The Sovereign
AI Brief.

In 1960, France detonated its first atomic bomb in the Algerian Sahara. Charles de Gaulle had spent the decade prior arguing that American security guarantees couldn’t be trusted in a serious crisis, that no superpower would risk its own cities to defend a junior ally, and that France needed weapons it owned outright. Six years later he pulled France out of NATO’s integrated military command, in a doctrine called the force de frappe.

The argument was simple: if you don’t control the means, you don’t control the outcome.

Sixty-five years on, the same argument has been ported to a different substrate. Replace warheads with GPUs, megaton yield with FLOPS, missile silos with hyperscale data centres, and you arrive at what governments are now calling sovereign AI.

The premise is the same; the failure mode this time round will be different from de Gaulle’s — and worse.

The World Economic Forum projected in February 2026 that AI-dedicated infrastructure spending will pass $400 billion a year by 2030, growing 10 to 15 percent annually. The Center for a New American Security tracked 139 sovereign AI projects in its index published in late April 2026, with 23 new infrastructure projects logged in Q4 2025 alone.

France committed €109 billion in February 2025 under Emmanuel Macron’s announcement, with a stated target of 1.2 million GPUs deployed by 2030. Macron framed it directly: “This is our fight for sovereignty, for strategic autonomy. We want our cloud, we want our data centres, we want our computing capacities.” Mistral closed €1.7 billion in September 2025, with Dutch lithography monopolist ASML taking €1.3 billion for an 11 percent stake and becoming the company’s largest shareholder. Canada’s Sovereign AI Compute Strategy comes in around $2 billion. India’s IndiaAI Mission is roughly $1.25 billion. The UAE and Japan, between them, account for over two thirds of all publicly disclosed sovereign AI investment, according to the CNAS index.

The numbers are real; but the word “sovereign” attached to them is doing more political work than the current technical stack can back up.

Roughly 70 percent of the projects in the CNAS index involve at least one foreign technology partner, and four-fifths of those involve a US company. Most “sovereign” stacks are running on Nvidia silicon, AWS or Azure infrastructure, and US-trained foundation models with a thin national wrapper bolted on.

Sovereign means something specific; it means owned outright, operated domestically, controllable without external permission, and able to keep running if a partner walks away. By that test, almost nothing tracked under the current sovereign AI banner qualifies.

Saudi Arabia announced an AWS cloud region in 2025 for delivery in 2026, with Amazon committing more than $5.3 billion. Microsoft said in February 2026 that customers could run cloud workloads from its Saudi Arabia East data centre region starting Q4 2026. Google Cloud and the Saudi Public Investment Fund announced a $10 billion AI hub partnership in May 2025. In the UAE, Microsoft and G42 announced a 200 megawatt expansion through Khazna Data Centers in November 2025, part of a $15.2 billion commitment running through 2029. Oracle deployed the region’s first OCI Supercluster in Abu Dhabi powered by Nvidia Blackwell GPUs.

The Gulf is buying American hyperscale capacity at industrial scale, while labelling the result a national capability; Riyadh and Abu Dhabi get the megawatts, the jobs, the strategic narrative, and the photo ops at signing ceremonies. Seattle, Redmond, Austin, and Santa Clara get the dependency.

Then March 2026 happened. Iranian drones struck AWS facilities in the UAE and Bahrain. Reuters reported that an affected UAE insurance platform had to seek regulatory approval to shift workloads outside the region because local rules required insurance data to remain domestic. This was the first time in modern conflict that commercial hyperscale data centres became kinetic targets. As the Associated Press noted, cloud reliability is engineered against component failures, not missile strikes.

What sovereign AI doctrine treats as critical infrastructure has now been targeted as such.

The American answer is Stargate, and it’s already coming apart.

In January 2025, Donald Trump announced Stargate at a White House press conference with Sam Altman and Masayoshi Son standing at his side. The promise was $500 billion in AI infrastructure over four years, with $100 billion deployed immediately. OpenAI’s announcement called it “a strategic capability to protect the national security of America and its allies.”

By April 2026, the Financial Times was reporting that “Stargate” had become an umbrella term covering a series of bilateral deals, with one insider telling them the concept was “completely outdated.” OpenAI paused the UK Stargate project, citing energy costs and regulation. The Norway facility in Narvik was handed to Microsoft after OpenAI couldn’t agree on offtake terms with Nscale. The 600 megawatt Abilene expansion got scrapped. The joint-venture structure was abandoned in favour of bilateral deals with Oracle, AMD, Broadcom, and Nvidia. UK AI Minister Kanishka Narayan told the FT that “the only thing that has changed since the moment of those commitments has been the financing environment for OpenAI.”

Then on April 28, 2026, the Wall Street Journal reported that OpenAI had missed its internal monthly revenue targets for early 2026 and missed its goal of 1 billion weekly active users by year-end 2025. The company has signed more than $1.4 trillion in infrastructure commitments with Oracle, Microsoft, Amazon, Google, CoreWeave, and the Stargate consortium. Its current monthly revenue is around $2 billion. The math, as Bloomberg’s evening newsletter phrased it that day, requires OpenAI to keep doubling its revenue plan for years, then keep doubling again.

The largest private capital expenditure concentration in American history is now anchored to OpenAI’s demand projections, and OpenAI just missed its quarter.

US export controls have been tightened and loosened repeatedly since 2022. Nvidia’s H20 chip was approved for export, then effectively blacklisted by Beijing on security grounds in late 2025. Huawei built its Ascend 910C chips partly with two million chip dies sourced from TSMC through shell companies before US regulators caught the violation. Huawei now plans to ship 750,000 units of its 950PR AI chip in 2026, with CUDA compatibility specifically designed to break Nvidia’s software lock-in.

The Council on Foreign Relations published an assessment in December 2025 estimating that the best US AI chips are five times more powerful than Huawei’s, with the gap projected to widen to seventeen times by 2027. Huawei’s own roadmap, the CFR analysts noted, shows its 2026 chip will be less powerful than its current best chip. Beijing has committed over $50 billion to semiconductor self-sufficiency through 2030. The DeepSeek R1 release in early 2025 was widely called a Sputnik moment for AI in Washington and prompted a quick policy review.

China’s state strategy assumes the gap can be closed through industrial policy combined with scale. The American strategy assumes the gap can be maintained through export controls and capital concentration. Neither side trusts its own position. Both are spending accordingly.

Germany operates the JUPITER exascale supercomputer through EuroHPC. The EU’s InvestAI initiative aims to triple AI compute capacity by 2027 and is funding a network of public AI Factories that give universities and startups subsidised access to large compute. AxeleraAI, a Dutch inference chip company, received a €61.6 million grant from the European Commission to build a domestic accelerator. Switzerland released Apertus, a public foundation model trained on European data. South Korea announced a partnership with Nvidia in late 2025 to deploy more than 260,000 GPUs across “sovereign clouds” and AI factories.

These programs are smaller than the Gulf or Stargate buildouts, and their backers have tended to admit what the larger players obscure. Europe’s published documents repeatedly note that NVIDIA still controls 80 to 90 percent of the AI accelerator market and that diversification is a multi-decade project. France’s “cloud de confiance” program, governed by ANSSI’s SecNumCloud certification, is explicitly designed around legal protection from US extraterritorial surveillance laws like the CLOUD Act. The Europeans are buying American compute and writing legal moats around it.

The decade will tell us whether the legal moats hold under stress.

The physical exposure of compute infrastructure has become a hard military variable. The March 2026 strikes on AWS facilities are a precedent; any country whose AI dependence rests on a foreign hyperscaler’s regional data centre has now learned, publicly, that the facility can be hit and the rerouting can be regulated against. The UK’s reported eight-to-ten-year wait for grid connections, flagged by the WEF in February 2026, doubles as a measure of how slowly sovereign capacity can actually be built once a country decides it needs it. Power availability and land permitting have become the binding constraint, with chip supply secondary.

The capital structure is fragile in ways that previous infrastructure buildouts weren’t. The AT&T comparison repeated since the WSJ Stargate piece is instructive. AT&T’s buildout was backed by a regulated monopoly with predictable cash flows over multi-decade time horizons. The Stargate buildout is backed by a private company whose 2027 revenue plan is, charitably, optimistic. A 10 percent revenue miss in 2026, multiplied across vesting schedules, becomes a solvency problem by 2028. Vertiv, Eaton, Schneider Electric, Caterpillar, Cummins, and three regional utilities are all now indirectly exposed to one company’s growth curve. The American industrial base has effectively underwritten OpenAI’s quarterly revenue.

The doctrine itself is mostly theatre. The CNAS Sovereign AI Index showed 70 percent of “sovereign” projects rely on foreign partners. Nvidia still controls 80 to 90 percent of the AI accelerator market by analyst estimates. The wrapper is political. The technical work is still ahead. The French nuclear analogy works as rhetoric and falls apart at the supply chain. France in 1960 owned its fissile material end to end. France in 2026 doesn’t own a single advanced fab.

Some of these projects may become actually sovereign over time — through the same pathway nuclear programs eventually did, indigenous fabrication, custom silicon, trained domestic operators, and a domestic energy base sized for the load. Most will remain hosted dependencies in a national branding exercise. The split between the two is going to define a lot of what state power means in the 2030s, and the press releases coming out of Riyadh, Paris, Washington, and Tokyo today are not a reliable guide to which side any given country ends up on.

The countries that admit they’re dependent, and use the next five years to build substitution capacity at the chip level, at the energy level, at the operator level, at the model level, will end up holding the upper hand. The countries that announce sovereignty in 2026 and keep buying Nvidia in 2030 will end up holding the bills.

De Gaulle’s force de frappe took fifteen years to build, ate a serious share of French defence spending, and was politically expensive at home and at NATO, but it produced an arsenal France still maintains in 2026. It worked because the people running it understood, from the start, that a sovereignty doctrine you don’t pay for is a sovereignty doctrine that doesn’t survive contact with a serious adversary.

The current sovereign AI brief, judging by its press releases and its supply chain, hasn’t absorbed that lesson yet. The bills come due before the decade does.