Coinbase cuts AI costs by swapping to Chinese models
Coinbase is quietly swapping its AI infrastructure from Western labs to Chinese models like GLM 5.2 and Kimi 2.7. An automated routing system now selects the best-performing and most cost-effective model for each request, and aggressive caching has pushed successful response rates from five percent to sixty percent. Despite a surge in token usage, the company has halved its AI spending.
Behind the routing curtain
An internal optimizer constantly weighs model cost, latency, and accuracy before routing queries. During periods of high demand, cheaper Chinese models take the load, while premium Western options handle complex reasoning tasks. The result is a balance between performance and budget that still meets customer expectations.
What it means for Western labs
The move underscores pricing pressure on Western AI providers as open-weight and regional alternatives mature. With Coinbase demonstrating material cost savings, other enterprises may accelerate evaluations of non-traditional model families to stay competitive. The shift does not signal an outright abandonment of Western partners, but it highlights a pragmatic shift toward multi-sourcing in a tightening budget environment.
Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.