Carmen Li's Plan to Build a Futures Market for Compute
Carmen Lee, CEO of Compute Exchange and Silicon Data, discusses her efforts to build futures markets for GPU compute on the CME, drawing parallels to oil commodity markets. She explains how GPU price indices are constructed, who the natural buyers and sellers of compute futures would be, and addresses challenges like GPU performance variance and fungibility. The conversation covers spot markets, refurbished chips, residual value calculations, and the broader question of whether AI represents a bubble.
Summary
The podcast episode features Carmen Lee, founder and CEO of Silicon Data (an index provider for GPU indices) and CEO of Compute Exchange (a spot marketplace for GPU procurement), recorded live at an Odd Lots show in New York. She was previously connected to Don Wilson of DRW, who was discussed in an earlier episode about building a compute exchange.
Carmen describes two distinct but related ventures: Silicon Data, which has developed GPU price indices now available on Bloomberg Terminal and is launching GPU futures and options on the CME pending regulatory approval; and Compute Exchange, which facilitates spot, reserve, forward, and refurbished GPU contracts. The CME partnership represents a major milestone after roughly two and a half years of index development.
On market structure, Carmen draws an extended analogy to oil markets. NeoCloud providers who own GPU hardware are likened to oil producers (like Shell) who are naturally long and want to hedge revenue volatility by shorting futures. Enterprises and AI startups who consume compute are naturally short, like airlines hedging fuel costs, and want to lock in prices. She notes the shift from on-demand to reserved and forward contracts as GPU scarcity has increased, with on-demand pricing swinging from $3 to $9 per hour depending on supply-demand dynamics.
A significant technical challenge discussed is GPU non-fungibility. Carmen references a paper published at the GPGPU conference with Jefferson Lab showing 38% performance variance for the same A100 chip across different providers and data centers. Compute Exchange addresses this through independent GPU benchmarking (described as a 'Carfax for GPUs'), verifying chips on FLOPS, memory bandwidth, tokens per second, and SLA before delivery.
The index methodology is described in detail. Rather than simple averaging, Silicon Data collects six months of historical trading data from over 100 data sources, ingests 150,000+ traded prices daily, normalizes them across chip characteristics and configurations, and calculates settlement prices. The resulting daily volatility for A100 and H100 indices runs around 20-30%, described as a healthy commodity volatility range, even though individual chip configurations at specific geolocations can show 80-100% volatility.
Carmen shows a chart of GPU rental rates for B200, H100, and A100 chips. She notes B200 prices launched high, dipped, then rose above launch prices — a steeper-than-expected demand signal. H100 prices rose ~8% in the last three months, and A100 prices (the oldest chips) rose 10-15%, indicating broad supply-demand tightening across generations.
On refurbished GPUs, Carmen describes a residual value framework using discounted cash flow analysis of forward contracts to determine break-even and resale value. She cites data showing second-year H100 resale value at approximately $0.85 on the dollar and third-year at $0.84 — a relatively shallow depreciation curve compared to automobiles.
Regarding CME futures settlement, Carmen confirms they will be financially settled like traditional commodity futures, not physically delivered. She notes Compute Exchange experimented with Polymarket contracts for GPU prices, which she described as starting organically when someone listed her product without consent, eventually leading to a collaboration.
On the AI bubble question, Carmen declines to offer price guidance as an index provider but frames the question through cash flow analysis at the machine level — whether forward contract revenue, discounted back, supports the purchase price of hardware. She notes that signed forward contracts provide some visibility into demand but acknowledges overbuilding risk and the lag between announced data center investment and actual GPU availability.
Key Insights
- Carmen Lee argues there is 38% performance variance for the same A100 chip depending on the provider and data center, which she calls the 'GPU lottery' problem — meaning GPU compute is fundamentally non-homogeneous and cannot be simply averaged for index construction.
- Lee claims that Silicon Data normalizes over 150,000 traded prices daily from 100+ data sources against chip characteristics to produce settlement prices, resulting in a daily volatility of 20-30% for H100 and A100 indices — which she considers a healthy commodity volatility range comparable to other physical commodities.
- Lee observes that B200 GPU prices followed the historical pattern of launching high then declining, but the slope of decline was less steep than expected for prior generations, and prices subsequently rose above launch levels — which she interprets as a demand-supply curve in a structurally different state than previous chip generations.
- Lee states that second-year resale value for H100 refurbished chips was approximately $0.85 on the dollar and third-year approximately $0.84, arguing this shallow depreciation curve contradicts widespread narratives about rapid GPU obsolescence.
- Lee draws an explicit analogy between GPU market participants and oil markets: NeoCloud providers are like oil producers (Shell) who are naturally long and hedge by shorting futures, while AI startups and enterprises consuming compute are like airlines — naturally short and wanting to lock in cost stability.
- Lee argues that even older A100 chips saw 10-15% price increases in the past three months despite being the oldest generation, which she uses as evidence of broad supply-demand tightening across all GPU generations, not just the latest hardware.
- Lee contends that announced data center investments do not translate quickly into available GPU supply because co-location, optic fiber, and other infrastructure constraints create significant delays between capital commitment and actual compute availability.
- Lee frames the AI bubble question at the machine level through discounted cash flow analysis of signed forward contracts rather than equity valuations, arguing that if forward contract revenues support the purchase price of hardware, there is no bubble at the compute infrastructure level regardless of what happens to AI company valuations.
Topics
Full transcript available for MurmurCast members
Sign Up to Access