The Software Test: AI Risk Inside BDC Portfolios
Software lending offers a concrete test of how private-credit risk can look manageable in current data while becoming more uncertain underneath. This chapter follows why business development companies (BDCs) built such large software exposures, how generative AI could affect the cash flows behind those loans, and why similar spreads or low delinquencies do not settle the question. The aim is to distinguish sector exposure from realized impairment and borrower-level diversification from concentration across lenders.
Why software became attractive to direct lenders
Software had many characteristics that made it appealing to direct lenders during the past decade: recurring subscription revenue, high margins, and relatively low capital needs. Those features could make cash flows appear predictable and leave more operating income available for debt service. They also fit a private-credit strategy built around multi-year loans to small and mid-sized companies. A lender could underwrite a business whose value rested more on recurring customer relationships than on a large stock of physical assets.
BDCs provide an unusually visible view of this allocation. They are U.S. private-credit intermediaries that generally lend to small and mid-sized firms, and the BIS estimates that they originate about one-fifth of U.S. direct loans. At the end of 2025, roughly 170 BDCs reported their holdings loan by loan each quarter to the Securities and Exchange Commission, covering about $550 billion in assets. That reporting does not make BDCs a census of private credit, but it makes their sector choices and shared borrowers easier to study than those of many other funds.
The software exposure is large within that window. BDCs had lent around $115 billion to software firms, about one-fifth of their lending and more than 80 percent of their technology portfolios. The exposure grew as lenders treated software’s historical revenue model as a source of dependable cash flow. The Federal Reserve’s report also describes software as the largest sector in private-credit portfolios, reflecting elevated private-equity activity. These observations explain the allocation; they do not prove that software borrowers were safer than every other borrower or that historical stability will persist.
Generative AI changes the uncertainty around that historical model in two related ways. First, an AI tool may substitute directly for an existing software product, allowing a customer to perform a task without paying for the prior product. Second, AI can lower development costs and reduce barriers to entry, allowing new competitors to offer similar functions or forcing incumbent firms to compete at thinner margins. A borrower may therefore retain customers and still generate less cash if pricing power erodes. The risk is not limited to a sudden collapse in revenue; it can be a slower deterioration in growth, retention, or margins.
The timing matters because a private loan can run for several years. A lender may originate a loan when the borrower’s reported earnings and renewal rates look strong, while the competitive effect of new technology develops before the loan reaches maturity. The loan’s contractual cash flows do not update instantly when the borrower’s long-term business model becomes less certain. Nor does a current payment tell us whether future refinancing will be easy.