Coinbase CEO outlines five tactics to halve AI spend while token use

Coinbase announced that its AI spending has fallen to roughly half of its peak level, even as token consumption climbs to one of the highest points in the company's history, thanks to five cost‑saving tactics unveiled by CEO Brian Armstrong on X【1】. The move aims to keep the firm’s AI infrastructure scalable without throttling engineers’ access to large language models (LLMs).

Cost‑cutting tactics in detail

Armstrong’s first lever is swapping default LLMs for cheaper open‑weight Chinese models—GLM 5.2 from Z.ai and Kimi 2.7 from Moonshot AI—rather than defaulting to premium offerings from Anthropic or OpenAI【1】. The second step routes each prompt to the most appropriate model based on task difficulty, letting “frontier” models handle planning while cheaper models handle execution【1】. A third measure improves inference cost by using more aggressive caching, and a fourth keeps context lean by starting fresh sessions when switching tasks【1】. Finally, the company makes every engineer’s token consumption visible, tying higher spend to higher impact expectations rather than imposing hard caps【1】.

Impact on spend and usage trends

Armstrong attached a graph showing token usage climbing to historic levels while AI spend dropped sharply, though the exact timeline isn’t disclosed【1】. The Decoder reports that the same routing and caching upgrades lifted the hit‑rate from 5 % to 60 % and cut Coinbase’s AI bill in half as token usage kept rising【4】. In a separate Business Insider post, Armstrong said the firm has kept costs “roughly flat” despite exponential token growth, and he forecasts that within 12‑18 months, 80 % of workloads will run on models that are 99 % cheaper than today’s frontier options【2】.

Industry context

Coinbase’s strategy mirrors a broader shift away from the “tokenmaxxing” craze, where firms previously encouraged unrestricted token consumption to showcase raw AI power. Instead, companies now impose usage caps or visibility rules to curb runaway costs. Armstrong’s approach aligns with moves by other tech firms—Lindy’s adoption of Deepseek v4 and Snowflake’s testing of Chinese models—adding pricing pressure on Western AI labs as they prepare for potential IPOs【4】.

What to watch

• AI spend vs. token usage: Monitor quarterly updates for any divergence between rising token counts and AI expenditure.
• Model adoption: Track the proportion of prompts routed to GLM 5.2 and Kimi 2.7 versus frontier models, especially as the 12‑18 month cost‑efficiency target unfolds.
• Caching efficiency: Watch for reported hit‑rate improvements (e.g., moving from 5 % toward the 60 % benchmark) in Coinbase’s internal AI performance reports.

By halving AI costs while allowing token usage to expand, Coinbase demonstrates a scalable model for crypto‑focused firms that need AI‑driven productivity without unsustainable spend. The open question remains whether the cost‑saving measures will sustain as AI workloads become more complex and demand higher‑end models.