Carlyle puts private credit's AI buildout at $1 trillion as investors seek caps

Carlyle Group's new white paper estimates that private credit may need to supply roughly $1 trillion to finance AI computing infrastructure, an amount equivalent to more than half of the private credit assets currently under management, and it names the danger at the top: failing to put clear limits on concentration in AI compute could prove to be "the biggest mistake of all."

Carlyle is not calling for retreat. "We're in a period where the revenue model to date is uncertain," Mark Jenkins, the firm's co-president and head of global credit and insurance, said in an interview. "In such an environment, it's really hard for us as credit investors to say, 'well, we're all in.'" What the firm says it wants instead is to take the risk in a balanced manner.

Balance, in this instance, is being defined by the firm's own investors. Speaking on Bloomberg TV, Jenkins said some of Carlyle's investors are coming to the firm and asking it to put limits in place on these allocations, specifically on the bank syndicated loan side, in the range of 8% to 10%. The scope of that request is worth reading twice. It covers one channel of AI-related financing, and it arrives from the buy side of the table, which makes concentration in the fastest-growing corner of private credit a term of the mandate rather than a matter left to the manager's judgment. A white paper can argue for discipline in the abstract; an 8% to 10% ceiling requested by the capital that pays the fee is discipline with a number attached.

The two halves of Carlyle's message sit alongside each other without contradiction. The $1 trillion is an argument for the size of the opportunity, and the concentration warning is an argument for restraint in taking it, made by the same executive in an interview and on Bloomberg TV.

The trillion dollars is not one trade. Private credit managers are being called on to finance a buildout whose related capital expenditure is expected to top $5 trillion through 2030, and the funding is taking a range of forms: data-center construction and power financing, loans backed by the chips powering the technology, lending to special-purpose vehicles. Carlyle's distinction from software is that credit risk tied to data centers and other AI-related assets is more speculative and more likely to be correlated with the broader economy, while many of the financing structures in use remain largely untested.

The software comparison is Carlyle's own. Software went through a similar boom between 2020 and 2022, accounting for about half of private equity deals over that period, and the paper's warning is that private credit firms now rushing into AI risk encountering the same concentration problems that have plagued lenders exposed to software companies.

One reason concentration will be hard to police is that the borrower's revenue is hard to place. Jenkins said it remains unclear where the eventual profits from AI will accrue, whether among chipmakers, data centers or the companies building applications on top, which means a lender underwriting a compute buildout is pricing collateral and structure more than a named stream of repayment. That makes this a program rather than a single vintage, and a manager that accepts an origination ceiling in the first year of a multiyear buildout gives up share in the rest of it.

An 8% to 10% ceiling agreed with the largest allocators changes the arithmetic of the next fund. Accept it and the manager gives up origination capacity in the market everyone wants to lend into; decline and it carries the refusal into the next fundraise with the same investors. A ceiling on one channel would also likely push the marginal dollar of AI credit toward the others, so the cap would limit the concentration without necessarily reducing the risk the paper describes.

The coverage does not say whether any AI compute debt sits inside the credit vehicles sold to wealth accounts, and nothing in it establishes that it does. If it does, an 8% to 10% discipline requested by the industry's largest institutional allocators would likely reach those investors as a fund document, because the leverage to ask for a number belongs to the buyers writing the biggest checks. That is inference, not reported fact. The advisor's version of the question is plainer arithmetic: several credit funds holding overlapping data-center paper amount to one exposure inside a client's account, whatever the sleeve labels say, and the coverage offers no figure for how much of that exposure the wealth channel carries.

Carlyle has given its own answer: lend, in balance, at a size its investors are helping to set. What that looks like across the industry is still open. If the 8% to 10% range turns up as a term in the next syndicated AI loan to clear, it has moved from a request toward a convention, and every manager's deployment math moves with it. If it stays inside private negotiations with the biggest allocators, the discipline is real and partial, and the capital that never got a seat at that table funds whatever is left, in whatever structure the paper would call untested.