Citadel Securities says signs of widening demand elasticity are emerging in the
AI market. Since end‑June, the usage‑weighted token price has fallen about 40%,
yet H100 GPU rental rates have rebounded from June lows and the top 1% of
AI‑spending firms saw per‑user spend rise roughly 49% MoM in July. Combined
revenue at hyperscale cloud providers reached $106.3bn in Q2, up about 43% YoY,
indicating new compute capacity is finding paying demand. Market metrics for AI
economics are shifting from raw token prices to cost per completed effective
task. Capable open‑source models can cut customer costs while preserving
near‑comparable outputs, expanding usage and supporting GPU and cloud demand but
weakening pricing power of closed‑source frontier models. Citadel cautions the
compute‑price market is still young and opaque; current high utilization does
not guarantee a multi‑year supply‑demand balance, and large efficiency gains
from architectural changes could reduce compute demand below levels implied by
existing capex.
