Sovereign AI: The Friday the US Pulled a Model, and Every Country Started Building Its Own

It was a Friday evening in June 2026. People outside the US who’d been using a certain AI model went through the same thing at roughly the same moment. A model that had answered fine the day before suddenly stopped. Not an outage, not maintenance. The US government had issued an order to the company, and the company took the model down that same day.

The models in question were Fable 5 and Mythos 5, the top-tier line from Anthropic — the company behind Claude, with these two sitting above the rest of the Claude family as its most powerful tier. The government judged that bypassing Fable 5’s safeguards could unlock serious offensive cyber capability, and invoked its export-control authority on national-security grounds. The scope of the order was unusual. The line wasn’t drawn at a country but at nationality. Inside the US or outside it, if you were a foreign national, you couldn’t use it. Even the company’s own foreign employees weren’t exempt. Since vetting users one by one by citizenship was effectively impossible, the company simply switched both models off for everyone — US citizens included.

The company publicly disputed the order. The bypass, it argued, worked only in one narrow case — asking the model to read a specific piece of code and patch a flaw — and competing models had the same capability anyway. Yanking a commercial model used by hundreds of millions over that, it said, was overkill. Right or wrong, one fact stood confirmed from the user’s side: the power switch for the model they’d been using sat in Washington.

The model was never something you owned — it was a faucet

Buy a chip and it’s yours. Block the exports, and the unit you already bought still sits in your warehouse. A model is different. Most people don’t buy and keep a model; they connect to someone else’s server and rent it. Day to day the water flows when you turn the tap, so it feels like yours — but the shutoff valve is in the supplier’s hand.

For a long time, plenty of countries and companies didn’t bother thinking this through. Use the cheap, high-performing American model and be done with it — why build your own? Self-development looked like a question of cost and practicality, nothing more. This episode changed that math. Model access had become a security question. Hand core infrastructure to a single foreign government’s judgment, and on the day that government decides otherwise, your industry stops.

The name for this realization is sovereign AI — AI sovereignty. It isn’t a new idea. Europe, China, and the Gulf have been growing their own models and infrastructure for years already. This episode didn’t invent a motive that wasn’t there; it mostly handed a justification to the people who’d been putting it off.

Sovereignty has layers too

“We’ll do AI ourselves” can’t be left as one lump. Depending on what you actually hold in your own hands, sovereignty splits into three layers.

At the top is data sovereignty: keeping your citizens’ data from crossing the border to foreign servers. Put a data center inside your own country and process it there, and the problem is relatively within reach.

In the middle is model sovereignty: whether you hold model weights you control. Building your own from scratch is the head-on route, but there’s a detour — take an open-weight model whose weights are public and tune it on your own data. So it sits at medium difficulty.

At the bottom is compute sovereignty: whether the chips to run that model are in your hands. It’s the hardest of the three. However well you build a model, without the high-performance GPUs to train and run inference on it, the model is useless — and the overwhelming majority of those GPUs come from one American company (Nvidia) and are, again, export-controlled.

And here a paradox surfaces. The harder you push for model sovereignty, the more you run into the compute problem. That’s exactly why the US reached from chip controls over into model controls. However tightly you hold the upper floors, if the foundation at the bottom belongs to someone else, your sovereignty stays half a thing.

The hand on the foundation

That compute is the real bottleneck shows up most clearly in what the rich countries are doing.

The two Gulf states are the obvious case. Saudi Arabia and the United Arab Emirates are pouring sovereign-wealth money into AI infrastructure. Saudi Arabia set up a state AI company and signed a large GPU supply deal with Nvidia; the UAE is building a vast compute cluster of its own. Yet the decisive lever on these plans sits outside. In late 2025, only after the US Commerce Department approved the export of Nvidia’s latest chips to the two countries did deals that had been stalled for a while finally clear. The money is there — but whether the chips come in is decided in Washington. The episode laid that structure bare.

China sits at the opposite pole. The US blocked advanced chips at the outset, so self-reliance became compulsory rather than a choice. The result is the most integrated homegrown ecosystem anywhere. Huawei makes its own AI chips, models like DeepSeek and Alibaba’s Qwen run on top of them, and a state cloud handles distribution. But even self-reliance has a ceiling. Design a chip and you still need to mass-produce it — and the manufacturing equipment for that is also US-controlled, so China can’t even fully meet the demand it created for itself.

Europe takes yet another road. France’s Mistral is being raised as, in effect, Europe’s home team, while several countries flag “digital sovereignty” and attach public money to their own infrastructure. Mistral, backed by large investment, is building a data center near Paris — and the chips going into it, too, are bought from Nvidia. Europe, the Gulf, China: different starting points, different strategies, and the same wall at the single point of the chip.

Korea’s ‘national champion AI’

Korea is in the game too. The government is running a project formally called the Sovereign AI Foundation Model program, nicknamed the “national champion AI.” The goal is to reach 95% of the performance of the world’s best models under its own power, backed through 2027 by a budget in the hundreds of billions of won along with GPUs, data, and people. As the name suggests, the core of the program isn’t performance alone but “how much of it you built yourself” — that is, originality, weighed right alongside the scores.

The format is unusual. Rather than picking one team and backing it, the program lines up several teams and runs a survival contest, evaluating and cutting every six months. It launched in August 2025 with five teams — Naver Cloud, LG AI Research, SK Telecom, Upstage, and NC AI — and about five months later the first evaluation produced an upset. Naver Cloud, widely regarded as the strongest in the country, was eliminated. Its scores were near the top, but it had leaned in part on a model built by someone else and so failed to clear the originality bar (NC AI was cut alongside it). One more team was added through a second open call to join the surviving LG, SKT, and Upstage, and four teams now compete ahead of the next evaluation.

The interesting part is the team that came in. It’s a place called Motif Technologies, spun out of Moreh, an AI-infrastructure software company. Moreh’s original line of work was optimization — making models run well across many kinds of GPUs without being locked to any one chip. Carrying that expertise, Motif built a small model with relatively few GPUs and in a short span of time, and that model was rated as beating much larger ones on some reasoning benchmarks. Into the spot Naver lost on originality stepped a team that put “designed from scratch, ourselves” front and center.

The compute problem runs underneath this program too. Through a supplementary budget, the government separately bought some ten thousand of Nvidia’s advanced GPUs to parcel out across this and other national AI projects. The resolve to build a model yourself leads straight into the question “so where do the chips come from” — and the structure of the program itself shows it.

The model that can’t be cut off

This episode quietly produced one winner. The open-weight model, whose weights are public.

A model you rent through an API is over the moment the supplier locks it; weights you’ve downloaded once can’t be clawed back. The US can throw the switch, and the model inside my own server keeps running. Meta’s Llama, France’s Mistral, China’s DeepSeek and Qwen all fall here. They can be isolated and run in a domestic data center, so the user picks up a degree of both data sovereignty and model sovereignty at once. It’s also the first thing a country that can’t manage full self-development reaches for — take an open-weight model and tune it on national data.

There’s no free lunch, of course. A good share of the most advanced open weights come out of China, so any organization that scrutinizes how data is handled weighs the origin too. The advantage — “isolate it on your own server and the data doesn’t leave” — comes bundled with the caution — “the provenance still makes me uneasy.” Part of why Korea is willing to spend big to build a model of its own is exactly this double discomfort about depending on someone else’s.

So what’s left

Sum up the sovereign-AI boom as “now everyone gets their own model” and you’re half right. What you want to do and what you can do are different things.

The biggest constraint is, again, compute. Almost every camp converges on the same Nvidia chip. The Gulf trying to buy it, Korea pushing it through state policy, China forced into self-reliance — all of them hesitate, in the end, in front of the chip. The frontier gap doesn’t close easily, either. Building a domestic model doesn’t suddenly produce the world’s best one, and a generation or two of distance still separates “good enough, homegrown” from “cutting edge.”

There’s a boomerang on the US side, too. The harder it cranks the controls, the more it pushes even allies toward self-reliance and a multipolar split — which, over the long run, can shave the world share of American models by its own hand. China’s path in semiconductors went exactly that way.

So the most plausible picture isn’t a future where everyone goes fully self-sufficient. The few will go all the way to full self-reliance, China-style; the majority will more likely hedge by mixing. Use a US frontier model but don’t get tied to one; lay an open-weight model in a domestic data center and keep it as a backup; grow a state model as a safety net for regulated industries. The point is that the question has shifted — from “which model is best?” to “if some model vanishes overnight, does our service still run?”

That Friday evening, the first thought to cross the minds of the people staring at a model frozen on the screen was probably the same one. If this happens again, what can I switch to? The question governments are now spending hundreds of billions to answer is, scale aside, exactly that one.

— tomte