Sovereign AI is not a flag you plant. It is a stack you build one owned layer at a time: energy, compute, models, chips. No country owns all four yet, not even the United States.
France is stacking three of them faster than almost anyone in the world.
Mistral is closing in on a 23 billion dollar valuation.
France’s nuclear grid generated 547.5 TWh in 2025, 95.2 percent low-carbon.
Indonesia, by contrast, is renting the top of the stack, 170,000 Nvidia chips, before it owns anything underneath. Here is the layer nobody controls yet, anywhere, and why the order you build in matters more than the flag you fly.
In 1994 I built my first website on a hand-me-down Packard Bell running Windows 95, in a bedroom close to Lyon France. I owned none of the stack under that machine. Not the chip, not the operating system, not the phone line it dialed into. I owned one thing: the code I wrote and the habit of shipping it. That single owned layer is the only one that ever compounded.
Everything since has been built on top of it. Countries are making the same bet right now, in public, and most commentary is reading it backwards.
Nintendo never tried to win the spec sheet
Nintendo has spent forty years refusing to compete on raw hardware power, and winning anyway. The Switch runs on a chip years behind its rivals. What Nintendo owns instead is the layer that compounds: the library, the characters, the control scheme nobody can clone. It never waited to own the fastest processor before shipping. It shipped what it controlled and let the rest catch up.
France is running the same play with AI. It is not waiting to own the chip layer before investing in the layers that compound: energy, national compute, and a model lab. That is not a consolation prize. It is the correct order to build in when you cannot yet own everything.
France’s energy layer is already owned, and it is the whole advantage
France’s nuclear grid is the sovereign AI advantage nobody planned for in the 1970s. The French grid generated 547.5 TWh in 2025, with 95.2 percent low-carbon output, mainly nuclear, according to a July 2026 analysis from TNW (source). Most countries racing to build AI infrastructure are fighting grid constraints and energy shortages right now. France is sitting on a surplus built for entirely different reasons decades ago.
That surplus is already attracting capital. Fluidstack has committed roughly 10 billion euros to build a nuclear-powered AI supercomputer on French soil, incorporating around 500,000 AI chips (source). SoftBank is backing a separate 75 billion euro data center program targeting 3.1 gigawatts by 2031.
Energy is the layer you cannot rent your way into fast.
France already owns it.
Jean Zay turned French compute from a research tool into national infrastructure
Jean Zay, France’s national supercomputer, just quadrupled its dedicated AI computing capacity (source). That is a government treating compute as sovereign infrastructure, not a line item. Paired with the Fluidstack-backed cluster, France now runs two owned compute layers instead of renting capacity from a foreign hyperscaler by default.
French AI startups raised 4.6 billion euros in the first half of 2026, up 65 percent year over year. Capital follows owned infrastructure. Nobody funds a factory built on a landlord’s terms at that pace.
Mistral is the layer almost no other country in Europe can claim
Mistral, the Paris-based lab, is reportedly closing in on a 23 billion dollar valuation (source). Its annual revenue passed 400 million dollars in February 2026, up from roughly 20 million a year earlier. The same week as this issue, Mistral shipped Robostral Navigate, its first physical-navigation model, and CEO Arthur Mensch confirmed a new open-weight flagship model heading into partner access, aimed at closing the gap with the largest US labs.
Owning the model layer means owning the thing that improves every week without renting someone else’s roadmap. France has it. Germany does not. The UK does not. This is the layer that separates a country with an AI strategy from a country with an AI budget.
The chip layer is the one nobody owns, and that changes the whole story
Nvidia now runs a strict new approval process for its Asian customers to keep chips out of China, and more than half of its past regional customers failed the new checks this month (source). New rules from late May 2026 require a license for any transfer of Nvidia’s most advanced chips to entities linked to China, closing the overseas-subsidiary loophole. Nvidia’s share of the Chinese AI chip market has gone from roughly 95 percent in 2023 to effectively zero on new shipments by mid-2026.
This is the layer that makes « sovereign AI » a misleading label for any single country, France included.
Even France’s nuclear-powered, Jean Zay-backed data centers run on chips subject to US export licensing for years to come. That is not proof France’s strategy is fake. It is the reason the sequencing France chose, own energy, own the model, rent the chip for now, is the smart order rather than a compromised one.
Indonesia is renting the top of the stack before owning the bottom
In Batam, Nvidia, Firmus Technologies and DayOne are building a 360 megawatt AI Factory, deploying up to 170,000 Nvidia chips by 2028, launched in Jakarta this month under the banner « Architecting Indonesia’s Sovereign and Scalable AI Future » (source). Indonesia’s AI market is projected to approach 11 billion dollars by 2030.
The investment is real. The sequencing is different from France’s. Indonesia is starting at the top of the stack, chips, before it has an owned energy advantage or a homegrown model lab underneath. That is not a failure. It is an earlier stage of the same build, and the order matters for what « sovereign » will actually mean by 2028.
Own vs rent, by layer
| Layer | France | Indonesia |
|---|---|---|
| Energy | Owned. 547.5 TWh, 95.2% low-carbon nuclear grid | Not yet a distinct AI advantage |
| Compute | Owned. Jean Zay (quadrupled capacity) + Fluidstack 10B euro cluster | Rented. 360MW AI Factory, Nvidia-supplied |
| Model | Owned. Mistral, ~23B dollar valuation, shipping frontier models | Not yet built |
| Chip | Rented. Subject to US export licensing like everyone | Rented. 170,000 Nvidia chips by 2028 |
Three owned layers versus one. That table is the whole argument.
Build on what you own. Rent the rest. Compound.
Build.
France did not wait to own the chip layer before investing in Jean Zay, nuclear capacity, and Mistral. It built what it could control now. Do the same with your own AI stack: do not wait for full ownership before you start. Build the process on today’s tools and let ownership of the harder layers, better vendors, better contracts, better leverage, catch up over time.
Advise.
If you sit on a board or advise a company, stop asking « are we sovereign. » Ask which layers you own: your data, your workflows, your trained processes. Ask which you rent: the underlying model, the compute, the chip. France can answer this question layer by layer. Most companies cannot answer it at all.
Compound.
Jean Zay’s capacity compounds. Mistral’s model improvements compound weekly. French nuclear energy, built decades ago for reasons that had nothing to do with AI, is compounding into an advantage nobody planned. The process you own and refine every week in your own business compounds the same way, long after any single vendor relationship has changed.
This is the same logic behind why we built Asymmetriq the way we did. Most companies rent their entire AI stack from day one, a chatbot subscription with no owned data layer, no owned workflow, nothing that compounds. Asymmetriq starts from the layer a business can actually own: its trained, documented AI process. Companies running it typically replace tooling that cost close to 4,700 dollars a month in scattered subscriptions with one owned system. Own what compounds. Rent what you cannot build yet. That is the French model and the Asymmetriq model, applied at two different scales.
Frequently asked questions
Q: Is sovereign AI actually achievable for any single country right now?
A: No, not fully. Every country, including France and the United States, still depends on Nvidia-controlled chip supply subject to export licensing. Sovereign AI in 2026 means owning the most layers you can, not owning all four.
Q: Why is France’s nuclear energy relevant to AI at all?
A: AI data centers are constrained by power, not just chips. France’s grid produced 547.5 TWh in 2025 at 95.2 percent low-carbon output, giving it spare, reliable capacity most countries currently lack for AI buildouts.
Q: Is Mistral actually competitive with OpenAI or Anthropic?
A: Mistral’s revenue passed 400 million dollars in February 2026, up twentyfold year over year, and it is shipping new frontier and robotics models on a fast cadence. It is smaller than the US labs but it is the only European lab operating at genuine frontier scale.
Q: Is Indonesia’s AI strategy actually behind France’s?
A: It is earlier-stage, not wrong. Indonesia is investing seriously in compute (170,000 Nvidia chips by 2028) but has not yet layered in an owned energy advantage or a national model lab, the two layers doing the most work for France.
Q: Why do most companies get their AI stack backwards?
A: Most rent every layer, model, workflow, and even their process for using AI, then wonder why nothing compounds. The fix is inverting the order: own your data and your trained process first, rent the model and compute underneath.
Q: What happens to French sovereignty if US chip export rules tighten further?
A: France’s owned layers, energy and models, become more valuable, not less. The chip constraint is shared by every country. The country with the most owned layers underneath absorbs a chip shock better than one renting the whole stack.
The verdict
Sovereign AI is not a country’s slogan. It is a ledger, layer by layer, of what you own versus what you rent, and France’s ledger currently reads better than almost anyone else’s outside the United States and China. By January 1, 2027, expect a Mistral model to be confirmed as the named production engine on at least one flagship French sovereign compute cluster, publicly documented. Expect no equivalent homegrown model claim yet for Indonesia’s Batam AI Factory. That gap will be the clearest public proof that owning the model layer, not just the building, is what sovereign AI means in practice.
Three questions worth asking this week: Which layers of your AI stack do you actually own today, and which are you renting? Are you waiting to own the whole stack before you start, or building on what you control while the rest catches up? If you ranked your AI dependencies by layer, data, workflow, model, compute, which one would compound the most value over the next three years?
That bedroom near Pierre Bénite taught me this before I had language for it. You do not get the whole stack on day one. You get one layer, you build on it, and you compound from there. France is doing exactly that, in public, at national scale. Build your own AI stack the same way.
If you want to build that owned layer directly with me, I run a small, hands-on version of this work: the AI Sprint, four live two-hour blocks, one on one, to deploy Claude inside your business rather than talk about it. Three spots a month, 2,900 dollars. Details at remybigot.pro/sprint. If you want to talk it through first, book a call at remybigot.fr/call.