When Blue Owl Capital capped redemptions at 5% across two of its largest private credit funds in April, after investors sought to pull more than $5 billion, the headlines treated it as a single story: Private credit is in trouble because of AI.
That framing isn't wrong, but it is incomplete in a way that matters. The private credit industry has been roiled by concerns that it is overexposed to the software industry, an area under pressure over fears of disintermediation from artificial intelligence. But it is also being reshaped by a separate force: Private credit is financing the AI buildout itself, through debt structures that have no real historical precedent.
These are related concerns, but they are not the same concern.
Problem one: Legacy software loans under AI pressure
The first problem is the older and more familiar one. Beginning in 2020, private credit lenders concluded that enterprise software companies were ideal borrowers. The reasons were straightforward: fat profit margins, stable customer bases, and reliable revenue from license subscriptions. Private equity firms loaded up on software acquisitions financed by direct loans, often at high leverage ratios.
The scale of this lending is substantial. According to the Bank for International Settlements, outstanding loans to SaaS firms increased from about $8 billion in 2015 to over $500 billion, or 19% of total direct loans, by the end of 2025. An analysis by credit research firm Octus of 155 public and private BDC portfolios found that software companies represented about 29% of investment cost as of September 2025. That figure is higher than the roughly 20% estimate from Jefferies because Octus counts software-adjacent companies that some BDCs classify under other labels.
The concern now is that AI is eroding the competitive positions of many of these borrowers. Loans originated prior to 2024 did not contemplate AI as a meaningful business risk, according to an assessment by institutional investment consultant Prime Buchholz, potentially resulting in underpriced credit exposure. At the Milken Institute Global Conference this week, Davidson Kempner Capital Management CIO Tony Yoseloff warned that AI-driven change could further weaken recovery rates in leveraged loans tied to software companies. He noted that average recoveries on first-lien debt have been below 40 cents on the dollar over the past five years, with software assets likely to perform even worse due to their limited tangible collateral.
Marathon $MPC Asset Management CEO Bruce Richards has been even more pointed. Richards argues that AI-driven pricing disruption, combined with eight to ten times leverage ratios and near-zero free cash flow after debt service, makes a 15% default rate in direct lending software portfolios a mathematical certainty in 2027 and 2028 when maturity walls hit.
Not all observers agree with that timeline, but the directional concern is broadly shared. UBS estimated in January that 25% to 35% of private credit portfolios face elevated AI disruption risk, with the most acute exposure in technology and business services. This is the problem driving most of the redemption headlines. Investors in semi-liquid private credit vehicles cannot see the underlying exposure clearly.
No major fund currently discloses the split between infrastructure software (which is AI-resistant), vertical SaaS (variable), and horizontal application software (highest AI-displacement risk), and the AI-displacement-risk metric does not exist as a disclosed category. When investors cannot quantify a material risk, the rational response is to reduce exposure, which is what is happening.
Problem two: GPU-backed debt for AI infrastructure
The second problem sits on the opposite side of the AI trade. Instead of loans threatened by AI, this involves loans made to build AI's physical infrastructure, through financing structures that did not exist three years ago. CoreWeave is the defining example. In August 2023, CoreWeave secured a $2.3 billion debt financing facility led by Magnetar Capital and Blackstone by using Nvidia $NVDA's H100 GPUs as collateral, the first time that H100-based hardware had been used as collateral.
Since then, the company has scaled this model aggressively. CoreWeave's total debt exceeded $21 billion as of the end of 2025, up from under $8 billion in 2024. In March, it closed an $8.5 billion delayed draw term loan facility, the first GPU-backed financing to receive an investment-grade rating. And this month, CoreWeave launched a $3.1 billion broadly syndicated GPU-backed loan, its fifth such financing, led by Morgan Stanley $MS and MUFG, expanding the potential investor base beyond private placements. GPU loans are typically backed by two things: a customer's contract to use the chips, creating a cash-flow stream to service the debt, and the value of the chips themselves.
The rating on a GPU loan typically reflects the credit standing of the customer. When the customer is Meta $META, that produces an investment-grade rating. When the customers are OpenAI and Cohere, both unrated, the debt is expected to earn ratings in the BB range, with Fitch already assigning BB+. The risk embedded in these structures is distinct from the software loan problem. GPU rental rates have already fallen 50% to 70%, shrinking collateral value even as repayment schedules reference depreciating hardware. The financing structures used to build AI infrastructure embed assumptions about residual values, utilization rates, and counterparty stability that are not well-supported by available data.
One can be bullish on AI and still skeptical of the specific debt instrument financing its construction. The BIS estimates that outstanding private credit to AI firms has exceeded $200 billion and could reach $300 billion to $600 billion by 2030. The terms of these loans do not differ markedly from those to other sectors in maturity or rate spreads, but the loans are on average substantially larger ($169 million versus $90 million).
Why the distinction matters
Conflating these two risks has practical consequences for anyone evaluating private credit exposure, whether as an allocator, a fund investor, or someone with a target-date retirement product that holds corporate bond funds. The legacy software loan problem is a credit quality issue. The question is whether borrowers can continue servicing debt as their business models come under AI-related pressure. The resolution depends on whether AI adoption erodes revenues at PE-owned software companies faster than lenders can work through restructurings. The GPU-backed infrastructure debt problem is a collateral and structure issue.
The question is whether a novel asset class, GPU compute capacity secured by long-term contracts, can sustain the financing assumptions built into it as hardware depreciates, chip generations turn over, and customer concentration remains high. A private credit fund with 25% software exposure and zero neocloud exposure has one risk profile. A fund financing data centers through asset-level GPU debt has a different one.
As PitchBook noted, while software loans make up a large share of some of the largest asset managers' private credit investments, the sector's troubles tell only part of the private credit story. The headlines will keep saying "private credit."
The two problems inside that phrase require two separate assessments.