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

Episode Summary

Executive Summary: The episode explores whether GPU compute can become a tradable commodity like oil. Carmen Lee of Silicon Data/Compute Exchange explains efforts to build GPU price indices and futures with CME, and how spot, forward, and refurbished GPU markets work. The discussion focuses on benchmarking, fungibility, volatility, hedging use cases, data collection, and the emerging market structure around AI infrastructure.

Main Topics: GPU compute as a new commodity market (Priority: 5/5): The hosts and Carmen Lee frame GPU capacity as a potentially tradable commodity, with futures and spot markets designed to let buyers and sellers hedge compute price risk much like energy markets. Silicon Data’s GPU indices and CME futures (Priority: 5/5): Lee explains Silicon Data’s role as an index provider for GPU pricing and the planned launch of financially settled GPU futures/options on CME, pending regulatory approval. Compute Exchange spot marketplace (Priority: 4/5): Compute Exchange is presented as a spot marketplace for GPU procurement, including reserve, forward, and refurbished contracts for neoclouds, startups, enterprises, and inference providers. Fungibility, benchmarking, and basis risk (Priority: 5/5): A major challenge is that GPUs are not perfectly interchangeable across chip type, configuration, and location. The firm uses benchmarking and normalization to create an index that can serve as a hedging instrument with known basis risk. Volatility and market participants (Priority: 4/5): Lee argues GPU prices now show enough volatility to justify hedging and that both long and short participants—providers, banks, enterprises, and traders—can use these instruments. Refurbished GPUs and residual value (Priority: 3/5): The conversation covers the resale and refurbishment market for GPUs, including residual value calculations and how chip depreciation compares to cars and traditional capital equipment. AI bubble and demand uncertainty (Priority: 4/5): The episode closes on the broader question of whether AI infrastructure is in a bubble, with Lee emphasizing that GPU-level cash-flow analysis is more concrete than broad narratives about AI valuations.

Key Arguments: GPU capacity is scarce and economically important enough that a futures market could help buyers and sellers hedge price risk. Because GPUs differ by chip, configuration, provider, and geography, any usable index must normalize raw pricing data rather than simply average prices. There is enough price volatility in GPU rental rates to support a derivatives market; Lee cites daily volatility in the 20% to 30% range for A100 and H100 after normalization. A hedging instrument is only useful if it correlates closely enough to real-world compute costs, so basis risk must be measured and disclosed. Spot, reserve, and forward contracts already show the compute market is maturing toward standard financial-market behavior. Speculators and institutions are necessary for liquidity, even if the original purpose of the market is hedging. Refurbished GPUs have meaningful residual value, so the market should also price used hardware and remaining useful life. The practical bottlenecks to GPU supply include not just chips but servers, colocation, optics, and deployment timing, which affects price and availability. The bubble question is less useful than asking whether future cash flows from compute assets justify current prices. Prediction markets or alternative venues are not viewed as threats; the priority is getting a real market with transparent pricing and hedging utility.

Data Points: GPU index launch timing: Launched on Bloomberg Terminal in 2025 - Silicon Data says it launched the world’s first GPU indices on Bloomberg Terminal. CME product launch: GPU futures/options in a couple of months, pending CFTC approval - Lee says CME-listed GPU financial products are planned soon. Data history used in indexing: Six months of historical trading data from over 100 data sources - Used to identify drivers of price differentiation and normalize GPU prices. Live pricing ingestion: Over 150,000 trader prices ingested every day - Input into the index settlement-price calculation. Global pricing dataset: 8 million pricing points globally - Lee describes the scale of pricing data collected across compute markets. Number of data stores: Around 200 data stores - Sources feeding the pricing/index methodology. GPU performance variance: 38% performance variance for the same chip - A100 testing with Jefferson Lab showed large variation even for identical chip models. A100/H100 daily volatility: Around 20% to 30% - Lee says normalized index volatility falls in a healthy commodity-like range. Some chip volatilities: 8% to over 100% - Different GPU configurations and geographies can produce very different raw volatilities. B200 on-demand price trend: Price rose above initial launch levels after an earlier decline - Lee says B200 supply/demand dynamics tightened faster than expected. H100 price trend: Up about 8% in the last three months - On-demand NeoCloud H100 pricing rose recently. A100 price trend: Up about 10% to 15% over the past three months - Older A100 GPUs also saw renewed price increases. GPU residual value, year 2: About 85 cents on the dollar - Lee cites resale value for H100 after one year in a prior analysis. GPU residual value, year 3: About 84 cents on the dollar - Lee says value remains relatively stable after the first year decline. Inference token price: $2.21 per million tokens - Mentioned as a related benchmark for model/token usage economics. Reservation/contract types: On-demand, reserve, forward, refurbished - Compute Exchange offers multiple procurement and trading structures.

Pivotal Quotes: "What if compute prices keep going to go up?" — Tracy Alloway / Joe Weisenthal (recalling prior episode): Used to frame the original thesis for GPU indices and compute futures. "We actually call it GPU lottery." — Carmen Lee: Explaining why verification and benchmarking matter when buying off-market GPUs. "If we can't do what you said, then we fail at our job." — Carmen Lee: On the necessity that a GPU index/futures contract meaningfully tracks real compute costs for hedging.

Implications: GPU compute is evolving into a financialized infrastructure market. If CME-listed futures gain traction, AI buyers and providers could hedge price risk more efficiently, while price discovery and benchmarking may become central to how compute is bought, sold, and valued.