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

PodcastOdd LotsCarmen LiJun 15

The podcast outlines how heterogeneous GPU SKUs and rapid software improvements create shifting demand and economic risk across hardware owners. Market participants and Silicon Data are moving to standardize compute into tradable contracts, with a CME partnership proposing cash/API‑settled futures and options to improve hedging and liquidity. Structural hurdles persist: lack of fungibility, capex lead times, volatile spot rents (H100 +8%, A100 +15%), competing indices with 20–30% daily volatilities, and the need for specialized trading desks and credit protections. Secondary markets and a Bloomberg‑tracked LM token that doubled since December 1 add alternate price signals for participants.

Trading Strategies AI Infrastructure Risk Management Market Structure Market Volatility Crypto Markets

Preview · 16 more points in the app

Read the Full Summary

Key Takeaways

Silicon Data has built GPU price indices and announced a partnership with CME to list cash/API‑settled GPU futures and options using existing CME workflows to reduce operational friction, but practical fungibility issues (location, data‑center infrastructure, configuration) and measurable basis risk mean index construction and settlement methodology will determine whether these contracts work as true hedging tools.

On‑demand spot GPU prices have become both higher and more volatile—H100 rental rates rose roughly 8% and A100 about 15% over the past three months—driving many buyers into reserved, forward or refurbished contracts on spot marketplaces while capex and deployment lags prevent new hardware from quickly relieving tight supply.

Different GPU SKUs (e.g., A100 vs H100) are not homogeneous and software advances (model compression/optimization) can shift which chips are usable within months, producing distinct index behaviors (A100/H100 composite daily volatilities around 20–30%) and increasing the need for specialized compute trading desks and robust index governance to avoid speculative rather than hedging outcomes.

TECHNOLOGY

GPU SKU Risk — GPUs are heterogeneous: SKUs such as A100 and H100 serve different use cases and are not perfectly correlated. Software advances—model compression and optimizations—can make a chip usable for a model within months, shifting demand across SKUs and altering hardware economics.

“There's a lot of context where number 1 GPU, it is not homogeneous product. Number 2, you have to understand the use case of the A100, H100…”—

MARKETS

Compute Commodity — Market participants are exploring standardizing scarce GPU capacity into tradable unit contracts—for example, an H100 priced per hour. That would enable futures and options, letting speculators and hedgers express views on per‑hour compute costs.

“It's essentially like, okay, GPUs were we know they're very important for the AI boom, etcetera…”—

Fungibility — Identical chips behave differently in practice: location, data‑center infrastructure and configuration change delivered performance and price. That lack of fungibility complicates standardized compute contracts.

“Or 1 chip the same chip at 1 data center might not be equal to the same chip at a different data center.”—

CME Partnership — Silicon Data, an index provider, has built GPU indices and announced a partnership with CME to launch GPU derivatives. It says its indices are on Bloomberg and expects CME‑listed GPU options pending regulatory approval.

“So I'm the founder CEO for Silicon Data. So that's the index provider for GPU indices. We recently announced partnership with CME…”—

Live Show Intro

Hosts introduce the live Odd Lots show and set up the topic of whether GPU capacity can become a tradable commodity, previewing their guest.

Silicon Data & CME

Carmen Lee explains Silicon Data's role as an index provider and announces a partnership with CME to launch GPU futures and options pending regulatory approval.

Who Buys Compute

Discussion of market participants — new clouds, AI startups, enterprises and inferencing providers — and the tradeoffs between on‑demand, reserved and forward contracts.

GPU Performance Variance

Carmen describes verification, the 'GPU lottery' performance variance across identical chips, and how Compute Exchange provides pre‑delivery verification and transparency.

Index Construction

How Silicon Data normalizes hundreds of thousands of trade points across sources to build indices, manage basis risk, and produce a hedging instrument correlated to real‑world prices.

Data, Market Structure

The panel covers data licensing, competing indices, market adoption, and how banks and trading desks might provide the speculators and liquidity needed for a healthy futures market.

Pricing, Refurbish & Bubble

Conversation on GPU rental rate trends, refurbishment economics and residual value, chip lifespans, supply constraints, and how to assess whether an AI investment boom is a bubble.