SAN FRANCISCO — Flo Crivello’s startup, Lindy.ai, develops AI-powered “assistants” to manage emails and calendars. Initially, Lindy depended heavily on Anthropic’s advanced AI models. However, Crivello soon realized that Anthropic was the startup’s biggest expense, surpassing payroll costs for more than two dozen employees and rent.
Faced with mounting expenses, Lindy switched entirely to the Chinese AI model DeepSeek-V4 last month. According to Crivello, this decision was straightforward since DeepSeek-V4 was ten times cheaper, saving the company millions. Artificial intelligence has become one of the fastest-growing costs for U.S. businesses, prompting many firms to consider shifting from pricey American models to more affordable Chinese alternatives.
While American companies like Anthropic, OpenAI, and Google are leaders in cutting-edge AI development, experts say Chinese models lag six to 12 months behind in capabilities. Nevertheless, China has excelled in open-source AI models, which are freely available for download and adaptation. “Right now, the open-source scene is completely dominated by the Chinese,” Crivello said, noting that many founders in the AI space are contemplating or have already switched to Chinese models.
The issue of rising AI costs isn’t limited to startups. Uber CEO Dara Khosrowshahi mentioned on the Invest Like the Best podcast that Uber exhausted its annual AI budget within a quarter, compelling adjustments. (NPR couldn’t confirm if Uber uses Chinese models.) Airbnb CEO Brian Chesky also noted last year that Alibaba’s Qwen model provided good, fast, and cheap solutions.
Access to Chinese models, often available on platforms like Hugging Face and GitHub, raises political sensitivities. However, companies like Featherless provide access to over 30,000 AI models, including those from China. Eugene Cheah, Featherless’ CEO, indicated that the popularity of Chinese models stems from their cost-effectiveness, despite not being frontier models.
Cheah likened this choice to selecting between a Ferrari and a Honda: “While a Ferrari may offer luxury, a Honda delivers scale.” Some firms are content using models “N-1,” where N indicates frontier models, as the performance gap diminishes over time.
OpenRouter reported rising use of DeepSeek, increasing from 9% to nearly 20% since January. Models from Chinese companies like MiniMax, Xiaomi, and Tencent have also gained traction. Some users self-host these open-source models, while others use paid hosting services such as Featherless, ensuring U.S.-based data retention.
Victor Su-Ortiz from MiniMax highlighted cost-per-token as a decisive factor for adopting open-source Chinese models. Many repetitive tasks can be executed by these models, yielding substantial savings. For high-volume coding tasks requiring repetition, models like MiniMax M3 excel at one-tenth the cost.
Not all companies find Chinese models adequate. Comment.io co-founder Jon Gordner emphasized the importance of quality. He explained that his company invests in Anthropic and OpenAI models since both are subsidizing user subscriptions, offering discounted tokens—although such offers might vanish as these firms approach public listings.
Lead economist Ara Kharazian from Ramp, a firm that aids businesses in automated spending management, believes American AI firms will respond to competition by adjusting pricing strategies or launching high-quality open-source models.
Gordner remains skeptical. He predicts U.S. AI companies could raise prices as they strive for profitability, potentially driven by the pressures of public listings. “At some point,” Gordner said, “the music’s going to stop.”