On October 5, 2026, CME Group plans to list the first two US-regulated compute futures contracts on NYMEX, pending regulatory review: Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures. Each one is a cash-settled bet on one month of hourly GPU rental cost for a specific chip, confirmed in CME's press release. For the first time, the cost of renting a GPU is a hedgeable financial instrument, the same way oil, natural gas, or wheat has been for decades.
That matters for anyone planning multi-month AI infrastructure spend, not just trading desks. If you've watched your H100 bill swing 10-40% in a matter of weeks with no way to lock in a rate beyond whatever your provider offers, a public, tradable price index changes the conversation. This post covers what the contracts actually track, why the volatility behind them got bad enough to justify a futures market, what today's gpu cloud pricing actually looks like, and where a futures contract helps you and where it doesn't.
What CME's Compute Futures Contracts Actually Track
CME Group's Global Head of Energy and Environmental Products, Pete Keavey, framed the launch in market terms: "Just as oil fueled the 20th century economy and evolved from spot trading into a global derivatives market, our futures contracts will now turn compute into a standardised, tradable commodity that will provide global businesses with a reliable, regulated venue to manage price risk," he said in CME's announcement. The comparison isn't just marketing language. Oil went from a fragmented, regionally-priced commodity to one with a public benchmark (WTI, Brent) precisely because buyers and sellers needed a shared price to hedge against. GPU rental is at the point oil was decades ago: real demand, real volatility, no shared reference price.
Two Contracts, Two Chips: Silicon Data H100 and B200 Rental Index Futures
The two contracts at launch cover NVIDIA's H100 and B200, the two chips with the deepest rental markets right now. Each contract represents one month's rental cost for the underlying GPU, based on an index of hourly rental prices that index provider Silicon Data publishes. Silicon Data is backed by trading firm DRW, which is the kind of institutional backing CME needs behind an index before regulators and institutional traders will touch it.
CME isn't the first mover here. Architect Financial Technologies already runs perpetual futures on GPU and DRAM rental rates through its Bermuda-regulated AX exchange, in partnership with index provider Ornn Data. CME's contracts are dated (they expire on a fixed schedule, unlike a perpetual) and cash-settled on NYMEX, which makes them the first compute derivatives listed on a US-regulated venue rather than an offshore one.
How the Underlying Index Is Built and Settled in Cash
You can't deliver a GPU the way you deliver a barrel of oil. A contract holder doesn't get an H100 shipped to them at expiration; they get a cash payment or obligation based on where the Silicon Data index landed relative to the price they locked in. Settlement happens financially: "traders exchange cash based on the benchmark result. They do not deliver a physical processor or a reserved server," as one analysis of the CFTC's review of the contracts put it. That same piece is blunt about what the contract can't do: "A holder cannot use the contract to obtain an H100 during a capacity shortage. The instrument addresses price exposure, not operational access." That distinction is the whole ballgame for anyone evaluating whether this replaces a reserved rental contract. It doesn't.
The launch also isn't a done deal yet. The CFTC has opened a public comment period examining a genuinely open question: does computing power meet the legal bar to be a financial commodity at all? Regulators are weighing index-manipulation risk, whether the market has enough liquidity, and what happens if a small number of participants can move the index. October 5 is CME's target, not a guarantee.
Why GPU Price Volatility Created Demand for Hedging
A futures market doesn't get built for a commodity that trades in a flat, boring range. It gets built when price swings are large enough, and frequent enough, that buyers and sellers both want a way to lock in a number ahead of time. H100 rental pricing has had exactly that kind of year, and much of it traces back to the same GPU supply crunch that has kept Hopper and Blackwell capacity tight since late 2025.
The Swings on Record: H100 Rates From $1.70 to $2.35/hr in Five Months
H100 one-year rental contract pricing rose nearly 40% in five months, from a low of $1.70/hr per GPU in October 2025 to $2.35/hr per GPU by March 2026, according to SemiAnalysis. That's not a one-off spike either. Between December 9, 2025 and January 6, 2026, H100 hourly rates jumped 10% in four weeks, from $2.00/hr to $2.20/hr, the largest short-term move Silicon Data had tracked since mid-2025, while A100 and B200 pricing stayed flat over the same window. That last detail is worth sitting with: this wasn't a market-wide repricing. It was a targeted move on one chip, which is exactly the kind of risk a generic hardware budget can't hedge against but a chip-specific futures contract can.
Zoom out further and the picture gets more interesting, not less. H100 cloud rental rates have fallen from $8-10/hr in early 2024 to roughly $1.80-3.50/hr in Q2 2026, a 64-75% decline over that period, despite the sharp short-term spikes layered on top. So the long-term trend for a team budgeting a year out is down, but the month-to-month path is anything but smooth. That combination, falling on average but spiking without warning, is precisely the volatility profile that makes a hedge worth paying for.
No Public Reference Price Meant No Way to Hedge
Silicon Data's CEO, Carmen Li, put the underlying problem plainly: "For years, two companies buying the exact same GPU capacity could pay wildly different prices, with no benchmark comparison available," she said in CME's launch announcement. You can't hedge a price you can't observe. Oil futures work because there's a public spot price every trading desk agrees on. GPU rental never had that. Every provider quotes its own rate, negotiated bilaterally, often under NDA for enterprise deals, which means there was no shared number two counterparties could write a contract against. The Silicon Data index is CME's attempt to manufacture that shared number, the same way Platts and Argus manufacture reference prices for physical commodities that don't trade on a single central exchange.
GPU Cloud Pricing Today: The Volatility a Futures Market Is Reacting To
A quick answer if you're just checking rates: on Spheron right now, H100 runs $2.64/hr on-demand and $2.04/hr on spot; B200 runs $9.36/hr on-demand and $5.34/hr on spot. Those numbers move with availability, sometimes within the same day, which is exactly the kind of variance a futures contract is designed to let a buyer plan around instead of just absorb.
Live On-Demand and Spot Rates Across Providers Right Now
| GPU | Spheron On-Demand | Spheron Spot |
|---|---|---|
| RTX 4090 | $0.53/hr | N/A |
| A100 80GB | $1.43/hr | $1.15/hr |
| L40S | $0.96/hr | N/A |
| H100 | $2.64/hr | $2.04/hr |
| H200 | $4.22/hr | $3.31/hr |
| B200 | $9.36/hr | $5.34/hr |
Pricing fluctuates based on GPU availability. The prices above are based on 27 Aug 2026 and may have changed. Check current GPU pricing → for live rates.
Zoomed out across the whole market rather than one provider, on-demand H100 80GB pricing spans roughly $2.19 to $11.06/hr depending on provider, contract terms, and region, as of August 2026. That's a 5x spread for the same chip. For a full side-by-side across Spheron, AWS, Azure, GCP, and the major neo-clouds, the GPU cloud pricing comparison tracks live rates across every major provider and updates as they move.
Why Fragmented, Provider-by-Provider Pricing Is Exactly What Futures Contracts Standardize
The GPU rental market is fragmented by design, not by accident. "Two rentals involving the same GPU model can still differ in network performance, memory configuration, processor access, software support, geographic location, and contract duration," notes one breakdown of why compute doesn't behave like a normal commodity. A provider might rent you a single GPU, a full server, or a networked cluster, and those aren't interchangeable products even though they're all "an H100." That's the same reason vendor-financed pricing structures distort the picture further: our breakdown of NVIDIA's neocloud backstop financing covers how a provider's rate can reflect embedded financing terms with NVIDIA rather than pure supply-and-demand for the hardware. A futures index has to average across all of that noise to produce one number, which is both the point of the exercise and its biggest weakness: the index is a proxy, not the price you'll actually pay for your specific workload on your specific provider.
The market this index is trying to standardize is not small. GPU rental is estimated at $52.04 billion in 2026, up from $34.62 billion in 2025, and projected to reach $198.74 billion by 2031, a 30.73% CAGR. A market growing that fast, with that much year-over-year price movement, is exactly the environment financial markets try to build hedging instruments around.
What This Means for Teams Budgeting Multi-Month GPU Spend
Who Actually Uses These Contracts: Hyperscalers, AI Labs, Trading Desks
The buyer profile for a NYMEX futures contract is not a Series A startup with a Spheron account. It's hyperscalers and large AI labs with multi-hundred-million-dollar annual compute budgets who need to protect a forecast, and it's trading desks and funds that want exposure to AI infrastructure costs without owning hardware. Clearing a futures position requires a futures commission merchant relationship, margin accounts, and someone on staff who understands derivatives, none of which a lean engineering team building on rented GPUs is going to stand up for a hedge on a $50,000/month compute bill.
What a Futures Desk Can Do That a Procurement Team Can't (Yet)
A futures desk can take a view on where H100 or B200 pricing is headed six months out and put capital behind that view without ever touching a GPU. A procurement team, by contrast, is stuck negotiating bilateral deals with actual providers, one contract at a time, with no public price to anchor the negotiation against. That asymmetry is the gap CME is trying to close: give procurement teams (eventually, once liquidity and access broaden) the same kind of public reference price that a commodities trader already has for oil or wheat. Right now, most AI teams still get their price signal from a provider's published rate card, not from a derivatives exchange.
Hedging vs Just Locking In a Reserved-Rate GPU Contract
What a Futures Contract Protects You From That a Reserved Rate Doesn't, and Vice Versa
A futures contract protects against price risk in the abstract: if the Silicon Data H100 index rises above the rate you locked in, the contract pays out the difference in cash, regardless of whether you ever rent a single GPU. That's useful if your exposure is financial, for example if you're forecasting compute cost for a board deck or need to smooth a P&L line. What it doesn't do is guarantee you a GPU. If your real problem is availability, not price, cash from a futures payout doesn't get you capacity when a provider's pool is sold out.
A reserved-rate contract with an actual GPU provider does the opposite. It locks in both a price and access to hardware, because you're contracting for real capacity, not a derivative on a price index. It doesn't protect you from paying more than the market average if rates fall after you sign, and it's a bilateral deal you have to negotiate yourself rather than a liquid, publicly-traded instrument. The table below breaks down the trade-off:
| Factor | Compute Futures | Reserved GPU Contract |
|---|---|---|
| Protects against | Price index moving against you | Price and availability |
| Settlement | Cash, against a published index | Actual GPU-hours delivered |
| Access requirement | Futures/clearing account | Direct provider contract |
| Best suited for | Hyperscalers, labs, trading desks | Teams that need the hardware, not just the hedge |
| Available today | No, launching Oct 5, 2026 (pending review) | Yes |
Why Most Teams Under Enterprise Scale Are Better Off With a Reserved or Spot Strategy Today
If you're not running a treasury desk, a reserved contract or a well-managed spot strategy protects you from the same volatility a futures contract targets, without the clearing account and without waiting on a regulatory review that hasn't concluded yet. Our GPU cost optimization playbook covers how to combine on-demand, reserved, and spot pricing tiers to smooth out exactly the kind of monthly swings SemiAnalysis and Silicon Data documented above, using tools that already exist rather than a market that's still pending approval. If you're earlier in the process and still working out how to evaluate providers on more than price alone, the AI infrastructure buyer's guide is the place to start, and what GPU cloud actually is if you're newer to renting compute in the first place.
Compute futures are a real and useful development, but they're a macro hedge for institutions with capital markets access, not a procurement tool for a team renting GPUs to ship a product. Watch the CFTC review and the October 5 launch date. In the meantime, the tools that actually protect your budget, reserved rates, spot pricing, and multi-provider comparison, are the ones you can use this quarter.
Whatever CME's futures market does to the index, your actual bill still comes down to the rate on your invoice. Compare H100 on Spheron and B200 GPU pricing against the rest of the market before you sign a reserved contract or budget for the next quarter.
Frequently Asked Questions
Compute futures are cash-settled derivatives contracts that let a buyer lock in a price for GPU rental costs ahead of time, instead of paying whatever the spot market charges when the compute is actually rented. CME Group and Silicon Data plan to list the first two, Silicon Data H100 and B200 Rental Index Futures, on NYMEX on October 5, 2026, pending regulatory review.
Each contract represents one month of hourly rental cost for the underlying chip, H100 or B200, based on an index of GPU rental prices that Silicon Data publishes and updates. The contracts settle in cash against that index rather than delivering a physical GPU or a reserved server, since compute capacity can't be shipped or stored the way a barrel of oil can.
Technically yes once the contracts launch and clear regulatory review, but in practice futures desks and clearing accounts are built for hyperscalers, AI labs with dedicated finance teams, and trading firms, not for a startup renting a handful of GPUs. Most teams below enterprise scale get more practical price protection from a reserved-rate or spot contract with their GPU provider.
A reserved contract locks in a price and guarantees you can actually get the hardware, because you're contracting directly with a provider for capacity. A futures contract only protects the price: it pays out cash if the index moves against you, but it doesn't reserve you a single GPU, so it does nothing if your real constraint is availability rather than cost.
