If your strategic view of the global AI model landscape treats the US and China as the only meaningful producers and the rest of the world as consumers, the production landscape is broader than that framing captures. Cumulative AI models released by region: United States 1,618, China 849, Europe and Central Asia 666, East Asia and the Pacific (excluding China) 330, North America (excluding US, primarily Canada) 125, Middle East and North Africa 74, South Asia 21, Latin America and the Caribbean 2. The US-China duopoly that dominated 2020-2022 has expanded into a multi-polar landscape with Europe as a meaningful third bloc and emerging regional hubs producing visible model output.

The trajectory observations matter as much as the absolute counts. Three structural data points:

The first: US cumulative model output grew from 237 in 2018 to 1,618 in 2025, roughly 7x growth. China's model output grew from 151 in 2022 to 849 in 2025, more than 5x growth in three years. The Chinese acceleration since 2022 is the steeper trajectory. Both countries continue to expand, but the rate at which China is closing the absolute gap is meaningfully faster than the rate at which the US is extending it.

The second: Europe and Central Asia grew from 127 to 666 models over 2018-2025, with the United Kingdom (229 models) and France (141) as the leading contributors. Canada (125 models, captured in the North America region figure) is the fifth-largest country producer overall. The European bloc is now producing roughly 41% of US output, a meaningful share that did not exist as a comparable bloc in 2020.

The third: East Asia and the Pacific (excluding China) grew from 39 to 330 models, an 8x expansion. The bloc includes Japan, South Korea, Singapore, Taiwan, and other regional producers. The growth rate is the steepest of any region.

Beyond the top blocs, regional emerging hubs are producing visible if smaller outputs. Chile's Latam-GPT initiative, the UAE's Falcon model series, Singapore's SEA-LION, and several India-based model initiatives represent intent toward regional model sovereignty. The Middle East and North Africa region reached 74 models. South Asia (largely driven by India) reached 21. Latin America and the Caribbean 2.

The methodological note: the Epoch AI dataset captures publicly reported model releases, including smaller and less prominent ones. This differs from the notable-model dataset, which applies narrower criteria such as state-of-the-art performance and high citation counts. The broader count here is more representative of the expanding base of model development. The narrower frontier count is more sensitive to frontier-level changes. Both views are valid; the broader view captures the multi-polar expansion that the frontier view doesn't.

Three structural implications follow for organisations engaging with the global model landscape.

The first implication: model sovereignty is now a more variegated landscape than 2020-2022 framing suggested. Countries that "couldn't produce their own AI models" three years ago are now producing meaningful national or regional model output. The UAE's Falcon, Chile's Latam-GPT, India's various model initiatives represent not just symbolic sovereignty but operational model availability for domestic use cases. Strategic plans that treated model selection as a choice among 4-5 US and Chinese options now have access to a substantially broader set.

The second implication: the regional models often serve specific national or linguistic needs that the leading US and Chinese models don't address well. Language-specific models (e.g., Latam-GPT for Spanish/Portuguese, SEA-LION for Southeast Asian languages), domain-specific models trained on regional data, and culturally-aligned models can outperform leading frontier models for specific use cases even when their general capability is lower. The model selection framework needs to engage with this specificity rather than treat model selection as primarily a frontier-capability decision.

The third implication: the open-source diffusion is a separate but related dynamic. AI-related GitHub activity shows open-source development diffusing more broadly across regions, even where the absolute scale and capability asymmetries persist. The regions that produce few proprietary models often participate substantially in open-source AI development. The combination of limited proprietary model output and meaningful open-source contribution produces a regional AI development position different from what the proprietary-model-count alone suggests. The strategic implication: open-source participation is a different sovereignty layer than proprietary production, and many regions are stronger on the open-source dimension than the proprietary dimension.

The trajectory through 2026-2030 suggests continued expansion of the multi-polar pattern. China's growth rate will continue to be steep. Europe will continue expanding through both UK/France/Germany-led production and EuroHPC-supported initiatives. East Asia (Japan, Korea, Singapore, Taiwan) will continue at high growth rates. The Middle East (UAE specifically) and India will likely close meaningful share. Latin America and sub-Saharan Africa face the steepest gap-closure challenges but have visible initiative.

The methodological caveat: the regional model counts are conservative estimates for regions with less systematic model documentation and reporting. The growing ecosystem of smaller and language-specific models in sub-Saharan Africa, for example, is not fully represented. The actual model output in trailing regions may be higher than the recorded count suggests. The trajectory direction (more models, more regions) is robust to the measurement caveat; the absolute counts are less so.

For organisations selecting AI models for global deployment, the planning implication is that the model selection space has broadened. The 2022 default of "select among 4-5 leading US models or 2-3 leading Chinese models" is now too narrow. The 2026 selection includes leading models from at least three regions (US, China, Europe), regional and language-specific models from multiple jurisdictions, and a growing set of open-source models that may match or exceed proprietary alternatives for specific use cases. The selection framework that engages with this breadth produces better-fit selections than the framework anchored on the 2022 default set.

The trajectory: the global AI model production landscape is becoming more multi-polar. The US lead in absolute count is narrowing in relative terms. China is closing absolute gap at a steep rate. Europe is producing meaningful third-bloc output. Regional emerging hubs are producing visible if smaller output. Open-source diffusion is broader than proprietary production. The strategic anchor for 2026 should recognise the multi-polar landscape rather than continue the duopoly framing that the data has moved past.

Sources

  • Primary: Stanford AI Index 2026, Chapter 8 (Policy and Governance) 8.3 — hai.stanford.edu/ai-index/2026
  • Model release tracking: Epoch AI, 2026 — publicly reported AI model releases by region, 2018–2025
  • Cross-reference: Stanford AI Index 2026 Chapter 1, 1.5 — open-source diffusion analysis
  • Regional model examples referenced: UAE Falcon series; Chile Latam-GPT; Singapore SEA-LION; UK and France model production; Japan/Korea/Singapore/Taiwan regional production