Coinbase's CEO outlined 5 strategies to keep AI spend low without limiting tokens

In a recent post on X, Coinbase's CEO Brian Armstrong shared insights on how the cryptocurrency exchange is managing to keep its AI expenditures in check while maximizing token efficiency. Armstrong pinpointed five strategic approaches aimed at balancing innovation with cost-effectiveness. The first strategy involves selecting more efficient default large language models (LLMs) for engineers to use. Armstrong noted that Coinbase is testing Chinese LLMs, which offer a more affordable alternative to premium models from leading American AI firms like Anthropic and OpenAI. He mentioned the company's exploration of open-weight models such as GLM 5.2 and Kimi 2.7, created by Chinese AI developers Z.ai and Moonshot AI, respectively. Armstrong's second strategy focuses on directing prompts to the most suitable models based on the complexity of tasks. He illustrated that while a top-tier model may be necessary for planning, it may be excessive for execution. He further emphasized that this process could be automated, relieving engineers from the burden of manual model selection. The third tactic outlined is the implementation of improved caching techniques to cut down on inference costs. Following that, Armstrong advised maintaining a lean context by initiating new sessions when shifting between different tasks. Finally, he stressed the importance of transparency in AI spending, allowing engineers to track their token usage. This transparency is expected to encourage more impactful contributions from those who invest more in AI. Accompanying his post was a graph illustrating the trends in token usage and AI spending at Coinbase. It revealed that while token consumption has surged to record levels, overall AI expenditures have plummeted to nearly half of their previous peak. Armstrong clarified that the intention isn't to limit usage but to create a robust infrastructure that supports sustainable growth. This initiative comes on the heels of Coinbase's decision to downsize its workforce by 14%, a move partly influenced by the evolving role of AI in the workplace. Armstrong remarked on the remarkable efficiency gains achieved by small, dedicated teams leveraging AI, stating that tasks which once required weeks are now being completed in mere days. His approach reflects a broader industry trend, moving away from previous practices of rampant token usage towards more responsible and monitored consumption.