Most Countries Have Already Lost the AI Race. Here Is What They Do Next

Somewhere in Abu Dhabi, a gigawatt of computing power is rising out of the desert, financed by Gulf capital, filled with American chips, and wired into a data centre ecosystem run out of California. It will be, for a while, the largest such facility on Earth. The United Arab Emirates paid dearly for it and negotiated hard. And yet the uncomfortable truth, laid bare in a new Chatham House study, is that the UAE does not own the thing it built. Not really.

That paradox sits at the heart of what a team of researchers in Chatham House’s Digital Society Programme call the problem of sovereign AI. Their central claim is bracing: for every country on the planet that is not the US or China, complete technological independence in artificial intelligence is simply off the table.

The authors, led by Francisco Javier Varela Sandoval, spent much of 2025 interviewing the people actually trying to build this stuff, from Brussels to Bangkok. What they found was a world of middle powers, dozens of them, all reaching for the same prize and all constrained by the same hard limits. You can measure the gap in a single figure. By March 2025, the United States accounted for roughly three-quarters of the world’s AI supercomputer performance. China held about 15 per cent. Everyone else, put together, scrapped over the remainder.

So the question the paper poses is not the obvious one. It is not “how does a country win the AI race?” The race, for most, is already lost. The better question, the authors argue, is far more interesting: on whom should you choose to depend, and on what terms?

To answer it, they break AI down into eight building blocks. Three are foundational: data, compute (the raw processing muscle), and the frontier models themselves. Five more make up the surrounding scaffolding, things like energy, talent, infrastructure and public trust. Almost nowhere does a middle power hold all eight. The trick is knowing which ones you can plausibly grab, and which you must borrow.

Four ways to be almost sovereign

From this comes the paper’s spine: four pragmatic pathways, none of them a clean victory. A country can specialize, carving out one indispensable niche in the global supply chain. It can align, throwing in its lot fully with Washington or Beijing in exchange for access and protection. It can share, pooling sovereignty with like-minded neighbours to punch above its weight. Or it can hedge, cherry-picking bits and pieces from a spread of foreign suppliers while building muscle at home.

The specializers are perhaps the most instructive, because a couple of them have already pulled it off. Take the Netherlands. The Dutch firm ASML is, at present, the only company on the planet capable of making the extreme ultraviolet lithography machines needed to print the most advanced chips. That is leverage, the kind that gives a small country a seat at very large tables. Taiwan’s TSMC, the world’s dominant chip foundry since 1987, has a similar chokehold, though it comes freighted with the ever-present anxiety of what China might one day do about it.

Alignment, meanwhile, is the UAE’s game, and the researchers are careful to frame it as a choice made from strength rather than weakness. Abu Dhabi traded away its dreams of full independence, yes. But it used its capital, its cheap energy and its geopolitical position to bargain for privileged access to American technology. A valued partner, in the paper’s reading, rather than a vassal.

Then there is the cautionary tale that haunts the whole document. In 2014, Google bought a promising London AI lab called DeepMind for somewhere north of $500 million. At the time it looked like a triumph, proof that British science could grow world-beating companies. A decade on, the authors suggest, it should be read rather differently: as a country nurturing something extraordinary and then failing, utterly, to hold on to it. The talent flowed to America. The sovereignty flowed with it.

The clock nobody can read

What gives all this an edge of urgency is a variable no one can pin down: how long until AI becomes something categorically more powerful. The paper does not take a firm view on whether artificial general intelligence arrives in two years or twenty, sensibly, because the experts themselves are all over the map. But the timeline matters enormously. A short runway to that moment hands the advantage to the superpowers who are already sprinting. A longer one gives everybody else room to build. Middle powers are, in effect, being asked to place bets on a clock they cannot read.

And there is a deeper reframing lurking underneath the whole analysis: For decades we’ve measured a nation’s clout by its economy, its army, its diplomats. This work suggests a reshuffling is underway, in which the capacity to shape technical standards, govern who gets access, and manage your own dependencies becomes a form of power in its own right.

Whether any of the four pathways delivers is something we won’t know for years. None guarantees sovereignty. At best they offer a measure of agency in a system where nobody outside two capitals gets to be fully in charge. The prize, as the paper frames it, is to keep sovereignty from becoming the privilege of a couple of giants, and to make it instead something the rest of the world negotiates for, together, one hard bargain at a time.

Frequently Asked Questions