AI is a rapidly growing business expense. Some companies are cutting costs by switching to cheaper Chinese AI models.
Imen Ben Youssef/Hans Lucas/AFP via Getty Images
hide caption
toggle caption
Imen Ben Youssef/Hans Lucas/AFP via Getty Images
Subscribe to NPR first A morning newsletter with the articles you need to start your day.
SAN FRANCISCO — Flo Crivello’s San Francisco-based startup Lindy.ai creates an artificial intelligence “assistant” to manage email and calendars. Initially, the company relied heavily on Anthropic’s top-of-the-line AI models.
But Crivello said one thing became clear after meeting after meeting with treasurers. “So far, our biggest expense has been human expenses,” he said. “It’s more than just payroll.”
For more than 20 employees, it’s more than a paycheck. More than the rent. Above all. So last month, Crivello announced that Lindy had migrated 100% of its traffic to the Chinese AI model DeepSeek-V4.
“It was exactly 10 times cheaper,” he said, adding that the company saved millions of dollars. “So it was a very simple business decision.”
Artificial intelligence became one of them, but it wasn’t. of — Costs are increasing fastest for U.S. businesses. But for many companies, this is a necessary but expensive double-edged sword. To survive, more and more companies are switching from American-made models to cheaper Chinese-made AI.
In the race to create the best AI models, American companies like Anthropic, OpenAI, and Google are leading the way. Experts say the Chinese model is six to 12 months behind in terms of functionality.
But China has carved out a niche for open source models that can be freely downloaded and adapted. “The current open source scene is completely dominated by the Chinese, not even close,” Crivello said.
He said all the founders he knows in the AI space are considering switching to the Chinese model or have already done so.
And rising AI costs aren’t just an issue for startups. Uber CEO Dara Khosrowshahi said this last month: Invest like the best Podcast. “We’ve used up our AI budget in a quarter, essentially all year, and that’s forcing us to adjust,” he said.
(Uber did not respond to NPR’s request for information about whether it uses the Chinese model.)
Bloomberg reported that Airbnb CEO Brian Chesky said last year that the company relies on Alibaba’s Qwen model, which is “better,” “faster and cheaper.” Perplexity and Nvidia also use Qwen.
like ferrari or honda
Many companies are wary of promoting the use of Chinese models due to political considerations, but the models are widely available on AI Model Hubs Like Hugging Face, it goes through model aggregators and inference providers based outside of China on the code hosting platform GitHub.
They include Featherless, a San Francisco-based company that provides access to about 30,000 AI models. Founder and CEO Eugene Chia said Chinese models are popular, even if they are not “frontier” or top-of-the-line models.
“It’s like the difference between driving a Ferrari and a Honda. You can have the best luxury car, or you can have a big functioning Honda,” he said.
“In fact, many open source AI groups are perfectly fine with N-1, where N is the frontier,” he continued. “Because as the gap continues to narrow, at some point you start to question whether it actually matters.”
For many people like Lindy, that doesn’t matter. AI Honda is perfect.
OpenRouter, another platform where startups can access various AI models, reported that DeepSeek usage in China has increased from about 9% to nearly 20% since January. The use of models from Chinese companies MiniMax, Xiaomi and Tencent is also increasing.
While some users download and self-host open source China AI models, many use the models via paid AI hosting companies like Featherless or OpenRouter, so their user data is stored in the United States.
Victor Su-Ortiz, head of global product marketing at Shanghai-based MiniMax, attended a recent conference for AI engineers in San Francisco. Companies pay for the use of AI models in tokens, or units of AI work. Suortis said it all comes down to cost per token.
He said that when compared to leading AI models, “many repetitive tasks can be performed with models that have comparable performance but at a much lower cost per token.” “And this is essentially why we brought these open-grade models to the United States.”
He said companies are moving away from “token maxing,” using AI as much as possible, to reducing costs by limiting usage, switching to cheaper models, and routing different types of AI work to different types of models.
For example, it could improve the performance of state-of-the-art models for research and “deep inference,” Suortis said. “But if you’re routing repetitive, high-volume coding tasks, one of our models, especially the MiniMax M3, offers much better performance at just one-tenth the cost.”
Saving a few dollars isn’t worth it for everyone
For some companies, the Chinese model is still not enough. Jon Gordner is CEO and co-founder of Comment.io. Comment.io was founded a few weeks ago and is developing a Google Docs-like product for programmers and AI agents.
“We need to make the best software possible as quickly as possible, and for us it’s not worth it to save a few bucks on a cheaper model if we have to spend an extra couple of weeks fixing mistakes,” he said.
Gordner said his company gets value from the Anthropic and OpenAI models, in part because both companies subsidize users to acquire customers. He said he is currently offering tokens at a deep discount on a monthly subscription, but that likely won’t last forever.
“It would make much more sense for us to start evaluating the Chinese model and the open source model,” he said.
Ara Kharazian is principal economist at Ramp, a company that helps businesses track, manage, and automate spending. The company has insights into AI spending, and Karazian said U.S. companies will continue to adapt, potentially holding down prices or introducing high-quality open source models to beat Chinese rivals.
“The rise of these Chinese models speaks to the fact that companies are looking for something today that the American model companies are not offering,” he said. “The only reason I’m bearish on the Chinese model is because I expect the US model companies to respond competitively.”
Comment.io’s Gordner isn’t so sure. He thinks big U.S. AI companies may need to start charging more That’s likely the case for AI, as the pressure to prove profitability increases as it nears going public. Both Anthropic and OpenAI have submitted confidential documents to the U.S. government in preparation for their eventual initial public offering.
“At some point, the music will stop,” Gordner said.
Anthropic is a financial supporter of NPR.
