The Lender Pullback: $296B of Hyperscaler Debt, Then Banks Start Saying No

Market Insight · Tuesday, November 10, 2026 · Ad Valorem · AI-generated content

Video companion: The Syndicate Room EP54 (data-center vote & private-credit episode — the closest shipped companion to this story). AI-generated illustrative visuals. The full written note below carries the detail and sources.

For a year, the question about the AI build-out was whether there were enough chips, enough power, enough cooling. The answer was yes, but only if you borrowed. On Monday, the question flipped. It is no longer can we raise the money? — it is will anyone take the other side?

Reuters reported on November 9 that banks and credit investors are upping scrutiny of AI-related debt deals and showing reluctance to increase exposure, even when Anthropic or Google backstop the risk. That is a meaningful shift from October, when Apollo, Blackstone, Goldman, BofA, Citi and Morgan Stanley were lining up to syndicate deals into the tens of billions. The trigger was Thursday's selloff — the Nasdaq 100 lost $2.5 trillion in one session, Nvidia fell 9% — plus the midterm results, which put data-center oversight back on the policy agenda.

The numbers behind the pullback

• $296B of hyperscaler bonds already issued in 2026. Goldman's credit strategy team (Amanda Lynam) tallied Alphabet, Amazon, Microsoft, Meta and Oracle at $296B of year-to-date bond supply, up from $108B in all of 2025. That is 18% above Goldman's own $250B full-year forecast for 2026, and roughly one-third of the five names' capex is now debt-financed. (Goldman via MarketWatch, Nov 5, 2026.)
• $67B raised in a single week. The five hyperscalers completed five bond transactions worth $67B between late October and the election. Goldman now expects another $440B of direct hyperscaler supply in 2027, plus ~$300B of project finance — a $740B pipeline, excluding neoclouds, chipmakers and utilities. (Goldman, Nov 5, 2026.)
• Spreads have moved. BofA's Yuri Seliger put the spread on 10-year hyperscaler debt at ~50 bps wider year-to-date, versus ~20 bps for high-quality industrials not tied to AI. That 30 bps of relative widening is the first clean measure that the market is pricing AI-credit risk as a separate asset class, not a tech premium. (BofA via Reuters, Nov 9, 2026.)
• Bond deals are being shelved. Nebius, which explored a high-yield bond issue in September, decided in October not to go ahead. It instead took a $3.5B senior secured loan facility from Citi — a one-year bridge, not a permanent capital structure. Lenders are now stress-testing cash flows and contract terms, including the conditions under which a large technology customer can reduce or terminate commitments. (Reuters, Nov 9, 2026.)

Who is exposed, and how

The pullback is not uniform. It is sharpest at the margin — neoclouds and chip-financing vehicles — and then it works its way up the stack.

• Neoclouds (CoreWeave, Nebius). CoreWeave paid $640M in second-quarter interest — roughly five times its adjusted operating income. Its 2032 notes yield more than 16% and its 5-year CDS is past 1,155 bps. Nebius shelved its high-yield bond and took a one-year Citi bridge. Both are now financing through private credit or bridge facilities rather than the public bond market. (Yahoo Finance / Reuters, Nov 9, 2026.)
• Chip-financing SPVs. Broadcom's $63B deal to finance custom AI chips for Anthropic (BofA, Blackstone, Citi, Morgan Stanley) is the template: a special-purpose company buys the chips, Anthropic leases them, Broadcom backstops residual value. Apollo and Blackstone are also working a $50B+ financing for a Broadcom–OpenAI chip program. SpaceX is in talks for ~$40B to buy Nvidia Rubin chips, with Apollo leading and PIMCO participating. None of these is signed; all of them depend on the same lender appetite that is now cooling. (WSJ / Bloomberg / Reuters, Nov 8–9, 2026.)
• Hyperscalers themselves. Alphabet, Microsoft, Oracle and Meta still have investment-grade access, but Goldman notes the issuance is increasingly "early lock-in" of 2027 financing. That is a signal that even blue-chip borrowers no longer expect the window to stay open. (Goldman, Nov 5, 2026.)

Why this matters now

Because the bottleneck on the AI build-out has moved from the data center to the balance sheet. For two years the constraint was physical — chips, power, cooling, permits. The market priced that scarcity into every neocloud, every power deal, every frontier-lab round. Now the constraint is financial: lenders are deciding, deal by deal, how much AI credit they will hold in 2027. When the answer is "less," the cost of capital for the build-out rises, the timeline slips, and the equity and bond valuations of every company that needs to borrow to keep growing have to be re-underwritten. That is the story sitting in the spread data, the shelved bonds, and the $296B number — and it is a story that has to be positioned for, not predicted.

The practical read

You cannot buy a specific AI credit pool, and you should not try to pick which neocloud's bridge loan restructures next. What you can do is underwrite the theme deliberately:

• Map the stack by credit quality. The hyperscalers are still investment-grade; the neoclouds are high-yield or private; the chip-financing SPVs are structured to look like a credit product but carry operating risk from the underlying chip program. The discount to each layer is different, and the risk is different. The opportunity is in the gap between what the market is pricing at each layer and what the structure actually supports.
• Decide, at what price, you exit if the credit and the equity turn together. A CoreWeave 5-year CDS stepping from ~1,155 bps toward 1,400 bps, a second neocloud shelving a bond, and a hyperscaler spread widening another 25 bps would be that signal. Set the level now, before the next credit event moves it.
• Position before you predict. The read is not "AI is a bubble." It is that the build-out's cost of capital is rising, the lender is now the counterparty, and the window to raise at 2025 pricing is closing. The discount to any exposure priced as if 2026 spread data did not exist is the opportunity — and the downside is the same position marked against a slower top line, a wider spread, and a bridge loan that does not roll.

Watchlist

• Goldman's hyperscaler debt tally (next update): $296B → where by December. The 18% over-forecast gap is the starting point; the trend is the signal.
• BofA hyperscaler spread vs. IG industrial spread: 30 bps of relative widening is the baseline; 50 bps or more confirms the re-pricing.
• Neocloud bridge-loan renewals: Nebius's $3.5B Citi facility is due in a year; CoreWeave's next refinancing is the real test. Watch for extension, PIK, or a step-up in margin.
• Chip-financing SPV closings: Broadcom–Anthropic $63B, Broadcom–OpenAI $50B+, SpaceX $40B. If even one of these slips a quarter, the template is under stress.

Sources / further reading

• Reuters — "Banks, investors grow wary of AI debt, spooked by market selloff" (Nov 9, 2026)
• MarketWatch — "Data centers run on debt. That was a problem for bonds this year — and it's only going to get worse." (Nov 6, 2026, citing Goldman credit strategy)
• Yahoo Finance — "Broadcom, Oracle and SpaceX tap private credit for AI hardware: report" (Oct 7, 2026)
• Yahoo Finance — "Banks that raced to fund neoclouds like CoreWeave are now hesitating" (Nov 9, 2026)
• Bloomberg — "Apollo Says It Will Play Defense in Lending for AI Buildout" (Nov 3, 2026)
• Ad Valorem — "The Bottleneck Moved From Chips to Ballot Boxes: 40 of 41 Measures and the Private-Credit Wall" (Nov 9, 2026)
• The Syndicate Room EP54 — "Voters Rebuff Data Centers: 41 Ballot Measures & the Private Credit Squeeze" (Nov 7, 2026)