More US startups and companies use cheaper open AI models from China

In short: More Silicon Valley startups and some large US companies are picking lower-cost AI models made in China, especially models that can be downloaded and run on their own systems.

What's going on

Developers in the US are increasingly using Chinese-made AI models such as Alibaba’s Qwen, DeepSeek, Moonshot’s Kimi, and Zhipu’s GLM. These models are often described as “open weight,” which means the core files that make the model work can be downloaded (like getting the full recipe, not just ordering the dish). That lets teams run the model on their own computers or cloud accounts, and adapt it for their specific needs.

Some of this use is now public. Airbnb’s CEO said the company prefers Alibaba’s Qwen over ChatGPT for a customer service chatbot because it is “fast and cheap,” according to reporting. Other users, including investors and developers, have also described switching from US-based AI services to Chinese models for cost and performance reasons.

The scale is harder to pin down. Qwen has been reported as one of the most downloaded model families on Hugging Face, a popular site for sharing AI models, with hundreds of millions of downloads and many spin-off versions. But viral claims like “80% of startups are using Chinese models” are not backed by clear public data.

What to watch

This shift is likely to increase pressure on US AI companies to lower prices. It also raises questions for US policy makers about data security and national security, even when companies run these models outside China. Rules could change quickly, so companies may limit these models to low-risk tasks until there is more clarity.

Source: NYTimes