CME to Launch GPU Computing Power Futures on October 5, Tracking H100 and B200 Rental Costs

As investment in artificial intelligence infrastructure continues to expand, GPU computing power is gradually evolving from an infrastructure cost within the AI industry chain into a financial asset that can be priced, traded, and hedged.

CME Group and GPU market data company Silicon Data announced on Tuesday their plan to launch two computing power futures contracts on October 5, which are currently pending regulatory approval.

The two contracts are the Silicon Data H100 Rental Index Futures and the Silicon Data B200 Rental Index Futures. They will be listed on CME and subject to the relevant rules of the New York Mercantile Exchange (NYMEX).

The contracts will track the H100 and B200 GPU rental price indices compiled by Silicon Data, using hourly GPU rental costs as the core pricing metric.

This means that market participants may soon be able to use futures instruments to hedge against fluctuations in AI computing power prices. CME stated that the new products aim to help traders, financial institutions, AI developers, and cloud service providers manage the price risks associated with the rapidly growing computing power market.

CME has previously referred to computing power as "the new oil of the 21st century," stating that computational resources are becoming an independent emerging asset class.

GPU Rental Prices Are Becoming an Important Cost Indicator for the AI Industry

For companies involved in AI model training and inference, leasing GPUs directly from cloud providers or specialized computing power service providers is an important way to acquire computational capabilities, in addition to capital expenditures such as purchasing GPUs and building data centers.

Therefore, changes in hourly GPU rental prices can, to some extent, reflect the supply and demand status of computing power.

Silicon Data currently publishes daily GPU rental price benchmarks, covering major AI accelerators such as the H100, A100, B200, and AMD MI300X. It forms standardized prices by collecting data from cloud service providers, hyperscale cloud vendors, managed data centers, and the private leasing market.

The B200 is one of the core products of NVIDIA's Blackwell architecture, offering higher computational performance compared to the previous-generation H100. As Blackwell's production capacity and deployment are still in the expansion phase, B200 rental prices are currently influenced by both tight supply and strong demand for AI training.

Silicon Data stated that the B200 rental price index reflects standardized hourly prices from various computing power supply channels.

The Significance of the Futures Market Lies in Establishing a Public Forward Price Discovery Mechanism

Currently, the GPU leasing market is highly fragmented, with significant price differences existing among different cloud providers, computing power service providers, and lease terms. Companies also lack mature tools, similar to crude oil futures, to lock in future computing power costs.

When CME and Silicon Data first announced their partnership in May this year, they stated that the new contracts aim to help AI companies and cloud service providers manage the risk of computing power price fluctuations. Silicon Data subsequently launched GPU forward price curves, covering models such as the H100, B200, and A100, with tenors extending up to 36 months.

For computing power suppliers, if they expect GPU rental prices to decline in the future, they can use futures for hedging; for AI companies that need to lease large numbers of GPUs, futures may provide a tool to lock in future computing power costs.

Meanwhile, financial institutions and traders can also participate in price trading based on their judgments regarding AI demand, GPU supply, and data center construction cycles.

This also means that the financialization of the AI industry is extending further from chips, data centers, and related stocks to the underlying computational resources themselves.

If CME's products launch successfully and generate sufficient liquidity, H100 and B200 futures prices could become important market indicators for observing the supply and demand of AI computing power, data center utilization rates, and returns on AI infrastructure investments.

However, there is a key difference between computing power and traditional commodities: GPU computing power cannot be stored like oil or gold. Once computing power sits idle, its corresponding hourly service value disappears. Therefore, the pricing mechanism, liquidity, and hedging effectiveness of GPU futures will differ significantly from those of traditional commodity futures.

A recent study on the pricing of AI computing power assets also pointed out that because computing power is non-storable, the no-arbitrage pricing relationships found in traditional commodity futures cannot be directly applied to the computing power market.

Therefore, if these two contracts are listed as scheduled on October 5, their significance may lie not only in adding two new futures products but also in establishing a public financial benchmark for AI computing power, which previously lacked a unified market price.