Which Countries Make the World's Most-Used AI Models?

The United States and China, and almost nobody else. In May 2026, US companies trained an average of 27 of the 50 most-used models on OpenRouter each day and Chinese companies trained 20. That is 54% and 40% of the places, 94% between them. The remaining places went to France, with about one model, and to anonymous preview models whose developers had not been disclosed.

Seventeen months earlier the list looked more varied. In January 2025 the two countries held 76% of the places, France held 20% and Canada 3%. The combined US and Chinese share climbed past 90% in May 2025, reached 94% in October 2025 and has stayed above 93% in every month since. Its highest point was 97% in March 2026.

The 94% in May 2026 is therefore the latest figure, not a record. It dipped from 97% in April mainly because anonymous preview models held more places that month, about 1.4 on average against 0.2 in April, not because any third country came back.

Chinese companies trained about 5 of the 50 most-used models in January 2025, 10% of the places. By February 2025 they had already passed France, and by May 2026 they held about 20 places, or 40%. China's share quadrupled in 17 months.

The rise came in two clear steps rather than a steady climb. Between June and August 2025 China's average count went from about 11 to about 17, the largest rise in the series, taking its share from 23% to 35%. It then stayed between 32% and 36% from August 2025 to March 2026. A second step came between March and April 2026, from 36% to 41%, the highest share China has reached.

Neither OpenRouter's rankings nor Our World in Data's analysis say why. The data records where the most-used models come from, not why developers chose them, so the timing of the two steps is a pattern in the list rather than an explanation of it.

Yes, on this list, though by much less than before. US companies held 66% of the top 50 places in January 2025 and reached their highest share, 69%, in May and June 2025. After that their share fell to 58% by August 2025 and to 54% by May 2026, about 27 models a day.

The gap between the two countries narrowed from 56 percentage points to 14. Counted in models, the US went from about 28 more than China in the daily top 50 to about 7 more. China has not overtaken the US in any month of the series.

The US decline and the Chinese rise mirror each other closely. Between June and August 2025, the US lost about 10 points while China gained about 12, and between March and April 2026 the US lost about 5 while China gained about 5. Most of what China gained on the list in those months, the US lost.

France was the third country on the list, and in January 2025 it was a strong third. French models took about 10 of the 50 daily places, 20% of the list and twice China's share at the time. France's share then slid, to below 10% from May 2025 and to 2% from March 2026, about one model a day. Our World in Data credits France's place on the list to Mistral AI, whose NeMo model is the one that remains.

Canada left the list entirely. Our World in Data notes that Cohere's Command R models represented Canada early in 2025. They held 3% of the places in January 2025, fell below 1% by April and were gone from June 2025.

Put together, every country other than the US and China held 23% of the places in January 2025 and 2% in May 2026. No other country appears in the series at all. The models whose developers were not disclosed, mostly anonymous previews, never held more than 3%.

This is one platform. OpenRouter serves developers who build on AI models through its interface. People chatting in the ChatGPT, Claude or Gemini apps, and companies calling a lab's own API directly, are not in these figures, so the list shows what is used on OpenRouter, not what the world uses.

It counts places, not usage. A model in 50th place counts as much as the model in first, so a country with many mid-ranked models can hold a larger share than a country with one dominant model. Our World in Data assigns each model to the country of the company that trained its base weights, so a model adapted elsewhere still counts for the country where it was first trained.

The figures are measured, and they end in May 2026. Our World in Data published the series as a chart, so each value here was read from it and is accurate to about 0.4 points. The analysis was published on 30 June 2026 and does not cover the months since.