Coinbase CEO Brian Armstrong announced that the company reduced its AI spending by nearly 50% after testing open-weight models as alternatives to more expensive closed-model providers. The disclosure signals a growing trend of cost discipline among major crypto firms investing in artificial intelligence infrastructure.
Armstrong shared the savings figure on X, framing it as a result of experimenting with open-weight models across internal AI workloads. The claim, which drew coverage from Yahoo Finance and Business Insider, lacked a breakdown of which specific workloads or models were involved, and no independent verification of the savings has been published beyond Armstrong’s own statements.
Open-weight models allow companies to run AI on their own infrastructure, avoiding per-token fees charged by closed-model providers. For Coinbase, which relies on AI for compliance screening, customer support, and code generation, the cost difference can be substantial. However, the CEO’s suggestion to explore open-weight systems—including Chinese-origin models like Zhipu’s GLM 5.2—has raised security and geopolitical concerns. GLM 5.2, which uses an MIT license and can be self-hosted without sending data to external APIs, offers competitive pricing ($1.40 per million input tokens vs. $5 for Anthropic’s Opus 4.8) and strong coding benchmarks. Yet the model falls under China’s National Intelligence Law, prompting U.S. lawmakers to investigate cybersecurity risks and raising alarms after Anthropic disclosed a massive campaign to steal its capabilities via Chinese AI platforms.
Armstrong emphasized that the key to controlling AI costs lies in “better defaults, routing, and caching.” The savings are purely operational and do not affect Coinbase’s product lineup, trading fees, or revenue. Details such as the baseline AI spending, the exact models replaced, and the measurement timeframe remain undisclosed. The story reflects broader industry pressure, with Uber exhausting its 2026 AI budget by April and Meta implementing spending controls.
