Startups cut AI bills with lower-cost Chinese models
NPR reports that some U.S. firms are moving AI work to Chinese open-source models as token costs become a top expense.
By James Whitfield · Staff Writer
4 min read
Some U.S. companies are shifting AI workloads to Chinese models to lower fast-rising computing bills. The move shows how price is becoming a major factor in the contest between American AI leaders and lower-cost rivals from China.
NPR reported that Lindy.ai, a San Francisco startup that builds AI assistants for email and calendars, moved all of its traffic last month to DeepSeek-V4, a Chinese model. Founder Flo Crivello told NPR the change cut model costs by a factor of 10 and saved the company millions of dollars.
Crivello said Lindy had previously relied heavily on Anthropic’s top models, and that Anthropic had become the company’s largest expense, ahead of payroll for more than two dozen employees, rent and other costs. He described the switch as a straightforward business choice because of the savings.
U.S. model makers including Anthropic, OpenAI and Google remain ahead in the race for the most capable systems, according to experts cited by NPR, who put Chinese models roughly six to 12 months behind on performance. China, however, has gained strength in open-source models that users can download, modify and run through hosting services.
Crivello told NPR that Chinese groups dominate the open-source AI field and said founders he knows in AI are either considering Chinese models or have already adopted them. Many companies are cautious about publicizing such use because of political sensitivities, NPR reported.
The cost pressure is not limited to startups. Uber CEO Dara Khosrowshahi said on the Invest Like the Best podcast that Uber used up what was effectively its annual AI budget in one quarter and had to adjust, according to NPR. NPR said Uber did not respond to questions about whether it uses Chinese models.
Bloomberg has reported that Airbnb CEO Brian Chesky said the company used Alibaba’s Qwen model last year and described it as good, fast and inexpensive. NPR also reported that Perplexity and Nvidia have used Qwen.
Chinese models are available through AI model hubs such as Hugging Face, code platforms such as GitHub and model-routing or hosting companies outside China. Featherless, a San Francisco company that provides access to about 30,000 AI models, said Chinese options are popular even though they are not usually the top-performing systems.
Eugene Cheah, Featherless’ founder and CEO, told NPR many users do not need the most advanced model for every task. For high-volume uses, a less capable but dependable and cheaper model can make more business sense.
OpenRouter, another platform that lets companies access different AI models, said DeepSeek’s share of use on its service has risen from about 9% to nearly 20% since January, NPR reported. OpenRouter also reported increases for models from MiniMax, Xiaomi and Tencent.
Some companies download Chinese open-source models and host them themselves. Others use paid hosting and routing services such as Featherless and OpenRouter, which NPR reported can keep user data in the United States.
Victor Su-Ortiz, who works on global product marketing at Shanghai-based MiniMax, told NPR the key issue for many customers is cost per token, the unit used to price AI work. He said repetitive tasks can often run on lower-cost open-weight models, while more advanced systems may still be better for research or complex reasoning.
Su-Ortiz said companies are moving away from using as much AI as possible and toward cutting usage, choosing cheaper systems or sending different tasks to different models. He said MiniMax M3 can handle some repetitive, high-volume coding work at about one-tenth the cost of leading models.
Not every company is ready to switch. Jon Gordner, CEO and co-founder of Comment.io, told NPR that his startup needs the strongest software it can get quickly, and that a cheaper model would not be worth it if it caused weeks of extra repairs.
Gordner said his company currently gets value from Anthropic and OpenAI models because subscriptions provide discounted access, but he expects that pricing may change. He said Comment.io would be more likely to assess Chinese and open-source models if those discounts fade.
Ara Kharazian, lead economist at business spending firm Ramp, told NPR that Chinese models are gaining attention because businesses want something American providers are not yet offering at the same price. He said U.S. model companies may respond by controlling prices or releasing stronger open-source options.
NPR reported that Anthropic and OpenAI have filed confidential paperwork with the U.S. government tied to possible future initial public offerings. Gordner told NPR that pressure to show profits could push leading U.S. AI companies to charge more over time. NPR disclosed that Anthropic is a financial supporter of NPR.
This story draws on original reporting from NPR.