The Credit Wall

The Thesis

On 31 August the ten-year Treasury paid 4.75%. On 2 January it paid 4.19%. Over eight months the cost of borrowing for a decade rose 56 basis points, the two-year rose 87, and the thirty-year rose 39. Every maturity that matters got dearer, and the front of the curve moved most.

The price of risk went the other way. Moody’s Baa corporate yield sits 1.54 points above the ten-year, and 2026’s average to date is 1.65. The dot-com capex peak averaged 2.28. The housing boom averaged 1.73. The year before the financial crisis averaged 1.85. The 2021 melt-up averaged 1.95. Every prior boom charged more for credit risk than this one does.

That gap is the wall. Not a shortage of money — there is a great deal of money. The wall is its shape. Lenders want more to part with cash for ten years and less to accept that a borrower might not pay them back, and the largest capital programme in corporate history is being financed into exactly that configuration.

Then there is the part the index cannot show you. Aggregate spreads say lenders are relaxed. The coverage record, deal by deal, says something else: a bank hedging the loans it just wrote, two of the largest private-credit buyers declining paper, a syndication that would not clear, and loans changing hands below what they were written for. Both things are in the record and both are true. They are true of different instruments, and that is the finding this page is built on.

This page tracks how the AI trade became a credit trade, who is lending, and where the record shows the lending straining. It does not forecast a credit event, and nothing here is a view on any security.

The Evidence

Start with scale. In October 2025 JPMorgan put AI-linked debt at $1.2 trillion — 14% of the US investment-grade market, ahead of US banks at 11.7%. A sector that barely borrowed became the market’s largest single block, passing the industry whose entire business is borrowing.

The flow behind that stock accelerated through 2025 and again in 2026. S&P Global counted data centre debt issuance nearly doubling year over year to $182 billion, with Meta alone raising $62 billion since 2022 and about half of that in 2025. Across the whole US investment-grade market, companies sold $1.7 trillion of bonds against the $1.8 trillion record set in 2020; Goldman put AI at roughly 30% of it. By May 2026, hyperscaler unsecured supply alone had reached $155 billion for the year — more than 45% above the whole of 2025 with seven months still to run. Morgan Stanley expects issuance near $570 billion, and has estimated that of $2.9 trillion of infrastructure needed through 2028, hyperscalers can fund $1.4 trillion from their own cash. The remaining $1.5 trillion has to come from somewhere.

The corpus-only measure is how hard the record itself is working on this. Eight semantic queries over the article archive return 389 pieces on AI borrowing across its whole history. There were 6 in 2024 Q1 and 78 in 2026 Q2 — thirteen times as many in nine quarters. And 249 of the 389, close to two-thirds of everything ever written on the subject, landed in the last four quarters.

Now the split. In the same four months, three of the largest names in technology sold bonds into ferocious demand: Meta took $96 billion of orders for $25 billion of paper, Nvidia took about $85 billion in its first sale since 2021, and Alphabet took roughly $115 billion — 4.6 times covered — while telling the market it now plans to issue twice a year. That is not a market refusing to fund AI.

In the same season, a CoreWeave-tied data centre sold $900 million of five-year notes at par to yield 7.5%, and published no order book at all. The difference between those two experiences is not the industry. It is the instrument, and what stands behind it.

Debt written against the buildout itself is where the record shows friction, and it shows it repeatedly:

• November 2025. Deutsche Bank, having lent billions to hyperscalers, explored hedging that exposure — including by shorting AI stocks. A lender buying protection against the thing it had just financed.
• February 2026. Trading in credit derivatives on individual tech borrowers kept rising on concern they were borrowing too much. Separately, developers began seeking ratings on unbuilt facilities to unlock capital that would not otherwise come.
• March 2026. KKR, Blackstone and others turned down data centre debt over insufficient insurance against risks including natural disasters. The buyers with the most capital, declining.
• April 2026. JPMorgan and other banks struggled to spread the risk of billions lent against data centres leased to Oracle in Texas and Wisconsin — the same sites behind the record $38 billion sale six months earlier. Origination worked; distribution did not.
• May 2026. Lenders explored private sales of data centre debt, with some banks seeking to offload Oracle-linked loans at a discount. Paper changing hands for less than it was written for is the plainest row in this record.
• August 2026. The buildout was described as stretching major lenders across financing, insurance and underwriting of what the reporting calls a novel asset class.

Of the 36 dated financing events in this record between June 2025 and August 2026, eleven are a lender stepping back rather than a borrower stepping up.

Two structures deserve naming because they change what the reported numbers mean. The first is the vehicle. By December 2025 an accounting of debt moved off balance sheets put Oracle at $66 billion, Meta at $30 billion, xAI at $20 billion and CoreWeave at $2.6 billion. The obligation is real; it is not on the page a reader checks.

The second is collateral. The CoreWeave model — pledge the chips, house the loan in a vehicle — spread through 2026. AMD backstopped a $300 million Goldman loan so Crusoe could buy AMD chips. Nvidia agreed to rent back unused GPUs from young cloud providers for a share of revenue. In July a lender took inference-specific silicon as security for a $400 million loan. SoftBank borrowed $10 billion against its OpenAI shares — a stake whose value is set by the funding rounds the borrowing helps finance.

Each of these is sensible on its own. Together they mean the collateral behind a growing share of this debt is the buildout itself, valued by the boom it is funding. The collateral ladder sets the nine dated steps out in order.

The Companies

Oracle is the clearest borrower in the record. It sought $15 billion of bonds in September 2025, had banks assemble a $38 billion data centre debt sale in October, sold $25 billion in an eight-part deal in February 2026, and closed a further $16 billion on a Michigan campus in April. It also carries the largest off-balance-sheet figure in the record. Its name appears on both sides of this page: on the deals that priced, and on the loans banks later tried to sell at a discount.

CoreWeave supplied the template everyone else borrowed. Chips as collateral, contracts as collateral, obligations in vehicles — the mechanics the FT traces back to it. The 7.5% print on a site tied to its name is the cost of that structure once it leaves an investment-grade balance sheet.

Meta is the borrower the market likes most: $30 billion in October 2025, $25 billion into $96 billion of orders in May 2026, a BlackRock-led $12 billion site sale in July, and $30 billion moved off balance sheet. It is running the full instrument set at once.

Nvidia occupies the strangest position. It backstops customers’ borrowing, rents back their unused hardware, and in June borrowed $25 billion itself — the most cash-rich company in the trade choosing debt. A separate signal on Nvidia financing its own demand traces the loop from the vendor’s side.

Salesforce is the control case, and the most useful one. It sold $25 billion in March 2026 and investors demanded steep concessions, pricing at a significant premium — for a deal funding a buyback, not a data centre. When the market charges more for the borrower with the least AI exposure in this group, the premium is not being applied to compute.

Alphabet has normalised it. Twice-yearly issuance, $115 billion of demand in August, and a 100-year sterling bond floated in February. A century of maturity, funding a demand case argued in years.

The Lenses

Dalio on where this sits. The template in Big Debt Crises is not about whether borrowing is wise but about what the borrowed money produces. Credit is good when it funds activity that generates enough income to service it, and the failure is never the borrowing — it is misjudging the return, repeated widely enough to become a bubble. Phase two of his archetypal cycle is specifically the one where vendors finance their customers and the lending turns circular. That is a description of the mechanics on this page, not a claim about which phase we are in; the coverage record can date the deals, and cannot date the top.

Keynes on the number underneath. Building continues until the expected return on the marginal asset falls to the rate of interest, and of those two the expected return is the forecast. His sharpest point for this trade is that expecting cheaper money later lowers the return on equipment built now, since it must compete with later equipment financed more cheaply. The hyperscalers moved their server-life assumptions in the opposite direction, adding about $10bn to reported profits across four companies.

Minsky on which borrowers are exposed to the market. Minsky sorts financing units by the relation between cash commitments and expected receipts, and the useful distinction is not severity but what each one is exposed to. A hedge unit expects receipts to exceed commitments in every period; it can be hurt by costs or by revenue, but not by the financing market. A speculative unit’s income still covers its interest — the shortfall is on principal, met by rolling maturing debt, which he calls the short financing of long positions, and which leaves the unit exposed to rates and credit standards as well as to its own business. A Ponzi unit’s financing costs exceed its income, so interest is capitalised and the face amount of the debt grows. He is explicit that this last is a cash-flow classification and that the intent “is not necessarily to cheat”.

Read against those definitions, construction debt that has to be refinanced when a building completes is speculative by construction: its exposure is to the refinancing market, not to whether the data centre works. That is the exposure April’s failed syndication and May’s discounted sales actually describe. The separate observation that the collateral itself keeps moving further from an income statement is ours, not his — the collateral ladder sets it out.

Mehrling on what an index spread is a price of. The money view starts from an observation that reframes this whole page: price “is first of all a matter of market liquidity”. A quoted spread is not a reading taken off the world; it is produced by dealers willing to make a two-way market, and it exists only where they are. Mehrling’s lesson from 2008 is that funding liquidity does not automatically become market liquidity — the dealer system is the thing that translates one into the other, and it can stop.

That answers the obvious objection to this page, which is that investment-grade paper and single-asset project debt always price differently and always have. True, and beside the point. The instruments where distribution failed are not priced tighter or wider in the index; they are not in it at all, because paper nobody will quote has no continuous price to contribute. So the aggregate is not the other half of a comparison. It is a measurement of the part of the market where market-making still works, which is why it can sit near a forty-year low in the same season a syndication does not clear.

Shiller on why nobody objects. A new-era narrative spreads socially, and makes the familiar risk premium look excessive for this case, which is different. Where Mehrling explains what the index can and cannot measure, Shiller explains why the people inside it are comfortable.

Graham on what to check. The instruction is to read the instrument, not the issuer’s story. Two of the securities on this page carry the same three letters of industry association and almost nothing else in common: one is a claim on Alphabet’s consolidated cash flows, the other on a single building leased to a tenant whose own lenders are selling. The record is unusually clear that the market knows the difference.

What Moved

• 2025-06-03xAI sells $5B of debt at a double-digit rate. The earliest row here where AI credit is priced in double digits.
• 2025-10-07JPMorgan puts AI-linked debt at $1.2T — 14% of the investment-grade market, past US banks.
• 2025-10-24A record $38B data centre debt sale assembled for Oracle sites in Texas and Wisconsin.
• 2025-11-05Deutsche Bank explores hedging its data centre exposure, including by shorting AI stocks.
• 2025-12-25Off-balance-sheet totals published: Oracle $66B, Meta $30B, xAI $20B, CoreWeave $2.6B.
• 2026-02-04Banks market $56B+ of construction loans secured on leases Oracle has not begun paying.
• 2026-02-20AMD backstops a $300M loan so a customer can buy AMD chips — its silicon as collateral, first known case.
• 2026-03-12Salesforce pays a significant premium on $25B, for a buyback rather than capex.
• 2026-03-17KKR and Blackstone decline data centre debt over insufficient insurance.
• 2026-04-25JPMorgan and others fail to syndicate the Oracle-leased site loans.
• 2026-05-04Banks seek to sell Oracle-linked loans at a discount.
• 2026-06-16Nvidia borrows $25B on $85B of demand — its first bond sale since 2021.
• 2026-08-07Alphabet raises $25B on ~$115B of peak demand and commits to issuing twice a year.
• 2026-08-26The buildout reported as stretching major lenders across financing, insurance and underwriting.

What to watch, in the record rather than in the price. Whether investment-grade order books stay three to five times covered on the next hyperscaler deal. Whether any Oracle-linked loan sale prints at a stated discount rather than a reported one. Whether the collateral ladder takes another step down from inference chips. Whether the ratio in this record shifts — eleven lender-pullback events out of thirty-six today — and whether a syndication that failed in April succeeds on a second attempt. Each is dateable, and each will appear in the coverage record before it appears in an aggregate spread.

Rates data on this page comes from the US Treasury’s daily par yield curve and from FRED, both public and free; every figure can be checked against the same two sources. The credit risk premium is Moody’s Baa yield over the ten-year, which has forty years of history. The high-yield and investment-grade option-adjusted spreads quoted elsewhere on this site carry only three years here, and are described that way wherever they appear.

Sources

Ray Dalio, Principles for Navigating Big Debt Crises — the cycle template, and the vendor-financing phase. See our reading notes.

Hyman Minsky, Stabilizing an Unstable Economy — the hedge/speculative/Ponzi classification, taken from the text rather than from the familiar summary of it.

Perry Mehrling, The New Lombard Street — the money view: a price is a matter of market liquidity, and the dealer system that supplies it can stop. Robert J. Shiller, Irrational Exuberance (2nd ed.) — new-era narratives and the social spread of a valuation regime. John Maynard Keynes, The General Theory of Employment, Interest and Money — liquidity preference, and why a rate can resist falling. Milton Friedman and Anna Schwartz, A Monetary History of the United States — the long record of credit conditions turning before output does.