Is the AI buildout competing with my financing?
For allocation, yes. For availability, not yet. Goldman Sachs Research tracks nearly $500 billion of AI-related debt issuance so far in 2026, of which the large cloud and compute operators are only about 40 percent, and describes AI-related supply as the dominant theme in the credit markets (Goldman Sachs Exchanges, 5 August 2026).
The reason a founder or a sponsor should care about a number set in the investment-grade bond market is that the borrowing does not stop there. Goldman's credit strategy team describes the same theme drawing on high yield and on the leveraged loan market, and expects private markets to do what it calls the heavy lifting for whatever the public markets cannot absorb. Those are the markets a middle-market loan is written in.
The scale is not the unusual part. Credit markets have funded large releveraging waves before, in pharmaceuticals, in telecommunications, in bank capital. What Goldman calls unique here is the multi-year nature of it, and the concentration: one theme, running for years, rather than one issuer doing one acquisition and then going quiet.
It is worth being precise about what the firm is and is not saying. Goldman states twice that it is not concerned about access to capital for this theme. Its argument is that the conversation shifts to where the risk gets taken and at what price. For a borrower, that is the whole point. Price and place are what you negotiate.
How much of the credit market has this already taken?
About a fifth of new supply, in two years. AI-related issuance was $10 billion of US investment-grade supply in 2024, or 1 percent of the year-to-date total. In 2025 it was $108 billion, about 7 percent. In 2026 it is running at about 18 percent. Zach Ablon, who runs Goldman's credit sales desk, puts the high yield share at about the same.
The duration figure is sharper than the headline. About 40 percent of all fifteen-year-and-longer investment-grade issuance in 2026 has come from AI companies or from companies funding the AI theme. The composition of the index has moved with it: on Goldman's account Amazon is now the highest duration weight in the investment-grade index, from twentieth a year earlier, and Google has moved from eighty-sixth to eighteenth.
The borrowers are not short of cash. The large operators issued $108 billion globally in 2025 and $194 billion so far in 2026 while still holding cash on balance sheet, which Goldman reads as getting ahead of a multi-year investment cycle rather than as distress. Its framing is a waterfall of capital: internal cash flow, then equity, then debt, working down the depth of each market in turn.
That framing is the one to hold, because it tells you the direction of travel. When the top of the waterfall is exhausted, the flow moves down. Investment grade is near the top. Private credit is further down.
| Measure | Reading | Period |
|---|---|---|
| AI-related debt issuance, all markets | Nearly $500bn | 2026 year to date |
| Large operators' share of that issuance | About 40% | 2026 year to date |
| Large operators' debt issued globally | $108bn | Full year 2025 |
| Large operators' debt issued globally | $194bn | 2026 year to date |
| AI-related share of US investment-grade supply | 1% | 2024, on $10bn |
| AI-related share of US investment-grade supply | About 7% | 2025, on $108bn |
| AI-related share of US investment-grade supply | About 18% | 2026 year to date |
| AI share of 15-year-and-longer investment-grade issuance | About 40% | 2026 |
| Capital expenditure that is debt financed | 27% | 2025, actual |
| Capital expenditure that is debt financed | About 33% | 2026, expected |
| Capital expenditure that is debt financed | 35% | 2027, expected peak |
| Further investment-grade capacity at US bank issuer levels | About $510bn | Case study, qualified as possibly generous |
| Global private-markets dry powder, all categories | $4.5trn | As stated, August 2026 |
| Project finance and data centre transactions expected | $300bn | 2027, above direct issuance |
“A borrower who has never issued a bond still borrows from the same institutions that buy them. When one theme takes a fifth of new supply in the deepest market and the same share in the next one down, the question is not whether your loan gets funded. It is where in the queue it sits and what the lender charges for the place.”
Why does an investment-grade issue matter to a company that will never issue a bond?
Because the investment-grade market has a ceiling, and Goldman's view is that the ceiling, not the borrowers' balance sheets, is the binding constraint. Taken to two times net leverage or three times gross, both ordinary for investment-grade ratings, the large operators could carry something in the ballpark of $2 trillion of additional debt. Goldman does not think that limit binds.
What binds is absorption. Goldman's case study is the three largest issuers in the US investment-grade market, which happen to be US banks: on the Bloomberg index none carries more than $175 billion of index-eligible debt and none is more than 2.3 percent of the index. Bringing the large technology issuers up to bank levels would create roughly $510 billion of further capacity in that market alone.
The firm qualifies its own figure, and the qualification matters more than the number. Institutional investors push back that they already hold the theme in equities, which tempers their appetite for the same theme in credit, and that long-dated paper is not counted dollar for dollar against short-dated paper. High yield indices, unlike investment grade, sometimes cap single-name exposure at 2 or 3 percent outright.
So the question becomes which market takes the rest. Goldman's answer is private markets first, then non-US bond markets, then new structures. It sizes global private-markets dry powder across private credit, infrastructure, real estate and private equity at $4.5 trillion, and notes that the category lines are blurring: data centre financing is now being written as private credit, some as venture debt, some inside private equity. On top of the direct issuance, it expects $300 billion of project finance and data centre transactions in 2027.
What has this done to the price of a middle-market loan?
Nothing yet, on the published evidence. The broadest measure of what a performing private loan actually costs, Houlihan Lokey's Private Performing Credit Index, had the weighted spread at 576 basis points at 30 June 2026, the lowest quarterly reading since that series began in 2017 (Houlihan Lokey, Private Performing Credit Index, Q2 2026). A market being crowded out does not print its tightest spread on record.
The strain, so far, is showing up in the AI trade itself rather than in the market around it. Goldman's own AI leadership basket has gone from tights of 74 basis points to nearly twice that over twelve months. Insurance demand for the long end has thinned: in the first quarter a large AI-related financing might draw fifteen insurance accounts with orders above $50 million into the thirty-year tranche, and by the end of the second quarter that count was about half.
New issue concessions tell the same story in the price a borrower pays to get a deal done. Goldman puts them at two to three basis points before the recent indigestion, and as wide as twenty basis points on one very large deal. In high yield, seventeen of twenty-three data centre joint-venture deals are now trading wide to the yields they were originated at.
Read those three together and you have the honest position. The AI trade is being repriced. The middle market is not, yet. What connects them is the private-markets pool Goldman expects to absorb the overflow, and the fact that the deals arriving in it are large, lumpy and mostly five-year.
What should a borrower do about it before 2027?
Finance into the window rather than after it. Goldman expects the debt-financed share of the large operators' capital expenditure to rise from 27 percent in 2025 to about 33 percent in 2026 and to peak at 35 percent in 2027, which puts the heaviest year of supply after the current one, with $300 billion of project finance on top of it. A facility that has to be refinanced in 2027 is competing for attention in the year the wave is largest.
Ask your lender the question directly. What share of new commitments in the last two quarters went to digital infrastructure, and against what return? A private credit fund writing data centre paper at investment-grade-adjacent spreads is making a different portfolio decision than one writing sponsored middle-market unitranche, and a borrower is entitled to know which fund it is talking to before it accepts a term sheet.
Do not read the $4.5 trillion as a promise. It is a global figure across four strategies with different mandates, and Goldman's own point is that the money will come from different pockets, not that every pocket is open to every borrower. Capital that can be deployed into a hyperscaler joint venture at scale is not automatically capital that wants a $40 million unitranche with a maintenance covenant.
And watch the double derivative rather than the level. Goldman's own answer, when asked what would signal the theme wobbling, is the point at which capital expenditure stops rising as fast, not the point at which it falls. A borrower does not need a view on whether the buildout works. It needs to know that the market it borrows in has committed a fifth of its new supply to one theme, and to fix its terms while the spread on its own credit is still at a series low.
As of September 2026
Sources: Goldman Sachs Research, via Goldman Sachs Exchanges, How AI Debt Is Reshaping Credit Markets, published 5 August 2026, with Amanda Lynam, head of Credit Strategy Research, and Zach Ablon, head of the credit sales desk in Global Banking & Markets, read from the published transcript, for nearly $500 billion of AI-related debt issuance in 2026 to date and the hyperscaler share of about 40 percent; for hyperscaler debt issuance of $108 billion globally in 2025 and $194 billion in 2026 to date; for AI-related issuance of $10 billion and 1 percent of year-to-date US investment-grade supply in 2024, $108 billion and about 7 percent in 2025, and about 18 percent in 2026; for the high yield share running at about the same level as investment grade; for about 40 percent of fifteen-year-and-longer investment-grade issuance in 2026 coming from AI companies or companies funding the AI theme; for Amazon moving from twentieth to the highest duration weight in the investment-grade index and Google from eighty-sixth to eighteenth; for debt-financed capital expenditure of 27 percent in 2025, about 33 percent expected in 2026 and 35 percent expected in 2027 as the peak; for direct hyperscaler supply of about $250 billion expected in 2026 excluding project finance; for the ballpark of $2 trillion of incremental debt capacity at two times net or three times gross leverage and the statement that this is not the binding constraint; for the US bank case study, none of the three largest US investment-grade issuers carrying more than $175 billion of index-eligible debt or representing more than 2.3 percent of the index on the Bloomberg index, and the resulting $510 billion of further investment-grade capacity together with the firm's own qualification that the comparison may be generous; for high yield indices capping single-name exposure at 2 or 3 percent where investment grade does not; for global private-markets dry powder of $4.5 trillion across private credit, infrastructure, real estate and private equity and the blurring of those category lines in data centre financing; for $300 billion of project finance and data centre transactions expected in 2027 above and beyond direct issuance; for participation from the high yield and leveraged loan markets and the observation that the deals are chunky, multi-billion and mostly five-year; for the AI leadership basket moving from tights of 74 basis points to nearly twice that over twelve months; for insurance participation in the thirty-year part of large AI financings falling from about fifteen accounts with orders above $50 million in the first quarter to about half that by the end of the second; for new issue concessions of two to three basis points widening to as much as twenty basis points on one very large deal; for seventeen of twenty-three data centre joint-venture deals trading wide to originated yields; and for capital expenditure growth slowing, rather than falling, being the signal the firm would watch. Houlihan Lokey, Private Performing Credit Index, Q2 2026, for the weighted average spread of 576 basis points at 30 June 2026 as the lowest reading since the series began in 2017; a companion article on this site carries that index in full. No published series measures AI-related issuance as a share of private credit origination, or the price effect of that issuance on a middle-market loan, and neither is asserted here. The reading of the waterfall as a queue, the three questions for a lender and the timing argument for financing ahead of 2027 are ours. Companion articles on this site cover why leveraged loans now yield more than high yield bonds, why lenders holding record undeployed capital concede structure before price, and how the energy and digital buildout is reaching middle-market services businesses on the equity side.