Coinbase turns to Chinese AI models to reduce operating costs
U.S. crypto exchange Coinbase has reportedly started using the Chinese AI models GLM 5.2 and Kimi 2.7 in an effort to lower AI-related spending. According to the report, the company combined model replacement with operational optimization and ultimately cut its AI bill by half. For infrastructure-heavy companies, this is a notable signal that model selection is increasingly being driven by cost-performance tradeoffs rather than brand preference alone.
Cost optimization came from routing, caching, and context engineering
The reported savings were not achieved simply by switching vendors. Coinbase also used automatic routing to match different tasks with different models, improved cache hit rates to reduce repeated compute costs, and refined context engineering to lower unnecessary token consumption. In practice, that means the cost reduction came from both cheaper model access and more disciplined orchestration at the application layer.
Other U.S. companies are making similar choices
The same report says that companies including Lindy and Snowflake are also moving toward lower-cost model options to manage AI expenses. This suggests a broader procurement shift across the industry: enterprises are becoming more willing to diversify away from a small set of premium model providers if alternative systems can deliver acceptable results at significantly lower cost. As that behavior spreads, pricing discipline becomes a central competitive factor.
AI pricing models are being reshaped
The broader implication is that the AI industry’s pricing structure is being redefined. As enterprises adopt multi-model routing, optimize cache reuse, and focus on token efficiency, premium pricing becomes harder to defend unless it is matched by clearly superior output or workflow value. In that environment, major players such as OpenAI and Anthropic are increasingly exposed to direct price competition. For the market, the Coinbase example highlights how quickly enterprise AI spending strategies can change once cheaper and operationally viable alternatives emerge.
