What did Carlyle's paper on financing AI say?
That private credit is about to take on a far larger sector bet than software, and should cap it before the buildout runs rather than after. Carlyle says bank analysts estimate private credit will finance roughly $1 trillion of AI compute, more than half the industry's current assets under management.
The paper, by Jason Thomas and Mark Jenkins, was published on 1 October 2026. Its funding stack, from Carlyle's analysis drawing on bank estimates including J.P. Morgan's of June 2026, puts the total need at $5.5 trillion. Internally generated cash covers $1.0 trillion and investment-grade bond markets $2.1 trillion. The remainder falls to equity, structured products, leveraged finance and $1.4 trillion of alternative capital, the share earmarked for private credit.
The paper's reason for the shift is plain: with the largest technology companies' internally generated cash flow nearly exhausted, further growth in AI spending needs far more outside money. A companion article on this site covers how much of that is already arriving through the public bond market.
| Funding source | Trillions of US dollars |
|---|---|
| Organic cash flow | $1.0 |
| Incremental equity capital | $0.4 |
| Structured product markets | $0.3 |
| High-grade bond markets | $2.1 |
| Leveraged finance markets | $0.4 |
| Need for alternative capital | $1.4 |
| Total funding needed | $5.5 |
Why does Carlyle point to software lending as the warning?
Because direct lenders built a book that looked diversified and was not. On a broad definition of software, Carlyle's analysis of Preqin data puts the sector at about half of all private equity deals between 2020 and 2022, with a comparable share of deal value in the years after.
Lenders liked it for the same reason sponsors did: recurring revenue that seemed sticky and unconnected to the economic cycle. The paper argues the cash flows proved correlated after all, not through the economy but through a shared risk that AI would make products obsolete.
The strain shows in refinancing. Carlyle's analysis of PitchBook LCD and Morningstar LSTA US Leveraged Loan Index data finds that by June 2026 software borrowers had reduced their 2028 maturities by 5 percent this year, against 42 percent for the rest of the loan market. It singles out platforms underwritten at more than 15 times recurring revenue. A companion article on this site sets out the size of that software maturity wall.
“Sellers tend to think the buyer sets the price. In a leveraged deal the lender sets a good part of it, because the debt available decides what the equity cheque has to cover. When lenders decide they hold enough of a sector, every company in it feels that, including the ones that are performing well.”
Why does a lender's concentration limit matter to a seller?
Because, on Carlyle's own argument, lenders' underwriting rules help decide which companies sponsors buy and what they can bid. The paper says credit investors are not prisoners to deal flow, and that the software boom could not have run as far if creditors had widened spreads for concentration risk.
Read from the other side of the table, that is a statement about price. On our reading, a sponsor's bid for a company is built on the debt it can raise against it. If lenders cap their exposure to a sector, they offer less leverage or charge more for it, and the bid falls even when the company itself has not changed.
The paper explains why it expects limits to bite harder on AI than they did on software. It calls AI-related lending far more speculative, with the size of the market and how it will be divided between hardware, infrastructure, models and applications still unsettled. It notes that the structures in use, from GPU-backed loans to lending to special-purpose vehicles, are largely bespoke and that workouts remain untested.
It also expects AI credit to move with the wider economy. Citing research projecting more than $5 trillion of AI-related capital spending through 2030, against net US fixed investment of about $950 billion a year, it argues that a downturn in AI spending would spill into construction, energy and capital equipment.
What should an owner or a sponsor do with this?
Ask lenders about their sector books, not only their terms. How much of the portfolio sits in software, in AI infrastructure and in the industries that supply it, and whether a concentration limit is in place or coming. A lender near its limit in a sector will be a weaker bidder for the next loan in it, whatever its headline appetite.
Look beyond the data centre itself. The sectors Carlyle names as exposed to an AI downturn, construction, energy and capital equipment, are also where lenders setting AI limits may count exposure. On our reading, an owner in those supply chains should expect lenders to ask how much revenue depends on AI spending, and should have the answer ready.
If the business is software with debt due in 2028, start now. On Carlyle's figures, software has refinanced a fraction of what the rest of the market has. Its June 2026 work on software maturities found that the tool lenders respond to most is real paydown, funded by new debt or fresh equity, rather than fees or a higher spread.
If the business sits well outside the crowded sectors, that is worth something. On our reading, a lender capping AI and software exposure needs other places to lend, and a steady company in an unrelated sector is what diversification looks like on its books. That case is best made to the lender directly, in the financing process.
As of October 2026
Sources: Carlyle, Jason Thomas and Mark Jenkins, "Financing AI without a SaaS repeat", white paper, published 1 October 2026, read first-hand on carlyle.com and in the linked PDF, for: the estimate, attributed in the paper to bank analysts, that private credit will finance roughly $1 trillion of the AI compute buildout, more than half of the industry's current assets under management; the Figure 3 funding stack of $1.0 trillion organic cash flow, $0.4 trillion incremental equity, $0.3 trillion structured products, $2.1 trillion high-grade bonds, $0.4 trillion leveraged finance and $1.4 trillion alternative capital against $5.5 trillion needed, from Carlyle's analysis drawing on bank estimates including J.P. Morgan's of June 2026; the statement that the largest technology companies' internally generated cash flow is nearly exhausted; software's share of about half of private equity deals between 2020 and 2022 on a broad definition, from Carlyle's analysis of Preqin data as of April 2025; the 5 percent reduction in software 2028 maturities this year against 42 percent for the rest of the market, from Carlyle's analysis of PitchBook LCD and Morningstar LSTA US Leveraged Loan Index data as of June 2026, and the reference to platforms underwritten above 15 times recurring revenue; the argument that underwriting guidelines partly determine the sectors sponsors target and the prices they can bid; the description of AI-related lending as more speculative, its bespoke structures and untested workouts; the more than $5 trillion of AI-related capital spending through 2030 projected by research reports cited in the paper, against net US fixed investment of about $950 billion a year, from Carlyle's analysis of Bureau of Economic Analysis and US Census Bureau data as of June 2026, and the spillover to construction, energy and capital equipment; and, from the paper's summary of Carlyle's June 2026 work on software maturities, real paydown as the most effective tool. The rounding difference in the table is our arithmetic. The companion articles cited for AI bond issuance and the software maturity wall carry their own sources. The reading of what a sector limit does to a buyer's price and the guidance to an owner or a sponsor are ours.
Corrections: factual errors are corrected on the page and the correction dated. Write to admin@kadenwoodgroup.com.
This position sits within our debt advisory practice.