Is Private Credit Underpricing AI Risk in Software?
Why a $500 Billion Loan Exposure May Be More Fragile Than It Looks

Is Private Credit Underpricing AI Risk in Software? Why a $500 Billion Loan Exposure May Be More Fragile Than It Looks

Software once looked like the ideal private-credit collateral: recurring revenue, low churn and high margins. Generative AI is now challenging those assumptions faster than most loan books can reprice them, and by the first quarter of 2026, the theory has started becoming tangible.

When the private credit industry built its software franchise, the logic was seductive: enterprise software companies enjoy recurring revenues, low churn, asset-light balance sheets and predictable cash flows that behave almost like bonds. Business development companies and direct lenders poured capital into the sector throughout the low-rate era, financing leveraged buyouts at enterprise-value multiples exceeding 20x EBITDA. The scale of that bet has since become clearer. Outstanding loans to SaaS firms grew from roughly $8 billion in 2015 to over $500 billion, representing 19% of total direct loans, by end-2025, with a third of all private credit funds now carrying exposure to the sector. Among BDCs specifically, software accounts for around 17% of investments by deal count.

Then came generative AI. Within months of ChatGPT's public release in late 2022, the question shifted from how fast software companies would grow to whether AI would simply replace them. By February 2026, Morgan Stanley was warning that an AI-led software sell-off was spilling into credit markets. Global software stocks slid roughly 30% from their 2025 highs. JPMorgan began re-marking loans tied to software companies. The resulting hard question for every private credit investor: were lenders adequately compensated for a risk that didn't exist when the deals were written? In this essay, underpricing means that loan spreads, covenant protections, and portfolio valuation marks may not yet fully reflect the risk that generative AI weakens recurring revenue durability, compresses margins, and lowers recovery values for software borrowers

The Software Bet: How Direct Lenders Got Here

The concentration in software lending reflects the market conditions of 2020 to 2023. Near-zero interest rates compressed yields across asset classes and pushed investors toward private credit in search of premium returns. Software buyouts reached $256 billion in 2021 alone, over one-fifth of all leveraged buyouts that year, and direct lenders rather than banks supplied the majority of the debt. First-lien structures offered average loan-to-value ratios of approximately 50%, reinforcing the impression that lenders were well cushioned even in downside scenarios.

The key underwriting metric was annual recurring revenue. Unlike EBITDA-based lending in traditional sectors, ARR underwriting priced loans against contracted future subscription income, treating renewals as near-certain. This framework rewarded lenders handsomely in a world where software churn rates were low and enterprise IT budgets were expanding. What it did not price was the possibility of structural disruption, a shock that could compress ARR not gradually through competitive pressure, but rapidly through technological substitution.

Four Ways AI Rewrites the Credit Rulebook

Generative AI threatens the assumptions embedded in software loan books across four distinct dimensions.

Erosion of pricing power. AI-enabled tools allow enterprise customers to automate tasks previously supplied by expensive SaaS subscriptions. When customers can replicate core functionality using an AI agent at a fraction of the per-seat cost, software vendors lose the pricing leverage that made their ARR so attractive to lenders. For credit investors, the implication is straightforward: if software vendors face greater pressure on pricing, recurring revenue may prove less resilient than lenders originally assumed.

Shorter competitive half-lives. The terminal value of a software franchise was historically assumed to be durable. That assumption is being directly challenged. Concerns around AI-driven disruption hit the software sector hard in early 2026, triggering a sharp repricing in companies that had previously been viewed as having durable competitive advantages. Even industry participants are conceding the point: the Ares Capital CEO acknowledged on a February shareholder call that AI would disrupt many software companies with single-function products that produce content or analyse and visualise data. AI-native development tools such as Lovable and Replit further suggest that, in parts of the application layer, software creation is becoming faster and cheaper. Competitive advantage may erode more quickly than lenders previously assumed.

Rising costs to stay competitive. Even where revenue remains relatively stable, software companies face meaningful cost pressure. Competing in an AI-first environment requires continued investment in model integration, proprietary data pipelines, computing capacity, and governance infrastructure. These expenditures compress margins well before revenue declines become visible. For lenders, weaker margins reduce free cash flow available for debt service even when headline revenue appears resilient.

Uncertain recovery values. Software lenders typically rely more on enterprise value than on hard collateral. If AI structurally compresses software-sector valuation multiples, as the public equity sell-off suggests, recovery rates in a default scenario could fall well below the levels assumed in original underwriting. UBS has warned that in an aggressive disruption scenario, default rates in U.S. private credit could climb to 13%, significantly higher than stress projections for leveraged loans and high-yield bonds, estimated at around 8% and 4% respectively. This is particularly consequential in software lending because recoveries depend on the durability of future earnings rather than tangible assets that can be readily sold.

Emerging Signs of Credit Reassessment

The most important structural problem is that private credit portfolios are slow to reflect reality. Loans are held at par on quarterly marks with minimal secondary market trading. When the AI disruption thesis was first raised in early 2026, the counterargument was that stress hadn't yet appeared in actual credit data. That counterargument is now weaker.

The most consequential development since then is Apollo. In March 2026, Apollo Global Management informed investors that withdrawal requests for its ADS BDC had reached 11.2% of total outstanding shares for the quarter. With over $1.5 billion seeking the exit, Apollo enforced a 5% quarterly redemption cap, meaning investors who sought to liquidate received just 45 cents on the dollar for the deferred portion of their requests. Software is the largest single sector in the Apollo Debt Solutions BDC, at more than 12% of the portfolio. Apollo was not alone. Blue Owl and KKR also implemented redemption gates during the same period. The dynamic has the hallmarks of a self-reinforcing feedback loop: news of increased withdrawal requests caused additional withdrawal requests, amplified by the opaque nature of these strategies.

Blackstone's BCRED is also showing strain. BCRED recorded a 0.4% loss in February 2026, its first monthly loss in three years, as the fund marked down loans including debt linked to SaaS company Medallia. JPMorgan's earlier decision to re-mark software loans signals that banks financing private-credit portfolios are reacting faster than the portfolios themselves. Reuters Breaking views offered another warning: software loans now constitute nearly a third of distressed credit trading below 80 cents on the dollar.

The headline credit data reinforces concern. Fitch Ratings recorded a private credit default rate of 9.2% in 2025, a record even before AI disruption has fully manifested in cash flows. Notably, Fitch recorded no software defaults in 2025. This absence may reflect genuine resilience, but it may equally reflect that stress has not yet had time to materialise: with 30% of software loans maturing by 2028, the reckoning may simply be deferred rather than avoided.

Emerging Signs of Credit Reassessment

The most important structural problem is that private credit portfolios are slow to reflect reality. Loans are held at par on quarterly marks with minimal secondary market trading. When the AI disruption thesis was first raised in early 2026, the counterargument was that stress hadn't yet appeared in actual credit data. That counterargument is now weaker.

The most consequential development since then is Apollo. In March 2026, Apollo Global Management informed investors that withdrawal requests for its ADS BDC had reached 11.2% of total outstanding shares for the quarter. With over $1.5 billion seeking the exit, Apollo enforced a 5% quarterly redemption cap, meaning investors who sought to liquidate received just 45 cents on the dollar for the deferred portion of their requests. Software is the largest single sector in the Apollo Debt Solutions BDC, at more than 12% of the portfolio. Apollo was not alone. Blue Owl and KKR also implemented redemption gates during the same period. The dynamic has the hallmarks of a self-reinforcing feedback loop: news of increased withdrawal requests caused additional withdrawal requests, amplified by the opaque nature of these strategies.

Blackstone's BCRED is also showing strain. BCRED recorded a 0.4% loss in February 2026, its first monthly loss in three years, as the fund marked down loans including debt linked to SaaS company Medallia. JPMorgan's earlier decision to re-mark software loans signals that banks financing private-credit portfolios are reacting faster than the portfolios themselves. Reuters Breaking views offered another warning: software loans now constitute nearly a third of distressed credit trading below 80 cents on the dollar.

The headline credit data reinforces concern. Fitch Ratings recorded a private credit default rate of 9.2% in 2025, a record even before AI disruption has fully manifested in cash flows. Notably, Fitch recorded no software defaults in 2025. This absence may reflect genuine resilience, but it may equally reflect that stress has not yet had time to materialise: with 30% of software loans maturing by 2028, the reckoning may simply be deferred rather than avoided.

Why Systemic Stress Has Not Yet Materialised

Despite these warning signals, several structural features temper the immediate threat of a disorderly credit event.

Senior secured protections remain robust. BXSL's portfolio, for example, was 97.6% first-lien with an average 50.5% loan-to-value at origination, implying that considerable equity would have to be destroyed before lenders are impaired. Private equity sponsors also have powerful incentives to support portfolio companies through turbulence, protecting their own reputations and future fundraising access.

The broader credit environment is not in distress. Investment-grade spreads remain near historic tights, and corporate balance sheets outside software are generally healthy. Credit impairment requires sustained earnings deterioration, which takes time even if AI is compressing margins. Outcomes will also vary sharply based on where a software company sits in the technology stack. Mission-critical infrastructure with proprietary data advantages faces a fundamentally different risk profile than single-function SaaS products that AI can plausibly substitute within a product cycle.

The liquidity picture, however, deserves more weight than it previously received. The suspension of redemptions across several large non-traded vehicles has exposed how appraisal-based valuations, limited secondary-market liquidity, and concentrated exposures in enterprise software can interact under pressure. The SEC has signalled it is closely monitoring the gating mechanisms of non-traded BDCs, and broader disclosure requirements are likely to follow. A wider redemption run could force funds to sell assets at depressed prices, amplifying stress well beyond the software sector itself.

What Investors and Regulators Should Do Now

The debate is not about imminent default cascades. It is about whether the spread earned on software loans today adequately compensates for technology-induced uncertainty that did not exist at origination, and whether the liquidity architecture of private credit can absorb a repricing that public markets are already executing in real time.

Stress-test cash flows for AI disruption. Traditional ARR-based underwriting assumes stable churn and pricing. Portfolio models should now incorporate faster customer erosion scenarios, margin compression pathways and lower recovery value assumptions. The absence of software defaults in 2025 should not be read as validation of legacy models. It should be read as a lag.

Scrutinise loan structures, sponsor incentives, and fund liquidity terms. Covenant flexibility, maturity profiles and equity cushion adequacy matter more than ever. So does the liquidity architecture at the fund level. The Apollo, Blue Owl, and KKR gating episodes demonstrate that retail investor behaviour can create redemption pressure that forces asset sales at exactly the wrong time, further compressing recovery values. Lenders should demand shorter tenors where possible, step-up coupons tied to revenue retention, and AI-specific performance covenants that trigger review rights before formal default.

Diversify away from intangible-asset concentration. Industry participants have noted that AI-exposed software is just the first fault line, with the real risk sitting across any highly-leveraged, rate-sensitive borrower whose business model was priced for free money. Portfolios with 25 to 30% software exposure carry meaningful sector-specific risk that warrants active rebalancing, not patient monitoring.

For regulators, the priority is transparency. Fitch's record default rate, JPMorgan's loan re-markings, and the Apollo gating episode all illustrate how risk can build unseen inside opaque, illiquidity-driven portfolios. More frequent valuation updates, clearer sector-level disclosure requirements and closer monitoring of interconnections between private credit funds and their bank financing providers are all now overdue rather than merely warranted.

Conclusion:

Based on the evidence available in early 2026, private credit lenders appear to be underpricing AI-related risk in their software portfolios. This is not because investors disbelieve that AI disruption will occur. Most acknowledge it as a serious trend. Rather, it stems from structural lags in marking illiquid loans and from underwriting models built for a world where software companies were assumed to have longer-lasting competitive advantages.

An immediate credit crisis is not the base case. Many software borrowers remain mission-critical, sponsor support is intact, and first-lien loans still provide meaningful protection. But the events of Q1 2026, including Apollo's gate, BCRED's first monthly loss in three years, industry-wide redemption pressure, and the first concrete loan markdowns, suggest that the gap between where private credit valuations sit and where public markets have already repriced is narrowing, and not comfortably. The real danger remains complacency: assuming that the durability premium historically earned on software ARR still holds when the technology underpinning that ARR is itself being disrupted. Illiquidity is not an excuse for analytical stasis. Dynamic stress-testing, sharper sector underwriting and deliberate diversification are the only credible responses to a tech shock that moves faster than a loan book can reprice.

Sources:

BIS Quarterly Review (March 2026); Bloomberg Opinion (March 2026); CNBC (February and March 2026); Financial Content / MarketMinute (March 2026); LPL Research (March 2026); Prime Buchholz (February 2026); Morningstar (March 2026); Corbett Road Q1 2026 Market Review; Reuters; S&P Global Ratings; UBS CIO; Alternative Credit Council; Blackstone BXSL and BCRED quarterly disclosures; Fitch Ratings; Octus BDC analysis 2025; Kroll–StepStone private credit benchmark.