Armstrong: Energy and Compute, Not Models, Will Cap AI Growth

Coinbase CEO Brian Armstrong predicts 80% of AI workloads will run on models about 99% cheaper within 12–18 months, with 20% remaining on frontier models.
On June 8, 2026, Coinbase CEO Brian Armstrong replied to investor Tommy Shaughnessy, arguing that energy and compute capacity — not further improvements in model quality — will set the upper limit for artificial intelligence growth. He predicted roughly 80% of workloads would move to models priced about 99% below current top-tier options within 12 to 18 months, while the remaining 20% would continue to run on frontier models.
Armstrong wrote that “demand for intelligence is near infinite” and described a commercial split: very low-cost models for routine tasks and high-performance models reserved for scientific research and advanced orchestration agents.
The exchange focused on metered API pricing and enterprise budgets. Shaughnessy highlighted companies hitting AI budget limits early in the year; one enterprise exhausted its full 2026 AI allocation months before year-end. He also noted open-source models such as DeepSeek V4 operate within the performance range of proprietary systems at roughly one-thirtieth the cost.
Coinbase has adjusted operations to reflect those pressures. Armstrong noted the exchange routes prompts to lower-cost models when quality trade-offs are acceptable, which has kept AI spending roughly flat even as token usage rises. Coinbase reorganized parts of its product and engineering teams in early 2026 to support agent-driven workflows and AI-native tooling.
Armstrong argued that as per-inference prices fall, scarcity will shift to physical infrastructure: data center power and specialized chips. He warned the bottleneck will move to the power and silicon required to run models at scale.
He cited industry figures showing strong venture investment in AI — global funding of $242 billion in Q1 2026 — alongside rising data center power demand and constrained capacity. Armstrong wrote that cheaper models and efficiency measures change where resource constraints appear but do not remove the underlying demand for compute and energy.








