Everything About AI
NextFin News - Jason Thomas, head of global research and investment strategy at Carlyle Group, warned on Nov. 10 that the financing model now underwriting the artificial-intelligence infrastructure buildout carries the same structural flaw that preceded the mortgage crisis: lenders are underwriting to the investment-grade credit of the financial sponsor rather than to the underlying project. The warning, delivered in a television interview, reframes the AI buildout from a technology story into a credit-cycle story — and puts private credit, which Carlyle estimates may be asked to finance roughly $1 trillion of AI compute, squarely in the crosshairs.
The stakes are large enough that the warning cannot be dismissed as a house view on a single sector. Carlyle's white paper, "Financing AI without a SaaS Repeat," published Oct. 1 and co-authored by Thomas and Carlyle co-president Mark Jenkins, estimates the AI compute buildout will require about $5.5 trillion through 2030. Internally generated cash flow covers roughly $1 trillion and investment-grade bond markets about $2.1 trillion. The residual falls to incremental equity of about $0.4 trillion, structured products of about $0.3 trillion, leveraged finance of about $0.4 trillion, and roughly $1.4 trillion of alternative capital — of which about $1 trillion is expected to land on private-credit balance sheets. That is an amount equal to more than half of the industry's current assets under management, a sector bet larger than anything direct lending has underwritten before, arriving just as the software loans that dominated the last cycle are showing the first signs of strain.
This is not a call to avoid AI credit altogether. Carlyle's own position is narrower and more useful: set explicit concentration limits before the buildout runs, not after, because underwriting guidelines shape deal flow rather than the other way around. The question is whether lenders will act on that advice while the music is still playing.
The Mortgage Analogy: Sponsor Credit, Not Project Credit
Thomas's core charge is specific. Project financing for AI development is frequently supported by the investment-grade credit quality of the financial sponsor, he said in a mid-September interview — meaning lenders are not necessarily taking a view on the underlying project at all. That is a close structural cousin of the pre-crisis mortgage model, where the apparent safety of a loan rested on the originator, the rating, and the house-price trend rather than on the borrower's ability to repay from income.
The mechanism matters because it determines where the loss lands. A data-center loan today is typically extended to a special-purpose vehicle, secured by GPUs and a long-term lease with a hyperscaler or an AI tenant. On paper the contract looks bankable: take-or-pay revenue, an investment-grade counterparty, real collateral. But the economics of the ultimate tenant — often an unprofitable AI company — are untested across a full cycle, and the collateral is a chip whose useful life and resale value are being estimated, not observed. Thomas has questioned the depreciation schedules embedded in these deals and the ability of loss-making AI tenants to honor multi-year commitments if funding conditions tighten.
In other words, the loan is priced as if the sponsor's balance sheet is the backstop. If the sponsor's support proves contingent rather than contractual, or if the collateral depreciates faster than the loan amortizes, the structure has no second layer. That is the precise failure mode of 2008: not that every mortgage borrower defaulted, but that the instruments were sized to a collateral value that only held while the trend persisted.
The Software Precedent: Diversified Until It Wasn't
Carlyle's warning is not abstract; it is anchored in a precedent that is already playing out. Direct lenders built the last cycle around software, on the view that subscription revenue was sticky and uncorrelated to the economic cycle. Generative AI complicated that view — not through the macroeconomy, but through a shared technology shock that made entire product categories obsolete at once. The cash flows proved correlated after all.
The strain is visible in refinancing behavior. By June 2026, software borrowers had reduced their 2028 maturities by only 5 percent, compared with 42 percent for the rest of the loan market, according to Carlyle's analysis of PitchBook LCD and Morningstar LSTA data. The paper singles out platforms underwritten at more than 15 times recurring revenue — valuations that assumed perpetual growth, not obsolescence. A refinancing rate that is eight times slower than the rest of the market is not a timing issue; it is a statement that lenders are already pulling back from the sector that defined the last cycle.
The deterioration is showing up in the BDC complex, the most transparent corner of private credit. As of the second quarter of 2026, nearly 10 percent of total debt investments held by the largest business-development companies were tied to borrowers with at least one distressed debt position, according to PitchBook LCD. Software loans and second-lien tranches accounted for 89 percent of the year-over-year increase in distressed debt across those portfolios. Fitch Ratings put the private-credit default rate at a record 6.3 percent in August 2026, with 47 percent of recent default events involving maturity extensions, payment-in-kind features, or missed interest rather than traditional payment defaults — the signature of a market extending and pretending rather than recognizing losses.
Liquidity stress has followed credit stress. On June 4, Blackstone's flagship private-credit fund, with $82.1 billion of holdings, gated quarterly redemptions at 5 percent of net asset value against investor requests of 10 percent — a two-to-one shortfall that is the mechanical consequence of lending long against redeemable short. Blue Owl Capital, which invests heavily in technology loans, later moved to merge its non-traded BDC into its listed vehicle after a sharp increase in redemption requests, and its listed shares traded at a discount to net asset value during the period. The pattern is consistent: the damage surfaces first in the least liquid, most evergreen vehicles, where the mismatch between asset and liability is largest.
The Arithmetic of the Buildout Does Not Reassure Credit
The most arresting number in Thomas's recent commentary is not the $1 trillion exposure estimate. Speaking on a major business-news program on Oct. 12, he described an exercise in which the industry discounted roughly $2.5 trillion of announced AI backlog to present value against the actual cost of buying GPUs, constructing data centers, and delivering power and cooling — and arrived at only about $375 billion of net economic value. A backlog that shrinks by roughly 85 percent when marked to the capital required to fulfill it is not a modest shortfall; it is a statement about how much of the announced buildout can ever clear an economic hurdle.
"I think that's much too much given what's just been experienced."
Thomas said of the prospect that private credit could reach a 40 percent concentration in AI-related financing, a level implied by current growth expectations. He has argued instead for capping AI-ecosystem exposure at roughly 15 percent. The logic is asymmetric in a way that equity investors understand and credit investors cannot afford to ignore. A venture or growth-equity portfolio can lose six of eight bets and still win if two become Nvidia. A credit portfolio that loses six of eight loans suffers permanent impairment; the upside is capped at the coupon, the downside at the principal.
The collateral question sharpens the asymmetry. A GPU loan amortizes over a fixed schedule, but the useful life of the chip depends on architectural progress. A re-engineering of transformer models that reduces memory-bandwidth requirements would cut the value of today's high-bandwidth-memory-heavy hardware without any change in total AI spending — dollar capex stays on trajectory, but the hardware per unit of AI output falls sharply. That is obsolescence risk sitting inside a credit instrument priced as if the asset were a building with a 30-year life.
Second-Order Effect: The Lender Sets the Price
The under-appreciated consequence of Carlyle's argument is that credit availability, not AI demand, may become the binding constraint on the buildout. In a leveraged deal, the lender effectively sets part of the price: the debt a buyer can raise determines the equity check required, and therefore the bid. When lenders decide they hold enough of a sector, every company in it feels that — including the ones still performing well. A sponsor's offer for an AI infrastructure asset is built on the debt it can raise against it; if lenders cap exposure or widen spreads for concentration risk, the bid falls even when the underlying company has not changed.
Run one step further, this makes private credit the transmission channel through which an AI slowdown becomes a real-economy slowdown. Carlyle's own analysis notes that AI-related capital spending is projected at more than $5 trillion through 2030, against net U.S. fixed investment of about $950 billion a year. A contraction in AI spending would not stay inside data centers; it would spill into construction, energy, and capital equipment — the industries actually pouring concrete, building transmission lines, and manufacturing turbines. The capital cycle for AI now runs less through the Federal Reserve than through the credit committees of private lenders.
That is the second-order risk the market has not fully priced: not that AI fails, but that the financing model breaks first, and takes the buildout down with it. Equity can rotate out of losers into winners; credit, once concentrated, can only wait for the workout.
The Counter-Case, and What Would Break It
The strongest argument against Thomas's warning is straightforward. AI demand is real, and the contracts are not with unproven startups but with investment-grade hyperscalers — Microsoft, Alphabet, Amazon, Meta — on take-or-pay terms. Private credit is not 2008 mortgage origination; it is floating-rate, covenant-heavy, and held to maturity by sophisticated institutions rather than distributed to unsuspecting bondholders. Power, not demand, is the bottleneck, and scarcity of electricity is the most reliable collateral of all. On the senior layer of the stack, this case is credible.
It is weaker on the collateral and the tenant. Hyperscaler contracts can be renegotiated when capital-expenditure guidance rolls over — cloud providers have cut spending before, and "take-or-pay" has a way of becoming "take-and-defer" when the counterparty is your largest customer and your only exit. And while covenants can trigger defaults, they cannot create a liquid market for used GPUs if the industry collectively decides the hardware is obsolete. The 2008 parallel is not that every AI loan will default; it is that the risk was concentrated in instruments whose loss severity was discovered only during the workout, when the price of the collateral was set by a market that had already moved on.
This is the cyclical-versus-structural call at the heart of the warning, and it cuts both ways across time horizons. The cyclical leg is the near-term financing shock: if credit committees tighten, leverage falls and spreads widen, and the buildout slows — a mean-reverting move once capital returns to a sector that still grows. The structural leg is darker and does not mean-revert on its own: if the collateral itself — GPUs, power contracts, take-or-pay leases — is worth less than the loans booked against it, then the loss is permanent and the financing model must be rebuilt, not merely repriced. Carlyle's argument is that the structural leg is plausible enough that lenders should cap the cyclical leg before they are forced to confront the second.
The falsifying signal is concrete. Thomas's structural-risk thesis would be wrong if AI-capex-backed loan loss rates stay below 2 percent through 2027 while the secondary market for three-year-old GPUs remains liquid at more than 60 percent of book value — proving the collateral holds — and if private-credit AI exposure is held under 15 percent of industry assets under management by the end of 2027. Any two of those failing, and the mortgage analogy moves from warning to precedent.
What to Watch
In the short term, expect spread widening and equity pressure in the most AI- and software-exposed corners of private credit. As of the latest session, Carlyle's own shares traded at $41.57, down 1.75 percent, while the largest listed BDCs were mixed to weak — Ares Capital at $19.335, essentially flat, and Blue Owl Capital Corporation at $11.105, down 0.58 percent. The Invesco Global Listed Private Equity ETF fell 0.28 percent to $60.445. These are modest moves; the repricing, if Carlyle is right, has barely begun.
Over the medium term, the watch items are behavioral: whether major lenders announce formal AI concentration limits, whether leverage multiples and pricing tighten for AI sponsors, and how the 2028 software maturity wall gets refinanced — only 5 percent had been addressed by June, against 42 percent for the broader loan market. Over the long term, the question is structural: whether AI financing migrates from sponsor-dependent credit to project-cash-flow financing, where the asset and its contracted revenue, not the sponsor's rating, carry the loan.
Scenarios split cleanly. In the base case, lenders adopt concentration caps, deal volume slows, and the buildout proceeds at a lower, more sustainable pace with no systemic credit event. In the downside case, an AI demand miss coincides with a GPU obsolescence wave, loss rates exceed the 2 percent threshold, and evergreen vehicles gate redemptions more broadly. In the upside case, hyperscaler demand absorbs the incoming supply, collateral values hold, and private credit's discipline earns the spread.
The AI buildout is not short of demand. It may be short of a financing model that survives a demand miss — and that is a credit problem, not a technology one.
Explore more exclusive insights at nextfin.ai.
