In a parallel track of the AI arms race, Chinese open-source models are quietly reshaping the competitive landscape through download volume data.

According to Hugging Face's Spring 2026 report, Chinese open-source models now account for 41% of the platform's monthly and total downloads, surpassing American models. Recently, this data was highlighted again by Hugging Face CEO Clem Delangue, sparking widespread discussion on social media.

Hugging Face is currently the world's largest open-source AI model hosting platform, hosting nearly 3 million public models and 1 million public datasets, with a new code repository created on average every seven seconds. Half of the Fortune 500 companies have already deployed their own private or open-source models on the platform.

Furthermore, the top six most popular models on OpenRouter are all from Chinese institutions, including Tencent, Xiaomi, DeepSeek, MiniMax, and KNOWLEDGE ATLAS.

Beyond download metrics, changes on the user end are equally significant. On the AI model calling platform OpenRouter, the current top six most popular models are all from Chinese institutions. Anthropic's Claude Opus 4.7 ranks seventh.

KNOWLEDGE ATLAS recently released the open-source model GLM-5.2, positioned for agentic coding tasks, which performs at a competitive level with Anthropic's latest models in relevant benchmark tests.

Infrastructure data also confirms this trend. According to Vercel platform data, in June 2026, open-source models processed nearly one-third of the platform's AI requests, handling a large volume of infrastructure-intensive workloads, while closed-source models have gradually retreated to the high-cost, high-end tier.

Corporate Shift: From Renting to Owning

The core logic driving this shift is cost and control. Delangue stated that companies, after seeing the bills for scaling closed-source frontier models, are reassessing the value of owning their own AI. He noted that an AI or tech company would not want to outsource its core capabilities to a black-box API it cannot control.

Microsoft CEO Satya Nadella has issued a similar warning against single-vendor dependency. He argued that if learning flows only in one direction, economic value concentrates with the owners of the learning infrastructure, not the creators of knowledge. He advocates for dispersing learning infrastructure to every enterprise.

In an AI research report dated June 23, 2026, UBS Securities analyst Karl Keirstead's team pointed out that approximately 60% of enterprises have already implemented some form of AI spending limit, with a core action being to set up guardrails for token usage.

The Rise of Model Routing

Corporate countermeasures are systematically reshaping the AI industry chain's beneficiary landscape. According to the UBS report, "model routing" has become a core technical action for token optimization—assigning different tasks to different models. Only complex reasoning, critical code, and long-context analysis are routed to the most expensive models, while simpler tasks are shifted to lower-cost or even Chinese open-source models.

Price differentials are a direct driver. Taking Anthropic as an example, the output price for Haiku 4.5 is $5 per million tokens, Opus 4.5-4.8 is $25, and the highest-end model reaches $50—a tenfold difference in output token price from low to high end.

Chinese open-source models are entering corporate procurement under this backdrop. As described in a case study in the UBS report, a large global bank has locally deployed Alibaba's Qwen to balance the usage costs of high-end models like Claude. AWS Bedrock's model menu now includes MiniMax, Kimi, Qwen, DeepSeek, and GLM; Microsoft Azure AI Foundry also provides access to DeepSeek.

Can Open Source Determine the AI Race's Outcome?

Delangue's assessment is more direct. Citing his views, Delangue believes open source is an accelerator for AI leadership, suggesting that if China continues to lead in open source, it could lead in AI overall within a year or two.

Anthropic CEO Dario Amodei holds a different stance. He believes there is risk in releasing increasingly powerful open-source models, as they become difficult to control once released.

Delangue countered this, arguing that locking AI behind the doors of a few players does not make it safer. He added that transparency allows defenders to "patch cybersecurity risks that open-source models have already exposed."