The Carlyle Group’s new white paper warns the private credit industry to tread lightly when financing the AI boom. Learning from private credit’s overexposure to the Software-as-a-Service (SaaS) business, Carlyle flags concentration risks for the AI buildout as more severe.
The paper carries two senior names: Mark Jenkins, co-president and head of global credit and insurance, and Jason Thomas, head of global research and investment strategy.
From 2015 to 2025, direct lending to SaaS and software firms expanded rapidly from under $10 billion to more than $500 billion. The Bank for International Settlements puts the 2015 starting point at almost $8 billion and the end-2025 total above $500 billion, which works out to 19% of every direct loan outstanding.
BDC (business development companies) portfolios tell the same story, with more than 15% of their 2025 lending going to SaaS borrowers. About a third of all private credit funds have money out to the sector.
Same Bet, Bigger Table
Private credit’s AI infrastructure book is far smaller than that today. The projections are what worry Carlyle. The white paper estimates the AI compute buildout could require roughly $1 trillion in private credit, more than half the industry’s current assets under management. Morgan Stanley and JPMorgan analysts put the total buildout bill at $1.5 trillion to $3.2 trillion, and a Forbes analysis of the Carlyle paper notes that $1 trillion would equal 40% to 60% of global private credit AUM, depending on the benchmark used.
AI, the Disruptor, Comes With Its Own Liabilities
Private credit’s concentrated exposure to SaaS-related debt became an issue because of the rise of AI. Agentic AI tools are proving increasingly capable of complex multi-step workflows, which threatens the pricing power of legacy SaaS business models.
What was considered “sticky revenue” is now being reevaluated. Software stocks fell almost 30% between October 2025 and February 2026, and BDCs with heavy software exposure underperformed their lower-exposure peers by about five percentage points. Private credit worries resurfaced across the $3 trillion market as AI pressure on software earnings became the sector’s defining question.
However, private credit is simultaneously facing concentration risks in its financing of capital-intensive AI buildouts. Jenkins says investors need to be “very, very thoughtful” about counterparty exposure, and notes that seven or eight top-tier counterparties sit behind most of the underlying financings he sees. His framing is that Carlyle wants the exposure, with limits.
Investment Grade in Under Three Years
SaaS businesses are not the only ones facing technological obsolescence. A good portion of AI debt is backed by the GPUs (graphics processing unit) used in data center buildouts, and that paper has climbed the ratings scale fast. CoreWeave borrowed $2.3 billion at a floating rate near 15% in August 2023.
By March 2026 it had raised an $8.5 billion facility priced at SOFR plus 225 basis points, rated A3 by Moody’s and A (low) by DBRS, anchored by Blackstone’s credit and insurance arm. Investment grade is the threshold that lets pension funds and insurance companies buy in, since their own rules generally bar them from going much below it.
The Collateral Depreciates Faster Than the Loan
The arrangement relies on computer chips holding value over several years, while historically they have depreciated rapidly. A high-end accelerator can lose roughly half its resale value within three years. The CoreWeave facility runs to 2032.
Michael Burry argued in late 2025 that cloud companies were writing chips down over five to six years while the hardware loses most of its economic usefulness in two to three, and put the industry’s understated depreciation at about $176 billion for 2026 through 2028. The useful life of the collateral remains a judgment call.
Meanwhile, it remains unclear which layer of the value chain will capture sustainable, long-term profits. Hardware providers and hyperscalers are generating massive revenue today, but the long-term economic landscape is undetermined.
Where the Risk Goes to Hide
In August, Sycamore Tree Capital Partners published a note on financing the AI buildout. Trey Parker, the firm’s chief investment officer and co-founder, writes that the composition of AI financing has changed and that the risk “increasingly sits where it is less visible, less liquid and less able to reprice.”
The note cites Morgan Stanley figures putting global data center capex near $2.9 trillion through 2028, with hyperscaler cash flows covering about $1.4 trillion and private credit expected to absorb roughly $800 billion of what is left. Lease commitments across the largest technology companies approach $932 billion, most of it sitting outside reported debt.
Parker expects repricing to surface first in traded credit and second in the private secondary market, and argues that private credit’s illiquidity will defer the adjustment rather than spread it. Sycamore manages CLO portfolios with real software exposure, so the conclusion favors liquid credit, which is where the firm does business. Essentially Sycamore is telling clients to hold liquidity, wait for private credit’s AI financing to reprice, and buy the dislocation. The thesis depends on AI eventually delivering on its promises, and the data behind those projections is still subject to interpretation.
Making the Grade
Thus far, rating agencies have maintained positive ratings for AI infrastructure financing. S&P Global Ratings counted 54 rating actions on 37 issuers worldwide between January 2023 and May 2026 where AI was a main factor in the change, and most of them were upgrades or positive outlook moves. However, in a research note published in July, Moody’s warned that hyperscalers’ move into debt financing threatened their credit ratings, calling the spending “unprecedented.”
While Moody’s notes that Amazon, Meta and Alphabet have some of the strongest balance sheets in the world, the AI buildout is eroding free cash flow.
Lender, Lessor, Supplier, Guarantor
Meanwhile, capital injections into AI buildouts are also coming from the circular financing route. Anthropic’s IPO prospectus details an arrangement under which Broadcom will lend the AI lab up to $42 billion through convertible notes that turn into Anthropic equity. Broadcom is also the chip supplier, the lessor on a five-year lease worth $125.2 billion, and a partial guarantor of the senior debt. Anthropic’s own filing flags that combination as a “potential conflict of interest.”
The chips at the center of it are Google’s Ironwood tensor processing units (TPU), which Broadcom co-designs and manufactures. Broadcom CEO Hock Tan told analysts in September that Anthropic was “on track to become our largest XPU customer in 2027,” the label Broadcom uses for the custom accelerators it builds for its biggest clients.
A $42 billion senior secured tranche is now being syndicated by Bank of America, Citigroup and Morgan Stanley, funding chip purchases through a special purpose vehicle that leases the hardware back to Anthropic. Blackstone leads an $18 billion junior tranche and is committing $9 billion of its own funds.
The deal reinforces opinions held by the Wall Street skeptics, who see the reciprocal financial arrangement as unstable. Carlyle’s own language is restrained and pointed. The white paper argues that AI-related lending “has a far more speculative character” than software lending, because nobody yet knows the size of the ultimate pie or how it gets sliced between the hardware, infrastructure, model and application layers.
Practice What You Preach
Carlyle’s own fund has already been through the squeeze it is warning lenders about. Investors in the Carlyle Tactical Private Credit Fund (CTAC), whose holdings are detailed in its latest SEC filing, asked to pull about 15.7% of shares in the first quarter of 2026. The fund repurchased only the 5% it had offered. A Carlyle spokesperson said CTAC’s window closed after many peers had shut theirs, which left the fund “exposed to those looking for liquidity that didn’t get it other places.”
Investors in BDCs have put in historically high redemption requests this year. Bank of America research tracked first-quarter requests running from 7.9% at Blackstone’s BCRED to 40.7% at a Blue Owl technology income vehicle, with most funds capping withdrawals at 5% and prorating the rest.
Robert A. Stanger & Company reported that non-listed BDC sponsors met about $6.9 billion of requests in the first quarter against $4.9 billion of gross sales, the first time redemptions have outpaced fundraising in that market.
Where the Gate Actually Sits
The two structures differ less than investors assume. An interval fund like CTAC must offer to repurchase between 5% and 25% of shares each period, and it picks the number. Non-traded BDCs run share repurchase programs at the board’s discretion, with no legal floor at all.
In practice both landed on 5% and prorated everything above it, which is how a 15.7% request becomes a 5% exit. That is the liquidity trap, and it catches some investors off guard.
While the 2026 investor exit from interval funds and BDCs is largely attributed to concerns that AI will erode software companies’ earnings, Carlyle clearly sees similar concentration risks accompanying AI infrastructure financing. The white paper argues that AI capex is too large relative to the rest of the economy for its financing choices to stay contained inside the credit book of any single lender.
Someone Already Wrote the Limit Down
Carlyle is not trying to dissuade investors from financing AI. The firm is advocating for concentration limits and better terms that protect investors, set before the buildout peaks rather than after the defaults begin. The Forbes analysis of the paper reaches for the Office of the Comptroller of the Currency and its guidance that a concentration “can become a problem even when it has not proven problematic.”
Parts of the market are already moving. In late September, PGIM anchored the first CLO carrying an explicit ceiling on AI exposure, a $500 million deal issued by Allstate and arranged by BNP Paribas. Every loan in the pool gets scored for how closely it is tied to the AI industry, and once the pool crosses 15%, no more AI-linked debt goes in.
One capped CLO against a projected $1 trillion of AI lending is a rounding error. It is also the exact shape of the discipline Carlyle is asking for, and proof that a lender can write the limit into a document before the cycle turns instead of discovering it afterward.
Private credit learned the software lesson by living through it, and the bill arrived as a 15.7% redemption request and a 5% door. Whether the industry learns the AI lesson in advance will be settled in the syndication letters going out over the next two years.
—
Author: Tim Tolka, Senior Reporter
The editorial team at #DisruptionBanking has taken all precautions to ensure that no persons or organizations have been adversely affected or offered any sort of financial advice in this article. This article is most definitely not financial advice.
See Also:
Could AI Data Centre Financing Become a Systemic Risk? | Disruption Banking
Why Has Carlyle Still Not Closed on Lukoil’s Foreign Assets? | Disruption Banking
