Citadel Securities research note: 'Tokenomics' — the economics of AI compute

Citadel Securities has published a Global Macro Strategy research note titled "Tokenomics", authored by strategist Frank Flight, applying the economics of prices and scarcity to artificial-intelligence compute.

The note argues that adoption of frontier AI is becoming less about what the models can do in principle and more about the price and scarcity of the inputs required to run them at scale — compute, power, cooling, memory bandwidth and inference budgets, which it frames as real and binding constraints. It cites recent signals that running agentic and complex workflows is proving expensive, including Amazon removing its token leaderboard, Microsoft cancelling Claude Code subscriptions and multiple reports of unexpectedly large token bills.

Drawing on economic theory — that prices signal scarcity, create incentives for substitution and ration resources toward their highest-value uses — the note predicts a bifurcation between "frontier" and "everyday" AI usage: inference-intensive deployment concentrated among a narrower set of firms with the balance sheets, research depth and operating domain to justify the cost, while simpler, cheaper models become the more cost-effective pathway for the wider economy.

As supporting evidence it presents the Silicon Data LLM Token Expenditure Index, a daily benchmark of effective price-per-million-tokens across the actively traded LLM market. The index has recently declined, which the note reads as users substituting toward cheaper or more efficient models rather than as falling demand — in an elastic market, lower unit costs can unlock additional usage even as its composition shifts.

The note remains constructive on AI as a productivity-enhancing technology but cautions that the route to that value is likely to be more selective and cost-conscious than markets had assumed. It stresses that the most durable productivity gains so far come from embedding AI as a complement to human labour — coding assistants, customer-support copilots and knowledge-work tools — rather than from autonomous agents "running everything everywhere all at once", and warns that asset prices may periodically be forced to reconcile AI ambition with physical constraint.

Read more: Mo Shaikh on X (part 1) · Mo Shaikh on X (part 2)