U.S. Startups Shift to Chinese AI Models to Cut Costs

Several U.S. startups have moved part or all of their AI workloads to Chinese models after finding that American systems became their largest operating expense. Lindy.ai, based in San Francisco, previously relied on Anthropic models. Last month the company switched 100 percent of its traffic to DeepSeek-V4, which founder Flo Crivello said cost one-tenth as much and saved millions of dollars.

Airbnb used Alibaba's Qwen model last year, according to Bloomberg reporting. OpenRouter data show DeepSeek usage rising from 9 percent to nearly 20 percent since January. Models from MiniMax, Xiaomi, and Tencent also recorded increased traffic on the platform.

Cheah said many users treat Chinese models as reliable but lower-tier options, comparing them to a Honda versus a Ferrari. Victor Su-Ortiz of MiniMax stated that repetitive tasks can run at one-tenth the cost of frontier models. io co-founder Jon Gordner said his firm continues to use Anthropic and OpenAI models because fixing errors from cheaper systems would offset any savings.

Ramp economist Ara Kharazian said U.S. companies may respond by lowering prices or releasing stronger open-source models.