Private credit: software debt faces the AI test

Artificial intelligence could make some software cheaper and easier to replace while weakening the loans that financed it. BIS research documents substantial private-credit exposure to software. It does not establish a wave of defaults caused by AI. Understanding the risk means following subscription revenue into a lender’s balance sheet, then separating observed developments from scenarios that remain untested.

A software company can keep its customers, continue selling its product and still struggle to pay its creditors. Revenue only needs to fall faster than expenses while the debt stays in place. The product works. The financing becomes less comfortable.

This is a less obvious side of the debate over AI funding. The exposure extends beyond the prospective returns on new data centres. It includes established businesses that borrowed against the expected durability of their revenue. An innovation that helps their customers can reduce the money available to service their debt. That is an economic possibility, not a description of an industry-wide collapse already under way.

The documentary cutoff is 16 September 2026. Most aggregate figures refer to end-2025. The company case study uses accounts for 30 June 2026, released on 5 August, and a financing transaction announced on 4 September. Those dates describe different things.

The loans behind the subscriptions

Research published by the Bank for International Settlements on 14 September 2026 puts technology companies at 44% of outstanding US direct loans in its 2025 sample, up from 22% in 2010. Technology is a broad category here, extending well beyond software publishers or AI businesses. BIS study and methodology.

Private credit, as discussed here, consists of financing negotiated outside public bond markets, usually with non-bank lenders. A borrower deals with a fund or a small group of lenders rather than a dispersed group of bondholders. The contract can be tailored to the business. In exchange, the lender accepts an asset that may not be easy to sell quickly. The IMF explains this structure and its differences from bank lending and public markets in its April 2024 study.

Business development companies, or BDCs, provide a useful window into software exposure. These US financing companies disclose detailed investment information. Some are publicly traded; others are not. They are not the whole private-credit market, and their portfolios do not consist exclusively of directly originated loans. SEC overview.

A BIS bulletin published on 14 July 2026 estimates that BDCs had lent about $115 billion to software companies at end-2025, roughly one fifth of their lending. This is an exposure measure, not an expected loss or a bill attributable to AI. Bulletin No 128.

The 44% technology share and the roughly 20% software share should not be turned into a trend. The former concerns a broad technology category within a direct-lending sample; the latter concerns software within BDC portfolios. Different lenders, classifications and denominators cannot be made comparable simply by placing percentages next to each other.

Expected renewals become borrowing capacity

Consider a business-management application sold on an annual subscription. When a customer renews, the publisher receives another year of revenue without having to build the product again from scratch. Contracts spread across many users can make revenue relatively predictable. That predictability helps a lender assess debt service.

Several different measures still need to be kept apart. Revenue records sales under the applicable accounting rules. Cash receipts depend on when customers actually pay. Annual recurring revenue, or ARR, extrapolates a defined set of contracts or subscriptions over a year, using the company’s stated methodology. It is neither profit nor money already sitting in a bank account.

Against those sales stand salaries, hosting, distribution and the spending required to keep the product competitive. Then come interest, taxes, investment and changes in working capital, including the timing gap between customer receipts and supplier payments. A renewal is welcome. The lender still needs to know how much money it leaves after these outflows.

This distinction matters when financing relies on the value of an operating business. A NBER working paper revised in May 2026 describes direct lenders’ specialisation in this form of lending, including to intangible-intensive companies. Their expertise lies in assessing a business’s ability to generate income, rather than simply pricing machinery that could be sold.

There is a sound economic reason for doing this. A customer base, a product embedded in business processes and a team that can maintain it may have substantial value. But that value depends on continued operations. It is therefore sensitive to some of the same commercial changes as the cash used to pay interest.

A secured loan does not lock in a recovery price

The word “secured” can create an overly reassuring impression. Collateral gives a creditor rights over assets or the proceeds of their sale. Seniority determines its place in the repayment queue. Neither fixes the amount that will eventually be recovered. The NBER paper on direct lending centres on the distinction between going-concern value and liquidation value.

Imagine a company whose most valuable asset is a software product used by thousands of customers. A buyer may pay for the relationships, contracts and expertise. If that buyer expects fewer renewals or permanently lower prices, the bid falls even though the creditor’s legal rights remain intact.

The same shock can therefore hit a loan twice. Lower margins first weaken the borrower’s payment capacity. A lower sale value then reduces potential recoveries if restructuring becomes necessary. Revenue and collateral are not necessarily independent layers of protection. In this example, both partly depend on the same customers.

This does not make security worthless. Contractual protections may permit early intervention. A lender may negotiate fresh equity, alter maturities or arrange a sale. But the strength of a legal claim must be distinguished from the economic value of the assets behind it. The IMF discusses the importance of loan terms, monitoring and restructuring in its private-credit study.

AI does not have to replace the entire company

A vulnerability test does not require the sudden disappearance of software. Much smaller changes could matter.

A customer might buy fewer seats when some work is automated. A competitor might offer a similar function at a lower price. An incumbent might retain users but have to include more services in an unchanged subscription price. In each scenario, the lender cares about the revenue and margin left behind, not the number of impressive demonstrations an AI model can produce.

Nor do customer savings automatically become publisher losses. The incumbent may cut its own costs, improve the product, attract new customers or sell additional functions. Usage-based pricing may replace per-seat billing. The outcome depends on bargaining power, competition and the value of the service. It needs to be measured at company level.

The IMF’s April 2026 report identifies possible defences for established providers, including data, distribution, security requirements, regulatory constraints and the cost of switching systems. A convincing demonstration does not establish that a service can replace a production system used by a business.

The relevant question is therefore: does higher productivity strengthen the borrower’s margin, or is most of the benefit passed to customers through lower prices? Without evidence on that balance, there is a risk hypothesis rather than a solvency diagnosis.

A small revenue decline can exhaust the remaining cushion

Take an entirely fictional company, measured in monetary units. It collects 100 in revenue, pays 70 in operating costs and has 200 of debt carrying 10% annual interest. Interest is therefore 20. That leaves 10, before taxes, capital expenditure, working-capital movements and principal repayments.

Now let revenue fall to 90, with no cost reduction. The operating surplus declines from 30 to 20. It still covers interest, but the remaining cushion disappears. Revenue has fallen by 10%; the surplus before interest has fallen by one third. The debt is unchanged.

This break-even point is not a universal default threshold. Opening cash, shareholder support or a distant maturity could buy time. Conversely, the spending excluded from the example could make the situation more difficult. The exercise isolates one mechanism: a modest sales decline can have a much larger effect on the amount left after fixed costs and interest.

Adaptation can reverse the outcome. With revenue of 90 and a 20% reduction in the original cost base of 70, operating costs fall to 56. The operating surplus becomes 34, compared with 30 initially. After the same interest bill of 20, the company retains 14. Falling revenue alone therefore does not establish deteriorating credit quality.

l0g LAB · FICTIONAL SCENARIOS

How much is left to service the debt?

Starting point: 100 in revenue, 70 in costs and 200 in debt. Change the assumptions to follow the effects on cash and debt. Calculations run locally; inputs are not collected.

Scenario assumptions

Starting pointCommercial pressureAdaptationInterest deferred

Revenue decline

0 %

Percentage of the initial 100 in revenue.Reduction in starting costs

0 %

Percentage of the initial cost base of 70.Annual interest rate

10 %

Applied to opening debt of 200.Share of interest capitalised

0 %

Added to debt at year-end (PIK).

Revenue collected

100

Operating costs paid

70

Surplus before interest

30

Total interest

20

of which paid in cash

20

of which added to debt

0

Remainder after cash interest

10monetary units

Year-end debt

200monetary units

Total interest coverage

1.5times

Remainder after all interest

10monetary units

The surplus covers interest in this model. The remainder is not a complete free-cash-flow measure.

One-year model excluding taxes, capex, working-capital changes, principal repayments and opening cash. No intra-year compounding. Slider bounds are illustrative. No default probability is calculated.

The simulator separates revenue, costs, the interest rate and the share of interest capitalised rather than paid in cash. It represents no actual borrower. Revenue is assumed to be collected and operating costs paid within the year. Interest is calculated on opening debt; unpaid interest is added to debt at year-end, without additional compounding within the period.

Two distinct readings of the stock market

The bulletin published in July uses a loan sample ending in the fourth quarter of 2025. Its findings of benign loan performance and limited spread differentiation concern that window. Its equity-market chart extends further and compares price-to-dividend ratios, using two groups of 17 BDCs classified by historical software exposure. That is not a direct measure of the difference in their recent share-price performance. Bulletin sample definition and Graph 3.

A different BIS study, published on 16 March 2026, finds underperformance among BDCs more exposed to software as a service. Between 1 October 2025 and 5 March 2026, the high-exposure group underperformed the low-exposure group by around 5 percentage points. Its 37 BDCs are split according to whether their software exposure was above or below the median in the third quarter of 2025. Quarterly Review, Box B and Graph B1.C note.

The studies therefore use different indicators, groups and observation windows. Publication dates alone cannot reconcile them. We cannot splice them into a single series or conclude that all investors remained indifferent to software risk until summer 2026.

A fall in a BDC’s share price is not, however, a default by one of its borrowers. The price may reflect concerns about future losses, distributions, fees or the liquidity of the shares. Net asset value, or NAV, is a separate measure: the recorded value of assets less liabilities, divided by the share count when stated per share. The stock-market price can diverge from it. SEC explanation of publicly traded BDCs.

These signals are useful because they describe different stages. A market price responds quickly to expectations. A default records a failure to meet payment obligations or another contractual condition. Between the two, accounts may show markdowns or restructurings. Treating these as equivalent would turn an anticipated loss into a realised one.

Inside OTF’s accounts

Blue Owl Technology Finance Corp., traded as OTF, provides a documented case. It should be distinguished from its manager and from other Blue Owl vehicles. Its quarterly accounts are neither a representative sample of all software lending nor an independent audit of every borrower.

In its results release furnished with a Form 8-K on 5 August 2026, OTF reports:

Measure	31 March 2026	30 June 2026
Net asset value per share, US dollars	16.49	16.48
Non-accrual investments, % of portfolio at cost	0.3%	0.6%
Non-accrual investments, % of portfolio at fair value	0.1%	0.1%

Unaudited quarterly figures reported by OTF. Percentages measure positions at each date, not annual default rates. The portfolio contains several types of investment.

Non-accrual means interest is no longer recognised through normal accrual accounting, notably because collection has become uncertain. The cost-based percentage uses the relevant investments’ cost basis; the fair-value percentage uses their current estimated values. An investment that has already been heavily written down can make up a small part of the latter ratio. The lower number is not automatically the more reassuring one.

An almost unchanged quarterly NAV per share does not isolate the performance of software loans either. It combines different investments and reflects other movements, including distributions and share repurchases. It is not a shareholder’s total return.

Another detail requires following the money. OTF’s second-quarter tables record $25.224 million of PIK interest and $17.062 million of PIK dividends. Their sum, $42.286 million, equals 12.5% of the quarter’s $338.032 million in total investment income, according to our calculation. The denominator is recorded investment income, not portfolio size or net income. Statement of operations in the same results.

Payment in kind, or PIK, compensates an investment without an immediate cash payment. Interest can increase the amount of a debt claim; distributions can take the form of additional securities. These categories should be distinguished. Their presence does not establish borrower distress: the arrangement may have been part of the original contract.

For a lender, the useful questions are whether PIK was planned, whether it has just been negotiated to avoid a cash outflow and whether the final amount due remains repayable. For a reader of the accounts, the principle is simpler: recognised income need not mean cash received during the quarter. Timing differences can also affect non-PIK income.

The simulator’s “Interest deferred” scenario makes the distinction visible. With 75 of revenue and 70 of costs, the company generates 5 before interest but owes 20. Capitalising 75% of the interest leaves a cash payment of 5, balancing this part of its cash budget. Debt nevertheless rises from 200 to 215. The economic burden remains; settlement has moved into the future.

Different managers may be lending to the same businesses

Another concentration sits beneath the sector labels. At end-2025, about 60% of BDC software loan volume went to borrowers financed by at least seven distinct BDCs, according to the BIS. This is a volume-weighted figure, not a count of companies. Graph 4 and its definition.

For a borrower, having several lenders can diversify funding sources. For an investor spreading money across several vehicles, the same companies may reappear underneath each allocation. Seven vehicles do not guarantee seven independent sets of borrowers; some may also share a manager.

This does not multiply a debt by seven. Each lender holds a specified exposure. Any eventual loss is allocated according to the claims, collateral and repayment priorities. The issue is correlation: several investments can be affected by the same event at the same time.

Overlap also goes beyond identical borrower names. Two different software companies may depend on the same customer budget or pricing model. A portfolio containing many names can still share a sensitivity to falling prices. The sources used here do not provide a complete inventory with which to measure that sensitivity for every investor.

Funding structure determines how quickly stress spreads

A borrower loss and a liquidity problem at its lender are different events. The investment vehicle’s funding structure helps determine whether one can contribute to the other.

An investor selling shares in a listed BDC normally finds another buyer in the market. The vehicle does not automatically return the seller’s capital. A falling share price therefore does not, by itself, require a loan sale the next morning. Some non-traded vehicles, by contrast, offer periodic repurchases that are limited and may be suspended under their terms. The SEC emphasises these restrictions. This is not deposit-like liquidity.

A NBER working paper published in June 2026 examines semi-liquid private-credit funds. In its sample, repeated repurchase requests can exceed cash buffers and repayments from borrowers. Limits on withdrawals do not make the underlying assets immediately saleable.

Consider a fund receiving fewer new subscriptions while repurchase requests rise. It may draw down cash, use a financing line, sell investments or limit repurchases under the applicable rules. Each response has a cost or constraint. Defaults are not a prerequisite: expectations of losses can change investor behaviour before a borrower misses a payment.

Leverage adds another layer. In a fictional balance sheet, a vehicle owns 100 of assets financed by 50 of debt and 50 of equity. If assets lose 10 in value and debt remains at 50, equity falls to 40, a decline of 20%. This is arithmetic, not a description of any particular BDC.

Connections to banks and other creditors also matter. In research published on 12 March 2026 using end-2024 observations, the Office of Financial Research documents funding received by private-credit participants. Those exposures are neither realised losses nor lending exclusively to software. They identify potential transmission routes rather than proving that those routes will carry a shock.

The IMF’s April 2026 report provides an important boundary: rising defaults alone need not produce financial instability. Funding maturities, withdrawal requests, contractual protections and liquidity reserves affect the size of any spillover. A contagion argument needs to establish the links in the chain.

Evidence that could defeat the bearish scenario

The strongest counterargument is economic. Established providers may be well placed to integrate AI because they already possess customer relationships, data and distribution processes. If their costs fall faster than their prices, debt-service capacity improves. The adaptation example demonstrates this without requiring sales growth.

Borrower selection matters as well. A lender may avoid easily replicated functions, prefer stronger contracts or finance businesses with less debt. Software exposure alone does not establish an equal vulnerability across all loans. Assessing the borrowers is more informative than treating a sector label as a diagnosis.

Access to funding offers another test. On 4 September 2026, OTF announced a $150 million senior unsecured note placement, carrying 7.60% interest and maturing in 2032. The transaction is reported by the issuer, not an independent assessment of credit quality. It nevertheless establishes that this vehicle obtained term funding on those conditions. Transaction announcement.

It does not show that funding is easy throughout private credit. Nor can 7.60% be called cheap without an appropriate comparison. But, for this issuer on this date, the transaction contradicts a claim that capital markets were completely closed. Evidence that does not fit the adverse scenario belongs in the analysis too.

The questions the accounts need to answer

The next step is not to count AI announcements. It is to match revenue, margins, debt and payments for the same borrowers.

Customer loyalty needs a precise definition. Retention measured by customer count can remain high while revenue falls. Net revenue retention, which includes expansions and reductions within the existing customer base, tells a different story from a measure excluding additional sales. Comparisons must retain the published definition and a consistent population over time.

Adjusted earnings then need to be reconciled with reported accounts and cash receipts. Removing an expense from a performance measure does not necessarily remove it from a bank statement. Announced future savings may support a turnaround scenario, but they are not savings already delivered.

Finally, capitalised interest, restructurings, non-accrual investments and valuations need to be tracked with their denominators intact. Following a consistent loan cohort would reveal more than a headline ratio that changes as a portfolio grows or sells assets. Attributing deterioration to AI would also require distinguishing its effects from interest rates, an economic slowdown or excessive initial leverage.

The documents used here establish substantial software exposure, market concerns by early 2026 and identifiable transmission mechanisms. They do not establish an aggregate private-credit loss caused by AI. OTF’s accounts illuminate particular financial items; they do not substitute for an analysis of every loan held by OTF or its competitors.

The crucial evidence will emerge as contracts renew. Debt was extended because revenue looked durable enough to service it. AI can change that durability in either direction. Retained margins, actual cash receipts and the recoverable value of the business will provide the meaningful test.

Sources and method

The figures are original l0g designs based on published values or explicitly conceptual diagrams. No annual observations have been interpolated. The simulator uses fictional assumptions only and runs locally without collecting inputs. The PIK ratio is (24.969 + 0.255 + 13.657 + 3.405) / 338.032 × 100 = 12.509%, rounded to 12.5%. These amounts are in millions of dollars; the source accounts report them in thousands. Issuer publications are identified as such; promotional management comments are not adopted as independent conclusions.

Further reading: our private-credit guide, the analysis of AI debt and sovereign borrowing, and the definitions of BDC and PIK.

The company case uses the unaudited results furnished with OTF’s Form 8-K. It does not claim to reconstruct the full Form 10-Q or identify every borrower shared by multiple BDCs. BIS samples, OTF’s figures and research on semi-liquid funds remain separate populations.

01. BIS / BRI. Financing the digital economy: the role of private credit. 2026-09-14. US direct loans; broad technology classification. The 22% and 44% figures are portfolio shares, not default rates.

02. BIS / BRI. AI disruption in private credit: exposure to software firms in BDCs. 2026-07-14. $115bn and about one fifth of lending at end-2025; 60% of software loan volume goes to borrowers with at least seven BDC lenders.

03. NBER. The Lending Technology of Direct Lenders in Private Credit. 2025-11 ; revision 2026-05. Young Soo Jang, Dasol Kim, Amir Sufi and Xiangyu Chen. Going-concern-value lending; working paper.

04. BIS / BRI. Markets recalibrate amid shifting currents, Box B: Private credit’s software lending meets AI disruption. 2026-03-16. Roughly 5 percentage points of underperformance for high-SaaS-exposure BDCs. Different periods and market measures from the July bulletin.

05. Blue Owl Technology Finance Corp.. June 30, 2026 Financial Results, Exhibit 99.1 to Form 8-K. 2026-08-05. OTF NAV, non-accrual and PIK income. Income tables are in USD thousands. Issuer reporting, not an independent audit of borrowers.

06. Blue Owl Technology Finance Corp.. Closes $150 Million Private Placement of Senior Unsecured Notes. 2026-09-04. $150m, 7.60% coupon, maturity in 2032. Debt raised, not profit or net cash inflow.

07. SEC / Investor.gov. Publicly Traded Business Development Companies (BDCs). Accessed 2026-09-16. Structure of listed vehicles. A stock-market sale is not an automatic repurchase by the fund.

08. SEC / Investor.gov. Investor Bulletin: Non-Publicly Traded Business Development Companies (BDCs). 2024-12-13. Limited liquidity in non-traded vehicles; repurchase terms vary and may be restricted.

09. IMF / FMI. Global Financial Stability Report, April 2026, Chapter 1. 2026-04. Private credit, repurchases and limits to systemic transmission; potential software defences against AI disruption.

10. NBER. The Fragility of Semi-Liquid Private Credit Funds. 2026-06. Chuck Fang, Itay Goldstein and Yao Zeng. Liquidity needs from repeated periodic repurchases; no simulation result extrapolated to all BDCs.

11. Office of Financial Research. Measuring Counterparty Exposures to Private Credit. 2026-03-12. Bank and non-bank funding of private credit. These exposures are not losses or exclusively software loans.

12. IMF / FMI. The Rise and Risks of Private Credit, GFSR April 2024, Chapter 2. 2024-04-16. Structural background only. No 2024 market size is presented as a September 2026 observation.