CoreWeave, Oracle and the AI Debt Selloff: Everyone Is Borrowing, the Moat Sets the Rate

On Thursday, October 15, CoreWeave's unsecured bonds due 2031 traded near 79 cents on the dollar and the cost of insuring its debt for five years hit a record. On Friday a hotter September inflation print pushed Treasury yields up, CoreWeave closed down about 12% at $61.39, Nebius was down about 11% by the afternoon, and the Nasdaq lost 2.1%. At Friday's open TheStreet's James DePorre described the AI group as splitting between companies that can fund their buildout themselves and companies that have to borrow, with the borrowers trading like rate-sensitive stocks.

I think the instinct is right and the line is in the wrong place. Alphabet's free cash flow was negative in the second quarter, Amazon's was negative over the twelve months to June, Meta's was under $1 billion, and the big builders had sold about $220 billion of bonds by August. So nearly everyone in this cohort is borrowing. What separates them is how much the price of that money moved when AI credit got scarcer, and that depends on what the lender is really relying on.

Read down the third column and the size of the move tracks what sits behind the debt. The hyperscalers sold a record amount of bonds this year and their spreads widened by roughly ten basis points. Oracle's insurance cost hit a record but stayed in investment-grade territory. CoreWeave sold unsecured notes at 9.75% in April, with an add-on at a premium to face value, and six months later the market wanted 16% for the same paper.

I want to be careful about the contract-backed loan. In August CoreWeave borrowed against customer contracts at SOFR plus 5.5 points, which works out to roughly the same all-in cost as its April unsecured notes. So I can't yet show that contracts buy it cheaper money than its own name does. What I can show is that the unsecured market has repriced CoreWeave's name and nobody has priced a contract-backed loan since. The next one is the test, and it's part of the falsifiable claim at the bottom of this page.

Microsoft is the only one of the big four still funding its build from cash flow in the plain sense, and even it has $329 billion of leases waiting to start. A lease is a way of having someone else borrow for you, and sometimes the someone is a neocloud. When Nebius signed its roughly $17.4 billion, five-year GPU contract with Microsoft in September 2025, its filing said cash from the contract would pay for part of the build and that its obligations would start once it had secured the extra financing it needed.

So the self-funded versus borrowed split doesn't sort these seven names cleanly. It puts Alphabet, Amazon and Meta on the borrowing side next to CoreWeave, and the screen above shows lenders don't treat them that way.

I've grouped the cohort by what pays the lender back. Microsoft, Alphabet, Amazon and Meta borrow against businesses that were throwing off cash long before the AI build started. CoreWeave and Nebius borrow against the build itself. Oracle has a foot in each, and it's the name I'd watch most closely.

The scorecard doesn't carry the financing story on its own, since valuation moves these names around more than any bond does. The moat matrix is closer. I've used three pillars that say whether a company's existing cash flow is protected: scale economics, how embedded it is in customers' operations, and network effects. Several other pillars, like system of record and business logic, aren't applicable to a GPU landlord by definition, so including them would make the neoclouds look empty for reasons that have nothing to do with lending.

The top group is rated well across this grid, with scale economics carrying most of the weight: these companies buy chips, power and capital more cheaply than anyone renting capacity to them. That's what a bond buyer leans on when Meta sells a note maturing in 2066 and long-dated spreads across the group sit near 118 basis points. The data centre could earn nothing and the coupons would still get paid from ads, search, AWS and Office.

The bottom group is rated weak on the same pillars. CoreWeave and Nebius sell capacity a customer can buy elsewhere when the contract ends, and they buy hardware and power at a disadvantage to the companies they compete with. That describes what a lender can lean on, not how well the companies are run, and it's why their customers' contracts matter so much more to their lenders than their own names do.

Oracle is the reason I don't think the split is simply big versus small. Its database and business applications are about as embedded as enterprise software gets. If Oracle were borrowing to grow those, I'd expect its credit to trade like the top group's.

It's borrowing for OCI, which grew 121% in the quarter to August and sells capacity to the same frontier labs CoreWeave rents to. S&P estimated in July that about half of Oracle's backlog was OpenAI. The framework scores Oracle as a software company and doesn't rate OCI on its own, so this part is my reading: on the OCI side, Oracle has no more of a cost advantage than the neoclouds do. Lenders seem to be pricing Oracle as a blend. Its five-year CDS hit a record of about 288 basis points on October 15, roughly a quarter of CoreWeave's level, and S&P has it one notch above junk.

Oracle has also been filling the gap in ways the hyperscalers haven't needed to. It sold $19.9 billion of stock in the quarter to August, and it took $11.4 billion of customer prepayments that carry a financing component. A prepayment is the customer lending to Oracle, so the credit question moves one step down the chain to whether the customer can keep paying.

• Refinancing cost lands on the equity. CoreWeave's existing notes carry fixed coupons of 9% to 9.75%, so Thursday's bond prices don't raise this year's interest bill. They raise the cost of each new dollar, and it spent nearly four times its operating cash flow on equipment in the first half, so it has to keep raising money.
• The contract's term caps what a lender will underwrite. The DDTL 5.5 contracts average about three years. A lender can underwrite three years of those customers paying, but not what the GPU earns after that, which is when a neocloud's equity value is supposed to show up.
• The top group can slow down and the bottom group can't. If rates stay high, Alphabet or Meta can trim capex and the core business keeps paying. A neocloud that stops building stops growing into its debt, so it's exposed to rates on both the cost side and the growth side.

The strongest objection is that I've answered a lender's question and an equity investor asked a different one. Cheap debt says the coupons will get paid. It says nothing about whether the build earns its cost. Alphabet can borrow at a sliver over Treasurys and still destroy value if its cloud capacity ends up rented at prices that don't cover depreciation. The people warning about an AI bubble mostly mean that, and a bond spread can't settle it. If Alphabet's and Amazon's free cash flow stays negative into 2027 while AI revenue disappoints, shareholders lose even though nobody defaults. I accept that. My argument is narrower: a rate shock like this week's lands first and hardest on the shareholders of companies whose debt has nothing behind it but the build.

The second is that I'm calling it a moat when it's just leverage and size. Credit analysts price debt off coverage ratios and balance sheets, and CoreWeave's $35 billion of debt against a few billion of operating cash flow would explain its spread whatever its pillars said. Oracle's $125 billion of borrowings explains a lot of its CDS too. I think that's partly right, and it doesn't fully cut against me. A company's coverage ratio is high because its cash flow is large and steady, and that's what the pillars are measuring. But a reader could reasonably say I've relabelled a credit metric, and on Oracle in particular the leverage explanation is simpler than mine.

The third is that the split is circular. The top group's cheap build depends partly on the bottom group's expensive one. Microsoft's unstarted leases and the neoclouds' contracts with Microsoft and Meta are two sides of the same arrangement, and if CoreWeave or Nebius can't fund new capacity, the hyperscalers either build it themselves, which raises their capex and borrowing, or go without. A neocloud credit crisis wouldn't stay contained to neocloud shareholders. Crypto Briefing reported that insurance on Amazon's, Alphabet's and Meta's debt also widened to record levels this week, though it didn't publish figures. I don't think that moves their debt anywhere near junk, but it would make the self-funders look less self-funded still.

The fourth is the CoreWeave bull case. The company had about $104 billion of contracted revenue at June 30, it raised contract-backed money in August rated one to two notches below investment grade, and Nvidia's Jensen Huang said on October 15 that its cost of capital is actually coming down. Nebius is a useful comparison here. It raised most of its first-half money through convertible notes, which are cheap debt with an equity option attached, plus share sales and pre-funded warrants, held $8.0 billion of cash against $8.5 billion of debt at June 30, and still fell about 11% on Friday. That looks like the market selling a category rather than a balance sheet, and categories sold on fear do overshoot. If CoreWeave keeps signing contracts and raising contract-backed loans near August's terms, the unsecured bond price matters less than it looks.

The last is Oracle, where I could be wrong in either direction. ERP Today reported that most of Oracle's new AI contracts came through prepayment, bring-your-own-hardware or similar structures and won't need incremental capital from Oracle. If that holds, OCI's growth is funded largely by its customers and Oracle's credit should drift back toward its software peers. If OpenAI's ability to pay comes into question, it drifts the other way, and the database franchise ends up servicing debt raised for a business it doesn't protect.

For Microsoft, Alphabet, Amazon and Meta, I'd read a selloff on AI financing as a selloff on the wrong variable. They are borrowing now, and Alphabet's and Amazon's cash flow did turn negative, but the lender's claim sits on businesses the build doesn't put at risk. The thing to watch in their Q3 2026 reports is whether the core businesses keep growing, because that's what keeps their debt cheap.

For CoreWeave and Nebius, the equity sits behind lenders who are mostly underwriting their customers. The two things I'd watch are the price of CoreWeave's next contract-backed loan and whether it can raise unsecured money at all without a big jump in coupon. For Oracle, the CDS is the cleanest read on whether the market thinks OCI is being paid for by its customers or by the database.

Next time a company announces a big build funded with debt, I'll start with the lender's question and ask what would get paid back if the new thing earned nothing. This week the market asked it of CoreWeave, and the answer was a set of customer contracts and not much else.