AI Race by Country vs GDP: Who Leads Both, 1960 to 2026

Data pulled: 29 October 2026 (UTC). GDP runs from 1960 to 2025, the latest year the World Bank has published. AI models run to 28 October 2026, the newest release date in Epoch AI's database.

By Frankie C. · Senior Market Researcher, Mindber. Tracks 500+ AI/SaaS tools across SEA markets via the Mindber Innovation Index methodology.

How we assessed this: We pulled GDP in current US dollars for every country from the World Bank's World Development Indicators and downloaded the full Epoch AI models database. Then we counted notable and large-scale models by the country of the organizations that built them. Private investment comes from the Stanford AI Index through Our World in Data. This is analysis of public datasets, not hands-on testing, and every yearly number is in the CSV files linked under the chart.

The United States and China lead both races, and the AI race is far more lopsided than the economic one. The US produced 26% of world GDP in 2025. Organizations based there built 60% of the notable AI models released since January 2025, and American AI companies drew 83% of the world's private AI investment last year.

Our reading of the numbers below: a large economy buys a place in the AI race, not a lead in it. Germany, Japan and India are three of the six largest economies, and Epoch lists no notable model from any of them since January 2025. South Korea makes 1.6% of world output and built 13.

Press play to watch 66 years of GDP, then switch the measure under the chart to see the two AI races. The notable-model view runs from 1960. The large-scale view starts in 2019, the year the first model crossed its threshold.

Each AI bar names a model. It is the one with the most reported training compute from that country's latest year with a release, or the most cited where compute is unknown. Open the table under the chart to read any year as plain numbers.

The same race as a film of under two minutes, with the large-scale model race added:

1:47

The economy race: one leader, and China since 2010

The US has been the largest economy in every year of the World Bank's series, which starts in 1960. Its 2025 GDP was $30.77 trillion. China's was $19.50 trillion, and the whole world's was $118.35 trillion.

China's climb is the main movement on the chart. In 2000 it ranked sixth with $1.22 trillion. It passed Germany in 2007 and Japan in 2010, when it became the second-largest economy. Its share of world GDP went from 14.9% in 2015 to 16.5% in 2025.

The US gained share too, from 24.1% to 26.0% over the same decade. Both leaders grew faster in dollar terms than the world did. US GDP rose 1.68 times between 2015 and 2025, China's 1.73 times, and the world's 1.56 times.

Japan held second place from 1972 to 2009. Then it slid. Its 2025 GDP, $4.44 trillion, sits slightly below its 2015 figure of $4.53 trillion. Germany passed it for third place in 2023. Current-dollar GDP moves with exchange rates, so a weaker yen shrinks Japan's total even in years its economy grows. The World Bank puts Japan's real growth at 1.2% in 2025.

India is the other mover. It passed the UK for fifth place in 2022 and held it through 2024. In 2025 the UK edged back ahead by $47 billion, $4.00 trillion to $3.96 trillion. India's real growth was 7.6% that year, the fastest of the ten largest economies. China grew 5.0%, the US 2.2% and the world 2.9%.

The AI race has had five runners-up

The US has led the notable-model count every year since 1960. Through 28 October 2026, American organizations worked on 704 of the 1,088 notable models in Epoch's database, or 65%.

Epoch marks a model as notable when it meets one of a few tests. It needs more than 5,000 citations, a training run above $1 million, over a million monthly users, a state-of-the-art benchmark result, or equal historical weight. The full criteria sit on the dataset page. Many of the 1,088 are research models from university labs. The rest include the large language models behind today's chatbots.

Second place is where the history shows. The UK held it from 1968 to 1974. Japan held it, sometimes tied with the UK, from 1975 to 2005. Its bars from those decades carry Kunihiko Fukushima's Cognitron and Neocognitron, early ancestors of today's image models.

Canada took second in 2006 and kept it until 2017. Those are the deep learning years at the University of Toronto and in Montreal. AlexNet, the 2012 Toronto image model often credited with starting the deep learning boom, is on Canada's bar.

The UK came back in 2018, driven by DeepMind's work in London, from AlphaGo to AlphaFold. China passed it in 2023 and now holds second with 191 notable models. The UK has 89, Canada 57, France 50 and Germany 31.

The two leaders' labs are familiar names. Since January 2025, OpenAI, which makes ChatGPT, worked on 40 of the American notable models. Google DeepMind, behind Gemini, worked on 20. Anthropic, maker of Claude, worked on 16, and Meta AI on 9.

China's list is spread wider. Alibaba worked on 21 of China's notable models since January 2025. DeepSeek and Zhipu AI each have 7, ByteDance and Moonshot 6 each, and Tencent 4. Our month-by-month race of AI brands shows how much attention each of those labs drew outside China.

A share of GDP vs a share of AI

A dash means no notable model in the period. The model shares can add up to more than 100%, because a model built by organizations in two countries counts for both.

Put the two measures side by side and the AI race looks like the economic race with the middle taken out. The US share of notable models is 2.3 times its share of world GDP. China's is 2.0 times. Below them, only South Korea gets more than its economic weight, at 4.1 times.

A simpler way to read it is notable models per $1 trillion of GDP. Since January 2025, South Korea released 6.9 notable models per trillion dollars of output. The US released 3.9 and China 3.4. Canada, France and the UK each released one model, which puts them between 0.25 and 0.43. Germany, Japan, India, Italy, Russia and Brazil released none.

Singapore's two are a reasoning model from the startup Sapient Intelligence and MolmoAct 2, a robotics model built with the Allen Institute for AI. Two models make a thin base for a ratio. The AI market data section of our learn hub explains the other figures we track for AI tools and the companies behind them.

Where the private money goes

Private investment tells a different story from model counts. In 2018 Chinese AI companies raised 32.4% of the world's private AI money, against 48.8% for American ones. By 2025 China's share had fallen to 3.6% and the US share had risen to 82.9%.

The 2026 AI Index puts US private AI investment at $285.9 billion in 2025 and China's at $12.4 billion. Global private AI investment reached $344.66 billion, up 127.5% in a year. Those are current dollars. The shares in the chart come from Our World in Data's series in constant 2021 dollars, and a share doesn't change with the dollar basis.

The AI Index adds its own caveat, and it matters here. Private funding rounds miss China's government guidance funds, so the report says the figures likely understate China's total AI spending. That fits the model counts: Chinese labs kept releasing more models while their share of private money shrank.

The AI Index also found that generative AI took nearly half of all private AI funding in 2025. Our report on SpaceX's IPO and its AI lab follows the xAI money into the public market.

China releases more big models, the US builds the biggest

Count only the big models and China moves ahead. Narrowly. Epoch tracks large-scale models, those trained with more than 10²³ operations of compute, a cutoff some AI laws use. Chinese organizations released 82 of them in 2025, American ones 66.

The 2026 count reverses so far, 20 American to 12 Chinese. Treat that with care: 2026 has 36 large-scale models to date, against 165 in all of 2025, so most of the year's entries aren't in yet.

The top tier is still American. Epoch marks a frontier model as one of the five largest training runs at the time of its release. Since January 2023 there have been 27, and American organizations built 25 of them. The other two are Falcon-180B from Abu Dhabi's Technology Innovation Institute in September 2023 and Zhipu AI's GLM-4-Plus in August 2024. The latest frontier model in the database is OpenAI's GPT-6 Astra, from September 2026.

The split follows from how each side releases its work. Of China's 66 notable models since January 2025, 48 are open-weight models anyone can download. Of America's 120, 15 are. A lab that gives its weights away reaches users by spreading the model, and one that keeps them closed needs the largest model to sell access. Our guide to open-source AI explains what each license lets you do.

Benchmark rank is a separate question from model count. The AI model rankings history shows which lab held first place on the Arena leaderboard each month, and our Model Arena tracks the current boards.

South Korea: a small economy with a national program

South Korea built 13 notable models between January 2025 and October 2026, more than every European country combined. All 13 are open-weight. LG AI Research made six of them, from EXAONE Deep in March 2025 to K-EXAONE 2.0 in July 2026.

Five arrived within three days at the end of 2025. Epoch dates all five between 29 and 31 December: K-EXAONE from LG, A.X K1 from SK Telecom, Solar Open 100B from Upstage, VAETKI from NC AI and a HyperCLOVA X model from Naver. Those are the companies behind the five teams the government chose for its sovereign AI foundation model project.

The Ministry of Science and ICT runs that project as a contest. In January 2026, two of the five teams failed to advance. The ministry said Naver Cloud's entry relied on fine-tuning existing overseas models. Motif Technologies joined later. In the second round, published in August, SK Telecom scored 70.6 points, Upstage 69.9, LG AI Research 69.0 and Motif 65.8, according to Edaily.

Motif-3, A.X K2, K-EXAONE 2.0 and Solar Open2 250B all followed between June and August 2026. We read it as the clearest case in this data of a government program moving a country's model count. South Korea now has 28 notable models in total, close to Germany's 31.

Big economies with almost no models: Germany, Japan, India

Three of the six largest economies barely appear in the AI race. The gaps are long. Germany's last notable model in Epoch's database is DeepL's language model from July 2024. Japan's is rinna's nekomata-14b from December 2023. India has two notable models in total, and both came from teams spread across several countries.

None of this means these countries do no AI work. The notable list measures one thing: who trains the models that set records, draw millions of users or cost millions to train.

Europe as a whole shows the same slide. European organizations worked on 31% of notable models from 2015 to 2019, 25% from 2020 to 2022, and 12% in 2023 and 2024. Since January 2025 they appear on 2 of 200, Mistral's OCR model and Stability AI's SPAR3D.

The large-scale list is kinder to Europe. It has seven European models since January 2025, three from Mistral AI and the rest from Swiss, German, Ukrainian, Polish and Bulgarian research groups. Even so, that is 7 of 201. Talent pulls toward the biggest labs too: our story on Google's AI brain drain follows Noam Shazeer to OpenAI and John Jumper to Anthropic in a single week.

The data behind the chart

This table puts the three measures together for the countries that lead any of them. The yearly series are in the GDP CSV, the notable models CSV and the large-scale models CSV.

A model counts once for each country among its builders, as Epoch records it. Three notable models list only "Multinational" as their country and 19 list none, and we left those out. Hong Kong appears on its own line in both sources, so it stays separate from China here.

GDP is in current US dollars. It mixes real growth, inflation and exchange rates, which is why Japan can grow in yen and shrink in dollars. The chart keeps the 27 countries and territories that reach the top 12 on any of the three measures, and leaves out regional totals.

The 2026 bars count models published through 28 October. Epoch updates the database daily, so those numbers will rise as it adds more of this year's releases. Our bar chart race maker can load any of the three CSV files if you want to restyle the race or cut it to fewer countries.

What these numbers can't tell you

A model count isn't a measure of quality or use. One frontier model can matter more than a dozen research papers, and the notable list treats them the same. For how models rank on tests, use the benchmark boards, which name the source of every score.

Notability is Epoch's judgment. Its staff can add a model with "notability at least as great" as the written tests. We'd expect models from countries that publish less in English to be missed more often. We counted what Epoch lists, as of 29 October 2026.

Private investment leaves out public money, as the AI Index notes for China. It also leaves out what large companies spend from their own cash on data centers and chips.

None of these sources measures how many people use AI in each country, or how much AI adds to GDP. Both questions matter for an economy, and neither has a clean cross-country series yet.

Frequently asked questions

More from Mindber

The AI labs on these bars sell most of the tools in our large language model directory and the wider AI models category. The AI chatbot category lists the assistants built on top of them, from ChatGPT and Claude to DeepSeek and Qwen.