← Latent Minds Institute Economics strand · originally published under MO3 Research
Working Paper · No. 2026-07Latent Minds Institute

A Forward Market for GPU Compute:

Contract Design, Settlement & the Term Structure of Silicon

Draft · Tinker FreelyAuthor: Muhammad Zane Abdullah
01Why the market is missing 02The settlement index 03Contract design: build a ticket 04The term structure of compute 05Hedging a Neocloud book 06Margining & counterparty credit 07Bootstrapping liquidity
01

Why the market is missing

Roughly $11T of cumulative AI capex will be deployed by 2029, most of it debt-funded, against an asset whose rental price has no forward curve, no hedge, and no public mark.

Today, GPU rental prices are set in bilateral, confidential negotiations. A lender financing a cluster at 75% loan-to-value is short a five-year strip of compute prices with no way to observe, let alone hedge, that exposure. The result is a market that only clears one deal template: the 5 year, take or pay offtake backstopped by an investment-grade hyperscaler. Everyone else is rationed out, including inference providers who refuse to sign beyond 1 year and VC-backed labs who want short bursts of large capacity.

Vendor backstops (Nvidia's take or pay guarantees to Neoclouds) are a bridge: a central bank supplying liquidity until private markets can price risk themselves. The permanent solution is the same one every other capital intensive commodity found, from oil and power to memory and freight: a standardized derivatives market that separates price risk from operational risk, so each can be held by whoever prices it best.

The claim of this paper: a cash-settled forward on an hourly GPU rental index is sufficient to (a) let lenders hedge collateral value, (b) let Neoclouds run short-tenor books against long-dated debt, and (c) let compute buyers lock cost without locking capacity to one provider.
02

The settlement index

Derivatives need a reference price nobody can push around. We propose CIX, a family of daily indices per SKU (e.g. CIX-GB300), computed as a volume weighted trimmed mean of reported bilateral transactions, with three defenses against manipulation:

DefenseMechanismAnalogue
TrimmingDiscard top/bottom 10% of prints by price before averagingLIBOR reform / SOFR
Volume floorPrints below 64 GPUs or 720 GPU-hours excludedPlatts eWindow MOC
AttestationReporters are audited providers; misreporting forfeits index membership & fee rebatesExchange self-reporting regimes

Below, tinker with a toy version of the index: inject a manipulative print and watch how trimming absorbs it.

Index engine · CIX-GB300 (toy)LIVE
Naive VWAP
...
Trimmed index print
...
Manipulation impact
...
Cost to attacker/day
...

Attacker cost = volume × |manipulated price − fair price| × 24h. The trim makes moving the print expensive; attestation makes it terminal.

03

Contract design: build a ticket

The core instrument is a cash settled monthly forward: at expiry, the parties exchange the difference between the agreed forward price and the average of the daily index over the delivery month, multiplied by notional GPU-hours. No GPUs move. Cash settlement is the whole trick. It means a hedge fund, a bank, or a pension can take compute price risk without ever racking a server, and that is where real liquidity comes from.

Configure a trade and the ticket writes itself:

Trade builder
Forward price (model)
...
Notional value
...
P&L per ±$0.10 move
...
Initial margin (section 6)
...

Forward price is read off the live curve in section 4. Change the curve there and this ticket reprices.

Contract spec choices that matter: monthly average settlement (not a single day) kills expiry-day manipulation; per-SKU contracts with a published cross-SKU performance ratio lets old-generation contracts keep trading as hardware ages; the 730 GPU-hour lot size equals one GPU-month, small enough for inference startups and stackable for hyperscale.
04

The term structure of compute

Compute should trade in backwardation: forwards below spot, because everyone knows next year's silicon is cheaper per FLOP. The interesting question is how much.

We model the forward curve with three parameters. Depreciation drift (δ) is the expected annual decline in rental price as new SKUs land, the compute analogue of contango's storage cost with the sign flipped. Scarcity premium (χ) is a convenience-yield-like kink at the short end: when chips are the binding constraint, near-dated compute trades rich. Demand shock (s) tilts the whole curve when a frontier model release or a datacenter delay reprices everything at once.

Forward curve · CIX-GB300 · $/GPU-hrDRIVES section 3 & section 5
12m forward
...
36m forward
...
Curve shape
...
Implied 1y carry
...

F(t) = S·e^(−δt)·(1 + χ·e^(−3t))·(1 + s). Dashed line marks spot. The steeper the backwardation, the more a Neocloud earns for holding physical capacity and selling it forward: the "curve trade" that today has no instrument.

The 36-month forward is a market-cleared estimate of GPU residual value. Today, lenders guess residual value from depreciation schedules; with a curve, it becomes a tradeable number. Loan-to-value ratios, DSCR covenants, and lease pricing can all key off it.

05

Hedging a Neocloud book

A Neocloud finances a 6-year cluster with debt but rents to inference customers on 1-year terms. It is structurally long compute prices, a maturity mismatch identical to a bank borrowing long and lending short. The fix is identical too: sell forwards against the unhedged tail of the book.

Below, a 1,024-GPU cluster with $2.9M/yr of debt service. Choose how much of years 2 through 6 revenue to pre-sell in the forward market, then shock the market and watch who survives.

Hedge simulator · 1,024 × GB300 · 6y horizonUSES section 4 CURVE
6y revenue, hedged
...
6y revenue, unhedged
...
Worst-year DSCR (hedged)
...
Worst-year DSCR (unhedged)
...

Bars show annual revenue. Solid is the hedged book, outline is unhedged. The red line is debt service. Lenders require worst-year DSCR of at least 1.30×. Find the hedge ratio that keeps the covenant alive at a −40% path: that number is what replaces the Nvidia backstop.

Note the symmetry: in an up-market the hedged Neocloud gives away upside, which is exactly the revenue share it pays Nvidia under today's backstop. The forward market prices that same protection competitively instead of bilaterally.
06

Margining & counterparty credit

A derivatives market dies the first time a big counterparty defaults on an unmargined position. We adopt the standard central-counterparty (CCP) stack: initial margin sized to a 5-day, 99% price move; daily variation margin in cash; a mutualized default fund behind that. GPU prices are young and violent, so IM will start punitive. That is fine; it compresses as realized vol history accumulates.

Margin engine · daily VM on your section 3 ticketSIMULATED PATH
Initial margin
...
Worst daily VM call
...
Margin breaches (>IM)
...
IM as % notional
...

Grey path: daily mark-to-market of your ticket. Colored bars: variation margin calls. Band: ±IM. If daily moves routinely breach the IM band, the clearinghouse raises IM. Crank volatility and watch capital efficiency collapse. This is why the index (section 2) matters: no trustworthy daily mark, no daily VM, no market.

07

Bootstrapping liquidity

Markets are two sided. The sequencing that has worked historically (power, freight, memory) maps here:

PhaseInstrumentNatural longsNatural shortsExit criterion
0 · IndexPublished daily CIX per SKUn/a (data product; sell to lenders & LPs)20+ attested reporters
1 · OTC deskBilateral cash-settled forwards, house-intermediatedInference startups locking costNeoclouds pre-selling capacity$500M gross notional
2 · ClearedStandardized monthlies, CCP margining+ hedge funds, AI-token specs+ lenders hedging collateralOpen interest > 2% of physical
3 · OptionsCalls/puts on CIX; floor sales replace vendor backstopsNeoclouds buying putsVol sellers, insurersn/a

The startup wedge is Phase 0-1: the index is a subscription data business with immediate buyers (every lender in a $7T credit market needs a mark), and the OTC desk monetizes the matching problem that already exists between 1-year renters and 6-year financiers. The clearinghouse is the venture-scale prize, but the index is the moat. Whoever owns the settlement price owns the market.

The deepest consequence: once puts on CIX trade, Nvidia's backstop becomes a market-priced instrument anyone can sell: insurers, funds, even sovereigns. The central bank of AI gets to retire, which is precisely what a central bank is supposed to do.