CoreWeave's $35.6 Billion Debt Stack: What Holds It Up, and Where It Bends

CoreWeave closed June 2026 with $35.6 billion of debt principal and a $104 billion contracted backlog. Both figures are real and both get quoted endlessly. The more useful disclosure is neither. It is the price ladder inside the CoreWeave debt stack, where the same company borrowed at SOFR plus 225 in March and SOFR plus 550 in August.

The short answer. CoreWeave funds GPU capacity with secured, contract-backed loans rather than corporate debt, so its cost of borrowing tracks the quality and length of each customer contract. At 30 June 2026 it held $35.6 billion of debt against $103.7 billion of contracted backlog. The stack is solvent while contracts hold and refinancing stays open.

What a neocloud is actually borrowing against

A neocloud is a cloud company with one product. It rents accelerated compute, and it has no advertising business or software licences to pay for the build.

That single fact sets everything else. A hyperscaler funds data centres from operating cash flow and treats the debt market as optional. A neocloud has to borrow, because the asset arrives years before the cash it produces.

CoreWeave spent $14.1 billion on capital expenditure in the first half of 2026 against $4.7 billion of revenue. That is close to 3 dollars out for every 1 dollar in, and it has to be funded from somewhere.

The instrument of choice is the delayed draw term loan, or DDTL. A DDTL commits a lender to fund in tranches as equipment is delivered, so the borrower pays interest on drawn money rather than on the whole facility.

What secures it is the part worth reading closely. The collateral is not really the silicon. It is the customer contract attached to the silicon, and the quality of that contract sets the price.

The stack, facility by facility

CoreWeave reported $35.6 billion of principal debt obligations at 30 June 2026. That splits into $31.4 billion of recourse debt, which the parent stands behind, and $3.7 billion of non-recourse debt ring-fenced inside financing subsidiaries.

The stack is not one loan. It is a series of separately rated, separately secured facilities layered over a base of senior unsecured notes and convertibles.

Terms come from CoreWeave's own 8-K exhibits at each closing. Facilities DDTL 1.0 through 3.0 are privately held and unrated, so their pricing is not public. This is the disclosed portion of the stack, not all of it.

Servicing this is now one of the largest lines in the business. Net interest expense reached $640 million in the second quarter, against $267 million a year earlier, and $985 million across six months.

The money has become cheaper per dollar borrowed. On the 11 August call, CFO Nitin Agrawal said CoreWeave had "reduced our weighted average cost of debt by almost 300 basis points" over the past year, worth roughly $1.1 billion of annualised saving at the quarter-end debt load. The rate came down. The balance went up faster.

The spread ladder is the real disclosure

Three secured facilities, five months apart, same company, same balance sheet, same class of accelerator. The spread between the cheapest and the dearest is 325 basis points.

The lender is pricing the contract, not the silicon

DDTL 4.0 was secured against previously contracted cloud services with a large enterprise customer, ran to a March 2032 maturity, and became the first investment grade rated financing backed by high performance compute. Moody's put it at A3. Blackstone Credit and Insurance anchored it.

DDTL 5.5 is a different animal. It backs customer contracts averaging about 3 years inside a facility maturing in about 5, and the structure carries an explicit option to renew those contracts or re-lease the capacity when the initial terms end.

Read that clause again. The lender is knowingly funding two years of unlet capacity at the back of the loan, and charging 325 basis points more than in March for the privilege.

Both founders said as much in public. Brannin McBee said "lenders are now comfortable financing shorter-dated contracts", and CEO Michael Intrator described the facility as opening "whole new markets". That is a real commercial win. It is also the first time somebody has priced CoreWeave's residual value risk in the open.

The rating ladder inside a single company

Most companies have one credit rating. CoreWeave has a range, and the range is wide enough to be the story on its own.

Seven notches, top to bottom

At the top sits DDTL 4.0 at A3, which is investment grade. Below it, DDTL 5.0 and DDTL 5.5 sit at Ba2 and BB+. Below those is the corporate family rating, Ba3 at Moody's and B+ at S&P. At the bottom are the senior unsecured notes, rated B1 by Moody's and B by S&P.

S&P attached a recovery rating of 5 to those notes, indicating 10% to 30% expected recovery in a default. That is structural subordination stated plainly by a rating agency.

Hold the secured paper and you own a contract with a named counterparty and a claim on specific hardware. Hold the bond and you are behind all of it. Who ends up holding which is the subject of the piece on private credit's exposure to AI infrastructure.

Near-term cash against near-term maturities

Leverage ratios are slow indicators. The faster one is the next twelve months of cash.

Add the first two rows and $7.5 billion comes due inside a year, against $5.5 billion of unrestricted cash. A single guided quarter of capex is larger than both combined.

This is not evidence of distress. CoreWeave has drawn capital repeatedly and at improving prices. It is evidence of dependence, which is a different thing. The model needs the financing window open every quarter, not most quarters.

The backlog is the collateral, and it has a tenor problem

The bull case rests on one number. Remaining performance obligations stood at $103.7 billion at 30 June 2026, up 246% year on year, with more than $25 billion of further commitments signed early in the third quarter.

Backlog of roughly 3 times total debt is a genuinely strong position, and it is why the secured facilities price where they do. The question is when it arrives.

Concentration is improving but still high. Two customers accounted for 63% of first-half revenue, at 40% and 23%. S&P noted the largest customer's share of backlog fell from 85% to roughly 35% over the prior year, which is real diversification rather than a rounding change.

One obligation sits outside the debt figure entirely. The company disclosed $35.5 billion of executed but uncommenced lease payments running to 2029. That is roughly the size of the debt stack again, for the accounting reason covered in the analysis of capex and off balance sheet commitments.

Three stress points, ranked by how soon they bite

These are not equally likely, and the loudest risk in the coverage is not the first one to matter.

1. The interest line grows faster than the operating line

Third quarter guidance puts revenue at $3.45 billion to $3.6 billion, adjusted operating income at $200 million to $260 million, and interest expense at $860 million to $940 million. At the midpoints, interest absorbs about a quarter of revenue.

Nothing here breaks. It does mean revenue growth has to outrun growth in the debt that produced it, quarter after quarter, with no pause.

2. Renewal is a more realistic failure mode than default

A take-or-pay contract protects the lender for its term. What happens after the term is a market question, and DDTL 5.5 is the first CoreWeave facility that puts that question inside the loan.

If a 3 year contract on 2026 hardware does not renew at a price that services 5 year debt, the loss shows up as a weak renewal rate rather than a missed payment. The residual value assumption underneath it is the one examined in the piece on GPU depreciation and reported earnings.

3. The financing window is a market, not a contract

Every facility above was priced into a receptive credit market. DDTL 5.0 tightened 50 basis points during syndication because demand was strong.

Markets change faster than data centres get built. A quarter where new issue spreads widen while $6.2 billion of current maturities needs refinancing is the scenario worth modelling, and it rhymes with the 2000 telecoms pattern rather than anything from the software cycle. That comparison is drawn out in the six metrics separating the dot-com bust from this one.

Where this argument is weakest

The case above reads the stack from the risk side. Here is the strongest version of the other one.

The credit market is moving against the bear case

CoreWeave has cut its weighted average cost of debt by almost 300 basis points in a year. It won the first investment grade rating on compute-backed paper. S&P revised its outlook to positive in April 2026 while affirming B+. Those are not the fingerprints of a deteriorating credit.

There is also a floor under part of the demand risk. Nvidia is obliged to buy residual unsold capacity under a $6.3 billion order form running to April 2032, subject to delivery and availability conditions. That is a backstop, and one more strand in the circular financing pattern.

What this analysis cannot see

The covenant packages are not public. Debt service coverage tests, cash sweep triggers and cure rights sit in credit agreements that were not filed in full, and those terms decide what happens in a bad quarter.

Ratings are opinions, not measurements, and the agencies currently disagree about this company by a full notch. Anyone claiming to know where the stack breaks is working from the same partial file as everyone else.

Frequently asked questions

Where to start this week

Two things, both quick.

If you buy compute from any neocloud, ask one question at renewal time. What is the maturity of the facility financing the capacity you sit on, and does it run past your contract? A provider funding a 3 year commitment with 5 year money has an incentive to keep you that no marketing page will mention.

Then put the next two CoreWeave quarterly reports in your calendar and read exactly two lines. Guided interest expense against guided adjusted operating income, and the current portion of debt against cash. Everything else in the release is a growth story. Those two lines are the constraint on it.

Related reading