AI High-Yield Debt Surge: Leveraged Loans Fund GPUs

Most coverage of AI financing focuses on the largest, safest borrowers: hyperscalers with investment-grade ratings selling long bonds. A quieter shift is happening one tier down. According to figures attributed to Goldman Sachs and reported by Crypto Briefing on October 5, 2026, AI-linked leveraged finance issuance reached $88 billion year to date in 2026, against $20 billion in the same period of 2025. High-yield AI infrastructure supply reached $40 billion in 2026, already more than the $12 billion for all of 2025.

This piece looks at that high-yield debt segment: what the terms mean, who is borrowing, and why financing fast-ageing hardware with multi-year debt creates a distinctive risk. It complements our earlier coverage of off-balance-sheet AI debt, CoreWeave's credit default swap spike and SoftBank's junk bond sale, and does not repeat their details.

$88BAI-linked leveraged finance2026 year to date, vs $20B a year earlier

$40BHigh-yield AI infrastructure supply2026, vs $12B for all of 2025

$489BTotal AI-related debt2026 year to date, Goldman estimate

68%Share of new long-term Treasury borrowingJPMorgan's Cembalest, 10-year-equivalent terms

5.17%10-year Treasury yieldabout 1 point higher since January, per CNBC

Investment grade versus high yield, in plain terms

Bond ratings split borrowers into two broad groups. Investment-grade (IG) issuers sit in the BBB category or above at S&P, or Baa3 or above at Moody's. Everything below is sub-investment-grade, called high yield or "junk." The label is a statement about default risk, not a prediction, and the practical difference is cost and access: weaker credits pay higher interest, and some investors, such as many insurers and pension funds, face limits on holding them.

Leveraged loans are the loan-market cousin of high-yield bonds. They are bank-arranged loans to heavily indebted or lower-rated borrowers, usually floating-rate, sold to institutional investors in a syndication. A bond pays a fixed coupon, whereas a leveraged loan pays a spread over a benchmark rate, which is why coverage of such deals quotes the premium in percentage points. Together, high-yield bonds and leveraged loans are what bankers call leveraged finance.

Why does this matter for AI? Goldman Sachs credit strategist Amanda Lynam explained on an August 3, 2026 Goldman Sachs Exchanges recording that all hyperscalers are rated investment grade, but that the wider ecosystem is "filling in financing around" them, with much of that happening in high yield and increasingly in leveraged loans. Hyperscalers had issued $194 billion that year at the time, and Goldman put them at only 40% of nearly $500 billion in AI-related issuance. The rest came from other borrowers across data centers, chips, software and project finance.

AI-linked sub-investment-grade issuance: 2025 vs 2026

Billions of US dollars. Leveraged finance compares year-to-date periods. For high-yield AI infrastructure, the 2025 bar is the full year and the 2026 bar is year to date.

Source: Goldman Sachs figures as reported by Crypto Briefing, October 5, 2026. Hover a bar for values.

How large is the high-yield slice?

Goldman's own numbers on the segment come with a scope caveat. The August recording put AI-driven supply at about 18% of 2026 year-to-date supply in investment grade, up from about 7% in 2025 and 1% in 2024, and said the share in high yield was "about the same." JPMorgan Asset Management's strategists, writing on July 28, 2026, found that technology accounted for 18.7% of year-to-date high-yield new issuance while making up 8.6% of the high-yield index, so tech issuance is running at more than twice its index weight.

The deals also look different. Lynam noted that the issuance has been "chunky," in multi-billion-dollar sizes, which pushed up the average deal size of the US high-yield market, and that most structures had five-year maturities. Concentrated, lumpy exposure is a feature investors need to price: a few large borrowers can matter as much as a broad pool of small ones.

Aerial view of the Goldman Sachs Tower at 200 West Street in Lower Manhattan, seen from One World Observatory — 200 West Street, the Goldman Sachs headquarters tower in Lower Manhattan, seen from One World Observatory (photo taken in 2015). Photo by www.Pixel.la Free Stock Photos via Wikimedia Commons, CC0.

The duration mismatch problem

Duration is how long money is lent. The central structural question in AI credit is whether the life of the debt matches the life of what it financed.

A data center shell, a power connection and a long-term lease can plausibly support debt of many years, because the building keeps its use. The chips inside are different. GPUs are bought in volume, installed, and replaced as newer generations arrive, so the useful, saleable life of the collateral can be shorter than the loan that financed it. If a loan is repaid over five years but the hardware is worth far less after two or three, the lender depends on the borrower's contracts rather than the resale value of the assets. This is analysis rather than a reported figure, and actual depreciation schedules and resale values vary by chip and by buyer.

Goldman's Lynam made a related point in the August recording, saying chip financing would have lower duration than the data center financings that had come through investment grade. By contrast, Goldman noted that 40% of IG issuance with maturities of 15 years or more that year came from AI companies or companies funding the theme. In other words, the long end of the market is carrying data center and hyperscaler risk, while the shorter end is expected to carry chips.

The question is not whether the debt is large. It is whether the debt outlives the asset that backs it.
Analytical framing, not a quotation from a source

GPU-backed lending, CDS and ABS, briefly

Three instruments come up repeatedly in this segment:

GPU-backed loans. Borrowers pledge chips as collateral. CoreWeave, for example, announced a $3.1 billion loan facility for GPU-backed financing in 2026. The lender's recovery depends on what the chips can fetch if the borrower fails.
Credit default swaps (CDS). A CDS is insurance against default, with the premium quoted in basis points a year. Its price is a real-time read on credit risk, which is why moves in single-name AI borrowers get attention.
Asset-backed securities (ABS). ABS pool income-producing assets, such as leases or loans, and sell slices to investors. Goldman strategists have suggested structured credit, ABS in particular, could play a role once data centers are finished and generating stable cash flow.

What the market is signalling

Goldman's recording noted that, as the hyperscalers started to trade wider, high yield "hung in there" for a while, then began to follow the indigestion in IG. Goldman speakers cited 17 of 23 data center joint-venture deals trading wide of their original yields, and the BB high-yield index trading near 165 basis points over (the benchmark was not specified in the transcript), a level a Goldman colleague described as screening very rich. Those readings were as of early August and will have moved since.

On the broader rate backdrop, CNBC reported on September 27 that the 10-year Treasury yield was near 5.17%, about one percentage point higher than at the start of the year, and that SoftBank had sold $11.1 billion of junk bonds with a yield as high as 9.75% on the seven-year tranche. CNBC also cited a June JPMorgan estimate of $4.1 trillion of AI-related debt through 2030. Separately, JPMorgan's Michael Cembalest estimated that the five largest hyperscalers plus Nvidia had issued roughly $320 billion in 2026, or about $303 billion in 10-year-equivalent terms, equal to around 68% of new long-duration Treasury borrowing, as reported by Yahoo Finance. That figure covers investment-grade issuers, so it measures competition for the same long-bond buyers rather than the high-yield segment itself.

Reading the numbers carefully

Dates and scope differ. Goldman cited nearly $500 billion of AI-related debt in an August recording and $489 billion in the October coverage, while another Goldman estimate of about $300 billion refers to senior hyperscaler and chip supply. Compare like with like.
Secondary sourcing. The $88 billion and $40 billion figures are attributed to Goldman in press coverage and were not found on the Goldman page we reviewed, so treat them as reported rather than independently confirmed.
Growth from a small base. A fourfold rise in leveraged finance still leaves it well below the investment-grade volumes.

What to watch next

Whether new-issue spreads on AI-linked high-yield deals widen further, and whether more deals trade below issue price.
Maturity and collateral terms in chip financings, since these set the size of the duration gap.
How private credit and infrastructure funds, which Goldman says hold about $4.5 trillion of dry powder across private markets, absorb what public markets will not.
Rate direction: floating-rate leveraged loans reprice with benchmarks, so higher rates feed directly into borrowers' interest costs.

For teams building on AI rather than financing it, the practical takeaway is concentration: infrastructure providers carry their own balance-sheet risk. Staying portable across model providers, as a model-agnostic workspace like Metir allows, limits how much any one provider's financing conditions can affect your work.

None of this is investment advice. It is a description of how a fast-growing corner of the credit market works and what the available data says.

Sources:

Image credits

Hero: the New York Stock Exchange building on Wall Street seen from the steps of Federal Hall, August 2017, by Arild Vagen via Wikimedia Commons, licensed under CC BY-SA 4.0. It is a general venue photo, not tied to any transaction.
In-body: 200 West Street (Goldman Sachs Tower) seen from One World Observatory, by www.Pixel.la Free Stock Photos via Wikimedia Commons, released under CC0.