If you wanted to know what a three-year-old H100 server is worth, you’d expect a number to exist. It doesn’t. The trackers that follow used GPUs disagree, spanning roughly 45% of the new price (Hashrate Index) to 69% (Silicon Data), and neither figure is an actual sale.
Depreciation schedules and residual-value guarantees exist to name that number and back it. They don’t agree with each other either.
This article walks through the three places a value gets asserted: the credit agreement, the depreciation schedule and the residual value guaranty. Each one declines to price the silicon. By the end you’ll be able to say what a three-year-old GPU is worth, who guarantees it, and why a long reservation is the only assertion you can lean on, leaving what the collateral is really worth for you to decide.
What is GPU-backed debt, and how does a lender secure a loan against hardware a competitor can replace in eighteen months?
GPU-backed debt is a loan secured by GPU servers and the assigned customer contracts that pay for them, held in a bankruptcy-remote special-purpose vehicle. The lender underwrites contracted cash flow, because a competitor can replace the silicon in eighteen months.
The format began with CoreWeave‘s August 2023 $2.3B delayed-draw term loan, co-led by Magnetar and Blackstone, at a 70% advance rate against more than $3B of H100 servers, priced at 11% to 14% all-in, or 600 to 870 basis points over SOFR, as the structure has been documented. That supports $2.1B of debt, amortising to zero before anyone sells a used GPU.
CoreWeave’s filing states the mechanism: the facilities are collateralised with the pledged contractual cash flows, generally from investment-grade counterparties.
Two spreads sit in those numbers, and separating them matters. The 600 to 870 basis point spread belongs to the 2023 loan. The 225 basis point spread is a later, investment-grade facility, CoreWeave’s March 2026 $8.5B print borrowed against a hyperscaler, while weaker customers on the same silicon pay SOFR plus 450 to 550. The spread ladder is the tell.
What the agreement omits is the point. A term search finds “GPU” dozens of times in definitions but zero instances of appraisal, residual value, advance-rate reset, orderly liquidation or remarketing, as one close reading shows. The contract is the lender’s protection.
For a listed borrower the agreement sits as an exhibit to its 10-K and 10-Q filings; for most private operators it isn’t public, which is itself a diligence answer. The clauses that decide whether the hardware keeps running are cross-default, borrowing base, amortisation, change-of-control and enforcement mechanics.
The collateral question is answered by who is paying.
Why is depreciation running at roughly half of revenue, and why does six years versus four matter to a buyer?
The lender declines to price the chip, so the only number that looks like a price is the accountant’s.
A six-year fleet books an annual charge of one-sixth of gross cost, and at current capital intensity that lands near half of revenue: CoreWeave recorded $1.4B of depreciation on $2.575B of Q2 2026 revenue, Nebius $259.7M on $582.3M. Six years rather than four lowers the charge, lifts operating income and supports a higher valuation.
That ratio is a function of capital intensity, and fleet depreciation at about half of revenue is the industry’s default.
Both schedules sit within accounting rules; only one is likely to match the market. Nebius wrote servers down over four years, then extended to five in January 2026. Sixteen of seventeen public schedule changes since 2020 lengthened lives.
H100 rental rates went from about $7 to $10 an hour in early 2024, to $1.70 by October 2025, and back to $2.35 by March 2026, a round trip no schedule anticipated.
The secondary market is too thin to settle it. CCIR‘s three-year record totals 3,009 listings, 10,911 units and $26.3M, while Manheim processes five million-plus used-car sales a year, so the residual value assumption is unfalsifiable in public data. The thinness is the finding.
For your business the stakes are the same as the lender’s. A multi-year rate card fixes a price against hardware whose market rate may fall below the contracted rate before the term ends. You are buying the provider’s ability to keep serving at that fixed price after the market has moved.
Benchmarking exists, within limits. Silicon Data runs a daily H100 index with a 36-month curve, and ICE and CME list cash-settled H100 and B200 futures, with Ornn and Compute Exchange offering quotes without custody or clearing. Treat a forward curve from a young market as a reference point only, since two indices can disagree, recently $2.63 against $3.57.
With no market deep enough to settle the schedule, the natural question is who will put a floor under the value.
What is a residual value guaranty, and why guarantee the substation but carve out the chips?
A residual value guaranty promises a floor value at a future date, turning an uncertain asset into something a lender can size. NVIDIA’s $105B instrument on the 4.25GW SB Energy campus for OpenAI covers data-centre, power and transmission costs but excludes its own equipment, defined as guarantor property. The vendor guarantees the shell and leaves the chips out.
Disclosed in August 2026, the filing sets a narrow trigger: OpenAI insolvency or missed rent, and NVIDIA covers the shortfall between a guaranteed minimum value and whatever a re-let or sale recovers. That minimum value is defined as data-centre, power and transmission costs.
The floor that determines what NVIDIA actually owes isn’t disclosed, so the $105B cap sits above a private floor. NVIDIA declines to put a number on the chips.
The exclusion is rational: a shell and substation hold value for decades, accelerators don’t. Guaranteeing your own product’s resale value means betting against your own next generation.
NVIDIA’s $6.3B guarantee of CoreWeave’s unsold capacity through 2032, its $500B platform programme with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR, and Google’s TeraWulf and Cipher backstops all protect the financing. Broadcom’s support on the Anthropic vehicle does the same. None promises a running cluster in year four.
The guaranty is credit support for a lender. For a buyer, the question is what happens when the guaranty isn’t enough.
Conclusion
Nobody in the chain will underwrite the chip itself: the lender secures the contract, the accountant picks a number, and the guarantor backs the shell and power while carving out the accelerators. So who guarantees a three-year-old GPU? No one.
What you can lean on is a multi-year reservation, which fixes the depreciation assumption in place and prices the provider’s ability to keep serving at a fixed price. That is the wider neocloud credit picture in a sentence, and why the loan outlives the hardware is the next question.
Frequently Asked Questions
What is a three-year-old H100 server actually worth today?
There is no published price, so the honest answer is a range. Going-concern appraisals put a three-year-old H100 at roughly 50 to 70% of new, while orderly-liquidation estimates sit nearer 30 to 50%. The depreciation schedule’s face value is a third number again, and it is a bookkeeping figure rather than a market one. Nobody trades enough volume to settle which is right.
If NVIDIA guarantees the residual value, am I protected?
Not on the hardware. NVIDIA’s $105B instrument on the SB Energy campus covers the data centre, power and transmission costs, and expressly excludes its own equipment as Guarantor Property. What it protects is the financing stack, not your cluster. If the accelerators lose value faster than the schedule assumes, the guaranty does nothing for the tenant paying a fixed rate in year four.
Which depreciation schedule is right, six years or four?
Neither is demonstrably right, and that is the point. Both sit inside accounting rules, but on identical hardware six years lowers the annual charge, lifts reported operating income and supports a higher valuation. Sixteen of seventeen public schedule changes since 2020 lengthened useful lives. With a secondary market too thin to validate either assumption, the schedule reflects a management choice rather than a measured resale forecast.
Where can I track H100 and B200 rental rates and forward curves to benchmark a quoted multi-year rate?
Start with Silicon Data’s daily H100 index, which carries a 36-month curve, then cross-check against cash-settled H100 and B200 rental futures on ICE and CME, plus reference quotes from Ornn and Compute Exchange. Treat all of it as a reference point rather than a forecast: the market is young and thin, and two indices can disagree on the same hour, recently $2.63 against $3.57.
Where do I find a provider’s credit agreement, and what should I read in it?
For a listed borrower, the credit agreement sits as an exhibit to its 10-K or 10-Q filings. For private operators it usually does not exist publicly, which is itself a diligence answer. Read the cross-default clause, borrowing base and concentration limits, amortisation schedule, change-of-control provisions and enforcement mechanics. Those determine whether the hardware keeps running, not the appraised value at origination.
What is an advance rate, and why does 70% matter more than the GPU model?
An advance rate is the share of appraised value a lender will fund, so 70% against $3B of H100 servers supports roughly $2.1B of debt. It matters more than the GPU model because the borrowing base tests contracted cash flow, not resale value. That is why the spread tracks offtaker quality: SOFR+225 against a hyperscaler, SOFR+450 to 550 against weaker customers on the same silicon.
Who actually owns the GPUs if my provider is only a tenant?
In most leveraged structures the hardware sits in a bankruptcy-remote special-purpose vehicle, with the provider as borrower and operator rather than outright owner. Your contract is typically assigned to that vehicle, which is why the lender cares about your payment stream and not your workload. It also means a provider restructuring can leave your cluster running while ownership quietly changes hands above you.
If my GPU provider fails, does the lender seize the hardware?
Usually the lender keeps the contracts and the cash flow, not the racks. Because the collateral is the assigned offtake agreement inside an SPV, enforcement tends to mean a new operator stepping in to service the same customers, or a sale of the vehicle as a going concern. An orderly liquidation of individual servers is the outcome nobody wants, because the market cannot absorb them.
Is a long GPU reservation just a way to lock in a cheap price?
No, and framing it that way is the common mistake. A multi-year rate card fixes a price against hardware whose market rate may fall below that price before term ends. What you are buying is the provider’s ability to keep serving at that fixed price after the market moves, which depends on their capital structure rather than their generosity at signing.
Can I buy used H100s instead of renting, and would that be cheaper?
Possibly cheaper on sticker price, rarely cheaper on total cost. The used market is thin, with CCIR recording 3,009 listings, 10,911 units and $26.3M across three years, so finding matching configurations at scale is hard. You also take on power, cooling, networking, spares and refresh risk. A three-year-old card with no vendor support is a capital asset with an expiry date.
Why do two providers quote different rates for identical H100s?
Because you are not buying silicon, you are buying a credit. The same H100 can carry SOFR+225 behind an investment-grade hyperscaler and SOFR+450 to 550 behind a weaker customer, and that spread flows into the rate card. Power contracts, utilisation, financing cost and the remaining depreciation schedule vary too. Identical hardware is the least interesting variable in the quote.
Does GPU depreciation affect the price I pay, or only the provider’s accounts?
Both, eventually. The schedule sets the annual charge the provider must recover through the rate card, so a shorter useful life tends to mean a higher required rate today. It also signals how much runway the provider has before refinancing. When the market rate falls below the contracted rate, the depreciation assumption is what determines whether the provider can keep serving you.
