The AI buildout's debt didn't shrink. It changed names.
October's credit break didn't shrink the AI buildout's borrowing. It moved the signature onto the customer, the landlord, the lessor, and the contractor.
This fall set two records in the same three weeks. Companies that borrow on the strongest credit — investment-grade, the tier that almost never defaults — issued $1.88 trillion of bonds, the most ever, and five technology giants supplied $370.7 billion of it, much of it to fund AI, on oversubscribed deals [1]. Meta alone sold $15 billion of dollar bonds on November 2 and followed two days later with a first-ever euro sale of about $9.2 billion [2][3]. Across the street sat the other record: CoreWeave's bonds trading below 80 cents on the dollar at yields above 15%, the territory of high-yield, the junk tier where lenders charge double digits because they expect some of it back [4]. Goldman Sachs put a name on it.
We think the widening reflects a reassessment of the economics of the AI buildout, particularly for lower-rated borrowers that are more exposed to higher-for-longer rates and potential energy-cost pressures. — Goldman Sachs Private Wealth Management
Which raises a question worth holding onto: if the money never left, whose signature did it need to keep moving? Ask the banks, and they will tell you whose signature they stopped accepting. Lending to the buildout did not stop. Lending to its sponsors did [5]. A regional lender put the rule in plain words.
We’re not lending money to the people sponsoring the build of data centers. We are lending money to people who construct data centers. — Fifth Third Bank
Wells Fargo describes its own limits in the same register.
We have very hard limits in terms of what we’re willing to do, and what we’re not willing to do, and are sticking to that. — Wells Fargo
Start with the customer. The neoclouds — the companies that rent Nvidia chips by the hour to anyone training a model — have begun signing their own buyers on as capital. Nebius grew revenue 454% last quarter at 41% margins, and its finance chief put the first source on the record.
Our first sources of capital remain customer prepayments and operating cash flow — Nebius Group
More than $9 billion in customer prepayments is expected this year [6]. The contrast with CoreWeave is the sorting principle now at work: CoreWeave grew revenue 112% and holds a $104 billion backlog, but it sits on a debt-to-equity ratio of 8.94, meaning $8.94 of borrowed money behind every dollar of equity, and $5.74 billion in negative free cash flow [6]. The market has stopped sorting these companies by demand and started sorting them by leverage. In Regina, the debt took the landlord's name. The phone company will own the $1.7 billion campus, and CoreWeave and Cerebras will lease space inside it, with analysts warning the hangover is unchanged: the rent still has to come from the debt-stressed tenants [4]. The largest GPU financing on record shows the same trick at scale. In the $40 billion package Goldman Sachs and Apollo arranged for SpaceX on November 4, an Apollo affiliate buys the chips and leases them back for just over five years, matched to how fast the hardware loses value, so the debt never touches SpaceX's balance sheet; Goldman sells slices of up to $500 million to investors [7]. CoreWeave itself tried the same re-titling the day before the break: a $4.2 billion convertible, a bond that can swap into shares, aimed at cutting its interest bill [8]. None of this means the money stopped. Commercial-and-industrial lending at the 25 largest U.S. banks grew nearly 10% in a year, the fastest in a decade outside the pandemic [5]. Blue Owl, a big lender in this corner of the market, raised its earnings forecast as spreads widened [9].
We think the supply for compute will obviously be a little more challenged given the change of what we’re seeing and the fallout potentially here from a change in government. — Erik Bissonnette
Ed Yardeni made the no-free-fall case the same week: projects under construction get finished, states will court the infrastructure, and investment-grade debt was still flowing at record volume into the very buildout the selloff punished [10]. Some of this was in place before the break. Nebius's prepayment model and CoreWeave's convertible both predate the October 14 repricing that broke the straight debt [8][4]. So the repricing did not invent the new signature; it ratified it. Aon's chief executive supplied the hinge in a sentence.
If it's more difficult to get insurance, that means the financing structure is going to change completely. You can't access the debt markets. — Greg Case
Add one more set of co-signers: Meta has backstopped $41.5 billion of residual value on its Hyperion campus, and Nvidia has agreed to guarantee up to $108.5 billion of its own chips' future worth [11]. Even the pitch for new money now markets the absence of debt: Groq is asking for an $8 billion round by telling investors it is not a neocloud [12]. So the answer is a roll-call. The risk did not leave the system; it changed address: the prepaying customer, the landlord in Regina, the lessor buying SpaceX's chips, the $500 million slices in the syndicate, the contractor with the bank loan, the guarantor backstopping a chip's future price. The buildout still borrows everything it did before. It just no longer asks the sponsor to sign first.
- 1. US Corporate Bond Issuance Hits Record $1.88 Trillion
- 2. Meta Raises $15 Billion in Bonds for AI Infrastructure
- 3. Meta Platforms Raises Billions via Debut Euro Bond Sale
- 4. AI Infrastructure Debt Crisis Triggers CoreWeave Stock Plunge
- 5. US Bank Lending Surges 10% Amid AI Data Center Boom
- 6. CoreWeave and Nebius Group Report Surge in GPU Demand
- 7. Goldman Sachs and Apollo Arrange $40 Billion SpaceX AI Deal
- 8. CoreWeave Secures $4.2 Billion Bond Deal to Fund AI Growth
- 9. Blue Owl Forecasts Higher Earnings Amid AI Compute Risks
- 10. Ed Yardeni Predicts Resilient AI Capital Spending
- 11. Private Credit AI Buildout Creates Systemic Financial Risk
- 12. Groq Seeks Series B Funding at $8 Billion Valuation