Private Credit And Equity Face Warnings On AI Lending Concentration

Private Equity and Private Credit Interconnections

Rather than viewing deal flow as an external reality to which lenders passively react, the Carlyle authors explain that underwriting guidelines actively shape which transactions private equity sponsors pursue. As Carlyle points out, private lenders do not operate in a vacuum; their investment opportunity set relies on sponsored deal flow created by private equity firms seeking convex equity upside. Direct lenders’ willingness to relax underwriting criteria and widen credit boxes directly enabled private equity sponsors to bid higher valuation multiples and target more software acquisitions than they otherwise could have. As lenders pivot toward financing data centers, power generation, and GPU clusters, concentrating in AI risks compounding the very vulnerabilities now unspooling across legacy software portfolios.

Earlier in the summer, Sycamore Tree Capital Partners released a note explaining that AI infrastructure financing “has migrated from hyperscaler balance sheets into external credit markets at a scale and pace with few precedents. To date, the focus has been largely an equity debate: whether the return justifies the capital. The more consequential issue for credit investors is less examined: the composition of financing has changed, and the risk increasingly sits where it is less visible, less liquid and less able to reprice. When the cycle turns, repricing will surface first in traded credit, and second in the private secondary market.”

The SaaS Underwriting Trap and the Fallacy of Uncorrelated Cash Flows

For over a decade, private lenders treated SaaS investments as the ideal credit asset. SaaS companies offered scalable equity upside for buyout sponsors alongside stable, recurring cash flows for lenders. Direct lenders constructed tight “credit boxes” designed to insulate portfolios from cyclical economic downturns, operating under the assumption that software subscriptions were effectively non-discretionary.

Generative AI shattered that premise. The emergence of coding-capable large language models (LLMs) introduced technological obsolescence as a systemic, sector-wide risk. Rather than suffering from a temporary macroeconomic dip, legacy software companies suddenly faced secular headwinds—including reduced legacy subscription budgets, falling maintenance spending, and direct disintermediation. Because lenders applied similar underwriting models across the entire sector, what appeared to be a diversified pool of independent software loans turned out to be a highly correlated, concentrated bet on a single technological paradigm.

Major credit rating agencies have reinforced these concerns:

• Pricing Divergence and Maturity Walls: Fitch Ratings has highlighted a sharp divergence in loan secondary market pricing between software credits and non-tech sectors, warning that software borrowers face an imminent, challenging $200 billion+ “maturity wall” that traditional syndicated loan markets are increasingly reluctant to refinance.
• Underwriting Against Speculative Demand: Moody’s Investors Service has warned against underwriting specialized AI infrastructure loans against projected long-term compute demand rather than binding, established operational contracts. Moody’s noted that if tenant renewal rates drop as hardware rapidly depreciates, these facilities face heightened risk of sudden multi-notch credit rating downgrades and default.
• Capital Structure Impairment: S&P Global Ratings has noted that Distressed Debt Exchanges (DDEs) and secondary market markdowns in tech and software now account for a disproportionate share of total corporate credit stress, linking this trend to software vendors losing pricing power against AI alternatives.

Unlike cyclical distress—where lenders can extend loan maturities and wait for an economic recovery—software’s current distress is structural. Refinancing legacy software debt requires meaningful de-risking, actual loan paydowns, or fresh equity injections rather than simple “amend-and-extend” concessions.

Empirical Proof: PitchBook Data Confirms Private Credit Stress

The structural warnings detailed in Carlyle’s analysis are no longer theoretical. Market data, published by PitchBook LCD in an article titled “Nearly 10% of top BDCs’ investments tied to borrowers with a distressed position” by Sebastian Kian, acts as real-time, empirical proof of Carlyle’s thesis:

• Spreading Portfolio Distress: As of Q2 2026, nearly 10% of total debt investments held by the largest Business Development Companies (BDCs) were tied to borrowers with at least one distressed debt position.
• Surge in Markdowns: Total distressed debt (defined as positions with a Fair Value–to–Cost ratio below 80%) jumped 64% year-over-year to $6.8 billion, representing 5.2% of all BDC debt assets.
• Software as the Primary Driver: Software loans and second-lien tranches accounted for 89% of the year-over-year increase in distressed debt across these major portfolios.
• Contagion Across Capital Structures: Borrowers holding even a single distressed loan tranche proved highly vulnerable to broader operational degradation across their remaining capital structures.

This sharp divergence between software debt and the broader credit market validates Carlyle’s primary thesis: concentrated exposure to a single, rapidly evolving asset class creates severe tail risk when technological assumptions shift.

Why AI Infrastructure Credit Risk Is Harder and More Speculative

Transitioning from SaaS to AI compute infrastructure does not escape these concentration risks—it amplifies them. While SaaS at least offered predictable subscription revenue, AI infrastructure lending introduces far more speculative dynamics:

• Unsettled Value Distribution: It remains entirely uncertain how long-term profits will be divided among chip manufacturers, foundation model developers, infrastructure owners, and application layer software.
• Accelerated Depreciation & Hardware Obsolescence: Rapid GPU iteration and evolving data center architectures risk accelerating the physical and functional depreciation of compute assets, rendering multi-year debt assumptions obsolete long before loans mature.
• Macroeconomic Interdependence: Unlike SaaS non-renewals, which remained low during past recessions, AI capital expenditure accounts for a massive share of total economic growth. An AI downturn would trigger widespread economic spillovers across construction, energy, and capital equipment sectors, causing AI credit defaults to correlate directly with broader macroeconomic downturns.

Why Private Credit Must Learn from Bank Regulatory Guidance

Big banks have no excuse for concentrating in any one sector or company. For decades, they have been told by the Federal Reserve, the Office of the Comptroller of the Currency, the Federal Deposit Insurance Corporation, and state bank regulators that concentration is practically a venial sin, if not a cardinal one.

While bank regulators do not supervise unlisted private credit funds, the regulatory framework governing traditional financial institutions offers a vital blueprint for alternative asset managers. The Office of the Comptroller of the Currency (OCC) outlines explicit standards for managing portfolio exposure in the “Concentrations of Credit” booklet of the Comptroller’s Handbook. I have worked in thirty countries, where I have often seen the OCC’s Comptroller’s Handbook translated into local languages and widely distributed to both on- and off-site bank supervisors and lenders.

While written for bank examiners, the guidance provides crucial risk management principles that private credit firms would be well-advised to adopt:

• Common Sensitivity Risk: The OCC defines a credit concentration as any pool of loans sharing common characteristics or sensitivities to financial or technological developments. At a certain threshold, common factors can cause even soundly underwritten loans to perform identically, threatening the institution’s core capital.
• The Delusion of Historical Performance: The OCC warns examiners that “a concentration can become a problem even when it has not proven problematic in the past.” The historical stability of software cash flows blinded lenders to emergent obsolescence; applying the same complacency to AI infrastructure risks similar outcomes.
• Board-Level Limits and Stress Testing: The OCC mandates that institutions establish quantitative concentration limits, conduct portfolio-wide stress testing under adverse scenarios, and maintain clear contingency plans to reduce exposure when thresholds are breached.

Setting Concentration Limits Before the Boom Accelerates Further

The broader financial ecosystem—including rating agencies, the Basel Committee on Banking Supervision, and investment banks—corroborates Carlyle’s call for discipline. Analysts at Morgan Stanley and JPMorgan estimate that the AI compute buildout will require $1.5 trillion to $3.2 trillion in capital, with private credit expected to absorb over $1 trillion. Depending on the benchmark used, $1 trillion represents about 40% to 60% of total global private credit assets under management (AUM).

Carlyle’s message is not that lenders should avoid AI infrastructure entirely. The sector might offer valuable opportunities to capture complexity premiums across data center construction, power financing, and structured equipment leases. However, alternative asset managers must institute explicit, firm-wide concentration limits before the buildout reaches its peak, rather than scrambling to manage correlated defaults after the capital has already been deployed.

And remember what I have been writing about for years. Nothing exists outside of the banking system. If the stress in private credit worsens, banks, and possibly taxpayers, will more than feel the pain.