The Tape Is Live: CME's First GPU Futures Start Trading at Tonight's Globex Open

For decades, if you wanted to hedge the cost of the inputs to your business, the futures markets had a contract for you: crude oil for airlines, corn for feedlots, natural gas for utilities, lumber for homebuilders. But if your business was building AI systems, the single biggest input — GPU compute — had no hedge. You signed a long-term cloud contract, crossed your fingers, and hoped rental prices moved your way.

That changes with tonight’s globex open. CME Group, in partnership with GPU pricing specialist Silicon Data, is listing the first regulated compute futures contracts with the October 5 trade date. Two cash-settled products go live: the Silicon Data H100 Rental Index futures and the Silicon Data B200 Rental Index futures, each tracking the hourly rental price of the Nvidia chips that power the overwhelming majority of today’s AI training and inference workloads.

What exactly is trading

Each contract represents a month’s worth of rent for one GPU — roughly 730 GPU-hours priced in U.S. dollars and cents per GPU-hour. The H100 contract settles against Silicon Data’s H100 Rental Index, a daily benchmark compiled from actual transacted rental prices across cloud and GPU-marketplace providers. The B200 contract does the same for Nvidia’s next-generation Blackwell part.

The design deliberately mirrors energy markets. CME houses the products in its energy division (listed on NYMEX), not equities, and the press materials lean on the analogy hard: compute is “the processing power that fuels the AI economy,” and the contracts “bring much-needed hedging and investment vehicles to businesses looking to manage the cost of compute.”

The two contracts serve two distinct constituencies:

• Natural shorts — data center operators, GPU owners, and cloud providers who hold physical compute and want to lock in future revenue or protect against falling rental prices as new capacity floods the market.
• Natural longs — AI labs, enterprises with large training budgets, and inference-heavy businesses that need to cap the cost of future compute and want protection against price spikes when everyone trains at once.

Why this matters now

The timing is not accidental. Three forces have converged to make a compute derivative market viable for the first time:

1. Compute spending has reached commodity scale. AI infrastructure capex has climbed into the trillions-of-dollars-annually conversation, and GPU rental has become a deep, liquid-enough spot market. A benchmark needs real transaction volume behind it, and the H100 rental market now has it — dozens of providers, publishable daily marks, and enough churn that the index means something.

2. Volatility is brutal and getting worse. GPU rental prices have whipsawed as demand surges with each frontier model announcement and supply arrives in lumpy, multi-thousand-rack increments. Silicon Data’s own research tracks the H200-to-H100 rental premium doubling between May and July 2026. Buyers signing 12-month cloud commitments have been flying without instruments.

3. Financing needs price discovery. The wilder edge of the AI financing boom — GPU-backed loans, securitized compute, yield-bearing “compute bonds” — has been criticized precisely because there was no independent forward curve to mark against. A regulated futures curve gives lenders, auditors, and rating agencies a reference price. CME’s own framing points at this: “more reliable tools for valuation” for AI builders, cloud providers, and investors.

The hedging case, concretely

Consider an AI company planning a training run next quarter that will consume roughly 50 H100s around the clock for a month. Under the old regime, its options were to reserve capacity now at whatever the cloud provider demanded, or gamble on spot pricing later. With the futures listed, it can buy H100 Rental Index futures to cap its effective cost: if rental prices spike before the run, the futures gains offset the higher spot bill; if prices fall, it loses on the hedge but buys the actual compute cheaper.

The mirror trade works for a neocloud holding thousands of H100s. Worried that Blackwell ramp and new entrants will crush H100 rental rates by Q2? Sell futures, lock in the revenue, and sleep.

None of this is without basis risk — the index tracks a market-wide average, not any individual provider’s quote sheet, and your actual cloud contract may move differently than the benchmark. That gap is the classic friction of every young futures market, from jet fuel to propane. It narrows as liquidity deepens.

Context: a crowded race to financialize compute

CME is first to the regulated party, but it is not alone. ICE has compute futures of its own in the works, prediction-market operator Kalshi has compute contracts, and a crop of startups is building OTC forwards and swaps on GPU time. Silicon Data, for its part, has spent months laying the groundwork — publishing the first GPU forward curve and positioning itself as the Platts of compute, the benchmark provider whose prints the entire derivatives stack settles against.

The prize is large. Every commodity that ever financialized — oil, gas, electricity, freight — grew a derivatives notional market many times the size of the physical one. If AI compute follows even part of that path, tonight’s listing is the first tick of a market that could eventually route hundreds of billions in risk.

What to watch

• Opening liquidity. New contracts often start thin; the tell will be whether market makers quote tight spreads within the first sessions and whether open interest builds past the curiosity stage.
• Who shows up first. Fast adoption by GPU holders (shorts) with labs and enterprises (longs) taking the other side would validate the natural-participant thesis. If it’s only speculators, the hedging story is still theoretical.
• The curve’s shape. The first genuine forward curve for GPU rental will be read eagerly by anyone underwriting data center debt — contango or backwardation will instantly become a talking point for AI-capex bulls and bears alike.
• Regulatory periphery. The contracts self-certified through CFTC review; expect attention on how index manipulation would be policed as volumes grow.

For now, the simplest summary: as of tonight, the price of thinking — at least the price of the silicon that does it — is a tradeable commodity with a settlement ticker, contract months, and margin requirements. How long from first trade to liquid market is the open question. But every commodities story starts with a first print, and this one lands tonight.