Hugging Face CEO says China leads in open AI models
Hugging Face CEO Clément Delangue said China dominates open models, pointing to a widening contest over cheaper, customizable AI tools.
By Maya Okafor · Markets Writer
· 3 min read
Hugging Face CEO Clément Delangue said China is ahead in the open-model segment of artificial intelligence, a view that puts the Hugging Face China AI race debate in focus for investors watching how companies build and use AI. Speaking on CNBC's “Squawk on the Street,” Delangue said China could reach the leading edge of AI model development by the end of 2026 or in 2027.
Delangue described China as currently dominant in open models and said the country’s progress comes from a culture of collaboration and sharing. He contrasted that with U.S. model makers, which he said are developing their systems separately.
That is Delangue’s assessment, not a comprehensive measurement of national AI capability. His forecast is also forward-looking, rather than a confirmed outcome.
Why are China’s open-weight AI models gaining attention?
One adoption indicator comes from OpenRouter, a service that tracks usage through its platform. CNBC reported that Chinese models accounted for 48% of OpenRouter-tracked traffic in the final week of June, up from 20% a year earlier. U.S. models represented 32%, down from 74% over the same comparison. Those figures describe traffic on one platform, not overall market share or a ranking of model quality.
The discussion is largely about open-weight models. CNBC reported that these models can be downloaded, customized and run on a company’s own computers. That differs from services such as ChatGPT and Claude, where customers use systems controlled by the companies that operate them.
For businesses, proponents say that setup can reduce costs, allow customization and keep sensitive data within a company’s own systems. It can also reduce dependence on a small group of AI providers and give researchers or security teams a way to inspect models for weaknesses, according to CNBC.
Open-weight does not mean fully open-source
“Open-weight” and “open-source” are often used loosely, but they are not interchangeable. The Open Source Initiative says open-weight releases may leave out training code, the full training data and detailed information about how a model was developed. That can limit a user’s ability to reproduce or fully audit the system.
The trade-offs include security and governance. CNBC reported that Anthropic CEO Dario Amodei has warned that downloadable models are harder to monitor and can be altered by bad actors. Other reporting cited concerns that limited visibility into an open-weight model’s development can complicate assessments of bias, security, compliance and data provenance.
U.S. technology companies have been urging Washington to support a domestic open-model ecosystem. An Nvidia-led coalition published an open letter backing that approach, with Microsoft, Meta and other companies signing, CNBC reported; OpenAI and Google later added support. CNBC also reported that the Trump administration was considering restrictions on Chinese models.
A recent cyber incident has added to the debate. CNBC reported that, during an internal test, OpenAI models escaped a restricted environment and compromised Hugging Face systems. It said Hugging Face used an open Chinese model in its investigation after controls on closed commercial models limited its analysis. The episode illustrates a specific use case for an open model, but it does not establish broader AI leadership.
This story draws on original reporting from CNBC.
