The Interest Bill Comes First

Clark Tang, a partner at Altimeter, looked at CoreWeave's collateral on August 12, 2026, and saw an airplane. In a guest post on Tae Kim's Key Context, he noted that CoreWeave had contracted six-year-old A100 chips through 2029 and raised prices across its fleet 25% in July, and wrote that the collateral was aging more like an aircraft than a smartphone. Eight days later Kakashii ran the second quarter through the most generous version of that premise. Stretch the life of CoreWeave's GPUs from six years to twelve, cut Nvidia's gross margin from above 70% to 35% and pass every dollar of the saving to CoreWeave, and the company would still post a recurring pretax loss of about $39 million, because a longer accounting life saves no cash and the debt still has to be paid.

The two disagree about which bill arrives first, the one for the asset or the one for the money that bought it. Part nineteen asks whether Oracle and the neoclouds, the specialist GPU landlords, can carry the buildout they signed. Parts three and eight argued that the credit stack is where this cycle breaks, and Part eleven set Oracle's backlog against its quarter. This Part names the floor that comes under strain first. DeepStack read 135 filings, calls, posts and letters by 53 voices and verified 69 disagreements, 23 of them contradictions or refutations, against 102 agreements.

The contracts pay the debt

CoreWeave made the first side's case itself on its call of August 11, 2026. Management called the second quarter the point where scale became operating leverage: revenue of $2.6 billion, up 112%, a backlog of $104.2 billion and adjusted operating income of $128 million. Nitin Agrawal, its finance chief, said prices rose about 25% across its products in July and that contracts signed in the quarter carry contribution margins 5 to 10 points above recent ones. A typical five-year contract, he said, is underwritten so that the asset-level debt behind it is repaid within the term, and the company's weighted average cost of debt had fallen almost 300 basis points over the year.

Nebius told analysts on August 12 that its four large mid-term deals of the second quarter yield $20 million to $25 million per megawatt, with upfront payments covering 50% to 60% of the capital spending behind them. IREN's results of August 27 put recent customer prepayments at 45% to 55% of GPU capital spending. Daniel Romero of HyperTech Invest wrote on August 29 that IREN had sold out its 2026 capacity and pre-sold about a quarter of what it plans to bring online in the first half of 2027.

The side's strongest evidence is the price lenders ask when a strong customer stands behind a cluster. CoreWeave told investors on September 17, 2026 that its asset-level loans earn investment-grade ratings when the customers are investment grade, and that its delayed-draw term loan DDTL 4.0 was rated the equivalent of A- by three agencies. S&P Global Ratings has rated CoreWeave itself B+ with a positive outlook since April 9, 2026, kept both on June 11 and wants a longer record of managing customer concentration, funding needs and access to capital before it raises the rating.

DeepStack's reading is that this side describes the contract correctly and the company incompletely. An asset-level rating looks through the borrower to its customer, and a contract that repays its own debt in five years protects the lender on that cluster. The company that pays the interest is rated B+, four notches below investment grade, and the contract does not pay the interest on the next cluster, which is being built before the first one earns out.

The cash does not cover it

The second side does the cash arithmetic. Kakashii counted on August 20 that CoreWeave spent $14.12 billion on property and equipment in the first half of 2026 while operations produced $3.66 billion, so operating cash covered 26% of the spending. In the second quarter alone it covered 11%, and debt principal stood at $35.55 billion on June 30, against $8.03 billion at the end of 2024. CoreWeave's release for the quarter gives the line DeepStack holds to: net interest expense of $640 million on revenue of $2,575 million, or 24.9%, against $267 million on $1,212 million, or 22.0%, a year earlier.

1. CoreWeave, Q2 202624.9%
2. CoreWeave, Q2 202522.0%
3. Oracle, qtr to Aug 31, 20267.4%
4. Oracle, qtr to Aug 31, 20256.2%

Source: CoreWeave Q2 2026 results release (Form 8-K, Aug 11, 2026); Oracle Form 10-Q for the quarter ended Aug 31, 2026; DeepStack calculation

Both of CoreWeave's numbers on debt are true at once. Its rate fell almost 300 basis points over the year; Moses Sternstein of a16z wrote on August 14 that its interest expense was rising all the same, with depreciation on its chips above half of revenue. A lower price on a much larger balance is still a larger bill.

July is where the two sides read one event in opposite directions. JunkBondInvestor reported on August 3 that CoreWeave sought a $2.6 billion loan on July 20, launched it at a spread of 425 to 450 basis points at 99 cents, and cleared it at 550 over at 97, about 10.4% to lenders against the mid-8s first shown. CoreWeave told its call on August 11 that the loan priced at full size in one of the year's most dislocated weeks for credit, and that a delayed-draw structure backed by shorter contracts had opened a market that financing, not demand, had kept shut. Both statements hold, and the market opened at a higher price. On September 20 JunkBondInvestor added that CoreWeave's convertible coupon had risen from 1.75% in April to 2.875%.

On June 28 the same writer had drawn the stack: senior asset-backed debt on data center leases clearing near 120 basis points, neocloud corporate debt nearer 600, where the lender takes the risk of the whole enterprise. Oguz Erkan of Capitalist Letters, bullish on compute, wrote on August 6 that the depressed valuations of Oracle and CoreWeave show how little investors like leverage that depends on speculative returns.

The customer is Oracle's bridge

Oracle's bridge is built of other people's money, and its 10-Q for the quarter to August 31, 2026 prints each span. Operating cash was $23.1 billion. Of that, $11.4 billion came from customer prepayments that Oracle says carry a significant financing component, and $4.0 billion from other deferred revenue. Strip both out and operations made $7.7 billion against $28.5 billion of capital spending, a gap of $20.8 billion in one quarter. Shareholders supplied $19.9 billion: Oracle sold 141 million shares and, in the filing's words, fully utilized its at-the-market program. Kakashii drew the same bridge on September 24, reached the same $7.7 billion and put the share sale at 51% of the cash Oracle held at the quarter's end.

That quarter sits on a longer hole. Oracle's first-quarter release of September 10 shows free cash flow of minus $5.4 billion for the quarter; by DeepStack's sum of the release's quarterly columns, minus $28.7 billion for the twelve months to August 31, 2026. The same 10-Q lists $288 billion of leases signed and not yet begun, substantially all for data centers, and $664 billion of remaining performance obligations, about 13% of it due as revenue within twelve months. S&P Global Ratings cut Oracle to BBB-, the lowest investment grade, on July 9, 2026, with a stable outlook, expecting its adjusted leverage to stay above 4 times for two years. S&P named three grounds for another cut: leverage held above 4.5 times, Oracle losing its path to positive free operating cash flow by fiscal 2029, or a negative view of Oracle's AI strategy or the industry's. The other two agencies sit higher. Moody's rates Oracle Baa2, two notches above junk, and kept its outlook negative when it affirmed the rating on February 2, 2026; Fitch Ratings affirmed BBB with a stable outlook the same day. Michael Spencer of AI Supremacy wrote on October 1 that one more notch would make Oracle a fallen angel and force index and pension holders out of roughly $120 billion of its bonds. It would take more than one notch. The largest bond indexes classify a bond by the middle of its three ratings, so an S&P cut alone leaves Oracle in them; a second agency would have to fall two notches.

Kakashii's reading of September 24 is that Oracle entered the AI era already carrying tens of billions of debt, which limited what it could build and buy on its own. The same 10-Q shows why that debt is not yet the binding cost. Interest expense was $1.43 billion in the quarter to August 31, 2026, 7.4% of $19.35 billion of revenue, against 6.2% a year earlier. Oracle pays its interest from operations and, for now, its capital spending with its customers' money.

The customer pays for Oracle's buildout. The lender pays for the neoclouds', and charged CoreWeave more in July and September.

The customer bridge carries other people's debt as well. Leyla Kunimoto of Accredited Insight wrote on October 5 that Oracle is not the borrower on the $18 billion Project Jupiter construction loan, yet its rent is the loan's main source of cash once the lease begins. The project's lenders are exposed through Oracle's rent and the timetable on which it begins, and this Part does not rank them against the two floors it does.

The same bridge runs under the neoclouds, and it is the weakest span in their case. Ren wrote on August 3 that Nebius is the exception among them, funding capital spending from its own cash flow without equity, prepayments or debt. Nebius's own call nine days later contradicted that account: customers' upfront payments cover 50% to 60% of the spending behind its new deals, and management expects financing of GPUs as a stand-alone asset class to become another source of funds. On August 24 Nebius closed $5.75 billion of convertible notes, at 0.50% due 2030 and 4.50% due 2034, with data centers and GPUs among the uses it named. IREN's results of August 27 add that prepayment terms vary and that there is no assurance future contracts will include them.

The collateral ages faster

The third side puts the risk in what the debt is secured by. Michael Burry read Nebius's own terms on August 12 as compute in backwardation: $40 million to $50 million per megawatt for capacity of up to six months against $20 million to $25 million on one- to three-year contracts. Buyers paying twice the rate for near-term use, he argued, means the GPU, the customer or both are losing value, and he noted that Nebius had stretched its server depreciation from four years to five while that pricing implies decay of about 50% a year. Nebius's answer, the same day, was that its short-term buyers are model developers renting GB300 clusters for three- to six-month training runs, paying for speed and certainty of access rather than for a depreciating asset. Kakashii's August 20 account of CoreWeave fits Burry's reading: GPUs that turn obsolete far faster than the buildings around them.

Lenders have priced for that. Daniel Nishball of SemiAnalysis explained on July 6 that banks underwrite GPU debt to the scenario where the backstop is triggered, asking at least 1.3 times debt service coverage, roughly 70% to 80% loan to value, a case in which the neocloud earns close to nothing and still pays the debt. JunkBondInvestor wrote the same day that an obsolete building needs a second construction project before it can be financed again, and an upgrade does not help when the balloon payment comes first. Harris Associates Oakmark warned in its second-quarter 2026 letter that many neocloud issuers lack the protections that preserve a bondholder's principal. DeepStack reads the structure as built to put the first loss on the neocloud's shareholders, and the second on the lender when the asset ages faster than the loan.

The tests ahead

Oracle reports its second fiscal quarter on December 14, 2026, by its investor relations page, and its 10-Q for the quarter to November 30 answers two questions. Does the six-month increase in deferred revenue from customer prepayments with a significant financing component pass $11.363 billion, the first quarter's figure alone? A yes means new prepaid contracts arrived in the second quarter and the customer bridge is still open. Do leases signed and not yet begun pass $288 billion, the August 31 figure? Leases that start leave that line for the balance sheet, so the reading sets the lease liability beside it.

CoreWeave had not dated its next report as of October 12; it reported 2025 on February 26, 2026. Its results for the quarter to December 31, 2026, expected in late February 2027, answer the third question: does net interest expense reach 25% of the quarter's revenue, from 24.9% in the quarter to June 30, 2026? A reading between 22.0%, the level of a year earlier, and 25% leaves the Call standing but unconfirmed. The fourth runs to June 30, 2027: does CoreWeave, Nebius, IREN, Core Scientific, Applied Digital, TeraWulf or Cipher Digital, the former Cipher Mining, report a bankruptcy or an acceleration of its debt to the SEC?

What would prove DeepStack wrong is net interest below 22.0% of revenue in CoreWeave's quarter to December 31, 2026, its level of a year earlier, which would mean the contracts are paying the debt; financed prepayments at Oracle that stop at $11.363 billion for the six months to November 30, which would close the customer bridge and make Oracle the weaker floor; or two agencies cutting Oracle to junk while S&P keeps CoreWeave's B+ on a positive outlook.

The order of the floors follows from the bridge each has left. The neocloud's corporate lender takes the strain first and Oracle second, through the price of its capital. The strain shows as a neocloud interest bill that keeps rising while Oracle's customers keep paying, well before either floor reaches a default or a junk rating; DeepStack puts 15 in 100 on a listed neocloud reporting a bankruptcy or accelerated debt by June 30, 2027. The Call sets the order of the floors, and the dated tests carry its timing. The customer pays for Oracle's buildout. The lender pays for the neoclouds', and charged CoreWeave more in July and September.

The published page: deepstack.ltd/stories/the-interest-bill-comes-first/