ResearchTrading

Compute credit is tightening. Residual value hedges are here for borrowers and lenders.

Compute credit is tightening, and the unpriced residual now sits in everyone's cost of capital. A floor on the compute rate prices it, for new deals and existing books.

Four market figures. Oracle: its five-year default protection widened 30 basis points in one week, to a record near 300bp. CoreWeave: its June bond went from a 9.625% coupon at issue to a 16% yield at the October peak. About $450 billion of AI-related debt this year, on the same handful of offtakers. And the new underwriting question: what is the hardware worth if it comes back?

If you’ve been told your advance is lower than in August, that you need more equity, or that funding is harder to find, you’re not alone. If you’re a lender wondering how to hedge the residual value of a GPU book, you’re not alone either.

The cost of insuring the debt of Oracle, an investment-grade hyperscaler, rose about 30 basis points in a week to a record, with Amazon, Alphabet and Meta at record wides too, and CoreWeave's June bond, issued at 9.625%, has traded as high as 16%. As Stanley Lee wrote in The compute credit cycle just turned, lenders now underwrite the operator and ask what the hardware is worth if it comes back. There is no market answer yet, so that unknown sits in everyone's cost of capital. A floor on the compute rate prices it, on new deals and on books already written.

How the residual raises everyone's cost of capital

Four steps, left to right. Step 1, rates and spreads widen: Treasuries are above 5%, and AI credit trades at record wides. Step 2, highlighted as the step the floor fixes: the residual has no mark, because no one can price the hardware if it comes back, or hedge it. Step 3, lenders charge for the unknown: wider spreads, lower advances and tighter covenants. Step 4, costs rise for everyone: operators post more equity, buyers pay more, and existing books lose value.

The crunch puts hedgers on both sides of the same rate

Stanley's note called this cycle healthy, and this is part of why: the same squeeze that makes lenders, lessors and operators want a floor makes buyers want a cap. A hedging market needs exactly that, two groups who each want the other's risk.

One rate, two sides, split by the LCI compute rate in dollars per GPU-hour. If rates rise, buyers pay more: compute buyers buy a call, which caps the budget if suppliers cannot close their financing and capacity gets scarce. If rates fall, the asset is worth less: lenders buy a put to cover the gap between re-let recovery and the loan, lessors buy a put to protect the book residual at lease end, and operators buy a put to floor capacity rolling off contract and post less equity.
Fig. 1 — One rate, two sides

Each hedge references the LCI for its own chip: A100, H100, H200, B200 or B300.

A floor moves the advance from 65% to 85%

A floor caps what a lender can lose if compute rates fall, so it can lend more. On an illustrative $15M cluster, it lets the lender advance 85% instead of 65%. The operator puts in $3.0M less equity.

Two charts. Left, with no numbers: what a lender recovers on a defaulted loan, against the compute rate when the hardware is re-let. Without a floor, recovery falls with the rate. With a floor, it stays flat below the strike, and the shaded area between the two is what the floor pays. Above the strike both lines rise until the loan is repaid. Right, funding an illustrative $15M cluster: without a floor, a 65% advance, $9.75M of debt and $5.25M of operator equity; with a floor, an 85% advance, $12.75M of debt and $2.25M of equity.
Fig. 2 — The floor caps the lender's downside, so it can advance more

Illustrative. The put covers 40% of the loan, struck out of the money. Premiums are set by the liquidity provider for each deal.

For the operator, the floor pays for itself: cheaper funding saves about $0.45M a year, and the floor costs about $0.17M a year ($0.51M upfront over three years).

Run your numbers, send the ticket

Make a free account on liquidcompute.com and turn your advance, residual, breakeven or budget into a quotable ticket in the Hedge Converter. Our primer, Hedging compute rates, walks through all four.

Hedge one deal or a whole book: a deep out-of-the-money put sized to a portfolio of GPU loans or leases caps the loss in a sharp drawdown for a small running cost, the way a credit fund buys index protection.

No premium? A forward or swap locks the rate at no upfront cost, in exchange for the upside.

Talk to us

If a financing has stalled on the advance, the residual or the equity check, a floor is usually the missing piece, and we can get one priced in days. If you're a lender with a GPU book, we can put a mark on it and a hedge around it. Either way, write to lending@liquidcompute.com while capital is still available.

Sources: Oracle and hyperscaler default protection, Reuters, “AI credit worries send default insurance to record high,” October 15, 2026, citing LSEG and S&P Global Market Intelligence. CoreWeave, Treasury and AI-debt figures, Stanley Lee, The compute credit cycle just turned, Liquid Compute, October 14, 2026.

Published by Liquid Compute Research for informational purposes only. This note is not an offer, solicitation, recommendation to lend, arrange financing or transact, and is not investment, legal, tax or accounting advice. The options described here are over-the-counter transactions between the hedger and third-party liquidity providers, available only to eligible contract participants, and are not listed or cleared on any exchange. Liquid Compute is never a counterparty to these transactions and does not take positions on its own index. Examples are illustrative, not real trades. The derivatives exchange is pending CFTC designation as a DCM and registration as a DCO.

Written by Zachary Frederick, Senior Research EngineerFollow on LinkedIn