Western Companies Rush to Chinese AI Infrastructure as Inference Costs Drop Up to 95%

A cost-driven migration wave from Western AI infrastructure to Chinese large language models is accelerating, according to an investigation by Chinese media outlet National Business Daily (NBD). Cryptocurrency exchange Coinbase, AI startup Lindy, Canadian law firm Substance Law, and UK edtech platform ExpertEdge have all shifted core workloads from US models to Chinese alternatives.

Coinbase set Zhipu GLM-5.2 and Moonshot Kimi K2.7 as default models for its engineers through an internal LLM gateway, cutting AI spending by nearly half. US startup Lindy switched entirely from Anthropic Claude to DeepSeek-V4 because its "API costs exceeded total employee salaries," reducing inference costs by approximately 95% and saving millions of dollars annually. Lindy's CEO reported stable performance metrics with only a minor increase in end-to-end latency within acceptable range.

OpenRouter data reveals that Chinese AI models now occupy the top six positions on its leaderboard, with DeepSeek-V4-Flash holding the number one spot for seven consecutive weeks and GLM-5.2 breaking into the top five. US companies' token consumption of Chinese AI models on OpenRouter has risen from 4.5% in the first half of 2025 to a peak of 46%. DeepSeek's traffic share surged from 1% to 17% in just a few months.

The driving force is a compelling cost-performance ratio. Chinese model prices are 60% to 90% lower than leading US models from Anthropic and OpenAI, while the performance gap has narrowed to just 1% to 4%. Zhipu's GLM-5.2 has narrowed the gap with Claude Opus 4.8 to 1% to 4% on coding and long-context benchmarks, making it a viable alternative for approximately 90% of everyday enterprise applications.

Beyond cost, supply chain independence is a growing concern. In June, a US export ban temporarily took Anthropic's Claude Fable 5 offline globally, highlighting the risks of relying on externally controlled models. Open-source models allow enterprises to run inference on their own hardware, reducing exposure to geopolitical or commercial decisions.

Industry experts believe this migration wave is still in its early stages. Former Hugging Face Asia-Pacific ecosystem lead Wang Tiezhen noted that when users can access near-frontier model performance at significantly lower costs with better data privacy guarantees, the shift from closed-source to open-source models will continue to gain momentum. Shanghai University of Finance and Economics professor Hu Yanping observed that the collective rise of Chinese AI model providers has created a diverse ecosystem serving a wide range of enterprise scenarios.

Brookings Institution scholar Kyle Chan noted that Chinese model prices are typically a fraction of US competitors while performance approaches frontier levels, with the overall technology gap estimated at 6 to 9 months. For 90% of routine enterprise applications, this gap is fully offset by cost advantages. The open-source model is evolving from a "technical choice" into an "industrial strategy," with expanding ecosystem capabilities becoming a long-term competitive advantage in the global AI race.