Carmen Lee is the founder and CEO of Silicon Data, the index provider for GPU indices, and CEO of Compute Exchange, a spot marketplace for GPU procurement. She recently announced a partnership with CME to launch the world's first GPU futures and options, pending CFTC approval.

The conversation explores the emerging market for tradable compute derivatives, building on previous discussions about Don Wilson's DRW investment in creating liquid markets for GPU capacity. Lee explains how her companies are addressing the challenge of turning heterogeneous compute resources into standardized, tradable financial instruments.

The discussion covers the technical challenges of creating fungible GPU indices, the variety of market participants from AI startups to traditional enterprises, and the volatility patterns that make compute futures economically viable for hedging purposes.

Building the World's First GPU Futures Market

Silicon Data launched the world's first GPU indices on Bloomberg Terminal in 2024 and will launch GPU futures and options on CME in a couple months, pending CFTC approval

The company has been developing GPU indices for 2.5 years, collecting over 150,000 daily pricing points from 200+ data sources globally

"We normalize the traded prices based on different characteristics of the model itself and normalize to a base case, and then we do the math" - Carmen, explaining their settlement price methodology

The Challenge of GPU Performance Variance

Research published at GPGPO conference with Jefferson Lab proved 38% performance variance for identical A100 40GB chips across different deployments

"We have a Plex for GPU, Carfax for GPU, depends how you look at it" - Carmen, describing their GPU verification service

Compute Exchange independently verifies GPU performance on flops, memory bandwidth, tokens, and other metrics before delivery to customers

Price differentiation occurs due to varying CPU, RAM, disk, geolocation, and memory bandwidth configurations even for identical chip models

Market Participants and Trading Patterns

Buyers range from AI startups spending millions on GPUs to traditional enterprises needing just a few nodes for inference deployment

"Most North American firms, they do a variety of combination of contracts. Obviously, on-demand gives you the most flexibility" - Carmen

Forward contracts are gaining popularity as users shift from unpredictable on-demand pricing that can fluctuate from $3 to $9 per hour

New cloud providers act as natural long positions (like oil producers), while compute users are naturally short and need hedging instruments

GPU Price Volatility and Market Dynamics

Daily volatility for A100 and H100 chips ranges from 20-30%, creating a "very healthy commodity volatility range" suitable for derivatives trading

Individual chip configurations at different geolocations show volatility ranging from 80% to over 100% before normalization

B200 prices initially dropped after launch but have risen above original launch prices, showing "demand-supply curve in a different stage"

H100 prices, despite being the oldest among the three chips, increased 8-10% over the past three months, indicating broad market tightening

Refurbished GPU Market and Depreciation Reality

Second-year H100 resale value holds at 85% of original price, with third-year value at 84%, contradicting fears of rapid depreciation

"I think my car depreciates way more than that, right? And I drive my car for whatever, 10 years" - Carmen, comparing GPU to automotive depreciation

Legacy L40s chips still command 40 cents per GPU per hour from hyperscalers, demonstrating sustained value for older hardware

ROI calculations for new cloud providers involve comparing latest chips versus refurbished options based on expected revenue generation and residual values

Trading Infrastructure and Market Access

CME futures will be financially settled using the same process as traditional oil settlement, with no physical delivery required

"The beauty of CME is you can do the same thing you're doing now trading CME products. The same process of processing margin" - Carmen

Any commodities broker that clients currently use will provide access to GPU futures through existing platforms and margin optimization systems

The company previously tested GPU price contracts on Polymarket with February and April settlements before the exclusive CME partnership

Resources Mentioned

shows 38% performance variance for identical A100 chips due to different configurations and data center locations Second-year H100 resale value holds at 85% of original price

rom 20-30%, creating healthy commodity-like trading conditions for hedging instruments Carmen Lee's research shows 38% performance variance for identical A100 chips due to different configurations and

Local LLM Inference Optimization A Comprehensive Guide to Quantization, Hardware Acceleration, and Efficient Private AI Deployment

- Carmen, explaining their settlement price methodology The Challenge of GPU Performance Variance Research published at GPGPO conference with Jefferson Lab proved 38% performance variance for identi

Programming Massively Parallel Processors A Hands-on Approach

because you mentioned really quick questions, so we actually call it GPU lottery. So we published a paper early this year at GPGPO conference with Jefferson Lab on GPU performances. Well, actually, s