Analyses

The US Bans Claude Fable for Caribbeans

Share via
On 12 June 2026, the US cut access to 2 Claude models for every foreign national — what that dependency really means for businesses in the Caribbean.
The US Bans Claude Fable for Caribbeans

🇫🇷 Lire en français : Les USA interdisent Claude Fable aux Guadeloupéens

I rushed. Fable 5, Anthropic’s most powerful model, was included in my subscription until 22 June — I wanted to get the most out of it before the deadline. Stunned: on 12 June, well before that date, the model had already vanished from my account. Dismayed when I understood why — the US government had just cut access for all foreign nationals, me included, on national-security grounds. Not a banned use: a nationality. For a business here that builds its operations on US tools, this isn’t a geopolitical anecdote: it’s a very concrete warning.


On 12 June, an AI — a computing capability — became a weapon  

  Important

An export-control directive doesn’t target what you do, but who you are: here, the blocking criterion is nationality, not use.

Let’s take the facts, because they’re precise. The US didn’t ban “AI”. It targeted 2 Claude models — Fable 5 and Mythos 5, the most powerful in the catalogue — under export controls. The other models remain available: the target is narrow.

What isn’t narrow is the criterion. The block doesn’t rest on some dangerous use, but on who you are: any foreign national loses access, wherever they are, including Anthropic’s own non-American employees. You could be a faultless developer in Pointe-à-Pitre: if your passport isn’t American, the door closes.

Anthropic is complying but contesting it publicly: it speaks of a misunderstanding tied to a narrow flaw it says is present in other models too, and says it is working to restore access. A reason eventually surfaced: according to Semafor and the Los Angeles Times, the measure was reportedly triggered by fears that a China-linked group had accessed Mythos, the other model targeted. Who’s right on the technical merits matters little. The fact that concerns us is elsewhere: a single foreign decision was enough to cut off, overnight, a tool used by millions of people — most of whom weren’t the target.

It’s nothing new: the cartoon keeps rerunning  

The scene is absurd — a consumer software product treated like military hardware. But it isn’t the first time it’s been played: it’s the same cartoon that comes back with every generation of technology, with new characters.

2000 episode: the PlayStation suspected of firing missiles. The PlayStation 2’s computing power — the famous “Emotion Engine” — and its encryption exceeded export-control thresholds. The Japanese government made the console subject to export approval beyond two units, fearing they could be clustered to guide missiles. Iraq reportedly imported thousands, triggering a military investigation. No weapons programme was ever proven: a living-room console reclassified as a weapon for its raw computing power alone.

1990s episode: encryption classed as “munitions”. In the United States, strong encryption was on the munitions list, on a par with missiles. When Phil Zimmermann put his PGP software online, in 1991, he was the subject of a federal investigation for “arms export”. His offence: giving private individuals a way to protect their communications. The first “individual-level” tussle over access to a digital capability.

You can see the pattern. A civilian computing capability becomes powerful enough to frighten a state, which reclassifies it as a weapon and decides who may use it. Console, encryption, AI: same plot, new episode. Each time the restriction looked ridiculous in hindsight and was eventually lifted — but in the meantime, those who depended on it paid the price.

Why it concerns us, here in the Caribbean  

  Warning

Dependency isn’t a problem as long as everything’s fine. It’s a switch someone else holds in your place — and can flip without asking you.

Ask yourself: what stops, in your business, if a US provider cuts you off tomorrow? Automated customer support? Content generation? Part of your code? If the answer is “a lot of things”, you’re holding the fragility I’m talking about — the same question as an Internet down for 72 hours, at the scale of your tools. A first step to pin it down: measure. I put a digital-sovereignty observatory to the test on kimoun.com — a sovereignty score already gives you a first picture of your visible dependencies.

This isn’t paranoia, it’s risk management. Building your whole business on a foundation someone else can pull away — political decision, price hike, change of terms — makes you a tenant, not an owner.

And for us, here, the fragility is structural. I’ve already written that value is already draining off the territory through US platforms, and that AI will amplify it. Technical dependency is its other face: 7,000 kilometres from Paris, and further still from Silicon Valley, we sit at the end of the chain. When they cut upstream, we’re the ones who feel the cold first.

And make no mistake: swapping masters isn’t a solution. Chinese models — Qwen, DeepSeek — now dominate open weights, but their tooling isn’t yet turnkey for a business, and Beijing is already tightening its grip: restricting the overseas travel of its top experts, pressuring its labs to monetise — even to close — once-open models. The real answer isn’t to trade one dependency for another — it’s to reduce the dependency itself.

Because this dependency isn’t only material. As Tariq Krim argues, cutting off Fable isn’t a mere switch: it’s a “geographic cognitive fracture”. The world keeps the everyday models but is stripped of the frontier intelligence — the very thing that multiplies the productivity of a firm, a practice, a software publisher. A foreign decision can, overnight, relegate you to second-tier AI while your competitors keep the best. It’s no longer about cost: it’s your productivity — and so your competitiveness — that a third party can throttle from afar, and with it your autonomy of thought.

Taking back control: it’s IT, not magic  

  Warning

Don’t go it alone hosting a model: it’s a craft. Running a frontier model on the desk next door doesn’t exist. What changes everything is the right use and the right provider.

Beware the misreading: the answer isn’t “I’ll install an AI on my own in the back office”. A small business won’t run a frontier model on an office machine, and the ones that fit on light hardware are far more limited. Promising otherwise would be lying.

But between “hand everything to a US giant” and “do everything yourself”, there’s a wide space. Hosting an AI for precise, well-scoped tasks is within reach of many businesses, provided you’re well supported. It’s IT: the job of an IT department in larger structures, of your local provider in a micro or small business.

There are many tools, and the right choice depends on the use. For sovereignty, a European open-weight model like Mistral is a better bet than a US model whose licence stays under its publisher’s control. You can host it on European infrastructure — or locally: Guadeloupe already has its own TIER III data centre, opened by CANAL+ Télécom at Jarry in 2022 and designed for data territorialisation; other local hosting projects are reportedly under study.

Aymé Macula Mbumba, CEO of CANAL+ Télécom, on data territorialisation (RCI Guadeloupe, 2022, in French).

I was part of Mediaserv’s technical management: I know this chain, from the cable to the server rack. Combined with the Caribbean Connect cable, this kind of facility opens the way to AI tools managed locally, by local players.

And there’s nothing theoretical about it. Recently, I set up for a Guadeloupean teacher a bespoke solution: his AI agents — Claude, ChatGPT — talk directly to all his lessons and documents. They live in Obsidian, versioned with git, on a server Kimoun manages for him in OVH’s data centre at Gravelines. He calls it his “second brain”. The AIs help him prepare lessons and track curriculum progress; but he owns his files, he decides who can access them, and he saves tokens by submitting to the models only what’s useful. A real European data centre, GDPR-compliant, but with a local contact — expert and available — rather than an anonymous desk in Silicon Valley.

Here’s the real point. Taking back control isn’t only about saving money: it’s a matter of competitiveness — and of not being unplugged by a decision you didn’t make. The PlayStation 2 and PGP eventually slipped free of those who wanted to lock them down; the question, for your business, is not to be the one that waits, unplugged, for the misunderstanding to be sorted out in Washington.

Frequently asked questions

No. The 12 June 2026 directive targets 2 specific Claude models, Fable 5 and Mythos 5; the other models remain available. So the problem isn’t that everything stops today. The problem is the precedent: a foreign decision can cut off a tool overnight, based on your nationality rather than your use. That’s the fragility to anticipate.

Not a frontier model on its own, no — it would be dishonest to promise that. But for precise, well-scoped tasks, hosting an AI is now within reach of many businesses, provided you’re well supported. It’s IT: the job of an IT department in larger structures, and of your local provider in a micro or small business.

For sovereignty, Mistral has an edge: it’s a European, open-weight model you can host on European or local infrastructure. Llama is also open-weight, but its licence remains under the control of Meta, a US company. The right choice mostly depends on your real use — a serious provider will help you decide.

It sharply reduces exposure, but it isn’t an absolute guarantee. It all depends on who controls the infrastructure, the software and the keys. The aim isn’t perfect sovereignty, which doesn’t exist, but to reduce the number of switches a third party holds in your place. Every block brought back under control is one risk less.

First list the AI tools you really depend on, and ask what stops if one of them is cut off. Then pick one or two simple, well-scoped uses you could move to a more controlled solution. Then talk to your local provider: it’s a classic IT project, run in stages, not a revolution.
Discuter sur WhatsApp