How Europe Is Counted in AI: Money, Compute, Models, Price

Abstract

This part sets out how each of the four scoreboards of Europe's AI position is built: venture capital, GPU-cluster compute, model output and inference prices. On money and compute Europe's shares are 6.1 percent of 2025 global AI venture capital as the EU27, about 12 percent once the UK, Switzerland and Norway are counted, and 5 to 6 percent of global GPU-cluster compute. On the record of what gets built, the widely quoted collapse in European model releases is in large part a property of the database it was read from, so the count was rebuilt from lab announcements under one rule applied to all three blocs: 1,290 full language models shipped in the United States, China and Europe between January 2023 and August 2026, of which Europe built 210 and Europe's commercial labs built 89. The rebuilt record still shows a fall, but confined to one half of Europe: those labs' share of the three blocs' commercial output fell from 22.0 percent in 2023 to 20.5, then 10.1, then 4.1 percent in the first eight months of 2026, while the rest of Europe's output, led by universities and national programmes, shipped within 5 percent of their best year. On price, Europe's best model is listed inside the band of China's two leading models while scoring far below both of them on the Artificial Analysis index. The same audit is applied to the money, where China's venture row omits the state and corporate totals whose AI share nobody publishes, and the US hyperscaler headline is total capital spending rather than AI spending. Five measurements do not exist as published figures. This part measures Europe's position without explaining it.