Which Countries Actually Have AI Capacity?

Tortoise builds the Global AI Index from 122 indicators across three pillars — implementation, innovation and investment — split into seven sub-pillars: talent, infrastructure, operating environment, research, development, commercial ecosystem and government strategy. For most of those indicators the index takes two readings. One counts total output, and that feeds Scale. One counts the same output relative to population or economy size, and that feeds intensity.

The two are weighted differently on purpose. Tortoise weights each relative indicator at one-third of its absolute twin, which makes the overall index roughly 75 per cent Scale and 25 per cent intensity — a change from the 67:33 split used in the fourth edition, made because access to computing power and capital increasingly decides AI outcomes.

So Scale is an answer to a specific question: how much AI is physically and commercially present in this country in total? It is not a measure of how well a country uses AI, how ready its institutions are, or how much of its economy the technology touches. Those are intensity questions, and they produce a different ranking. The World Bank, which plots both axes together, describes the pair as measuring a country's "AI readiness"; Tortoise's own word for what it measures is capacity, and the distinction is worth keeping.

One more thing about the numbers themselves. Tortoise's methodology states that the overall score and each sub-pillar score are normalised between 100 and the lowest original score, rather than stretching the floor down to zero, on the stated grounds that giving a country zero would wrongly imply no AI activity is happening there at all. That is why Ethiopia, last on all three of scale, intensity and the overall index, scores 2 and not 0. It is also why none of these figures is a mark out of a hundred: they are positions relative to whoever leads. The World Bank's own plot of the same data makes the point by accident, drawing its intensity axis out to 110 with Singapore's leading score of 100 sitting well short of the end.

The United States scores 100. China scores 57. The United Kingdom and India tie for third at 24. That means the gap between second and third place — 33 points — is larger than the gap between third place and last, which runs from 24 down to Ethiopia's 2 across eighty-one countries. Every familiar phrase about a global AI field applies to a range of 22 points; the top of the table needs its own scale.

Below the leaders the field compresses fast. France and Germany sit at 23, South Korea at 22, Canada at 20, and that is the entire 20-plus group: eight economies out of 83. Japan is 19, Saudi Arabia 18, Singapore 16, Spain 15. By the time you reach the twentieth-ranked country you are in single digits of difference between neighbours, and rank ordering stops carrying much meaning.

The tail is where most of the index lives. Sixty of the 83 economies score 10 or below, twenty-three score 5 or below, and the median score is 8. Thirty-one countries sit in the 6-to-9 band alone, which on a map means a third of the coloured world is a single shade apart from another third.

It is worth being blunt about what that shape rules out. There is no smooth global gradient here, no continuum from leaders through a middle tier to laggards. There is a peak of two, a shelf of six, and a floor that 70 per cent of the scored world stands on.

Singapore scores 16 on Scale, which puts it eleventh — behind Japan and Saudi Arabia, level with nobody in particular. On intensity it scores 100 and leads the world outright. Its overall index position, which blends the two at 75:25, is third. Three defensible rankings for one country in one edition of one index, and each of them answers a different question.

The inversion runs both ways. Israel is thirteenth on Scale at 14 and second on intensity at 74. Switzerland is joint twenty-fourth on Scale at 10 and fourth on intensity at 63. Luxembourg scores 9 on Scale and 53 on intensity; Estonia scores 6 and 39. These are small, rich, densely capitalised economies that look modest in totals and formidable per head.

Going the other way, India ties for third on Scale at 24 and ranks thirty-sixth on intensity at 19. China is second on Scale at 57 and twenty-first on intensity at 33. Brazil is 11 and 16. Indonesia is 8 and 11. Large populations divide large totals into unremarkable per-head numbers, which is the arithmetic the intensity measure exists to expose.

This is why the wording matters. A country high on Scale has a lot of AI in it. Whether that AI has reached its firms, its schools and its public services is a separate finding on a separate axis, and the two are not interchangeable in either direction.

The index scores 83 economies. Two of them, Taiwan and Hong Kong, are not UN member states, so 81 of the UN's 193 members are covered and 112 are not. On a world map that is not a scattering of small gaps — it is a continuous blank running from the Sahel across Central Africa, up through Central Asia, and down through Central America and the Caribbean.

Africa contributes twelve countries to the index: Algeria, Benin, Egypt, Ethiopia, Ghana, Kenya, Mauritius, Morocco, Nigeria, Rwanda, South Africa and Tunisia. The continent has 54. Central Asia contributes none — no Kazakhstan, no Uzbekistan, no Kyrgyzstan, Tajikistan or Turkmenistan. Central America and the Caribbean contribute none either, so the whole region between Mexico and Colombia is unshaded.

The reason is coverage, not judgement. Tortoise says it built the country list in consultation with experts and through literature review, starting from countries that had published national AI strategies because those offered enough data to rank. The fifth edition added 21 new countries on exactly that basis, so the blank is shrinking — slowly, and from a list that grew because governments published documents, not because capacity appeared.

The practical consequence is that grey cannot be read as low. A country with no score might have a functioning AI sector and no comparable statistics, or no sector at all; this map cannot tell you which. Anyone using it to argue about where AI is absent is reading the index's coverage decisions, not the world.

The World Bank's Digital Progress and Trends Report 2025, which is where this index reaches most policy readers, makes the argument for why a chart like this gets more lopsided rather than less. Digital technologies carry heavy economies of scale: once a platform or a model exists it can be served at close to zero marginal cost, and its value grows with the number of users. That produces winner-takes-most dynamics in which the largest markets breed the most competitive firms, which then attract more users, more data, more talent and more investment.

A large domestic market also lets firms test across millions of users and sectors without hitting regulatory or linguistic borders, and it justifies the two most expensive commitments in AI — frontier research and compute infrastructure. The report's own phrase for the result is that early advantages "can rapidly snowball into global dominance, as seen in China and the United States", while smaller and fragmented economies risk becoming "consumers of AI, not creators or shapers".

The compute figures underneath that argument are stark. In 2024 the United States hosted 50 per cent of the world's secure internet servers, other high-income countries a further 41 per cent, and everywhere else 9 per cent. On a per-capita basis the United States has roughly 200 times as many servers as a typical middle-income country and 20,000 times as many as a low-income one. As of June 2025 high-income countries hosted 86 per cent of the world's top 500 high-performance computing systems and 97 per cent of their capacity.

None of that is a law of nature, but it does explain why the Scale column is so top-heavy and why the top has been stable. Capacity of this kind is accumulated, not declared, and the countries accumulating it fastest are the two already at the top of the map.

A single Scale number flattens seven sub-pillars, and some of the countries in the middle of this map are made of contradictions. Saudi Arabia scores 100 on government strategy — the highest in the world, ahead of the United States at 83 — and 4 on talent. It has the most comprehensive published AI plan in the index and almost none of the people to run it, which is exactly the combination a strategy document cannot fix on its own.

Italy scores 100 on operating environment, the top score anywhere, covering trust in AI, practitioner diversity and AI in legislative proceedings. On development, the sub-pillar that counts open-source models and AI patents, it scores 2. India is the mirror image in a different direction: 90 on operating environment and 42 on talent — second in the world behind the United States on that measure — against 15 on infrastructure.

Further down, Rwanda scores 68 on operating environment and 2 on infrastructure. Kenya scores 68 and 5. Iceland scores 36 on infrastructure and 2 on government strategy. These are not rounding artefacts; they are countries doing one part of this well and having no realistic route to the others.

The floor is broader than the headline suggests, too. Thirty-three of the 83 economies score zero on development, twelve score zero on research and seven score zero on the commercial sub-pillar. For more than a third of the countries the index bothers to measure, the entire innovation half of the framework registers nothing at all.

The fifth edition of the Global AI Index was published on 19 September 2024, and it measures output from 2019 to 2024, discarding anything earlier. As of 21 September 2026 it is still the edition Tortoise serves from its own data page, and no sixth edition has appeared. The right way to describe these numbers is as a two-year-old snapshot, not as the current state of anything.

Two years is a long time in this particular field. The index was compiled before the 2025 and 2026 waves of national compute programmes, sovereign data-centre deals and frontier-model releases, and before several of the countries in the middle of this table announced capacity that would move them. It predates most of what the phrase "AI infrastructure build-out" now refers to.

What almost certainly has not changed is the shape. A 43-point lead at the top and a 33-point step down to third place is not the kind of gap that two years of announcements close, and the mechanism the World Bank describes — scale attracting the inputs that create more scale — points the other way. The precise scores here are stale. The structure they describe is the safer part to rely on.

Read as a ranking, this map says the United States and China have most of the world's absolute AI capacity and everyone else is arguing over the remainder. Read as a distribution, it says something sharper: that 60 of the 83 economies anyone has measured sit at 10 or below, and that the difference between ranks 20 and 60 is a handful of index points.

Read as a map, though, the finding is the grey. A hundred and twelve UN member states are not in this dataset, which means the most common condition for a country on Earth in this index is not low capacity but no measurement. That is a fixable problem, and a cheaper one than any of the four foundations the World Bank lists.

The thing the data cannot settle is the one people most want from it. Scale says where AI is; it does not say where AI is working. Singapore at 16 and India at 24 are close together on this map and nowhere near each other on the measure of how deeply the technology has penetrated either economy. Any sentence that treats a high Scale score as evidence of readiness has quietly swapped one axis for the other — and on this index those two axes very nearly disagree.