How Coinbase Slashed AI Costs in Half — And Why Armstrong Isn't Worried About Bitcoin

Coinbase CEO Brian Armstrong has revealed a strategy that allowed the company to cut its artificial intelligence expenditures by nearly 50%, even as token usage surged at an exponential pace. Rather than restricting employee access or issuing budget alerts, Armstrong credits the savings to smarter infrastructure decisions — a model he believes any organization can replicate.

Speaking publicly about the approach, Armstrong highlighted three core techniques driving the cost reduction. The first is intelligent model routing, which automatically directs tasks to the least expensive model capable of handling them adequately. This alone eliminates wasteful spending on high-end models for straightforward queries.

The second technique is aggressive output caching. By storing results from frequently repeated queries, Coinbase avoids regenerating identical responses and cuts redundant compute costs significantly. The third pillar involves transitioning routine workloads to open-weight models — cheaper, publicly available alternatives that perform just as well as frontier models for standard tasks where cutting-edge capabilities provide no measurable advantage.

Armstrong was clear that the goal was never to restrict AI usage across the company. Instead, the aim was to build an infrastructure layer capable of supporting sustainable, long-term scale. By decoupling consumption volume from cost growth, efficiency gains create room for usage to compound organically without triggering budget ceilings down the line. Notably, Armstrong did not disclose the actual dollar figures involved, but the implication is significant: a company that halves spending while usage grows exponentially has fundamentally changed its cost-to-output relationship.

This announcement builds on Armstrong's earlier commentary from early June, in which he argued that energy access and compute availability represent the biggest bottlenecks for AI adoption — not model quality. The new routing and caching data adds another dimension to that framework.

Shifting gears to cryptocurrency markets, Armstrong offered a notably calm take on Bitcoin's ongoing price correction. Rather than characterizing the current drawdown as a bear market, he described it as "barely even a winter — more like a cool breeze."

The numbers support a measured interpretation. According to River's historical data, the 2025–2026 cycle has seen Bitcoin decline approximately 53% from its October 2025 peak of $126,073. While that is a steep drop in absolute terms, it is the shallowest recorded bear market in Bitcoin's history. Previous cycles saw drawdowns ranging from 77% to 93%, with several lasting over 12 months.

In mid-June, Armstrong publicly predicted a $60,000 price floor for Bitcoin. However, on-chain metrics have yet to produce the capitulation signals that analysts typically associate with confirmed cycle bottoms, leaving a gap between price action and historical indicators that has persisted throughout this cycle.

Armstrong remains a vocal supporter of Bitcoin's four-year cycle thesis and has projected prices well above current levels by 2030. That said, the 500-day post-halving signal — a key benchmark followed by most cycle analysts — is not expected to trigger until November 2026, suggesting the timeline for a full recovery may extend further than Armstrong's optimistic framing implies.