AI Debt Bubble: Lenders Pull Back as $420 Billion Looms in 2027

Talk of an AI debt bubble has moved from stock analysts to the lenders themselves. Banks and credit investors are scrutinizing AI debt deals more closely and are reluctant to add exposure after the market selloff, Reuters reported on November 9. Some banks that were lining up to fund AI deals a few weeks ago now say they would hesitate even when a marquee technology company guarantees the loan.

The short answer to whether this is a bubble: borrowing has grown very fast, the price of that borrowing is rising, and some weaker borrowers are already changing how they raise money. The largest companies can still raise money, and far more issuance is planned. This piece walks through the numbers behind each of those statements.

$420B

Hyperscaler debt issuance Goldman Sachs Research projects for 2027

50 bp

Widening in 10-year hyperscaler bond spreads this year, per Bank of America

$88B

AI-related issuance by low-rated firms in 2026 through September, per Goldman Sachs

$400B

Estimate of Nvidia's total financing obligations cited by CNBC

How fast AI borrowing has grown

Start with the lower-rated end of the market, where the growth is most visible. AI-related issuance by low-rated firms reached $88 billion this year through September, according to Goldman Sachs figures reported by Reuters. For comparison, AI-related leveraged finance issuance was about $20 billion in the first 11 months of 2025, according to Neuberger Berman data cited in the same report.

Junk bonds alone tell the same story. AI infrastructure supply in the high-yield bond market reached $40 billion in 2026 by the end of September, against $12 billion for all of 2025, according to BNP Paribas data reported by Reuters.

The top of the market has changed just as much. Hyperscalers' share of investment-grade bond issuance rose from 1% to 10% in two years, in a market of about $15 trillion, CNBC reported on Wednesday. The banks are exposed through property lending too. JPMorgan had $29 billion of loans backed by other income-producing properties, a category that includes data centers, at the end of the third quarter. That is 12% of its commercial real estate loans, up from 8% three years ago, according to Reuters.

The price of AI debt is rising

Lenders have been asking for more compensation all year. The spread on 10-year hyperscaler debt, the extra yield over Treasuries, has widened by about 50 basis points in 2026, Bank of America credit analyst Yuri Seliger told Reuters. For high-quality industrial companies with no AI exposure, the widening over the same period was about 20 basis points.

Meta bonds issued earlier this year and SpaceX bonds are trading below par, the same Reuters report said. Further down the quality ladder the moves are sharper. The cost of insuring CoreWeave's debt against default rose 94% between October 1 and November 5 before settling, according to LSEG data cited by CNBC. An Iren convertible bond due in 2029 fell 23% between November 3 and November 5 as the company's shares dropped 26%.

Coupons were already high before the selloff. In September, borrowers rated just below investment grade were paying 9% to 10%, and lower-rated names faced 14% to 15%, analysts told Reuters. SoftBank paid 8.625% to 9.75% on notes it sold that month. Zenith Arc, a project company financing an Oklahoma data center, sold $2.25 billion of five-year secured notes in August with an 8.875% coupon at a price of 99.50. By late August the notes were more than seven points below their issue price, according to an investor letter cited by Reuters.

Who can still borrow, and who is changing plans

The pullback is uneven. Nebius explored a high-yield bond in September and did not proceed, a source told Reuters. In October it took a $3.5 billion senior secured loan from Citigroup instead. CoreWeave, whose bond yield topped 16% last week, said with its results on Tuesday that it would not use the unsecured debt market at current yields. It expects about $9 billion of customer prepayments and deposits over the next three months, as we covered in our report on CoreWeave's earnings.

Both examples point away from unsecured bonds and toward secured loans and customer money. That keeps projects funded. It also means lenders, not shareholders, hold the first claim on the most valuable assets if a borrower runs into trouble.

Politics adds a second source of doubt. The Nasdaq 100 lost $2.5 trillion of market value on Thursday, November 5, and lenders are also weighing the chance of tighter oversight of data centers after the midterm elections, Reuters reported. Baird Equity Research named Oracle and CoreWeave among the companies that could be hit by construction delays or moratoriums. We looked at that risk in our post on a possible data center moratorium.

The supply still to come

The caution arrives just as the industry plans to borrow more. Goldman Sachs Research projects global hyperscaler debt issuance of $420 billion in 2027, more than 60% above its estimate of $250 billion for fiscal 2026, according to Reuters.

More bonds sold into a market that is already demanding wider spreads usually means higher borrowing costs, smaller deals, or both. Our analysis of AI capex in 2026 shows why the borrowing is needed: free cash flow at several of the largest spenders no longer covers their capital spending.

Nvidia's guarantees tie the system together

Part of what makes AI debt hard to judge is that much of it leans on the same few names. Nvidia has made a series of commitments that help smaller AI companies borrow, CNBC reported on Wednesday. Critics call such arrangements circular, a theme we covered in our analysis of AI circular deals.

Regulators are paying attention. An International Monetary Fund report in October warned about growing leverage tied to AI, and Federal Reserve officials discussed the risk of a broad repricing of assets at their July meeting, according to CNBC.

There is a counterweight. Nvidia's market value was still about $4 trillion after its recent decline, and Van Nieuwerburgh told CNBC the company would have to get close to extinction before its guarantees came into question. Sheila Bair, a former head of the Federal Deposit Insurance Corp., told CNBC that Nvidia is not too big to fail in the way a bank is.

What the AI Bubble Index says about credit

The AI Bubble Index read 61.9 after Tuesday's close, in its Normal band, down from 63.3 a week earlier. The headline has eased as stock prices fell, but the parts of the index closest to this story have not. The systemic risk family reads 69.3, and the Fragility axis, at 69, sits above Euphoria at 62.4. Sentiment and attention is the hottest family at 95.5. The methodology explains how each family is built.

Put plainly, prices have cooled faster than the financing structure underneath them. That is the combination a credit pullback tends to expose.

How this compares with the telecom build-out

The dot-com era offers a useful comparison, and it is not the stocks. It is the telecom carriers that borrowed heavily to lay fiber for traffic that arrived years later than their debt schedules assumed. The network was real and so was the eventual demand. Many of the builders still failed, because debt has a calendar and demand does not. Our dot-com bubble timeline follows how that unfolded.

Today's largest AI borrowers are far stronger than those carriers were, and most have profitable core businesses. The comparison fits the smaller, more leveraged builders better than it fits the giants. Our guide to AI investment risks explains how leverage changes the outcome of a slowdown.