CME's compute futures: not a perpetual

On May 12, CME Group and Silicon Data announced the first compute futures — cash-settled against Silicon Data’s daily GPU rental indices, pending regulatory review, with launch later in 2026. The market it creates is what gets called, in derivatives shorthand, an index future on a non-storable underlying.

The right way to read this announcement is to ask two engineering questions and one structural one. What actually gets delivered? Why isn’t it a perpetual? And who arbs it? I’ll work through all three, because I think the press coverage has been too generous about how clean the hedging story is.

What gets delivered

Nothing physical. That part is forced by the underlying.

A GPU-hour is a consumed good, not a stored one. The moment you rent an H100 for sixty minutes, the supply on the other side is gone — it cannot be warehoused, rolled forward, or aggregated into a delivery contract the way a barrel of WTI crude or a bushel of corn can. There is no terminal warehouse for compute. The “spot” is itself a methodology, not a tradeable instrument.

So the contract is cash-settled against Silicon Data’s daily GPU rental indices. A trader who buys a contract at $2.50/GPU-hour and watches the index settle at$ 3.50 at expiry receives the $1.00 difference in cash. The settlement reference is whatever Silicon Data publishes on the relevant date — not a load delivered into a datacenter rack.

The Silicon Data index isn’t trivial, and that’s the point. Their methodology — refreshed daily, normalized for interconnect type, cluster scale, geography, and performance variance, drawing from a data network that covers more than 80% of the global H100 rental market — is what makes a clearable contract possible at all. Without the index, there’s no contract, because there’s no agreed-upon number to settle against. This is the same reason VIX futures couldn’t exist before VIX was published, and electricity futures couldn’t exist before regional reference prices were standardized.

A useful anchor for the basis-risk story below: in March–April 2026, the H100 Hyperscaler On-Demand Index traded in a $7.40–$ 7.52/hr band, while the H100 Neocloud Index sat at $2.43–$ 2.63/hr. The same H100 hardware, two different tapes, roughly 3× spread. That gap tells you most of what you need to know about how this contract will actually hedge real exposure.

The settlement mechanics, in detail

Three implementation choices buried in the fine print determine whether this contract is hedgeable at institutional scale or just a noisy speculation venue. The announcement didn’t spell them out, but you can predict most of them from how CME handles its other index-settled cash contracts.

The settlement window. Final settlement will almost certainly reference Silicon Data’s published index on a specific date — likely the last trading day of the maturity month, in line with how CME handles Henry Hub natural gas and PJM electricity. A single-day reference is cleaner administratively but concentrates manipulation risk into a 24-hour window. The alternative — averaging the index across the last N trading days (a settlement period rather than a settlement point) — is what CME uses on parts of its power and weather complex precisely because it dilutes the single-day attack surface. My bet: an arithmetic average over the final 3–5 business days, not a single-day snapshot. Anything tighter is asking for the GPU equivalent of a WTI roll squeeze.

TWAP vs. snapshot. Silicon Data’s daily index is itself already an aggregation — they don’t publish a tick. So at the index level the question isn’t “TWAP or snapshot” but “what does their daily methodology weight by?” If it’s volume-weighted across observed rentals, you get something closer to a true intraday TWAP rolled up to a daily print. If it’s a price snapshot at a fixed UTC time, you get something closer to a daily fix — easier to game for anyone with enough listing power on the supply side. The contract’s robustness inherits whatever Silicon Data’s daily methodology is; CME doesn’t get to patch a weak index after the fact. The index methodology is the contract’s load-bearing wall.

For grounding, here’s how CME’s existing cash-settled, non-physical contracts handle the same window question:

Product	Settlement reference	Style
Henry Hub natural gas	Last trading day, intraday-weighted	Single-day with intraday averaging
PJM electricity (peak)	Daily LMPs averaged over delivery month	Multi-day averaged
HDD / CDD weather	Sum of daily indices over the month	Multi-day summed
VIX	Single SOQ on expiry morning	Single point
Compute (predicted)	3–5 day arithmetic average of the Silicon Data daily	Multi-day averaged

The pattern: the deeper and more continuously priced the underlying, the more CME tolerates single-day references. Compute is the opposite of continuously priced — daily print, no spot tape — so the window has to do the work the underlying can’t.

Dispute handling and calculation-agent risk. When the published index is wrong, late, or based on data that later turns out to have been contaminated, who eats it? Standard CME practice for index-settled cash products: the exchange is the calculation agent of last resort, the index provider is the data source, and there’s a documented fallback — manual reconstruction by the exchange using best-available data, an emergency committee, and (for severe cases) settlement at a panel-determined level. The Silicon Data contract needs an analog. Worth reading the rulebook the day it drops — the looser the fallback language, the more counterparty risk you’re carrying against the exchange’s discretion. The tighter and more pre-specified, the more institutional the product feels.

The headline numbers — tick size, contract notional, initial margin — are second-order. The settlement window, the index methodology, and the dispute fallback are the product. Everything else is plumbing.

Why it isn’t a perpetual

This is the question I keep seeing asked, especially from people who came up through crypto. The short answer: CME doesn’t list perpetuals, and won’t.

US-regulated futures clear under the CFTC framework, which is a fundamentally dated-expiry regime. Every CME contract has a settlement date. Listings are a curve of monthly or quarterly maturities. Cash settlement happens at expiry against a published reference. That’s the architecture, and it isn’t going to bend for compute.

The perpetual swap — BitMEX’s 2016 invention, now the dominant product on Binance, Bybit, OKX, dYdX, and Hyperliquid — is a crypto-native structural choice that:

Has no expiry. The contract just keeps trading.
Anchors price to the index via a funding rate, charged every funding interval (8h is typical; Hyperliquid runs hourly).
Lives outside the CFTC dated-futures framework. None of the major perp venues are CME-equivalent regulated US exchanges.

CME’s compute future will look like its natural gas, electricity, or VIX contracts: monthly or quarterly maturities, fixed expiry, final cash settlement, no funding leg. You can stack a curve of them and synthetically approximate a perpetual position by perpetually rolling the front month — but you eat the calendar spread every roll, and the further out the curve, the more the basis to spot widens.

So: not a perpetual.

But — and this is the part the coverage misses — a compute perpetual is almost certainly coming on a crypto venue within twelve months. I’d put short odds on Hyperliquid being first, since it’s the venue most comfortable with novel index products. The mechanics are trivial: list a perpetual indexed to Silicon Data’s daily GPU price, take funding every hour against that index, no expiry. None of the CFTC dated-futures architecture applies, because the venue isn’t a CFTC-registered DCM.

When that happens, the two products will price differently — and that’s where the more interesting trading shows up.

What the Hyperliquid perp probably looks like

If you accept that some crypto venue lists a compute perpetual within twelve months — I do — it’s worth being concrete about what that contract has to look like to actually work. The interesting design decisions aren’t free; they’re forced by the structure.

Funding interval and rate cap. Hyperliquid runs hourly funding on its existing perps; a compute perp would inherit that cadence. Funding per interval is whatever pulls perp price toward the index, typically capped at something like ±0.05% per hour (≈ ±438% annualized at the cap). That cap matters more here than on BTC, because the underlying isn’t traded continuously. If the index only refreshes daily, perp price can drift far from the published number intraday with no funding force pulling it back. Expect either a higher-than-typical funding cap, or a more aggressive interval — funding every 15–30 minutes computed against a smoothed daily reference.

Index source — the actually hard problem. Silicon Data publishes daily. A perpetual needs at least an hourly anchor, preferably a streamed feed, for funding to mean anything. Three plausible paths:

Silicon Data publishes daily ─► Hyperliquid perp needs ≥ hourly anchor
                                            │
              ┌─────────────────────────────┼─────────────────────────────┐
              ▼                             ▼                             ▼
   (1) Silicon Data ships         (2) Hyperliquid builds         (3) Daily anchor +
       a sub-daily feed              its own GPU oracle             smoothed EMA
              │                             │                             │
   most defensible            cheapest oracle infra            cheapest legally
   highest ongoing cost       "your oracle is lying" risk      weakest robustness
              │                             │                             │
              ▼                             ▼                             ▼
   probable steady state         unlikely day-one path           likely week-1 shim

I’d take the over on (1) happening, with (3) as a transitional shim during the first few weeks.

Leverage and mark-price design. Compute price isn’t BTC. Realized vol on the H100 Neocloud index over the last year is real but not crypto-spot huge — single-digit daily moves are normal, double-digit days rare. That argues for higher max leverage at first glance, but the gap risk between daily index prints argues for lower — there is no continuously tradeable underlying to fade dislocations against. My guess: launch at 5× max, lift to 10× after a few months of data, never go to 50× the way BTC perps did. Mark price will be a function of order-book mid plus an index-anchoring term — same architecture as Hyperliquid’s existing perps, just with a slower-moving anchor.

Which contract first. Almost certainly one perp on the headline H100 blended index — not separate Hyperscaler and Neocloud sub-perps at launch. Derivatives liquidity concentrates; splitting volume across two contracts on day one kills both. The split, if it comes, comes later — once one contract has anchored a real book, the venue lists a second and you get a tradeable spread. That spread is the actual product the basis traders care about, but you can’t list it first.

Liquidity bootstrapping. Hyperliquid will pay for it — market-maker rebates, points-program weight on the new market, friendly onboarding for one or two prop firms. (DRW is the obvious candidate since they backed Silicon Data; regulatorily that’s awkward, so pick your favorite.) First three months will be thin, volatile, easy to push around. Don’t size a hedge into it in week one.

Pulling the predicted shape into a single spec sheet, against the CME contract:

Spec	CME compute future	Hyperliquid compute perp (predicted)
Expiry	Monthly / quarterly dated	Perpetual, no expiry
Settlement	Cash, end-of-period vs. Silicon Data	Continuous funding vs. index anchor
Funding mechanism	None	Hourly (or finer), capped near ±0.05%/hr
Index cadence	Daily	Needs ≥ hourly anchor (sub-daily feed or smoothed daily)
Max leverage	Initial-margin set by CME (likely modest)	Launch 5×, lift to 10×, never 50×
Counterparty	CME Clearing (CFTC-regulated DCM)	Hyperliquid (non-US, no CFTC oversight)
Mark price	Daily settle from cleared book	Order-book mid + index-anchor term
Day-one product	Headline H100 blended index	Same — split into sub-indices comes later
Liquidity bootstrap	MM program, FCM distribution	Points + MM rebates + prop onboarding
Primary user	Institutional hedgers, FCMs, basis funds	Retail, props, basis arb

When this product lands, the basis between the CME maturity strip and the Hyperliquid perp itself becomes the trade — same shape as the BTC perp / CME futures basis circa 2018–2021, where funding-rate carry was extractable for years by anyone willing to sit on both sides and roll. The compute version of that trade is what’s worth waiting for.

The closer analog is electricity, not Bitcoin perps

If you want a mental model for how CME’s compute future will behave, stop reaching for BTC perp. Reach for electricity futures and weather derivatives.

These products share the structural features:

	Storable?	Settlement	Term curve	Basis to physical
Crude (WTI)	Yes	Physical or cash	Contango / backwardation	Tight — pipeline + warehouse arb
Natural gas (Henry Hub)	Partial	Physical	Strongly seasonal	Moderate basis
Electricity (PJM, ERCOT)	No	Cash vs. regional reference	Spiky, seasonal, intraday	Wide basis to specific nodes
Weather derivatives (HDD, CDD)	No	Cash vs. NOAA temperature index	Pure exposure product	Always basis (your weather ≠ index)
Compute (this contract)	No	Cash vs. Silicon Data index	Expect spiky	Wide basis to your actual rental

Compute belongs in the bottom three rows. The absence of physical-spot vs. futures arbitrage is structural, not a transitional artifact of an immature market. Even mature electricity futures markets carry persistent basis to specific load zones. The market clears anyway, because hedgers don’t need zero basis — they need less basis than holding outright price risk.

The basis risk story is the whole product

I think this is the part the press coverage understates. The futures contract is one number per maturity. Your actual H100 exposure is many numbers, all moving differently.

Three concrete cases:

An AI lab on hyperscaler contracts. Your real exposure tracks the Hyperscaler On-Demand sub-index ( $7.40–$ 7.52 recently). If the listed future references a blended index, you’re carrying basis between the blend and your sub-index. Hedge ratio matters: under-hedge and you stay exposed; over-hedge and you become a directional speculator on the blend’s composition.

A neocloud arbitrageur. Your supply costs are Neocloud rates (~ $2.50). Your customer-facing rates are closer to hyperscaler ($ 7.40+). You’re already long the spread. The future lets you lock in one leg cleanly — but locking in the spread requires combining two contracts (if both sub-indices get listed separately) or eating residual basis.

A hedge fund taking a macro view. “AI capex peaks Q3, GPU rental rates compress 20% by year-end.” Clean to express through the futures — sell the strip. No basis problem because there’s no physical exposure to hedge against.

In year one, I’d expect the third group to dominate volume. Hedgers move slow, especially when the product is new, the basis is unfamiliar, and auditors haven’t agreed on hedge-accounting treatment yet. Macro and basis traders move first. Historical pattern for every new commodity contract: hedgers are the long-term reason the product exists, speculators are the short-term reason it has liquidity.

Who arbs it (and why pure arb is hard)

You cannot easily short physical GPU spot to arb the future. There is no shorting of compute — you can’t borrow GPU-hours and deliver them later. The closest synthetic short is canceling a forward rental contract you already had, which means you have to be a large GPU buyer to express the trade. Tiny set of participants.

Long arb is more accessible: buy GPU-hours forward on the physical market (a forward rental agreement with a cloud provider), sell the future, hold to settlement. But forward rental agreements are bilateral and illiquid — not standardized, not centrally cleared, conventions vary by provider. You wear counterparty risk and structural illiquidity for the round trip.

So pure futures-vs-spot arbitrage is shallow. The contract clears anyway because of:

Cloud providers selling forward capacity to lock in revenue
AI labs hedging training budgets ahead of a known compute campaign
Hedge funds expressing directional or term-structure views
Eventually, basis traders extracting carry from the gap between sub-indices

Price discovery in year one will be dominated by views, not by arbitrage. Expect the curve to be noisy and the term structure to do strange things until basis traders show up to clean it.

The compute commodity unlock

Step back. The structural significance of this listing is that compute becomes a commodity asset class. Not because the press release said so — because once you have a clearing benchmark, three things unlock:

Options on the future. Inevitable. Within 24 months you can buy a put on GPU rental rates to hedge a planned training run, or sell a call against your fleet to monetize upside caps.
Calendar spreads. As soon as multiple maturities list, the term curve becomes its own product. Backwardation in the front contract is a real signal about supply tightness. Contango in the back reflects expected capacity build-out.
Synthetic compute exposure for non-operators. A pension fund wanting AI infrastructure exposure without owning a datacenter can express it through the futures curve. Today they can only hold NVIDIA stock — a fundamentally different exposure (margins + competitive dynamics + balance sheet) than GPU-hour pricing itself.

The crypto-native perp, when it lists, will almost certainly be the higher-volume retail and prop venue. The CME contract will be the institutional one. They will price differently. The basis between them will itself become a trade — same playbook as the BTC perp / CME-futures basis, where double-digit annualized funding-rate carry was extractable for years by anyone willing to be long perp / short CME and roll.

The shape

If you’ve read the about letter:

A continuous score, a position, a measured edge.

CME’s compute futures are exactly that shape, applied to a new commodity. The score is Silicon Data’s daily index. The position is a contract on any listed maturity. The edge is whatever curve mispricing, basis dislocation, or directional view you’ve measured against an honest backtest before you press deploy. Same discipline that kept the TSFM ensemble out of UEFA production applies in this market too — the loss function here is just denominated in dollars per GPU-hour.

Short take: the announcement is a milestone, not because the CME contract itself will be wildly successful in year one — it’ll probably trade modest volume, dominated by macro flow — but because it locks in Silicon Data as the reference index and clears the path for the perpetual that follows on a crypto venue. That second product will be where the volume eventually lives, the same way it played out for BTC, ETH, and every other underlying that found its way onto perp rails after starting in dated-futures form.

Watch Hyperliquid.

— S.