Every four years, someone writes the same article: the US topped the Olympic medal table again, and also, if you divide by population, tiny Caribbean and Nordic nations quietly demolished everyone. Both tables are true. They’re just measuring different things, and depending on which one a headline picks, “who’s winning” can mean two completely different countries.
I’ve spent years around innovation measurement, which means I’ve spent years around exactly this kind of argument, just with GDP and patents instead of gold medals. Lately I’ve been watching AI eat the same discourse. Every “AI superpower ranking” I read seemed to be quietly picking a medal table without telling you which one.
So I built both tables myself, for AI specifically, and put them next to WIPO's Global Innovation Index (GII), the closest thing the innovation world has to an official leaderboard. One of those tables isn't a single stat, though. It's a composite indicator, meaning nine different measurements folded into one score per country. What I found is that the AI medal table and the traditional innovation medal disagree about entire countries.
What a composite indicator actually is, and why it’s a bit of a trick
A composite indicator rescales several different measurements onto the same 0–100 range and averages them into one score. The trick is that rescaling anchors the whole range to whichever country sits at the extremes. For example, the US’s training compute is roughly 15 times China’s, which is itself roughly 6 times the UK’s. So one outlier can flatten real differences between everyone else into a narrow band near the bottom.¹ The GII handles this by log-transforming any indicator skewed enough to trigger its own published outlier thresholds, and I mirrored that rule rather than inventing my own, but it dampens the effect, it doesn’t erase it. I’ve tried to ground each surprising result in the actual numbers behind it, so you can judge which is signal and which is closer to an artifact of skewed raw data.
The medal table
Start with the version everyone already assumes: raw AI output across 50 economies — the ones with the largest AI research footprint that also happen to be ranked in the GII.² I built a composite from nine indicators across three categories: compute and frontier models, research output, and open-source activity, all from data anyone can pull themselves.³
The US wins by far. It scores 85 out of 100, and its lead over second-placed Singapore (64) is bigger than the entire gap between Singapore and tenth place. If Olympic medal counts sometimes look like a two-horse race between the US and China, the AI table isn’t even that competitive. China comes in eighth.
Here’s the part that surprised me, though: the podium behind the US isn’t the “usual suspects” list you’d assemble from headlines about AI superpowers. Singapore, the UK, and Hong Kong round out the top five. None of them get there on frontier models or eye-watering compute (this isn’t a vague impression, the actual numbers behind it are stark). Hong Kong has roughly 465 GitHub developers per 1,000 people and Singapore about 355, against the US’s 95. Singapore publishes AI research at roughly 1,551 papers per million population, more than four times the US rate of 380. Once research and developer activity are measured per capita rather than as raw totals, small, dense, highly-educated economies mechanically score well on two of the three pillars, regardless of whether they host a single frontier lab.
AI-Innovation Index score, 50 economies
0 to 100, min-max normalised. Pillar contributions shown by colour.
The interactive version lives on my website.
The disagreement is the real story
I re-ranked the GII 2025 within this same 50-economy group and lined it up against the AI composite, country by country. The two rankings correlate reasonably well (Spearman’s rho of 0.80). Enough to confirm both are measuring some shared notion of “innovation capacity.” But 0.80 isn’t 1.0, and the gap between the two tables sorts countries into two very clean groups.
Countries that outperform their GII ranking on AI: India climbs 19 places, Saudi Arabia 17, Russia 16 — but they get there in three different ways. India has the second-most AI research papers and the second-most open-source developers of any country in the dataset, after only the US. That sounds impressive until you divide by population: per person, India's output is actually fairly modest. It has a lot of AI activity mainly because it has a lot of people, not because each researcher or developer is unusually productive. Saudi Arabia's gain is simpler: one number, the amount of computing power it has poured into training AI models, is bigger than France's or Germany's, thanks to direct state investment rather than a broad research base. Russia's rise looks the least like a real AI success story once you look closely: its AI score on its own is nothing special, it just climbs the ranking because its starting point on the GII was so low. That's less "Russia is quietly becoming an AI power" and more "there was nowhere to go but up."
Countries that underperform their GII ranking on AI: Sweden drops 17 places. Finland, Denmark, and Austria each drop 14. The pattern behind all four looks nearly identical: solid research and open-source activity, but almost nothing on frontier models and compute. Austria, for instance, has essentially no notable AI models and no meaningful training compute to speak of. Denmark and Sweden aren't far behind. None of these countries lack research talent, they simply don't host the handful of labs that train frontier AI systems, and this composite treats that absence as heavily as it treats the presence of one. Switzerland is the starkest case: first in the GII out of 139 countries, but eleventh here, because its tracked training compute is a tiny fraction of what the US alone has poured in. These are small, extremely well-run innovation economies with essentially no frontier-model infrastructure behind them.
AI-Innovation rank vs. GII 2025 rank (within the 50-economy universe)
Rank 1 = best, both axes. Diagonal = perfect agreement.
Want to play with it yourself? The interactive version, with every country and chart, lives on my website.
Neither table is lying to you, same as neither Olympic medal table lies. They’re just weighing “how much” against “how deep” differently, and once you see both, you stop being able to pretend there’s one honest answer to “which country is winning at AI.”
Two countries, two completely different training regimes
If you want the disagreement in miniature, look at China and India side by side, because they’re almost mirror images of each other.
China scores 87 out of 100 on compute and frontier models (genuinely dominant) but only 13 on open-source activity. Disclaimer: that second number is a good example of where this indicator can mislead rather than inform. The open-source score comes entirely from GitHub, but GitHub has been restricted in mainland China for years, so Chinese developers who’d normally show up there mostly don’t. Many of them work instead on domestic platforms like Gitee, which this composite has no way to see. So a low score here doesn’t mean China has little open-source AI activity, it means the main data source for that pillar can’t observe most of it. Read the 13 as a floor set by what the data can capture, not as a real measure of how much Chinese developers are actually building in the open.
India is close to the opposite. It’s thin on frontier compute (31) but strong on research volume and, especially, open-source activity (64). Where China’s AI strength looks top-down, a small number of major labs training frontier models, India’s looks bottom-up, an enormous number of individual developers building with AI tools rather than a handful of institutions building the tools themselves. Same “AI powerhouse” label, two completely different underlying stories, and a composite score would flatten that difference if you didn’t look at the components.
See any of the 50 countries broken down the same way, under the 'Country detail' tab on my website.
The medal table nobody can actually see
Two other numbers kept coming up while I built this, and I want to mention them even though I couldn’t put them in the composite, because the reason I couldn’t is itself informative.
WIPO’s own patent report says China accounts for roughly 70% of the world’s generative-AI patent families, about six times the US count.⁵ Meanwhile, Stanford’s AI Index has US private AI investment at $241 billion in 2025, more than twenty times China’s figure.⁶ Read those two numbers next to each other and you get a genuinely confusing picture: China is filing an overwhelming share of the patents, while the US is receiving an overwhelming share of the money and, per WIPO’s own citation analysis, the higher-impact research.
GenAI patent family publications by inventor location, 2014 to 2023
Private investment in AI, 2013 to 2025 (US$ billion)
Both charts, plus the full sourcing, are under the 'GenAI context' tab on my website.
I couldn’t build proper country-level indicators out of either number, because neither source publishes at the resolution I needed: the patent report only breaks out the top five inventor locations, and the investment figures are only published for the US, China, and Europe as blocs. That’s not a small caveat. It means two of the most-quoted AI statistics in the world (the ones that make it into every “who’s winning AI” headline) are only available as a five-country medal table or a three-region one. Everyone else is lumped into “rest of world,” which is a strange place for most of the planet to live in the data that supposedly describes it.
So, who actually wins the AI Olympics?
Depends which medal table you hand someone, and I think that’s the honest answer rather than a dodge. The raw composite says the US, by a distance nothing else in this data comes close to. The GII-comparison table says the real story isn’t who’s on top, it’s who moved: India, Saudi Arabia, and Russia rising on scale; Sweden, Finland, Denmark, Austria, and Switzerland falling because deep institutional strength doesn’t automatically convert into AI mass.
What I’d resist is the headline that picks one of these tables and presents it as the only honest way to count. The GII took decades to build a reputation for measuring “innovation” broadly; this composite measures something narrower and newer, on purpose, and the two are more useful read together than either is read alone. The full interactive version is on my dashboard, if you want to check whether your own country’s medal depends on which table you’re looking at.
The bigger thing I’d want you to take away, though, isn’t about rankings at all. It’s that a number like this is shaped as much by what I could measure as by what’s actually true. China’s low open-source score exists because GitHub is restricted there, not because Chinese developers build in the open less than anyone else, it’s a gap in what the data can see, not in what’s actually happening. The same goes for what isn’t in the composite at all: AI venture capital, talent flows, policy activity, the GenAI patent and investment numbers I mentioned earlier. None of those are missing because they don’t matter. They’re missing because I couldn’t verify them from an open source, which is a very different thing from them not existing. A single number, or even nine numbers averaged together, will always tell you a partial story shaped by what the person building it had access to. That’s true of this composite, and it’s worth remembering the next time a “which country leads AI” statistic gets thrown around without anyone asking where it actually came from.
A researcher’s disclaimer:
I know the limitations. The weights are equal because I decided they should be, not because a principal components analysis told me so, and there’s no bootstrap confidence interval around anyone’s rank, nor any test of whether that 0.80 correlation with the GII would survive once you controlled for GDP per capita. A referee at a serious journal would be right to hand this back covered in red ink before lunch.
But that’s rather beside the point of this column. I love rigorous research, it’s most of my day job, but a personal column and a peer-reviewed paper are held to different standards for good reason, and I’m not pretending this one clears the second bar. What matters here is that every number traces back to an open source, nothing is quietly imputed to paper over a gap, and the biases I know about are disclosed rather than buried in an appendix. That’s the bar I’m holding this column to, not a 200-page annex.
How I actually built the medal table, for anyone who wants to check my work
I’m not going to pretend a “which country wins at AI” number is self-evidently objective, so here’s what’s actually inside it. Three pillars, three indicators each, min-max normalized to 0–100 and averaged with equal weights, no discretionary thumb on the scale. Indicators that failed the GII’s own published outlier tests got log-transformed before normalizing, the same rule WIPO applies to its own data.⁷ Missing values are left out of an average rather than guessed at, again mirroring GII convention rather than inventing my own.
I also reran the whole composite using z-scores instead of min-max normalization, just to check the ranking wasn’t an artifact of my normalization choice. Ranks moved by a median of one position and a maximum of four. Reassuring, though I’ll admit “reassuring” is doing some work in that sentence, a robustness check tells you the method isn’t fragile, not that the underlying concept is right.
And there’s a longer list of things I didn’t include, on purpose, rather than approximate: AI venture capital by country (published only in aggregate or for a few economies), talent-flow data (not machine-readable at the country level), national AI policy counts (same problem). If a number couldn’t be independently retrieved and verified from an open source, it’s not in here, even where I suspect it would be interesting.
OECD/European Commission Joint Research Centre. (2008). Handbook on Constructing Composite Indicators: Methodology and User Guide. OECD Publishing. See also the widely cited critique of composite indices’ sensitivity to normalization choice.
WIPO. (2025). Global Innovation Index 2025, 18th edition. wipo.int/web-publications/global-innovation-index-2025
Epoch AI, Notable AI Models database (epoch.ai/data/notable-ai-models); OpenAlex works API, subfield “Artificial Intelligence” (openalex.org); GitHub Innovation Graph (github.com/github/innovationgraph). All retrieved 8 July 2026.
GitHub access has been restricted within mainland China since 2013, with intermittent blocking of specific paths continuing since; domestic platforms such as Gitee are not covered by the Innovation Graph.
WIPO. (2024). Patent Landscape Report on Generative Artificial Intelligence. wipo.int (GenAI PLR).
Stanford AI Index, private investment in artificial intelligence, via Our World in Data. ourworldindata.org/grapher/private-investment-in-artificial-intelligence
Outlier thresholds (|skewness| > 2.25, kurtosis > 3.5) as published in GII 2025, Appendix I.