High-Income Countries Take Most of AI Everything

Money is the most lopsided measure in the set. As of July 2025, high-income economies accounted for 91% of all venture capital raised by AI startups, upper-middle-income economies 8%, and lower-middle-income economies roughly 1%. Within that 91%, the United States alone holds 73% of the global total — its AI startups had raised more than $500 billion between them — leaving every other high-income economy to share the remaining 18%.

China is effectively the whole of the upper-middle-income 8%. Everything beneath it is a rounding exercise: the World Bank puts all other middle-income economies and every low-income economy together at under 2% of global AI venture funding, which is where the 1% on the lower-middle-income segment comes from.

Measured against output rather than population it still looks steep. High-income economies produce 64% of world GDP and attract 91% of AI venture capital, so the money is concentrating faster than the economies behind it. Funding is also the most forward-looking row here: it is a bet on what will exist in five years, not a record of what exists now.

The catalogue behind this row runs back to the 1950s and contains more than 900 notable AI models. High-income economies supplied 87% of them, upper-middle-income economies 13%, and lower-middle-income economies no measurable share. The United States alone accounts for 62% of lead contributors, other high-income economies for 25%, and China for the entire 13%.

That zero is worth reading precisely. India contributed two models and Argentina one — 0.2% and 0.1% of the total, which rounds to nothing on a chart drawn in whole percentage points. It means almost none, not none, and the difference matters when the number is being used to argue about capability.

Who builds these models has also changed. Before 2022, 51% of lead contributors came from academia against 47% from industry. Since 2022 that has flipped hard: close to 80% of notable models are now led by industry and only 20% by universities. A measure that once tracked where the research faculties were now tracks where the compute budgets are, which is a large part of why the high-income share is so durable.

Counting companies rather than capital narrows the gap only slightly. Of roughly 21,000 AI startups compiled by CB Insights as of July 2025, high-income economies host 86%, upper-middle-income economies 9%, and lower-middle-income economies 5%. The United States holds 45% of them and other high-income economies 40%, with the United Kingdom second worldwide on 6.8% — 1,456 companies.

China ranks third at about 6%, or 1,352 companies, and India fourth at around 4%. Between them those two account for most of what the middle tiers register at all, which is why the World Bank's own summary of this row is blunt: most low- and middle-income economies have few AI startups of any kind.

The distance between this row and the funding row is the interesting part. Lower-middle-income economies hold 5% of the companies and about 1% of the money. Founding an AI company is roughly five times more evenly distributed than financing one — a startup needs an idea and a registration, while a nine-figure round needs an investor willing to underwrite it from another continent.

Research is where the concentration finally loosens. Across 2000–24, high-income economies produced 54% of the world's AI scientific publications, upper-middle-income economies 33% and lower-middle-income economies 14%. Those three add to 101%: the source rounds each segment to a whole number, and this is the row where the rounding shows.

The lead is thinner in volume than it is in influence. Low- and middle-income economies together produced 46% of global AI publications but received only 31% of global citations. The United States accounts for just 26% of high-income academic articles yet 35% of all citations — a modest share of the papers and a commanding share of the attention they get.

This row is also not a story about the poorest countries. China produces 70% of the upper-middle-income total and India 68% of the lower-middle-income one, which works out at roughly 23% and 9% of world AI research respectively. Take those two out and every other middle-income economy on Earth shares about 14% of the world's AI papers between them.

Generative-AI patents filed between 2014 and 2023 are the single measure on which high-income economies lose, and they lose by more than two to one: 32% against 66% for upper-middle-income economies, with about 2% for lower-middle-income ones.

The 66% has one owner. The World Bank's own text states that China accounts for 66 percent of global GenAI patent filings during 2014–23, citing WIPO — so the upper-middle-income bar and China are the same number, and the other roughly fifty upper-middle-income economies contribute almost nothing to it. WIPO's own count puts China-based inventors at 38,210 of about 54,000 GenAI patent families over that decade, with Tencent, Ping An Insurance and Baidu the three largest applicants in the world.

Everyone else is small. WIPO records 6,276 families from the United States, 4,155 from the Republic of Korea, 3,409 from Japan and 1,350 from India, while Germany and the United Kingdom together account for under 3% of all generative-AI patent families. No low-income economy filed a single AI patent application anywhere in the entire decade.

A patent count is not a capability count. Filing is cheap relative to training a frontier model, filing strategy varies enormously between jurisdictions and between companies, and a patent family records what an organisation wanted to protect rather than what it shipped. This row belongs on the chart because it is real. It does not belong in a sentence claiming any one country leads AI overall, because the other four rows say the opposite.

The five bars are not one snapshot. Notable AI models run from 1950 to June 2025, research papers from 2000 to 2024, generative-AI patents from 2014 to 2023, and both startup measures are counts as of July 2025. A model built in 1997 and a patent filed in 2023 sit in the same picture — which is fine for describing a distribution and misleading for describing a moment.

The attribution rule matters more than it sounds. The World Bank's note on the figure states that where a model has multiple contributors, only the nationality of the first or leading contributor is counted. Modern AI research is heavily co-authored across borders, so a model with a lead author in California and half its team elsewhere registers entirely as high income.

Two of the fifteen segments are also not published numbers. The 1% on lower-middle-income funding and the 2% on lower-middle-income patents are drawn in the source chart but never labelled; both were obtained by subtracting the labelled segments from 100.

Both survive a check without being confirmed by one. The World Bank's text puts all other middle-income and low-income economies together at under 2% of AI venture funding, and WIPO's India count — 1,350 generative-AI patent families out of roughly 54,000 — works out at about 2.5%. Close to the derived figures, but not the same thing as a number the publication printed.

The patent row stops at 2023, and the two years after it were the largest in the field's history. WIPO reported in July 2026 that more than 56,000 new generative-AI patent families were published in 2024 and 2025 combined — more than the entire decade the chart covers — and that China-based inventors accounted for over 43,000 of them.

That is a wider margin, not a narrower one. Six of the world's top ten generative-AI patent applicants are now based in China. Japan climbed from fourth place to third on a 210% compound annual growth rate, with SoftBank the single largest applicant worldwide, while United States filings grew at 92% a year. The economies moving fastest are the ones already inside the top five.

None of that touches the other four rows, which is the point. Patent volume has accelerated almost everywhere without shifting who raises the money or who ships the frontier models. If anything, the chart's one reversal is becoming more of an outlier rather than the beginning of a pattern.

Strip the five measures back and the pattern is simple. High-income economies, holding 17% of the world's people, take 91% of the money, 87% of the frontier models and 86% of the companies. Their weakest row is research at 54%, and their only loss is a patent count that one upper-middle-income country wins single-handed.

The consequence is not that developing economies cannot use AI — they already do, at scale. It is that the tools they use are specified somewhere else. A model whose lead contributor sat in a high-income economy learns from that economy's data, languages and assumptions, and 87% of notable models were led from there. Adapting a system is a weaker position than designing one, and this chart is a picture of who gets to do which.

The row that reverses is instructive precisely because it is so narrow. It took a decade of coordinated state support, a very large domestic market and one specific measure — filings, not products — for a single upper-middle-income economy to beat the top tier at anything here. That is the scale of effort required to move one bar, and four bars have not moved at all.