June 30, 2026 – A quiet but industry-shaking shift is unfolding across Silicon Valley: U.S. tech firms are increasingly integrating Chinese open-source AI models directly into their production-grade infrastructure, as skyrocketing costs from leading domestic LLM providers push enterprises to seek far more cost-effective alternatives without sacrificing performance.
Leading the charge is major cryptocurrency exchange Coinbase, whose CEO Brian Armstrong took to social platform X over the weekend to announce that the company has rolled out an internal LLM gateway that sets two Chinese open-source models — Zhipu AI’s GLM 5.2 and Moonshot AI’s Kimi K2.7 — as the default large language models for its entire engineering team.
Armstrong did not disclose exact financial figures, but confirmed that even as the company’s token consumption continues to grow exponentially, the combination of switching default models, implementing smart request routing and enhanced caching mechanisms has cut Coinbase’s overall AI spending by nearly 50%. He emphasized that this cost-reduction framework is universally replicable for businesses of all sizes, requiring no compromise on AI usage limits for staff.
Notably, Armstrong revealed that 91% of Coinbase’s engineers had never hit their original token usage caps under the previous setup, so the optimization did not involve slashing employee token allowances. Instead, the company only redirected routine, high-volume tasks including code review, technical documentation summarization and internal knowledge base queries away from premium frontier models from Anthropic and OpenAI, to the two Chinese open-source alternatives.
Coinbase is far from an isolated case in this growing trend. Vacation rental platform Airbnb previously migrated its entire customer service LLM stack from OpenAI’s GPT series to Alibaba’s Qwen model family. Most recently, AI startup Lindy completed a full shift from Anthropic’s Claude to DeepSeek V4, after disclosing that its prior AI service bills had ballooned to exceed the total sum of its employee salaries. Snowflake’s chief executive has also publicly shared performance benchmarks showing that GLM 5.2 delivers comparable capabilities to Claude at a fraction of the operational cost.
This mass migration is clearly reflected on major third-party LLM aggregation platforms. On OpenRouter, a leading marketplace for AI model access, Chinese models have long secured top positions in the text model usage rankings, with DeepSeek, Xiaomi’s MiMo, MiniMax, Tencent Hunyuan and Zhipu GLM all ranking among the most widely adopted options for enterprise users globally.
