Citadel report warns frontier AI scaling hits real economic limits

A new note from Citadel Securities (Frank Flight, June 2026) warns that frontier AI's scaling era is hitting real economic constraints. Physical bottlenecks in compute, power, cooling, and memory are becoming visible on corporate bills. The report cites specific pain points: Amazon canceled its Claude Code subscriptions, and multiple firms report unexpectedly large token charges. The Silicon Data LLM Expenditure Index shows recent declines after earlier spikes, indicating users are already substituting away from the most expensive models toward cheaper, faster alternatives. This shift from frictionless experimentation to cost-curve discipline is now being validated by major financial institutions.

Citadel predicts a market bifurcation: heavy frontier-model usage will concentrate among a smaller set of firms solving genuinely hard problems, while everyday workflows shift to efficient models. Companies like Anthropic are seeing enterprise spend jump from $400K to $1.4M as real costs become visible. The report argues that teams with dashboards, budgets, and ROI tracking will have an edge. Tiered access, heavy monitoring, and aggressive experimentation with open-source or distilled models will become the norm. The subsidized 'try everything on the best model' phase is ending — and the $1M+ bill shocks are not isolated anecdotes but part of a broader transition.