Coinbase CEO Brian Armstrong published five strategies on Friday to manage AI spending as usage surges among the company's engineers. Writing on X, he emphasized that the goal is not to limit token usage but to make scaling more sustainable. The post arrives as companies grapple with ballooning costs from frontier AI models.
The most notable tactic involves switching default large language models (LLMs) to cheaper Chinese alternatives. Armstrong said Coinbase is experimenting with open-weight models like GLM 5.2 and Kimi 2.7, developed by Z.ai and Moonshot AI respectively, through its internal LLM gateway. This shift could reduce expenses compared to using Anthropic or OpenAI models by default.
Armstrong also expects tangible results from high-spending employees, tying budget flexibility to demonstrated outcomes. Engineers retain freedom to choose any model for specific tasks, but defaults shape routine usage patterns. The approach balances cost discipline with technical autonomy.
These measures reflect a broader industry challenge as AI adoption accelerates. Other tech firms have similarly sought to curb expenses without stifling innovation, though Coinbase's public playbook is unusually detailed. Armstrong's post suggests internal AI costs had become a management priority.
Critics might argue that relying on Chinese LLMs could introduce data security risks or compliance issues, especially for a regulated financial exchange. Armstrong did not address these concerns in his post.