Svmuu News: Brain Armstrong, CEO of Coinbase, posted on X stating that within Coinbase, the team is exploring how to keep AI spending stable while token usage grows exponentially—not through usage limits or frequent cost alerts, but by relying on better system design.
He noted that Coinbase optimizes AI usage costs primarily through five approaches: First is “better default models,” which allows engineers to freely choose models but defaults the system to lower-cost open-source pre-trained models (such as GLM 5.2 and Kimi 2.7), while still encouraging the selection of appropriate models for different tasks;Second is “smart routing,” where an automated system allocates the optimal model based on task type, cache hit rate, and model pricing, rather than relying on manual selection.
Third is “Cache Optimization,” which significantly reduces the cost of redundant computations by improving cache hit rates; fourth is “Context Streamlining,” which minimizes the consumption of invalid tokens when switching tasks and limits unnecessary tool calls; fifth is “Visual Transparency,” which allows engineers to clearly see their usage patterns without imposing restrictions to suppress usage.
Brian Armstrong stated that the goal is not to curb AI usage, but to build infrastructure capable of supporting exponential growth. Results from implementation show that, despite continued growth in token usage, overall AI spending has nearly been cut in half.