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 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:
| Defense | Mechanism | Analogue |
|---|---|---|
| Trimming | Discard top/bottom 10% of prints by price before averaging | LIBOR reform / SOFR |
| Volume floor | Prints below 64 GPUs or 720 GPU-hours excluded | Platts eWindow MOC |
| Attestation | Reporters are audited providers; misreporting forfeits index membership & fee rebates | Exchange self-reporting regimes |
Below, tinker with a toy version of the index: inject a manipulative print and watch how trimming absorbs it.
Attacker cost = volume × |manipulated price − fair price| × 24h. The trim makes moving the print expensive; attestation makes it terminal.
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:
Forward price is read off the live curve in section 4. Change the curve there and this ticket reprices.
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.
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.
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.
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.
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.
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.
Bootstrapping liquidity
Markets are two sided. The sequencing that has worked historically (power, freight, memory) maps here:
| Phase | Instrument | Natural longs | Natural shorts | Exit criterion |
|---|---|---|---|---|
| 0 · Index | Published daily CIX per SKU | n/a (data product; sell to lenders & LPs) | 20+ attested reporters | |
| 1 · OTC desk | Bilateral cash-settled forwards, house-intermediated | Inference startups locking cost | Neoclouds pre-selling capacity | $500M gross notional |
| 2 · Cleared | Standardized monthlies, CCP margining | + hedge funds, AI-token specs | + lenders hedging collateral | Open interest > 2% of physical |
| 3 · Options | Calls/puts on CIX; floor sales replace vendor backstops | Neoclouds buying puts | Vol sellers, insurers | n/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.