Confronted with surging AI bills, a growing number of U.S. startups are routing work from premium American systems to cheaper Chinese models, betting that “good enough” performance at a fraction of the price will keep them competitive. Lindy.ai, a San Francisco firm building AI assistants, says it cut costs by roughly 10x after moving fully to DeepSeek-V4, echoing a wider shift seen on model marketplaces such as OpenRouter and Featherless, where usage of Chinese systems is rising. Chinese providers have carved out a strong position in open-source “open-weight” models, letting firms self-host or keep data in the U.S. through third-party inference platforms to address privacy and political concerns.
Some companies still prefer top-tier U.S. models from Anthropic and OpenAI for complex reasoning and quality control, aided by temporary subsidies, but executives warn those discounts may wane as leading labs pursue profitability and potential IPOs. The calculus is increasingly task-based: reserve frontier systems for deep reasoning, while offloading repetitive, high-volume coding and agent tasks to lower-cost models. The competitive question now is whether U.S. leaders can hold pricing or ship compelling open-source offerings as China’s model quality gap narrows—against a backdrop of export controls, rising compute costs, and enterprise scrutiny of AI’s total cost of ownership.
Related articles:
— Open-source artificial intelligence
— Comparison of large language models
— Hugging Face
— Anthropic (company)
— OpenAI
