Compute Futures Arrive October 5. Do the Math • The Weighted Average

CME Group will list two compute futures contracts on October 5, pending regulatory review, turning GPU rental cost into something a finance team can hedge. The exchange’s announcement with Silicon Data specifies Silicon Data H100 Rental Index Futures and B200 Rental Index Futures, each contract representing “a month’s worth of rent” for the respective Nvidia GPU, listed under NYMEX rules. The announcement came 46 days before the listing date, which is roughly one budget cycle for anyone who wants a position on day one.

Put a number on the unit being traded. Lambda’s published on-demand pricing lists NVIDIA H100 SXM at $3.99 per GPU-hour and B200 SXM6 at $6.69. A 730-hour month at those rates is $2,913 for an H100 and $4,884 for a B200 — a 68% premium for the newer part, at one provider’s public list. That spread is the whole argument for two contracts instead of one: the generations are not fungible, and a hedge struck on the wrong chip is a bet, not a hedge.

Why a reference price changes procurement, not just trading

Silicon Data CEO Carmen Li framed the gap the contracts close: “For years, two companies buying the exact same GPU capacity could pay wildly different prices with no way to know who got the better deal. They will now have a benchmark to check that against.” That is a procurement claim wearing a derivatives costume. The immediate value for most operators is not taking a position; it is walking into a renewal with a published curve instead of a vendor’s assertion.

The arithmetic also exposes how crude most compute budgets still are. Engineering teams reason in dollars per million tokens because that is what the API bill shows; infrastructure teams reason in dollars per GPU-hour because that is what the cloud invoice shows; nobody reconciles the two. A contract denominated in GPU-months forces the translation. If a training run occupies 64 H100s for three weeks, that is 48 GPU-months, or roughly $139,800 at the $2,913 list-derived figure above — a number a CFO can hedge, approve, or refuse. The same run described as “three weeks of cluster time” is unhedgeable because it is not priced.

The timing is not accidental. CME’s Pete Keavey compared the launch to oil’s evolution “from spot trading into a global derivatives market,” and the announcement lands inside a broader financialization of compute: Nvidia signed memorandums of understanding with six asset managers to mobilize more than $500 billion of third-party capital for AI compute infrastructure, with Jensen Huang telling CNBC that “this is really the first time that technology chips have become an investable asset class”. Lenders underwriting GPUs need a mark; futures supply one. The Verge’s reporting on that financing push notes the awkward corollary — the same CEO said last year that “you couldn’t give Hoppers away” once Blackwell shipped in volume, and now describes A100-class silicon as having economic life “toward a decade.” A market that prices rental hours will settle that argument faster than any earnings call.

What could break the hedge

Four things. First, basis risk: the contracts settle against Silicon Data indexes of hourly rental prices, and your actual bill is a negotiated reserved-capacity contract with a specific provider, region, and interconnect. The index can move while your invoice does not, and vice versa. That is normal in commodity hedging and unfamiliar to most engineering organizations.

Second, liquidity. A new contract with no open interest is a quote, not a market. Judge these by volume in November, not by the launch press release, and remember the listing is still explicitly “pending regulatory review.” Commodity contracts routinely list and languish; the ones that work attract natural hedgers on both sides, which here means data-center operators selling forward and AI builders buying. Neither constituency has ever traded futures before, and the learning curve is a real drag on early volume.

Third, index construction. Silicon Data’s benchmarks aggregate rental and resale prices across clouds and brokers, which means the settlement price reflects a market that includes spot brokers whose quotes may not resemble what an enterprise pays under contract. Read the index methodology before treating the curve as a forecast of your own cost.

Fourth, the direction of the underlying. Rental prices for older parts have been rising on inference demand rather than falling on obsolescence, which is what makes the asset-class argument work at all. If inference supply catches up — or if efficiency gains cut hours per task the way the archive traced in the end of the token binge — a long hedge locks you into yesterday’s scarcity.

The operator call is narrow and concrete. If GPU rental is a material line item, ask your finance team to model a hedge against the H100 index for the portion of capacity you rent on demand, and use the published curve in your next reserved-capacity negotiation regardless of whether you ever trade. If you buy capacity through a neocloud lease, the index is a benchmark for the contract you already signed. And if your spend is mostly per-token API calls rather than rented hours, this changes nothing directly — though it prices the input under the layer today’s lead describes, where Stripe just paid a reported $7.5 billion for the meter that routes those tokens. Two floors of the same building became tradable in one week.

The evidence that would change the verdict: three months of open-interest data showing whether industrial hedgers or speculators dominate. If AI builders and data-center operators are actually on both sides, compute has become a commodity. If it is only traders, it is a product looking for a market.