nobody using the model had done anything wrong. The decision was made far away from their business. From one day to the next, a productive tool was no longer reachable.
This special edition isn't about Fable 5. It's about the question everyone using AI seriously should ask: who can switch off my AI – and what happens then?
The short answer
Use AI in a provider's cloud and you rent a capability that can disappear at any time – through a regulator, a price change, a discontinued model version, or a suspended account. Run an open model on your own hardware and you own that capability. It keeps running, whatever happens outside.
That's the core of local AI: not maximum capability, but control.
Three days, and the best model in the world was gone
Fable 5 was no niche tool. At launch it was the most capable widely available model, rolled out across exactly the cloud services that run a large share of business AI today. That's what made the dependency so invisible: you don't book some exotic software, you use “the cloud” – and you only realize you're a tenant when the landlord locks the door.
The trigger here was a government directive. But the pattern is older and bigger than any single model.
Someone else's cloud, someone else's control
The Fable 5 suspension is just one of several ways a rented AI capability vanishes:
Law and regulators. Export rules, directives, sanctions – as in the Fable 5 case. You don't get a vote.
Provider decisions. A model gets discontinued, the interface changes, the terms are revised. Your working process breaks because someone else cleaned house.
Price. What's predictable today can double overnight. With per-token billing, you carry the risk, not the provider.
Availability. Outage, throttling, a suspended account. Even without bad intent, your AI is then missing exactly when you need it.
The common thread: the capability lives on someone else's computer. You're a guest, not an owner.
What “local AI” actually means
Local AI means an open model – from the Llama or Mistral family, for instance – runs on your own hardware, in your own house or on a server you control. Three things follow from that:
The data stays in the house. No detour through a US cloud, no service reading along. That's also the strongest argument on data protection.
The version keeps running. The model you choose today won't be discontinued or swapped out overnight. You decide when to update.
The costs are predictable. You invest once in hardware instead of billing every token at a price that can change.
This isn't a special case – it's a deliberate architectural decision.
The honest part: what local AI is not
So this doesn't turn into a brochure, the uncomfortable part: an open local model is not as capable as the absolute frontier. A model running on your server won't beat Fable 5 or a top cloud model – not in raw capability, at least. It needs hardware, setup, and maintenance; those are real costs. And not every task justifies that.
But the decisive question isn't “which model scores highest?” It's: “which capability do I actually need – and can I afford to lose it overnight?” For a large share of office work – extracting documents, classifying text, drafting, searching internal knowledge – a local model is more than enough. And it's yours.
When local AI pays off
Three questions for an initial assessment:
- Do you regularly process sensitive or personal data that shouldn't leave the house?
- Would a sudden outage of your AI tool be a real problem for the business?
- Do the tasks run continuously and at volume, so that predictable costs matter more than pay-as-you-go?
Several yeses? Then local AI deserves a closer look. The answer, by the way, is often not either-or but hybrid: the sensitive and the always-on locally, the rare peak capability from the cloud. What matters is that the thing you depend on doesn't belong to someone else.
Frequently asked questions about local AI
Isn't a local model worse than ChatGPT or Fable 5?
In raw capability: yes. For most business tasks: no, it's enough. The difference is that a local model can't be switched off and your data doesn't leave the house.
What does local AI cost?
Instead of paying per use, you invest once in hardware and setup. Whether it pays off depends on the volume – a serious approach starts with one clearly bounded use case.
Do I need my own data center for this?
No. For many use cases in small and mid-sized businesses, a single capable server is enough. A data center is the exception, not a requirement.
Is local AI automatically GDPR-compliant?
It's a big step, because the data stays in the house. But compliance still comes only from the right architecture.
Bottom line
The Fable 5 suspension wasn't a fluke. It's the normal risk of renting capabilities instead of owning them. Local AI doesn't mean having the best model in the world. It means owning the model you depend on. Anyone who regularly works with sensitive data, or can't afford an outage, should know this math before it gets handed to them involuntarily.
Want to know whether local AI is the right fit for your business? We'll give you an honest first assessment instead of a sales pitch.
Let's talk about your project