Who Is Actually in the AI Race, and Who Is Not

Frontier development requires four things at once, and the requirement is conjunctive rather than additive. A country with three of them does excellent applied work. It does not train frontier models.

The first input is access to advanced compute. This is not only a question of money. Advanced accelerators are manufactured by a very small number of firms, allocated under supply agreements, and subject to export control regimes administered by a single jurisdiction. A country can be wealthy and still be unable to obtain them at the scale a frontier training run requires.

The second is capital patient enough to absorb years of losses. Frontier laboratories consume billions before generating meaningful revenue, and the capital that funds them cannot be conventional venture money operating on a five year return horizon. It must be strategic corporate capital, sovereign wealth, or public funding sustained across electoral cycles. Very few capital markets contain pools of that character.

The third is firm power at industrial scale. Training runs cannot pause when generation drops, so what matters is continuous electricity at predictable cost, not installed renewable capacity. This input has quietly become the binding constraint for a large number of otherwise capable countries, and it is the one their strategies most consistently ignore.

The fourth is researchers who stay. Producing them is not the difficulty. Retaining them against recruitment from a small number of extremely well capitalised laboratories is.

The United States and China hold the full set, and they hold it for structurally different reasons, which matters for how each is likely to be constrained.

The United States position rests on compute and capital. Accelerator design and the export regime governing distribution are both domestic, the concentration of senior researchers is the highest in the world, and the depth of capital willing to fund multi year losses has no equivalent anywhere. Its weakest input is power. Grid interconnection queues, local permitting and transmission constraints have become the practical limit on how fast new capacity comes online, and that is now a routine subject of corporate disclosure rather than a speculative concern.

China's position rests on energy, industrial capacity and coordination. Electricity generation and grid construction proceed at a pace no Western jurisdiction matches, manufacturing capacity across the physical supply chain is deep, state coordination aligns capital, procurement and research toward declared priorities, and the annual volume of trained engineers is very large. Its weakest input is access to the most advanced accelerators, which is a direct consequence of export controls, and its response to that constraint is the subject of the companion piece on the cost curve.

Neither position is fragile in the near term. Both are constrained in identifiable ways, and neither constraint is likely to be resolved quickly, because grid capacity and semiconductor fabrication both operate on decade long timescales.

A second group competes seriously on one or two inputs and builds accordingly. Describing this group as behind misreads what they are doing, because in most cases they are not attempting frontier development at all.

The United Kingdom holds exceptional research depth, a serious institutional position in safety and evaluation, and a government that has made the subject a priority. It lacks both compute scale and power capacity, and its strongest researchers are routinely recruited abroad. Its realistic and largely intentional position is standards, evaluation and applied work.

France is the most credible European case, and the reason is nuclear generation. Firm, low carbon, predictably priced electricity is the input most of Europe cannot supply, and France can. Combined with a genuine domestic laboratory and coordinated European policy support, it has the most defensible European claim to sovereign capability, still constrained by capital depth relative to United States peers.

South Korea and Japan are anchored in hardware. Memory and semiconductor manufacturing give both a structural position in the supply chain that does not depend on training a frontier model, and both have channelled effort into industrial and robotics applications where that hardware position compounds.

Israel produces applied artificial intelligence at a density no country matches relative to population, concentrated in security, defence, cyber and medical imaging. It has neither the power nor the compute for frontier training, and its ecosystem is not organised around wanting them.

India holds the largest pool of qualified engineering talent outside China, a growing domestic deployment market, and rapidly expanding data centre capacity. Its constraints are compute access and capital depth, and its trajectory is the most likely of any country in this group to change category within a decade.

A third group is converting resources it already has directly into artificial intelligence position, without developing the underlying capability. The United Arab Emirates and Saudi Arabia are the clearest cases, and the strategy deserves to be assessed on its own terms rather than dismissed.

Both hold the cheapest firm power in the world and sovereign capital accountable to no quarterly earnings cycle. Both are deploying that combination into large scale compute installations, equity positions across the supply chain, and aggressive recruitment of senior researchers at compensation levels domestic institutions elsewhere cannot approach. The result is real influence, real capacity, and a genuine seat in the conversation.

It is also structurally dependent, and specifically so. The accelerators are designed in one jurisdiction, manufactured in another, and supplied subject to an export policy neither country controls. The leading model weights originate elsewhere. Every party to these arrangements understands them clearly, which is why the negotiations are conducted at head of state level rather than as commercial procurement.

The most persistent error in national artificial intelligence strategy is the assumption that the constraint is people. In almost every case it is not.

Canada produced a disproportionate share of the foundational research behind modern machine learning and continues to train first rate researchers who are recruited into United States laboratories. Nigeria, Kenya, Brazil, Poland and Vietnam all produce more capable machine learning engineers than their domestic ecosystems can absorb. India produces them at enormous scale. The United Kingdom produces them and exports them.

Talent is the input almost every country has, and the one almost every country loses. The loss is not a failure of education policy. It is a rational response by individuals to a compensation and compute differential that domestic institutions cannot match, and no visa scheme, tax incentive or national institute closes a gap of that magnitude. Programmes premised on producing more researchers are therefore addressing an abundance rather than a shortage, while the actual constraints, which are power and patient capital, receive a fraction of the attention and none of the announcements.

For every country outside the first two groups, the useful reframing is that sovereign artificial intelligence does not require training a model.

It requires three things that are achievable. Owning the deployment layer, so that models run on domestic infrastructure under domestic law, which open weight models have made dramatically cheaper in the last two years and which is the single most consequential development for this group of countries in a decade. Owning the data, particularly public sector health, education, administrative and geospatial records, which is genuinely scarce, cannot be reproduced by a foreign laboratory, and is the one asset a mid sized state holds that the frontier laboratories cannot obtain elsewhere. And owning the regulatory regime, which sets the terms on which foreign models operate domestically and is the lever that has proven most effective in practice.

A country that holds those three has meaningful sovereignty. A country that spends its entire artificial intelligence budget attempting to train a frontier model it cannot serve, with power it does not have, will hold none of them and will not have the model either.

Four developments would change this assessment, and all are observable.

Grid interconnection timelines in the United States, because power has become the binding domestic constraint and a resolution there extends an existing advantage. Chinese domestic accelerator production yields, because a viable domestic alternative removes the one input China lacks. European firm power decisions, particularly on nuclear extension and new build, because that is the input separating France from the rest of the continent. And the rate at which open weight models are adopted for sovereign deployment, because that determines whether the third group of countries acquires real capability or remains a market.

None of these requires a forecast. All of them are reportable, and LUMINAIRE will track them. Readers should also read the companion reference on the cost curve pressure facing United States model developers, and the account of how leverage rather than technology destroyed a forty five billion dollar artificial intelligence fund this summer.