Carmen Li's Plan to Build a Futures Market for Compute

What It Covers

Carmen Lee, CEO of ComputeXchange and Silicon Data, explains how her two companies are building GPU price indices and futures markets, with a CME partnership pending FTC approval. The episode covers market structure, buyer types, price volatility data, chip depreciation, and the mechanics of financially settling compute contracts.

Key Questions Answered

• •GPU Price Volatility: H100 daily price volatility runs 20–30%, placing it within a healthy commodity hedging range comparable to oil markets. Individual chip configurations at specific geolocations show variance from 8% to over 100%, but index normalization produces the stable 20–30% figure that makes futures contracts viable for risk management purposes.
• •GPU Performance Variance: The same H100 chip model shows up to 38% performance variance depending on hardware configuration, data center location, CPU pairing, RAM, and memory bandwidth. Buyers using ComputeXchange receive independent benchmark verification across FLOPS, memory bandwidth, and token throughput before delivery, functioning as a Carfax-style transparency layer.
• •Chip Depreciation Reality: H100 resale value holds at approximately 85 cents on the dollar after one year, stabilizing thereafter. Older A100 chips still command active market prices and rose roughly 15% over the three months prior to recording, signaling demand-supply tightening even for legacy hardware no longer considered cutting-edge.
• •Hedging Framework — Long vs. Short GPU: Entities owning GPU infrastructure are naturally long and hedge revenue volatility by shorting futures. Entities consuming compute are naturally short and use futures to control cost volatility. This mirrors the oil market structure where producers short futures and industrial consumers go long to lock in predictable margins.
• •Index Construction Methodology: Silicon Data ingests over 150,000 trader prices daily from 200-plus data sources, normalizing each price point against chip configuration variables before calculating a settlement price. This regression-based approach, built on six months of historical trading data per chip class, enables basis risk calculation by geography, allowing users to quantify their local price deviation from the index.