CoreWeave‘s 9.75% coupon bond is now yielding 15.78%, and the cost of insuring its debt sits around 1,038 basis points, above the 1,000bp line the market calls distress. Those two numbers are a neocloud credit spread: the market pricing the counterparty risk inside your own three-year GPU reservation.
The diligence you already run, security posture, SOC 2 reports, reference checks, all answer whether a vendor can perform. A three-year reservation raises a second question: whether the vendor will still be operating in year three.
By the end you will be able to separate a coupon from a price and a price from a yield, and read the result as the credit signal behind the procurement decision.
Why does a GPU provider’s balance sheet show up in your procurement decision at all?
Because a three-year reserved-capacity contract is unsecured exposure to the provider for its full term. You are extending credit to the operator for three years, and size, brand and a large backlog do not repay a loan.
Credit agreements in this market often require the borrower to hold customer contracts sufficient to cover its repayments, so your contract is an input into the loan.
CoreWeave’s backlog is over $100 billion, and it is heavily concentrated: Microsoft alone made up an estimated 62% of revenue in 2025, and the three largest customers are expected to pass 85% of revenue in 2026. The company still carries more than $51 billion of debt and obligations. A backlog is a promise of future revenue, and when that promise is concentrated in three customers, lenders are really underwriting those three balance sheets.
Neocloud is the credit market’s label for a GPU provider. Your workloads sit above its credit agreement, and loan terms here are frequently cross-triggered, so a breach on one facility cascades to the rest. Your service continuity is tied to facilities you have never seen, and assessing provider failure risk before signing is the part vendors rarely volunteer.
Goldman Sachs counted $489 billion of AI-related debt issued in 2026 by late July, above the $322 billion for all of 2025, while the first facilities near maturity and the 2023-24 contract vintage rolls into renewal. You are not being asked to run a credit model, only to recognise one exists. That model runs on four numbers, and here is how to read them.
What does a 1,038 basis point credit spread actually measure, and why is 1,000bp called distress?
A credit spread is the extra yield, in basis points, an investor demands to hold a corporate bond over a matched-maturity Treasury. It is a price for default risk. At 1,000 basis points the implied default probability stops being a tail scenario, so the market calls it distress.
One basis point is 0.01%. A spread is the corporate yield minus the Treasury yield of the same maturity, the part of the return that is about the borrower.
The trap is coupon versus yield. CoreWeave’s 9.75% coupon was fixed at issuance and does not reset; the 15.78% is the yield to maturity, what the market now demands to hold the paper to 2031. Bond yields move inversely to prices, so the bond trades at a discount and the coupon did not change when the price fell.
A credit default swap is a second, independent price for default probability: the annual cost, in basis points of notional, of insuring against non-repayment. CoreWeave’s five-year CDS last stood at around 1,038 basis points, having touched above 1,140 intraday, up 16% week on week and 164% year on year. A spread of 1,000 basis points or above is generally a sign of credit distress, according to Gavan Nolan of S&P Global Market Intelligence. A 164% year-on-year move means the market’s assessment changed.
One name can move the whole category. Oracle’s five-year CDS rose to 282 basis points, up about 10% in a week, and S&P cut it to BBB-, one notch above high yield. Investment-grade-only mandates must sell on downgrade, so a rating action becomes forced selling, which raises the cost of capital for everyone in the group. Oracle carries $638 billion of remaining performance obligations against roughly $18 billion of project debt after issuing a force majeure notice on the 2.45GW Project Jupiter campus.
The Bank for International Settlements watches because AI-sector financing terms are poorly disclosed, the same asset can be pledged more than once, and the financing is circular, with chip makers taking stakes in the labs and neoclouds that buy their chips. That opacity compounds the vulnerability. The spread is pricing the collateral behind those loans as well as the borrower.
Is buying compute from an operator the same bet as holding its debt?
Roughly yes, with one difference that matters. A bondholder is paid behind the secured lender but ahead of the equity holder, and can sell, hedge with a credit default swap and diversify across issuers. A compute buyer sits behind the lender’s claim on the hardware, cannot exit quickly and cannot diversify. You are exposed to the same cash flows, but your position is worse.
Both positions are exposed to the same cash flows, whether the capacity gets built and whether the customer keeps paying. The question for a buyer is which risks travel with the contract.
The bondholder has three exits: sell, buy protection, or spread exposure across issuers. Your position is illiquid, undiversified and entangled. Data gravity, egress charges and workload integration make moving expensive, and you cannot split a training run across four providers to diversify. In a bankruptcy you cannot walk away either: termination clauses tied to a filing are generally unenforceable.
So here is the inversion. If you would not lend a provider money at 15.78%, do not accept a three-year rate card from it without pricing that number in. The spread is the market’s estimate of the discount you should demand. What it is pricing is ownership structures, collateral and the refinancing wall.
The 15.78% yield is the credit market pricing the position you already hold, and the number it demands of a bondholder is the number you should be charging on your rate card.
Every multi-year reservation is a credit position, and the yield is the market showing its work. Read the full neocloud credit picture, starting with who actually owns the GPUs behind the spread.
Frequently Asked Questions
Does a 1,038 basis point spread mean my GPU provider is going to fail?
No. A spread is a probability, not a prophecy. It says the market now assigns a materially higher chance of non-repayment than it did a year ago, and prices that chance into every dollar of the provider’s debt. A name can trade at distressed levels for quarters and still meet its obligations, or refinance and recover. What the number removes is your right to assume the risk is small.
Is a 1,038 basis point credit spread the same as a 1,038 basis point yield?
No, though they are related and often quoted together. A spread is the excess yield over a matched-maturity Treasury, so it isolates the part of the return that is about the borrower. A yield is the total return from holding the bond to maturity, Treasury base included. A 15.78% yield and a 1,038bp spread describe the same stress from two angles, but they are not interchangeable numbers.
Why would a GPU provider with a $103.7B backlog have to pay 15% to borrow?
Because a backlog is a promise of future revenue, not cash in the bank. CoreWeave’s $103.7B backlog is 72% concentrated in three customers, so the market is underwriting those three balance sheets rather than the headline total. Lenders also know the revenue only arrives if the capacity gets built and the customer keeps paying. Backlog size tells you about demand. It says nothing about the cost of financing it.
Why is everyone talking about AI data centre debt now?
Three things landed at once. Goldman Sachs counts $489B of AI-related debt issued in 2026 to date, the 2023 and 2024 contract vintage is rolling into renewal, and the first cohort of facilities is approaching maturity. Add thin disclosure about what collateral is pledged where, and the Bank for International Settlements starts flagging AI lending at roughly a quarter of recent corporate credit growth.
What actually happens to my workloads if my GPU provider files for bankruptcy?
The cluster usually keeps running at first. A filing triggers an automatic stay, and the operator’s incentive is to preserve revenue while it restructures or sells. What you do not control is what follows: contract rejection, an asset sale, or a wind-down that hands you back cold hardware on someone else’s timetable. Continuity of service is not a term you hold. Plan the exit before you need it.
Is a multi-year reservation really riskier than on-demand pricing?
Yes on credit, no on price. On-demand leaves you free to walk, but it exposes you to spot rate spikes at exactly the moment a provider is under stress, and it guarantees you no capacity. A reservation buys certainty of supply and a lower rate in exchange for a locked exposure to one balance sheet. The trade is only rational if the discount covers the credit risk you are taking.
Can I hedge GPU provider counterparty risk the way a bondholder hedges with credit default swaps?
Not really, and that asymmetry is the point. A bondholder can sell the paper, buy protection or spread exposure across issuers. A compute buyer can do none of those things at speed: data gravity, egress charges and workload integration make moving expensive, and you cannot split a training run across four providers to diversify. Contract terms are the closest thing you have to a hedge, so negotiate them like one.
What should I ask a GPU provider for before signing a three-year reservation?
Ask for the documents a lender would read: the latest audited accounts, the debt schedule and covenant list, any rating agency reports, and a plain answer on which facilities your contract sits inside. Then ask what happens to your reservation in a sale or restructure, who can assume the contract, and what notice you get if service stops. Slow answers are themselves an answer.
What is a take-or-pay clause and does it survive a bankruptcy?
A take-or-pay clause commits you to pay for reserved capacity whether or not you use it, which is what makes the contract bankable for the lender. In bankruptcy it becomes an executory contract: the provider can assume it, reject it, or assign it to a buyer. Rejection does not erase your claim, but it turns your service into a damages claim sitting behind the secured lenders.
What does a credit rating downgrade actually do to my provider?
It can raise the cost of capital within days, because investment-grade-only mandates are forced to sell on downgrade. Oracle is the worked example: 282bp on its five-year CDS, a BBB rating under pressure, and $638B of remaining performance obligations against $18B of project debt after force majeure on the 2.45GW Project Jupiter campus. Forced selling lifts yields, which lifts funding costs across the category.
How often should I re-check a provider’s credit signal after signing?
Quarterly at minimum, and immediately after any rating action, refinancing or covenant waiver. This is a fast-moving number: CoreWeave’s CDS widened 16% in a week and 164% in a year, which means an assessment formed at signing can be stale within a quarter. Put the spread, the rating and the maturity wall on the same review cycle as your capacity plan, and set a threshold that triggers a conversation.
