CME and Silicon Data Plan Futures Market for AI Compute
Silicon Data CEO Carmen Li told Bloomberg Technology that AI compute is becoming a commodity market large and volatile enough to require futures and options. She said Silicon Data’s planned work with CME would create a regulated hedging layer for GPU-price exposure, using Silicon Data’s indices to normalize fragmented pricing across chip types, locations and contract terms. Li argued that banks, data centers, cloud providers and AI companies need those tools because on-demand GPU prices can swing sharply and bottlenecks keep moving across the supply chain.
Compute futures are being built around a price signal that is already volatile
Carmen Li framed Silicon Data’s planned work with CME as market infrastructure for a commodity that has become financially material: AI compute. The proposed product would give participants a futures market for compute exposure, while Silicon Data’s existing data products and Compute Exchange provide the pricing and physical-market context underneath it.
The scale claim in Li’s opening was large. She said energy spend would “probably” exceed $10 trillion this year, then described the spot market around compute as a stack of commodity-like inputs: energy, colocation, and the servers themselves. Her point was that once those inputs become large balance-sheet exposures, participants need financial infrastructure to hedge and manage risk. Banks underwriting trillions of dollars of loans, in her example, need ways to manage their exposure as compute comes online and future price volatility emerges.
The pressure point in Li’s account is not abstract. She said on-demand GPU pricing can show “40% daily volatility,” while a Bloomberg chart of the Silicon Data H100 Index showed GPU rental costs rising from roughly the high-$1 range per GPU per hour in mid-2024 toward the mid-$2 range by May 2026. The chart was labeled “Silicon Data Index tracks GPU rental costs” and noted that prices had climbed in recent months.
Li argued that this kind of volatility requires the same category of tools already used in other commodity markets: futures and options that let exposed parties hedge rather than absorb price swings directly.
Everything is at a commodity level. You need financial infrastructure to hedge, to manage risk exposures.
The analogy she returned to was oil. Oil futures, in her example, have done “a wonderful job” helping people hedge oil fluctuations, including during the “Iran wartime.” Compute, in her view, is moving into the same class of exposure: large, volatile, and important enough that private contracts and spot purchases are not enough.
The market splits between physical capacity and financial exposure
Caroline Hyde asked who, beyond banks, would need to hedge compute. Li divided the answer into two groups: firms naturally long GPUs and firms whose costs increasingly depend on buying compute.
On the long side, she named data centers, neo-clouds, design houses, and fabs. These firms have exposure to GPU assets or GPU-linked capacity, and Li said they would want tools such as short futures or put options. On the buyer side, she pointed to AI startups and enterprises spending heavily on GPUs. Traditional SaaS companies, she said, historically had few major cost lines beyond people; AI companies now face GPU-level or token-level costs that can become a major operating line item.
That distinction also explains how Li separated Silicon Data’s existing Compute Exchange from the prospective CME product. Compute Exchange is a spot market for GPU resources, including forward contracts and reserve contracts for capacity. Its users, according to Li, sit in the underlying consumption layer, from AI startups to neo-cloud providers. Companies that want to reserve GPUs and receive physical delivery would use that venue.
The CME product would address a different need: hedging financial exposure to compute prices. Li described participants using longer-term contracts to lock in rates and avoid on-demand volatility. In her formulation, physical access belongs on Compute Exchange; futures and options belong in the financial hedging layer.
Silicon Data’s on-screen materials reinforced that division. Its website described “real-time price transparency & GPU performance data for traders, financial institutions, and builders.” A separate section for financial institutions described SiliconIndex as a benchmark for swaps, futures, and structured notes tied to GPU markets, and described SiliconNavigator data and forecasts as tools for managing GPU price volatility.
A useful GPU index has to normalize a fragmented product
Hyde pressed Li on whether the idea came from the lack of transparency across cloud and neo-cloud compute offerings, including providers such as AWS, Microsoft, and CoreWeave. Li said the market has some transparency, but “very little” unless a participant is inside the ecosystem and seeing neo-cloud transactions directly.
The opacity is not limited to today’s posted prices. Li emphasized reserve pricing and term-structure risk: long-term contracts can run three to five years, and market participants need visibility from on-demand pricing through the reserve curve. Without that, they cannot easily compare current costs with future contracted capacity.
Li also rejected the idea that GPU transactions can be averaged cleanly. GPUs are not homogeneous products. Prices vary by chip type, HBM configuration, geography, and other specifications. An H100 in one configuration and location is not automatically comparable with an H100 in another. Silicon Data’s index work, she said, is meant to normalize those differences into representative, trackable data.
The SiliconNavigator page shown on screen described daily refreshed rental and resale prices across clouds, brokers, and multiple GPU design houses, delivered through API or on-demand downloads. It also claimed up to eight years of historical data for trend analysis and planning. The visible product table included fields such as product name, memory bandwidth, TFlops, TFlops per watt, and low cost, underscoring Li’s point that the benchmark has to account for technical attributes as well as price.
Prices are rising because the bottlenecks keep changing
Asked what Silicon Data’s data show about the cost of compute, Li’s answer was direct: prices are going up. She said the increase began in December of the prior year and contrasted that with a common expectation that compute prices would keep falling, potentially toward zero over time. That is not what she sees in the data.
Her explanation was that the supply-demand curve and forward curve shift constantly as expectations change and new constraints appear. The bottleneck is not stable. Last year, she said, it was fabs. This year it is memory. In the future, it could be geolocation or colocation space.
That matters for the design of a hedge. Compute-price risk is not reducible to a single input cost, and the source of pressure can move across the supply chain. A buyer, data center, lender, or financial institution exposed to GPU economics may face different constraints at different moments, which is why Li argued for indices tied to broader market pricing rather than to one technical input.
Li presents the regulatory step as conventional futures-market plumbing
Hyde noted that the CME arrangement had been announced but remained subject to regulatory oversight. Li said Silicon Data is actively working with CME and reviewing the product with the CFTC and other regulatory bodies.
She characterized the proposed instrument as a traditional data futures product, not an exotic structure. Her comparison was again to existing commodity markets: oil, gas, and other asset classes already have futures built around market data and benchmarks. On that basis, she said she did not see a major concern.
Li also said Silicon Data launched what she called the world’s first GPU indices on the Bloomberg Terminal last year. Those indices cover prices ex-China, including Europe and the US, and normalize geolocation into single data points. Her stated goal is for participants around the world to hedge exposure using those indices, within that ex-China coverage.