Compute Capital Markets: Financializing the AI Supply Chain

Introduction

A Compute Capital Market (CCM) lets the producers and consumers of GPU time hedge and speculate on its price, similar to the way energy, metals, and agricultural commodities trade. The stack has four layers:

1. Spot and forward capacity platforms (neoclouds, GPU-as-a-service providers, and decentralized compute networks) deliver the physical GPU-hours.
2. Index providers convert fragmented rental quotes and printed trades into a single reference price per accelerator.
3. Exchanges list standardized, mostly cash-settled futures and options on those indices, meaning a contract transfers no hardware and instead pays the difference between the contract price and the reference index at expiry.
4. OTC dealers warehouse the basis risk that principals will not hold.

Compute demand is two-sided, the precondition for any hedging market. Producers (neoclouds and independent data centers) fear their inventory clears below cost. Consumers (inference platforms and the agentic application layer) fear compute will get more expensive. The common read holds that nonfungibility keeps both off any general exchange, since a buyer wants a named SKU in a named region rather than a basket, so the trade stays bilateral and the only exchange users are dealers hedging their book. That describes launch conditions, but understates how commodity markets form. Canonical benchmarks get made through trading, and reservations standardize as the curve deepens. The dealer-intermediated structure is not the end state, it is the seed of one.