Coinbase CEO Brian Armstrong revealed that to reduce costs in AI usage, they are routing queries to cheaper models. Armstrong explained that this approach keeps expenses roughly stable in some cases while maintaining efficiency despite a rapid increase in token usage.
Armstrong emphasized that although the newest models offer powerful features, they consume more tokens. According to him, users turn to these models only when the highest level of cognitive power is required.
The CEO predicted that within the next 12 to 18 months, 80% of workloads will operate on models that are 99% cheaper. Armstrong noted that the most advanced models will be used mainly in specialized areas such as scientific breakthroughs and managing AI agents.
In Armstrong’s assessment, the limiting factor in AI will increasingly be energy and processing power rather than better models. This approach marks a departure from the long-standing tech industry mindset of "use more tokens."
The post attracted attention from some tech figures. Marc Andreessen wrote that he found the approach interesting, while Hugging Face co-founder Julien Chaumond noted that model routing has recently grown rapidly.
Box CEO Aaron Levie said he found Armstrong’s figures somewhat extreme but agreed that AI usage will differentiate into high-end and high-volume tasks in the coming years. Harvey co-founder Winston Weinberg also considered "intelligence allocation" in AI to be becoming very important.
While the tech community previously favored approaches emphasizing higher token usage, recently a cost-focused perspective has gained strength. Around Y Combinator, entrepreneurs were encouraged to be relaxed about tokens, with some founders arguing that cutting tokens is unnecessary.
Glean co-founder Tony Gentilcore also found Armstrong’s view accurate. Gentilcore stated that technical circles already understand this picture, and the real uncertainty lies in financial markets assuming infinite scales for Opus prices.
