Sovereign AI: Why Every Nation Will Build Its Own Stack

For most of 2023 and 2024, the dominant story about artificial intelligence was a story of singularity. One model, then a slightly better model, then a model that could pass the bar exam. The frontier was a leaderboard, and the leaderboard had a single column.

That story is ending. Not because the models stopped getting better — they did not — but because the political economy underneath them has finally caught up. France funded a national champion. The UAE built compute clusters the size of small towns. Japan rewrote its copyright law to clear the path for domestic training data. India announced a sovereign LLM in three official languages. The European Union shipped the first comprehensive AI regulation, then immediately wrote the first national exemption to it.

Compute is the new oil refinery. Data is the new currency reserve. The countries that conflate the two will lose both.

I. Compute

The first thing every government noticed was that the supply chain for frontier AI was almost entirely American. The chips came from a single Taiwanese fab. The networking equipment came from two American companies. The hyperscale data centers were built and operated by three American clouds. The training runs were paid for by American venture capital, denominated in dollars, governed by American export law.

You can build a national AI strategy on top of that supply chain. Most of the world is. But you cannot pretend you have sovereignty over the result. The export-control regime that throttled high-end chips to China in October 2022 made that explicit, and every other country watched it happen.

What they concluded was that compute was now a strategic resource, on par with oil refining or rare-earth processing. And like those resources, the answer is not to build a complete domestic supply chain — that is a fantasy on a five-year horizon — but to build enough domestic capacity that the country has leverage in a crisis. France has framed its strategy in exactly these terms. So has the UAE. So, more quietly, has Singapore.

II. Data

The second thing every government noticed was that the data underneath the models was, almost without exception, English-language and American-internet shaped. A French model trained on Common Crawl will speak fluent French, but it will think in the cultural register of Reddit. A model that is supposed to serve as the substrate for civic discourse — for legal reasoning, for medical triage, for school curricula — cannot be trained that way. Or rather, it can, but the country it serves will pay a long, slow tax in the form of cultural drift.

This is the argument that finally got data residency taken seriously. Not the privacy argument, which has been around for a decade and which most economies were happy to ignore. The argument that landed was the cultural-substrate argument: if the foundational model is trained on someone else’s data, the country is renting its cognitive infrastructure.

A country that rents its cognitive infrastructure has not lost its sovereignty. It has just stopped exercising it.

The result is a quiet but consequential shift in data policy. Japan rewrote its copyright law to clear the way for domestic training corpora. The EU AI Act mandates training-data transparency in a way that is, in practice, a soft data-residency requirement. India is in the middle of building the largest public training corpus outside the United States. None of these are described as sovereignty plays. All of them are.

III. Law

The third thing every government noticed is that the existing legal regimes for AI — liability, accountability, harm — were designed around products. They do not work for substrates. A foundation model is not a product. It is closer to electricity. The question of who is liable when a model gives medical advice that injures someone is not a product-liability question; it is a question about the regulation of essential infrastructure.

Every major economy is now writing this regime more or less from scratch. They are doing it at different speeds and on different terms. The EU went first, with a horizontal regulation that classifies AI systems by risk. The UK went second, with a sector-by-sector approach that pushes the work to existing regulators. The US is, as of this writing, still arguing about whether to do it at all. China has already done most of it, quietly, through the Cyberspace Administration.

These regimes will not converge. They are products of different political traditions, different constitutional arrangements, different theories of what the state owes its citizens. The question for the next decade is whether they remain partially interoperable — the way GDPR and CCPA mostly are — or whether they fragment into incompatible blocs the way payment networks did in the 1990s.

What this means

For policymakers: stop framing the AI debate as a race. The race framing pushes you toward decisions that look strong in a press release and weak in five years. Frame it as infrastructure-building, with the same time horizon and the same humility you would bring to a national grid.

For founders: the addressable market for “the universal AI product” is shrinking, not growing. The market for “the AI product that is legible to a specific national regulator and culturally fluent in a specific market” is the one that is opening up. Almost every interesting product question of the next five years is downstream of that distinction.

For investors: the bet on a single dominant frontier lab is a bet on a world that is getting less likely with every quarter. The bet on the picks-and-shovels — eval, alignment, observability, regional deployment — is the bet on the world that is actually arriving.

For the rest of us: the next decade of AI will not be decided in any one company’s training run. It will be decided in a constellation of national stacks, each tuned to its own language, its own data laws, its own definition of harm. That is a more interesting world than the one we were promised. It is also a more fragile one. It is worth paying close attention.

Footnotes

1. I owe a lot of the framing in section II to a long conversation with the EU AI Act drafting team — see Episode 014 of the podcast.
2. The export-control framing draws on Chris Miller’s Chip War and a half-dozen working papers from CSIS.
3. I am deliberately not naming the “national champion” model that prompted the original draft of this essay. It is one of three.
4. A version of section III appeared first in a longer policy memo I wrote for a private workshop. The thesis there was narrower; this is the more honest version.