# Chinese Labs Now Train 20 of the World's 50 Most-Used AI Models

Apollo Global Management data puts China's share of the top 50 near 40 percent, and Chinese models processed 98 trillion tokens in June against 53 trillion for United States models.

- Published: 2026-07-13T05:34:23.811Z
- Canonical: https://polylog.news/ai/2026-07-13/chinese-labs-now-train-20-of-the-world-s-50-most-used-ai-mod
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news)

Deployed artificial intelligence is shifting toward China. According to figures compiled by Apollo Global Management and circulated by the channel [AI Post](https://t.me/aipost/7505), about 20 of the 50 most-used AI models now come from Chinese labs, roughly 40 percent, up from 5 at the start of 2025. Over the same period the number of United States models in that group fell from 33 to 28.

Usage data shows a wider gap. Among the 20 most-used models, Chinese systems processed [98 trillion tokens in June](https://tech.yahoo.com/ai/articles/china-ai-models-process-98-123726333.html) against 53 trillion for United States models. Chinese token volume rose 113 percent from May to June, compared with 43 percent for American models. Independent trackers including [Our World in Data](https://ourworldindata.org/data-insights/us-and-chinese-companies-train-almost-all-of-the-worlds-most-used-ai-models) confirm that United States and Chinese firms together train nearly all of the world's most-used models, leaving little share for anyone else.

The driver is licensing as much as capability. Chinese frontier labs, led by DeepSeek and the Qwen family, release downloadable open weights that enterprises and governments cut off from metered American application programming interfaces (APIs) can host themselves. On public benchmarks, the gap between the best open Chinese models and closed United States frontier systems has narrowed to a version or two, and for many production tasks that difference no longer decides which model a buyer chooses.

The numbers deserve caution. "Most used" weights consumer chat traffic heavily, and token counts reward cheap, high-volume serving rather than frontier reasoning. Apollo benefits from a persuasive narrative, and token accounting varies by who measures it. What is verified is directional and consistent across sources. Chinese models are gaining usage share faster than United States models are losing it, and Chinese token throughput has already passed American throughput.

## What this means

The competition is over distribution, not just capability. Chinese labs win share through open weights that bypass United States export limits on model access, and that erodes the pricing power of closed American APIs from OpenAI, Anthropic, and Google in every market where sovereignty or cost rules out a metered United States endpoint. Nvidia and the American compute stack still provide much of the hardware for this training, so the contest is over who captures the application and serving margin, not who buys the accelerators.

## What to watch

- Whether United States export policy extends from chips to model access itself, which would accelerate rather than slow the standardization on Chinese open weights abroad.
- Independent, deduplicated end-customer revenue behind the token counts, which separates genuine enterprise adoption from cheap consumer volume.