AI's $113B Debt Machine Just Hit a Wall in September

AI DEBT

Photo: Wall Street, New York. The lenders who raced to underwrite the AI boom are now offloading the debt — some at a discount. (Carlos Delgado, CC BY-SA 3.0, via Wikimedia Commons)

By Lalu Baidya · Published October 11, 2026

In June, Wall Street lent the AI buildout $113 billion in a single month. In September, it lent $23 billion. The money did not run out — the lenders lost their nerve. And in a turn that ought to worry everyone still cheering the boom, the same banks that underwrote the debt pile are now quietly shopping pieces of it around, some of them at a discount, just to get it off their books.

This is not the story we have told before on TechPulse USA. We covered KKR's warning that AI debt has swallowed 6% of America's bond market — that was about the stock of debt, the mountain already built. This one is about the flow, and the flow just stalled. It is also about the banks themselves running for the exits, and the Bank of England warning that nobody can quite see where all this money will land if things go wrong. Put the three together and September 2026 starts to look like the month the AI credit party changed character.

September's stall, in three numbers

The figures come from the Financial Times, which reported them on October 10, citing Morgan Stanley data that covers both public bonds and privately placed debt. Number one: $23 billion of global AI-related borrowing in September — less than half of August's total. Number two: financing had peaked at $113 billion in June. Number three: the drop is real, but Morgan Stanley itself cautioned against reading too much into a single month, attributing much of the slowdown to companies having already raised enormous amounts earlier in the year.

Fair enough. A borrower that gorged in June does not need to borrow in September. But the stall did not arrive alone, and the company it kept is what makes it interesting. Investors, it turns out, are getting noticeably more selective about who gets their money — and on what terms. Reuters reported back in September that bond buyers were demanding greater compensation for financing chips, data centres and computing infrastructure, and were reconsidering how much of their portfolios should ride on the same AI growth story. Portfolio managers told Reuters their concern centred on the scale and unpredictability of the borrowing, not on fears that the big technology companies would default. They wanted clearer evidence the investment would generate returns.

For the less-established borrowers, the questions are sharper and blunter: when do the customers start paying, how dependable is that income, and will it cover the debt? The market's answers, measured in basis points, tell the story of September better than any press release. Bonds from Amazon, Microsoft, Meta and Alphabet saw their spread over Treasuries widen toward 80 basis points, up from around 50 — a meaningful move for investment-grade debt from four of the most profitable companies on the planet. Order books thinned too: cover ratios on hyperscaler bond deals fell from roughly five times the amount on sale in February to about two times by July. Meta's $25 billion bond sale drew $96 billion in peak orders; a comparable $30 billion deal a year earlier had pulled in $125 billion. The appetite is still there. It is just a lot thinner than it was. (These market readings were reported by Startup Fortune, citing Bloomberg and Financial Times data.)

The banks that built the pile are now offloading it

Here is the part that should make you sit up. JPMorgan, Morgan Stanley, SMBC and MUFG — the same banks that spent the last year racing each other to underwrite data-centre loans tied to Oracle, Meta and the rest of the AI buildout — have been exploring significant risk transfers, according to the Financial Times. A risk transfer is Wall Street's polite machinery for letting a bank keep a loan on paper while shifting its riskiest slice to outside investors. The banks are doing this, the FT reported, because they are bumping up against internal limits on how much AI-linked exposure any single lender can hold.

In a normal month, nobody outside a risk desk has ever heard of a "significant risk transfer." This month they are front-page news. That alone tells you the scale of what is being financed — and how quickly the people closest to the money have started hedging their own bet.

The most telling case comes via Bloomberg: a group of banks including JPMorgan and MUFG has spent more than six months trying to distribute $38 billion in construction debt tied to a data-centre project in Texas and Wisconsin leased to Oracle. Six months. That is a very long time to sit on debt nobody underwrote with the intention of holding. Some lenders, Bloomberg reported, are now willing to sell pieces of the loans at a discount to non-bank buyers just to move them. When the people who originated the loans take a haircut to get rid of them, the market has stopped pretending the terms are fine.

Matthew Moniot, co-head of credit risk sharing at Man Group, told the FT the deal sizes involved are "out of scale to anything we've thought about, ever." Read that again. It is not a hedge-fund manager talking down a trade he wants no part of. It is someone whose actual job is pricing exactly this kind of risk, saying the numbers no longer fit any model he has ever used.

The Bank of England is watching — and it doesn't like what it sees

Central bankers do not usually spend their autumn meetings fretting about tech-industry financing structures. The Bank of England's Financial Policy Committee just did. The committee, which met on September 25 and whose conclusions the FT reported on October 1, found that rising AI-related debt issuance had increased risk exposure across capital markets. The numbers are dizzying: Morgan Stanley estimates cited in the minutes put global AI-related debt issuance at $450 billion from the start of the year through early September — more than twice the total for the whole of 2025.

The committee's worry is not really about $450 billion. It is about where that debt lives and who ends up holding the loss. As a growing range of investors and markets fund AI investment, the FPC assessed, exposure to losses from changes in earnings expectations has spread far wider. Downward revisions to earnings expectations and investor asset sales could, it warned, reinforce each other and accelerate price declines. We saw a dress rehearsal in July, when a sharp sell-off in AI and semiconductor shares pushed leveraged investors to sell concentrated positions, deepening the decline; some suffered substantial losses as the selling was accompanied by deleveraging.

Then there is the plumbing. The FT reported on September 20 that residual-value guarantees provided by Big Tech companies over the preceding year — promises that the future value of chips and data centres will stay above an agreed level, underpinning borrowing by the special-purpose vehicles that hold the assets — amounted to as much as $300 billion (reported via economy.ac). The Bank of England warned that the opacity of such financing, and "circular trading" between companies, combined with increased borrowing, could amplify losses if investment returns fall short. Translated from central-bankese: the debts are guaranteed by arrangements nobody outside the deals can see, between companies whose cash flows depend on each other. If the returns disappoint, nobody knows where the hole will open up.

The canary in the credit mine: Meta's Beignet bonds at 91 cents

If you want a single, tradeable readout of how this is going, watch Meta's Beignet bonds. Issued through a special-purpose entity to help fund Hyperion — Meta's vast AI data-centre project in Louisiana — the $27.3 billion of 6.581% senior secured notes maturing in 2049 were one of the largest private project-finance bonds ever sold. Morgan Stanley brought the deal to market in late 2025; PIMCO anchored it with roughly $18 billion and BlackRock absorbed more than $3 billion. The notes traded near 110 cents on the dollar at launch. On October 7, they fell to about 91 cents — a record low. The yield on them has risen to roughly 7.55%, about 2.3 percentage points above Treasuries. Meta's own credit default swaps have moved toward 100 basis points. (Reported by Finimize and TradingView/Seeking Alpha.)

Look at the structure and you will see why the market is squinting. The bond matures in 2049. Meta's lease on the data centre runs for four years. In between sits a building full of specialised equipment that loses value every year it sits in a rack, plus the risk that the AI business changes shape long before 2049. Equity in the joint venture is split between Blue Owl Capital (80%) and Meta (20%), which keeps the debt off Meta's balance sheet — but ratings firm S&P Global warned at issuance that problems during construction or operations could still end up landing on Meta anyway. This is the template, too: Morgan Stanley estimates roughly $3 trillion of similar off-balance-sheet financing and leasing structures across the AI build-out. If long-dated deals demand more yield, every new project gets more expensive to fund — just as hyperscaler spending is expected to run near $1 trillion this year and higher next, while elevated Treasury yields lift borrowing costs across the board.

We re-ran the dates ourselves this morning, and the pile-up is what makes this a story rather than a data point: the FT's September-flow numbers landed October 10, the bank-offload reporting broke in the same week, the Bank of England's warning surfaced October 1, and the Beignet notes hit their record low on October 7. Four signals, ten days, one direction. Markets that are merely pausing to digest do not usually produce all four at once.

Insurance on Big Tech is now the most expensive it has ever been

The credit-insurance market agrees. Credit default swaps — the cost of insuring against a company defaulting — on SpaceX widened to record levels last week on reports it is poised to issue another $40 billion in debt, IFR reported on October 9. It is not alone: AI-linked names from market darling Nvidia to the more heavily leveraged Oracle have seen default protection climb to all-time highs, as investors reassess the sector's unprecedented borrowing binge.

Andrew Sheets, global head of fixed income research at Morgan Stanley, told IFR that "there is a risk premium associated with that" — "that" being the fact that these companies are now embedded in the credit ecosystem in a dozen ways at once: funding capex directly through bonds, providing credit backstops to other facilities. "A lot of people in the market are increasingly exposed to these companies, either directly or indirectly," Sheets said. His analogy: bank credit spreads have historically traded "cheap" for their rating, because there is so much issuance, so many ways to be exposed, and a perception that if something goes wrong it could have systemic implications.

The sharpest edge is Oracle, whose $18 billion New Mexico data-center project hit the wall. Its five-year CDS spread hit 2.03 percentage points — an 18-year high, up from just 144 basis points at the start of 2026, according to market data cited by Startup Fortune. Total net notional CDS exposure on major tech names has hit a record $12.5 billion, up 500% since the second quarter of 2025, with Oracle alone accounting for $6.5 billion of it. That is not a market that believes the borrowing is trivial.

Our take: a stall is not a collapse — but this financing is shoddy

Let us be honest about what September's stall does not prove. It does not prove the AI buildout is fake, or that demand for compute is collapsing. Morgan Stanley's own reading — that the year was simply front-loaded — deserves its place. And the scale of what is still being planned is colossal: Normandie Research, writing up Morgan Stanley's estimates, puts data-centre spending at roughly $2.9 trillion over four years, with a financing gap of around $1.5 trillion that cash flow will not cover — about half of which is expected to come from private credit. (That last chain of estimates is reported by Normandie Research, citing Morgan Stanley, and should be read as a projection, not a fact.) When the numbers are this big, a single slow month is what a healthy market looks like while it catches its breath.

But nobody should mistake the structure for a healthy one. The September numbers exposed a financing model that is, in places, genuinely shoddy: debt maturing in 2049 backed by four-year leases, assets whose residual value is guaranteed by the very companies borrowing against them, risk passed around through instruments so opaque that the Bank of England cannot map where the losses would land. Nobody botched a single decision — that is the problem. A whole industry built a machine that assumed investors would keep showing up on the same terms forever, and for six months the machine ran at full speed. Then September arrived, the investors got choosier, the banks hit their limits, and the machine started choking.

What comes next is not a mystery, it is a negotiation. The buildout will continue — there is still enormous committed capital and genuine demand — but on terms that look much less friendly than the ones printed in June. Spreads stay wider, cover ratios stay thinner, and every new project gets priced like the risk it actually is. That is arguably what should have been happening all along. It just took $90 billion of vanished monthly issuance, four banks shopping their own paper, a central bank warning and a record-low bond price to make it happen.

Update, Oct. 11: the stall is over — October's mega-raises just doubled the pile

SpaceX is lining up $30 billion in investment-grade debt plus $10 billion in bank loans to buy Nvidia chips (Reuters, Oct. 8), and Broadcom is seeking $50 billion — weeks after September's $23 billion stall (FT/Morgan Stanley, via NewsNation, Oct. 11). Credit markets are pricing the risk: SpaceX's five-year CDS hit a record; deVere Group's Nigel Green put it bluntly: "The AI build out started on cash. It's increasingly running on credit, and credit changes the risk profile entirely." Meta's Hyperion bonds slid to a new record low near 91 cents on the dollar (Seeking Alpha, Oct. 7). The wall held for one month; October's issuers are testing it again. The same week, Waymo's $5 billion first-ever debt raise showed even cash-rich names joining the borrowing queue — while the Senate's answer: make data centres pay for their own grid.

Sources: Financial Times (via Al Bawaba, Startup Fortune and TechNewsday); Bloomberg (via Startup Fortune and Silicon UK); Reuters; Bank of England Financial Policy Committee (via Reuters and economy.ac); IFR; TradingView/Seeking Alpha; Finimize. Morgan Stanley data and estimates as cited in the outlets above.

By Lalu Baidya. Lalu Baidya covers AI infrastructure and the money behind it for TechPulse USA, from data-centre finance to credit markets. He writes about what happens when Silicon Valley's ambitions collide with Wall Street's balance sheets.

Photo: a server hall at CERN (file photo — illustrative of the data-centre hardware being financed, not an AI data centre). (Hugovanmeijeren, CC BY-SA 3.0, via Wikimedia Commons)