How do lenders judge GPU collateral, and how does GPU rent compare with the cost of owning the chip?

Short answer: A lender asks whether a GPU can rent for more than it costs to own, and how far rent can fall before it stops covering even the running costs. Glitch in the Credit measures the first with the collateral cushion: the rent for one GPU-hour divided by the full-cost floor, what an hour of rent has to earn to pay off the chip, its power and its data center space over a normal life. Below 1×, owners lose money on every hour rented. The cash-cost floor is power, space and running costs alone: below it, renting out a chip that is already paid for no longer covers what it costs to keep running. On Oct 30, 2026, in our mid case, H100 costs $2.34 an hour to own and $0.76 to run and B200 costs $3.03 an hour to own and $1.05 to run. Every input is open data: collateral.json, CC BY 4.0. It reports what the numbers show. It isn’t a forecast, and it isn’t advice.

What does a lender need to know about a GPU loan?

The chips are the collateral, so the questions are about the chip’s earning power:

• Does rent cover the full cost? That is the cushion. Above 1×, an hour of rent pays back the chip, its power and its space over a normal life.
• How far can rent fall? Owners usually keep renting below full cost, because the chip is already paid for. The cash-cost floor says where that stops.
• What is the chip worth if the borrower stops paying? This is the part we don’t model. We take resale value as zero, which makes our full-cost floors a little higher than they would otherwise be. Resale prices, loan terms, the borrower’s contracts and how much of the chip’s life is used up are loan-level facts that no public index shows.

What does the cushion look like today?

On Oct 30, 2026, rent against our mid-case costs. “Over cash” is rent divided by the cash-cost floor.

What goes into the floors?

The full-cost floor is what one rented hour must earn: the chip’s price spread over its life at the cost of capital (a loan-style yearly payment), plus the power it draws, plus its space in a data center, plus other running costs, all divided by the hours it is actually rented. The cash-cost floor leaves out the chip’s price, which is already spent.

In dollars per GPU-hour, mid case, from the inputs dated and sourced in collateral.json:

What moves the cushion?

• Rent. The only input that changes daily. It is third-party, and it is what the cushion is scored on.
• The chip’s price. The biggest part of the floor. A dearer chip needs more rent.
• Useful life and the cost of capital. A shorter life or a higher rate raises the yearly payment. Our three cases move life, cost of capital and hours rented together, the way a stress test would.
• Hours rented. Power and space are paid for every hour, rented or not, so an idle hour raises the cost of each sold hour.
• Power price, cooling overhead and space. Smaller parts, but they set the cash-cost floor.

The inputs, shown as cheap / mid / dear:

Useful life is not settled. Lives that companies have reported for their servers: CoreWeave 6 years (attributed, source); Meta 5.5 years (verified, source); Amazon 5 years (verified, source); Nebius 5 years (attributed, source). Each is a line in collateral.json.

How do I read the cushion?

• Well above 1×: rent covers the full cost with room to spare. Rent has to fall a long way before the full cost is uncovered.
• Near 1×: rent just covers cost. A small fall in rent, or a shorter useful life, puts the chip under water on a full-cost basis.
• Below 1× but above the cash cost: owners lose money on the chip but keep renting it. This is the zone to watch for borrowers with debt to pay.
• At the cash cost: rent only covers power, space and running costs. Nothing is left to pay the lender.

The collateral cushion feeds one of our 17 signals, GPU rental. It is scored on fixed lines: the cash cost scores 100, 0.85× scores 75, 1× scores 50, 1.2× scores 25, 1.5× scores 0, with straight lines in between. The lines don’t move with the cushion’s own history. On Oct 30, 2026 the GPU rental signal is calm and the collateral layer scores 25 (calm).

What can’t this tell you?

• Rents are index prices reported in the news, not what a financed data center actually collected. Utilization, contract terms and discounts are not in them.
• The cushion needs a rent and a cost on the same basis. Index prices are for a standard rental; a long contract can rent for less or more.
• Cost inputs are market-level, not any borrower’s. Hours rented, power price and space cost are assumptions we sourced, not a loan’s numbers.
• It doesn’t include resale value, loan terms or the borrower’s other debts.
• Of the 29 input records in collateral.json, 16 derived, 10 attributed and 3 verified. “Attributed” means taken from secondary coverage; “derived” means our own step from a source.
• Our B200 chip price sits below some October 2026 reseller listings, so the B200 floors may be on the low side. This is noted in collateral.json.

Where can I check it?

• Inputs and floors, each with source and date: collateral.json and collateral.csv (CC BY 4.0; credit “Glitch in the Credit” with a link)
• The signal page, with every input listed: GPU rental
• The method: how we price GPU collateral
• Rents are third-party and are not offered as downloads. Get them from the indexes linked in the table above.

Follow the debt. Not a timing tool. Not financial advice.

Sources: Glitch in the Credit collateral inputs and floors, Oct 30, 2026 (collateral.json); The rent sources linked in the table above; Each input’s own source, linked in collateral.json.