This coverage has traced how most of the world depends on a digital infrastructure it does not control — the chips, the cloud, the frontier models concentrated in a few American and Chinese hands. France has decided to do something about it, more aggressively than any other Western nation, betting billions on a national champion to claim a measure of AI sovereignty. The attempt is serious and well-funded. It is also shot through with a paradox that makes it one of the most instructive cases in this entire coverage: the harder France tries to be sovereign in AI, the clearer it becomes how much of that sovereignty still rests on foreign foundations.
Begin with the ambition, because it is real and large. At the Paris AI Action Summit in February 2025, President Macron announced €109 billion in public and private AI investment — framed explicitly as France’s answer to the scale of US programs. Most of it is going into data centers and compute on French soil. Macron framed the strategy as a “third way” in AI: not going it alone, but building alliances among European and other nations to counterbalance US and Chinese dominance. “This is our fight for sovereignty, for strategic autonomy,” he declared. “We want our cloud, we want our data centers, we want our computing capacities.”
The champion: Mistral
The clearest expression of the strategy is Mistral AI, founded in Paris in 2023 and now positioned as Europe’s strategic AI champion. What distinguishes Mistral from most European AI startups is that it does its own research and builds the underlying models itself, rather than building applications on top of American models. By early 2026, its models were deployed by enterprises across Europe, the US, and Asia, competing directly with the GPT-class models of OpenAI and Anthropic’s Claude in commercial markets. The state backing is unusually direct: France’s armed forces ministry awarded Mistral a framework agreement to run on French infrastructure with no US cloud providers; France and Germany announced a public-private partnership with Mistral and SAP for sovereign AI across public administration; and state investment institutions have joined its financing. This is industrial policy in the mold of past national champions — closer to South Korea’s backing of Samsung than to the venture-capital model of US AI firms.
The paradox, in hard numbers
Here is where the story turns, and where it earns its place in this coverage. For all the sovereignty rhetoric, Mistral’s infrastructure reveals an architecture of dependencies that partially undercuts the ideal — and the details are stark.
Mistral trains its models on tens of thousands of Nvidia GPUs. Nvidia controls over 90% of the generative-AI GPU market, and there is no credible alternative in 2026. France’s own chip initiative, SiPearl, is still at the prototype stage, with first processors expected only around 2027-2028 — too late for current needs. To secure compute, Mistral raised hundreds of millions in debt financing specifically to buy Nvidia chips — by one account, some 13,800 of them — and is building data centers around Nvidia’s Grace Blackwell systems. Even the company’s largest shareholder, after a roughly €1.5 billion investment, is not French but Dutch: ASML, the semiconductor-equipment maker, whose own machines are the indispensable upstream of the entire chip industry. A champion of European sovereignty, in other words, runs on American silicon, is anchored by a Dutch shareholder, and depends on a global supply chain it cannot replicate.
The vulnerability is not hypothetical. As one analysis put it bluntly, an American restriction on AI-chip access — an unlikely but not unthinkable geopolitical scenario, given the export controls examined elsewhere in this coverage — could cripple a company like Mistral within months. Mistral’s own CEO, Arthur Mensch, has been candid about the larger stakes, warning that Europe risks becoming a US “vassal state” within a couple of years if it does not build far more of its own infrastructure: “Whoever controls the chips, whoever controls the electrons, whoever has massive access to energy — that’s who wins.”
The deeper question: what is sovereignty, really?
This forces the question the whole coverage keeps circling. Is sovereignty an absolute — building everything yourself, from silicon to model — or is it something more modest and realistic: the capacity to decide, to have credible alternatives, to not be wholly captive to a single foreign supplier? On the absolute definition, France’s project is doomed, because no European country can replicate the entire stack alone. On the realistic definition, it may yet succeed: building sovereign compute on French and European soil, even with imported chips, gives European governments and firms a first credible option to develop and deploy AI without depending on US hyperscalers — a real gain in autonomy, if not a complete one.
The honest assessment is that France is buying partial sovereignty, and that this is not nothing. Hosting the compute, controlling the models, keeping data under EU jurisdiction and AI Act compliance — these reduce the leverage that foreign providers can exert, even if the chips underneath remain imported. Sovereignty, on this view, is a spectrum, not a switch.
Two readings, with comparable weight
The strategy admits two legitimate interpretations, worth presenting without tilting the scale.
Those who back the French approach argue it is exactly the right move: that the alternative — total dependence on US and Chinese AI — is worse, that partial sovereignty is real and worth the investment, that national champions backed by state resources are a proven model in strategic industries, and that someone has to start building European capacity or it will never exist. From this angle, Mistral and the €109bn plan are a credible, necessary bet, even with imported chips.
Those who are skeptical argue the sovereignty is more rhetoric than reality: that without domestic chips, sufficient energy, and a far larger talent base — France reportedly trains only a few hundred specialized AI engineers a year against a need for tens of thousands — the dependency remains structural, and “sovereignty” risks becoming a political talking point or sliding into protectionism. From this angle, real autonomy would require investing tens of billions in European silicon and talent over a decade, not just in data centers full of Nvidia chips.
It is not for this outlet to decree which reading is right. What can be stated is that both describe something true: France is genuinely building more autonomy than it had, and that autonomy genuinely remains incomplete and dependent at its foundation.
What this case reveals
What France and Mistral add to the coverage is the most concrete test yet of whether digital sovereignty is achievable or merely aspirational for any power short of the US and China. France is the hardest-trying case, the best-funded, the most politically committed — and even France cannot escape the foundational dependency on Nvidia’s chips and a global supply chain. If the country making the most serious bid for AI sovereignty still runs on American silicon, the lesson for everyone else is sobering: sovereignty in AI, for now, is a matter of degree, negotiated within constraints, not seized whole. The question is not whether to depend, but how much, on whom, and with what alternatives in reserve.
The verifiable fact is that France has launched the most ambitious AI-sovereignty push outside the US and China, that Mistral is its chosen champion, and that both depend on Nvidia chips, a Dutch top shareholder, and a global supply chain they cannot replace. Whether France achieves meaningful autonomy or a sovereignty that is more nominal than real will depend on decisions not yet made: on whether Europe invests in its own chips and talent at the necessary scale, on whether the partial sovereignty of hosting and models proves enough, and on whether geopolitical shocks expose or spare the underlying dependency. As in every story of this kind, what is decisive is not the ambition — which is genuine — but whether it is matched by investment deep enough to reach the silicon at the bottom of the stack.