“Sovereign AI” is 2026’s most invoked phrase in technology policy, and its most successful salesman is an American chip vendor whose sovereign business just passed $30 billion a year. Here is what the term actually means, who defined it, what Europe is spending on it, and the one question under oath that shows where sovereignty actually ends.
What is Sovereign AI?
Strip away the branding and the idea is old: a state should not depend on a foreign power for critical capability. Applied to AI, it means being able to train, run and govern the models your economy and government rely on, using infrastructure on your soil, under your law, staffed by your people.
The catch is that nobody agrees where sovereignty starts:
- NVIDIA defines it as “a nation’s capabilities to produce artificial intelligence using its own infrastructure, data, workforce and business networks”, a definition that begins, conveniently, with infrastructure.
- IBM splits it into four layers of control: data, operational, digital and infrastructure sovereignty, a framing that happens to sell governance software.
- Academic Robert Dale’s survey of the field defines it as AI built “free from dependency on foreign platforms or corporations” across the full lifecycle. He concludes that on that standard only two countries qualify, the United States and China.
Every other country buying “sovereign AI” is, on the strict definition, buying degrees of dependence. The interesting question is which degrees, and from whom.
Who Coined It, and Who Cashed It
The concept has European policy roots. Dale traces it to digital-sovereignty debates around 2020, in exactly the GDPR-shaped corner of Brussels thinking that produced the AI Act. But the phrase went global through a sales channel, not a parliament.
From late 2023, NVIDIA’s Jensen Huang began telling investors and governments that every nation would need its own AI production. At the World Governments Summit in Dubai in February 2024 he put it in a sentence built to be quoted by ministers: “It codifies your culture, your society’s intelligence, your common sense, your history. You own your own data.”
It worked, measurably. On NVIDIA’s February 2026 earnings call, CFO Colette Kress told investors the company’s sovereign AI business had “more than tripled year over year to over $30 billion”. She named Canada, France, the Netherlands, Singapore and the UK among the drivers.
Hold those two facts together and you have a paradox.
Sovereign AI means not depending on foreign technology. The world’s loudest advocate of it is the American company whose chips every sovereign project buys. France’s Mistral is building its flagship “independent AI stack” data centre near Paris on 13,800 NVIDIA GPUs. This is not hypocrisy on anyone’s part; there is currently no European alternative at that tier. But it does mean that “sovereign AI” as sold in 2026 is mostly sovereignty over the building, not the stack. Forrester’s Dario Maisto makes the same point when he explains that “local hosting does not protect workloads and data with regard to sovereignty concerns.”
The Sentence That Defines the Problem
If one moment justifies the whole debate, it happened in the French Senate in June 2025.
Anton Carniaux, director of public and legal affairs at Microsoft France, was asked under oath whether he could guarantee that French citizens’ data held by Microsoft would never be handed to the US government without France’s approval. His answer: “No, I cannot guarantee that, but, again, it has never happened before.”
That is the US CLOUD Act operating exactly as written, and it is the honest version of every “sovereign cloud” brochure. A hyperscaler region on European soil is still a company subject to American law. Whatever else sovereign AI means, this is the dependency it is trying to name.
The licence layer can produce the same result without a court order. As we reported this week, Meta’s Llama 4 Acceptable Use Policy withholds the rights grant for its multimodal models from EU-domiciled individuals and EU-headquartered companies outright. This is all while the same document’s military-use prohibition was waived for the US government and its Five Eyes partners in 2024. Access to “open” American AI turns out to be a permission, granted and withdrawn by geography.
What Europe is Actually Doing About It
The EU’s answer is money, announced in very large numbers. The AI Continent Action Plan of April 2025 promised to “mobilise” €200bn through InvestAI:
- Of that €200bn, roughly €50bn is public money. The other €150bn is private capital the Commission hopes will turn up.
- €20bn is earmarked for AI gigafactories, plus thirteen smaller AI Factories at existing supercomputing centres.
- The gigafactory plan has already shrunk once. Announced in February 2025 as up to five sites of around 100,000 chips each, by mid-2026 it was restructured into a phased, two-tier programme the EuroHPC info session still called “preliminary”, with tenders opening only this summer.
- The 76 expressions of interest trumpeted in 2025 have reportedly narrowed to around ten serious bidders, and first construction is targeted for 2027.
For scale: American hyperscalers will spend an estimated $700bn on AI infrastructure in 2026 alone. At model level, the sovereignty scoreboard we track in our map of European LLMs reads worse:
- Europe’s one commercial frontier lab, Mistral, runs on Nvidia and is partnered with Microsoft.
- Germany’s flagship champion, Aleph Alpha, agreed in April to be absorbed into Canada’s Cohere, a deal its own press material calls a “sovereign alternative to US AI labs”.
- Several national “sovereign” models are fine-tunes of foreign bases: Bulgaria’s BgGPT on Google’s Gemma, Hungary’s academic models on Alibaba’s Qwen.
None of this means the projects are worthless. It means the word is doing more work than the stack.
Verdict: A Real Problem Wearing a Sales Pitch
The dependency sovereign AI names is real, and Europe is right to care. But the phrase, as used in 2026, mostly describes procurement, not independence.
It was popularised by the vendor with the most to gain, it is invoked for projects that run on that vendor’s chips under foreign licences, and Europe’s flagship programme has already shrunk once before breaking ground. The honest test for any “sovereign AI” announcement is three questions: whose chips, whose base model, whose licence. When the answer to all three is “someone else’s”, the accurate word is not sovereignty. It is hosting.
Author: Akos Szima
See Also:
Frequently Asked Questions
The ability of a country to build, run and govern AI systems on its own infrastructure, with its own data and talent, under its own law, without needing a foreign company’s or government’s permission. Definitions differ on how much of the stack, chips, cloud, models, data, must be domestic to count.
The underlying idea comes from European digital-sovereignty policy debates around 2020. The phrase was popularised globally by NVIDIA’s Jensen Huang from late 2023, and NVIDIA now reports over $30 billion a year in sovereign AI revenue.
No. Data sovereignty is about where data lives and which law governs it. Sovereign AI is broader: it also covers who owns the compute, who trained the model, whose licence you build under, and who can switch it off. A model hosted in an EU data centre on American chips under an American licence satisfies data residency and almost nothing else.
Partially, and mostly in plans. GPAI rules apply since 2025, thirteen AI Factories are funded, and the €20bn gigafactory programme opened its tender in summer 2026 with first construction targeted for 2027. At the model level, Europe has one commercial frontier lab (Mistral, NVIDIA-powered, Microsoft-partnered) and a set of national models, several of which are built on American or Chinese base models.
Because it named the market and supplies it. Nearly every national AI programme, European, Gulf, Asian or Canadian, buys NVIDIA hardware, and NVIDIAs CFO reported sovereign AI revenue “more than tripled” to over $30 billion in fiscal 2026.
