AI-Innovation Index

A country-level AI footprint built entirely from open data, read alongside WIPO's Global Innovation Index. Like any composite built from the data that happens to be available, it has real limitations, covered in full in the Methodology & sources tab. I go further into what building this taught me about how data can shape a ranking in the companion piece: Who actually wins the AI Olympics?

AI-Innovation Index score, 50 economies
0 to 100, min-max normalised. Pillar contributions shown by colour. Hover for detail.

Copy this snippet — the chart resizes with its container.

Worth noticing: the United States doesn't just lead. Its 21-point gap over second-placed Singapore is larger than the gap between Singapore and tenth place. Measured directly, AI activity is concentrated to a degree that broad innovation composites rarely show. And the second tier is not who you might guess: Singapore, the UK and Hong Kong punch far above their size on per-capita research output and developer activity rather than frontier-model counts.

Where the index disagrees with the GII

Each economy's rank on the AI-Innovation Index (vertical) against its GII 2025 rank, re-computed within the same 50 economies (horizontal). Points below the diagonal do more AI than their innovation ranking suggests; points above have strong traditional innovation systems and a more modest AI footprint. If the two indices agreed perfectly, every dot would sit on the dashed line. They don't, and that gap is the whole point of this page.

AI-Innovation rank vs. GII 2025 rank (within the 50-economy universe)
Rank 1 = best, both axes. Diagonal = perfect agreement. Spearman rank correlation: .
Ranks ≥5 places higher on the AI index than on the GIIRanks ≥5 places lowerBroadly similar (within ±4 places)

Copy this snippet — the chart resizes with its container.

Largest disagreements
Positive delta = ranks higher on the AI-Innovation Index than on the GII.
Higher on the AI-Innovation IndexHigher on the GII 2025

Copy this snippet — the chart resizes with its container.

Worth noticing: the disagreements are not random. The economies that gain most, India (+19), Saudi Arabia (+17) and Russia (+16), are ones whose AI activity scales with sheer population size or deliberate state investment. The ones that slip most, Sweden (−17), Finland, Denmark and Austria (−14), are small, per-capita innovation powerhouses: superb traditional systems, modest AI mass. Switzerland, first in the GII, lands eleventh here. Neither reading is wrong; the two indices weigh size against depth differently, which is exactly why they are worth reading together.

Country drill-down

Pillar scores and the raw indicator values behind them. Missing values are shown as “—” and are excluded from pillar averages rather than imputed, the same no-imputation convention the GII uses.

Pillar scores
Compute & frontier modelsAI researchOpen-source AI
Raw indicators
Values as retrieved from source, before normalisation.

Copy this snippet — the chart resizes with its container.

Two profiles worth a look: China's compute-and-models pillar (87) towers over its open-source score (13). That is partly real and partly because GitHub is restricted in China, so read that bar as a floor. India is the mirror image: thin on frontier compute (31) but heavyweight in research volume and especially open-source activity (64), an AI system growing bottom-up through developers rather than top-down through big labs.

The GenAI patent and investment context

Two series that frame the story but could not become composite indicators, because their sources simply don't publish more countries. WIPO's GenAI patent report releases the annual country split only for the top five inventor locations (together about 92% of all GenAI patent families; everyone else is lumped into “rest of world”). The Stanford AI Index publishes private-investment figures at this level only for the US, China and Europe. That thin coverage is itself part of the measurement-gap story, and the reason these charts are context here rather than inputs to the composite.

GenAI patent family publications by inventor location, 2014 to 2023
Source: WIPO Patent Landscape Report on Generative Artificial Intelligence (2024), Figure 16. WIPO, based on patent data from EconSight/IFI Claims, Orbit by Questel and PATENTSCOPE, April 2024.

Copy this snippet — the chart resizes with its container.

Private investment in AI, 2013 to 2025 (US$ billion)
Source: Stanford AI Index via Our World in Data (“Private investment in artificial intelligence”). Only US, China and Europe are published at this level, which is itself part of the measurement-gap story.

Copy this snippet — the chart resizes with its container.

Worth noticing: volume and value tell different stories. China accounts for roughly 70% of the world's GenAI patent families, about six times the US count, yet WIPO's own analysis finds US-affiliated research attracts far more citations, and US private AI investment ($241bn in 2025) runs at more than twenty times China's. Patents count inventions; they don't weigh them.

Methodology & sources

If you only trust a chart after reading the fine print, same here. Everything below can be reproduced from public endpoints; no commercial or paywalled data is used anywhere in this composite.

Framework

Three pillars, three indicators each (nine indicators), over a universe of the 50 economies with the largest AI research output that are also ranked in the GII 2025. The composite is the arithmetic mean of pillar scores; each pillar is the arithmetic mean of its available indicators.

  • Pillar 1: Compute & frontier models (Epoch AI, notable AI models database): notable AI models 2015 to 2025; notable models 2023 to 2025 (recency); total training compute (FLOP) across those models. Multi-country models are attributed fractionally across the countries of the developing organisations, following OECD.AI-style fractional attribution. Absence from the Epoch database is treated as a true zero, not a missing value, because the database is a census of notable models.
  • Pillar 2: AI research output (OpenAlex): publications in the “Artificial Intelligence” subfield, 2021 to 2025; highly-cited AI publications (≥100 citations, same window); AI publications per million population (World Bank population). Country attribution is by author institutional affiliation, following OECD.AI's methodology.
  • Pillar 3: Open-source AI activity (GitHub Innovation Graph): developers pushing to repositories under AI-related topics (basket of 31 topics, e.g. machine-learning, llm, generative-ai, pytorch), last four available quarters (2025 Q2 to 2026 Q1); GitHub developers per 1,000 population; AI-topic pushers per 1,000 developers (intensity).

Normalisation and aggregation, mirroring GII conventions

Indicators with |skewness| > 2.25 and kurtosis > 3.5 (the GII 2025 Appendix I outlier thresholds) are log-transformed before normalisation (this applied to ). All indicators are then min-max normalised to [0, 100] over the country universe. Missing values are excluded from pillar averages rather than imputed, mirroring the GII's no-imputation policy.

Known biases in the sources, read before quoting

GitHub Innovation Graph measures activity on GitHub, which is restricted in China; Pillar 3 therefore severely under-counts Chinese open-source AI activity (domestic platforms such as Gitee are not covered), and China's composite score should be read as a lower bound. OpenAlex has documented English-language and coverage biases that can under-represent non-English research systems. The highly-cited indicator (≥100 citations, 2021 to 2025) mechanically favours earlier publications within the window. Epoch's “notable model” inclusion criteria involve editorial judgement. None of these caveats is hidden by the normalisation. They are the price of using only open, verifiable sources, and you deserve to know them before quoting a number.

Robustness

Re-computing the composite with z-score instead of min-max normalisation moves country ranks by a median of position(s) and at most . The Spearman rank correlation with the GII 2025 (within-universe) is : high enough to show both indices measure innovation capacity, low enough to show the AI dimension carries independent information.

What this composite explicitly does not do

  • It does not claim to replace or improve the GII; it complements it on one dimension.
  • It does not measure “AI capability” in any deep sense; there is no model-benchmark layer, because benchmarks are model-level, not country-level.
  • It does not attribute frontier-model performance to countries: licences, training locations and firm nationality diverge in unstable ways.
  • It excludes indicators whose data could not be independently retrieved and verified at build time: AI venture-capital investment by country (published only via commercial feeds or for US/China/Europe aggregates), elite talent-flow data (MacroPolo tracker; not machine-readable), GenAI patents by country (published for only ~7 economies; shown in the Context tab instead), and national AI policy counts (OECD.AI inventory; not machine-readable in bulk).

Sources

Limitations

This composite has real limits worth knowing before quoting a number from it. The weights across pillars are equal because I chose them to be, not because a statistical test demanded it. Some indicators, like open-source activity in China, are undercounted simply because the underlying data source can't see everything (GitHub is restricted there), not because the activity isn't happening. And several things I'd have liked to include, like AI venture capital by country or talent flows, aren't here at all, because I could only find that data published in aggregate or for a handful of countries, not because they don't matter. A single number, or even nine averaged together, only ever tells you as much as the data behind it allows.