Hedging the GPU balance sheet

01The exposure, before the instruments

A hedge program designed around instruments rather than exposures buys what is listed instead of what is owed. Decompose the book first.

02Decision one — how the hedge settles

Cash, physical, or cash with an EFP option. For a lender the answer is unusually clear once the default state is included in the analysis.

The default state is what elevates the EFP option from plumbing to program design. On enforcement, the lender becomes an involuntary physical holder: it owns racks it cannot operate, in a market that just fell — which is why it is enforcing. A pure cash hedge pays out and leaves the racks to be liquidated into a broker market at distressed spreads. An EFP-capable position lets the workout desk do what every physical commodity lender does with seized inventory: deliver it against the short futures leg, converting a fire-sale into a priced, scheduled exchange at a published basis. The crude, gold, and gas precedents in our EFP work all generalise: the physical option is worth the most to whoever holds unwanted physical — and a defaulted-on lender is the definition.

03Decision two — what unit to standardise on

Chip-level contracts are where liquidity will concentrate. Rack-level SKUs are what the collateral actually is. The hierarchy resolves the tension — hedge the benchmark, quote the config as basis.

Borrower fleets are heterogeneous in exactly the dimensions a liquid contract must ignore: SKU (H100/H200/B200/B300), memory configuration, interconnect (PCIe vs SXM vs NVL rack-scale), cooling, and location. A contract specified to rack level matches the collateral but fragments liquidity across dozens of thin instruments — the failure mode our hierarchy work exists to prevent. A chip-level benchmark concentrates flow but leaves config basis unhedged.

Two technical points from the measurement work belong in the standardisation decision. First, rack-scale systems change the energy content of a GPU-hour: measured duty factors run 0.795–0.880 on H100 SXM, and no equivalent public measurement exists for NVL-class racks — so a chip-hour hedge against rack-hour collateral embeds an unmeasured power basis. Second, the generation spread is not a detail: it is the collateral depreciation curve. A lender short the benchmark but long three-year-old Hoppers is long the Hopper–Blackwell spread whether it means to be or not. The two-factor structure from our curve work (scarcity + generation, hedge R² rising from 74.8% to 81.1% when the generation leg is added) is, for a lender, the difference between hedging revenue and hedging recovery.

04Decision three — whether the index can be trusted

Every candidate settlement index has strong design claims and zero published calculations. For a lender, index due diligence is credit due diligence — the hedge is only as good as the print it settles on.

Our IOSCO assessment across the four administrators reduces, for a hedging lender, to one uncomfortable sentence: the index with the best data claim (Ornn OCPI, transactions-only) carries the most settlement weight and has published no methodology, contributor set, or revision policy — and its competitors are no more disclosed. Principles 7, 9, and 11 are not abstractions here; they are the difference between a hedge and an unsecured bet on an administrator's judgment.

05The power leg — why ERCOT and Dominion books are different loans

Two identical fleets, one in ERCOT West, one in Northern Virginia, are not the same credit. The settlement data says so with numbers.

Dominion (DOM zone) example. Data-center alley's own zone carries the highest flat-block power price in our eight-location sample — $55.06/MWh over three years, against $32.92 at N. Illinois — and a winter-morning price structure: its most expensive average hour is 7 AM in winter ($118/MWh, p90 $231), not the summer evening. A NoVA borrower's opex line is structurally higher and its stress scenario is a January morning, when inference load (day-heavy, +2.1–5.2% shape premium at DOM) is exactly coincident. Underwrite the power passthrough on flat-block DOM and the model is wrong twice.

ERCOT example. The ERCOT hubs carry the largest shape premiums in the sample — up to +9.4% over flat-block at 40% utilization — and the premium peaks precisely in the ramp phase. For a lender this is not a curiosity: the ramp is the loan's riskiest period. A newly energised borrower is at peak leverage, minimum revenue, and — per the load-shape work — maximum exposure to the gap between the flat-block power hedge its PPA desk bought and the load-weighted price it actually pays. The covenant implication is direct: power-hedge requirements in credit agreements should be specified against the measured shape (or Nodal's hourly strip, listing 31 August), not a 7×24 block, and should tighten during ramp rather than after stabilisation.

06Three more decisions the program cannot skip

Beyond the principal three — briefly, because each has bitten a commodity lender before.

• Tenor and the stack-and-roll basis. Loan tenors run 2–5 years; the only live compute curve is Kalshi's weekly-to-quarterly ladder, and the pending futures will list short. The program is structurally stack-and-roll, which means it is short calendar-spread risk in a market with no calendar-spread history. Size the hedge ratio down for roll risk rather than pretending term coverage exists.
• Margin liquidity against illiquid collateral. A short hedge pays variation margin when rental prices rally — precisely when the collateral is appreciating but not liquefiable. The metals-lender margin spiral applies unchanged: reserve committed liquidity for hedge margin equal to a 2-sigma rally, or the hedge becomes the default trigger.
• Venue and counterparty regime. The same operator runs venues in two regulatory regimes on two indices; a program can diversify administrator risk and venue risk simultaneously by splitting the core short across the CFTC-regulated pair. Offshore perpetual liquidity is real but funding-rate basis and counterparty profile make it overlay, not core.

07The program, assembled