CME Group and DRW Launch GPU Compute Futures — Shanghai Designs AI Token Derivatives in Parallel
On May 12, CME Group and Silicon Data announced they will launch compute futures contracts later this year, pending regulatory approval. Silicon Data — the GPU market intelligence startup backed by trading firm DRW — will provide the underlying price indices. The contracts will track on-demand GPU rental rates in real time, giving cloud providers, AI startups, and financial institutions a standardised instrument to hedge compute cost exposure.
The Intercontinental Exchange, owner of the NYSE, is separately partnering with Ornn — an eight-person startup — to build a competing GPU futures product. BlackRock CEO Larry Fink described compute futures as “a new asset class” at a recent conference. DRW founder Don Wilson, whose firm seeded Silicon Data, said compute will become “the largest commodity in the world.”
The Numbers Behind the Market
The urgency is real. NVIDIA H100 rental prices rose 63% in a single month, per Ornn’s spot pricing data. NVIDIA Blackwell B200 rental costs have doubled since February. GPU spot rates for H100 range from $1.40 to $4.27 per hour depending on provider and region. Prediction market Kalshi assigns a 40% chance that B200s clear $5.91 per hour by the end of this week.
Nine cloud hyperscalers are forecast to spend $830 billion on AI infrastructure in 2026. CFOs dealing with soaring token budgets and GPU shortfalls need hedging instruments. The compute market “looks a lot like oil in the 1970s: a bottlenecked resource begging for financialisation,” wrote analyst Dave Friedman.
The CME product will be based on Silicon Data’s GPU benchmarks — the first daily reference prices for on-demand GPU rental rates. The goal is to do for compute what West Texas Intermediate did for crude oil in the 1980s: create a standardised benchmark that enables price discovery, liquidity, and risk transfer at scale.
China Takes a Different Angle
While US exchanges are targeting GPU rental costs, the Shanghai Futures Exchange is designing contracts tied to AI inference tokens — the per-call pricing units for AI services like GPT-5.5 ($5 per million input tokens, $30 per million output tokens) rather than the compute layer beneath them.
China’s position is not arbitrary. Daily token usage in China surged 1,000-fold between January 2024 and March 2026, reaching over 140 trillion tokens per day. Many Chinese AI model providers have rationed access due to compute shortages. HashKey Group chairman Xiao Feng described tokens as “the digital fuel or raw material powering AI models” — a direct consumption metric for hedging purposes.
The Shanghai design is preliminary and subject to regulatory review. Chinese brokerage Baocheng Futures estimated compute derivatives could launch in three to five years. The fragmented nature of China’s current compute market is the main structural obstacle.
Why This Market Is Hard to Build
The central problem with compute futures is that the biggest participants benefit from opacity. Meta and Google, which consume compute at scale, are better served by bundled, negotiated contracts than by transparent spot prices. Without commercial users — the drillers, refiners, and shippers that made oil futures work — price discovery degrades into speculation.
CoreWeave, whose co-founders ran energy hedge funds before pivoting to GPU rentals, does not have a trading desk oriented toward compute futures. Silicon Data’s Carmen Li acknowledged the company’s pricing indices rely on spot data from smaller “neoclouds” like Crusoe and Nscale, which lack the scale of hyperscaler contracts. The big bilateral deals between Google, Meta, and NVIDIA do not enter the price discovery mechanism.
The CME product is also narrower than its ambition suggests. GPU contracts will track rental rates on commodity hardware in the spot market — a useful hedge for neoclouds and AI-native startups that buy on short contract terms, but not a direct instrument for frontier model operators locked into long-term infrastructure deals.
Key Numbers
- CME + Silicon Data compute futures: launch pending regulatory review, announced May 12, 2026
- H100 spot prices: $1.40–$4.27/hr range, up 63% in the past month
- B200 Blackwell rental costs: doubled since February
- China daily token usage: 140T+ as of end-March, up 1,000-fold since January 2024
- GPT-5.5 token pricing: $5/M input, $30/M output
- Nine hyperscalers 2026 capex: $830B forecast
- DRW-backed Ornn: 8 employees, inked ICE deal for competing GPU futures product