Coinbase CEO Brian Armstrong stated on 上週五 (26日) that the cryptocurrency exchange has set Chinese open-source AI models GLM 5.2 and Kimi 2.7 as the default large language models for internal engineers. Armstrong reported that Coinbase cut AI spending by nearly half through routing optimization and caching improvements, while token usage continues exponential growth. This deployment reflects a broader trend among U.S. technology companies quietly integrating Chinese open-source AI models into production infrastructure to reduce costs and scale applications.
Coinbase Implements Three-Layer Infrastructure to Cut AI Costs
Armstrong attributed the cost reduction to a three-layer infrastructure overhaul. The first layer is "smart routing," where the system preprocesses prompts and automatically assigns tasks to the most suitable and economical model based on cache hit rates and model pricing. The second layer is "aggressive caching," which increased LibreChat's cache hit rate from 5% to 60% by requiring all requests to be cache-aware. The third layer is "streamlined context," which recommends opening new sessions when switching tasks and narrowing file scope to reduce wasted tokens.
Armstrong emphasized that the approach is not about suppressing usage but about scaling AI adoption. He described the method as key to achieving sustainable expansion of AI usage, stating that any enterprise can adopt this model to allow engineers free use of any quantity of tokens and models without setting a cost ceiling, while linking usage to business impact.
Chinese Open-Source Models Handle Routine Tasks in Multi-Model Strategy
The two Chinese open-source models are primarily deployed for routine task scenarios. For tasks requiring complex planning, engineers can still select frontier models. In the code review process, Coinbase employs a multi-model parallel strategy, allowing different models to cross-verify output results to maintain quality standards.
Armstrong noted that as costs for top U.S. model services continue to climb, the cost-effectiveness advantages of Chinese open-source models are gradually changing global technology companies' AI deployment strategies.
FAQ
What did Coinbase announce on 上週五 (26日) regarding AI models?
Coinbase CEO Brian Armstrong announced that the company set Chinese open-source AI models GLM 5.2 and Kimi 2.7 as the default large language models for internal engineers. Armstrong stated that this change, combined with routing optimization and caching improvements, reduced AI spending by nearly half while token usage maintains exponential growth.
How does Coinbase use Chinese AI models in its operations?
Coinbase deploys GLM 5.2 and Kimi 2.7 primarily for routine task scenarios, while engineers can still select frontier models for tasks requiring complex planning. In code review, the company uses a multi-model parallel strategy where different models cross-verify output results to maintain quality standards.
