Private credit fell in love with software for understandable reasons.

A good software company appeared to offer almost everything a lender could want.

Recurring subscription revenue.

High gross margins.

Low physical capital requirements.

Predictable customer renewals.

Growing free cash flow.

And, perhaps most importantly, a business model that seemed unlikely to change dramatically during the life of a loan.

That last assumption is now being tested.

Artificial intelligence is not merely changing how software is built. It is changing which software customers need, how many employees need seats, what customers are willing to pay for, how quickly competitors can reproduce features and whether an application remains a product at all.

Private credit has therefore encountered a risk it was not originally designed to price:

The Numbers Are Already Too Large to Ignore

The Bank for International Settlements estimates that outstanding direct lending to SaaS companies increased from less than $8 billion in 2015 to more than $500 billion by the end of 2025.

That represented approximately 19% of total direct loans.

Roughly one-third of private-credit funds had lent to SaaS companies.

The publicly traded BDC segment provides an unusually useful window into this otherwise opaque market.

BIS research published in July 2026 estimates that BDCs had lent about $115 billion to software companies, representing roughly one-fifth of their lending and more than 80% of their rapidly expanding technology portfolios.

The more surprising finding is what lenders have not done.

The BIS says uncertainty around software revenues caused by generative AI had not yet resulted in meaningful differentiation in the terms of these loans.

Loans to AI-related businesses were larger, yet spreads and maturities remained broadly similar to other private loans.

Average maturity was around 4.4 years in the BIS dataset.

That is the mismatch.

A Four-Year Loan Can Outlive a Software Moat

Credit analysts traditionally spend enormous effort estimating whether a company can meet interest and principal payments.

They model leverage.

EBITDA.

Cash conversion.

Covenant headroom.

Customer retention.

Default probability.

Those models work best when the underlying business evolves more slowly than the financing.

AI changes the clock.

Imagine lending to a software company in 2026 based on a four-year maturity.

The loan may still be outstanding in 2030.

But the borrower could face several generations of frontier models, open-source releases, autonomous agents and AI-native competitors before then.

A company does not need to disappear for the loan economics to change.

It may simply lose pricing power.

Its customers may need fewer seats.

A previously differentiated feature may become a commodity.

Implementation work may become automated.

Margins may compress.

Renewal negotiations may become harder.

The lender is therefore not merely underwriting default risk.

It is underwriting the half-life of the borrower’s economic moat.

ARR Is Recurring. That Does Not Mean It Is Durable.

Annual recurring revenue became one of the most powerful concepts in modern software investing.

And for good reason.

A customer paying every year is normally easier to underwrite than a customer who must be reacquired for every transaction.

But ARR measures recurrence.

It does not measure defensibility.

A subscription can renew every year right up until the moment a cheaper substitute changes the economics.

This distinction becomes critical in an AI environment.

Consider two software businesses, each with $200 million of ARR.

Company A owns deeply embedded regulatory workflows, proprietary data, mission-critical integrations and expensive migration requirements.

Company B sells a relatively horizontal productivity tool on a per-seat basis and depends on functions increasingly available inside general-purpose AI platforms.

The ARR number may be identical.

The credit duration is not.

The lender who treats those revenues equally is confusing a billing model with a moat.

The Software Credit Stack Needs to Be Rewritten

A better AI-era credit model should separate five layers.

1. Revenue recurrence

How much revenue contractually or behaviorally repeats?

2. Revenue duration

How long is that revenue economically defensible against technological substitution?

3. Workflow control

Does the software own an essential business process, or is it merely another interface sitting above data and models controlled elsewhere?

4. Data advantage

Does the company possess proprietary data that improves with use and cannot easily be recreated by competitors?

5. Switching friction

How costly is it for customers to move their data, processes, integrations, compliance procedures and employees somewhere else?

The last four may increasingly matter more than the first.

The Seat-Based SaaS Model Deserves Special Scrutiny

AI may also weaken one of software’s most successful pricing conventions.

Per-seat pricing assumes that the number of human workers using software is a reasonable proxy for the value delivered.

Agentic software complicates that relationship.

Suppose an accounting department falls from 100 employees to 60 because AI automates reconciliation and reporting.

A legacy software vendor charging per employee can become a victim of its customer’s productivity gains.

The customer may perform more work while purchasing fewer seats.

Alternatively, an AI-native competitor may charge for outcomes, transactions or completed work.

This means AI can disrupt a software company even when the underlying business function grows.

That is dangerous for credit models built around stable seat growth and renewal assumptions.

The Most Dangerous Software May Be the Software Closest to Labor

A useful way to think about AI disruption is to ask what a software company actually monetizes.

Some software monetizes infrastructure.

Some monetizes proprietary data.

Some monetizes regulation and compliance.

Some monetizes workflow coordination.

And some monetizes expensive human labor.

The closer the product sits to a task that a general-purpose model can perform directly, the more exposed the revenue may be.

This suggests a new private-credit hierarchy.

Public Markets Are Already Marking the Risk Faster

There is another clock mismatch.

Public equities reprice almost instantly.

Private loans do not.

BIS research shows software equities fell almost 30% between October 2025 and February 2026.

BDCs fell by roughly 10% over part of the same period, while their discounts to reported net asset value widened.

BDCs with greater software exposure subsequently underperformed those with lower software exposure.

That does not prove their loan books were impaired by the same amount.

It does reveal a timing difference.

The equity market can reprice the future today.

A private-credit portfolio may recognize the deterioration only gradually through quarterly valuations, covenant resets, amendments, refinancing events or realized defaults.

The Mark-to-Yesterday Problem

Mark-to-model is useful when assets rarely trade.

But rapid technological change introduces a particular problem.

The model is usually calibrated using historical relationships.

AI can change the relationships themselves.

Historical churn may no longer predict future churn.

Historical gross margins may not describe future pricing.

Past customer acquisition costs may not reflect competition from AI-native products.

A valuation model can therefore be mathematically precise and economically stale.

DN calls this the:

Mark-to-Yesterday Gap.

It is the difference between the speed at which economic reality changes and the speed at which the private asset’s reported valuation catches up.

Payment-in-Kind Can Buy Time Without Creating Cash

The Financial Stability Board has separately highlighted increased use of payment-in-kind arrangements in parts of private credit.

PIK interest allows a borrower to add some interest to the principal balance rather than paying it entirely in cash.

There are legitimate reasons to use it.

But economically it does something very specific:

it postpones the cash test.

If a software company is experiencing temporary pressure while investing in a successful transition, that flexibility can preserve value.

If the underlying business model is deteriorating, PIK can make leverage rise while the lender waits to discover whether the transition succeeds.

That distinction becomes particularly important with AI.

A cyclical revenue problem can recover.

A technologically obsolete business model may not.

AI Creates a Covenant-Lag Problem Too

Debt covenants are usually built around financial variables.

Leverage ratios.

Interest coverage.

Minimum liquidity.

EBITDA.

These are important.

But many are backward-looking.

A software company could still satisfy its leverage covenant while:

• new bookings collapse,
• net revenue retention deteriorates,
• customers reduce seats,
• AI usage migrates outside the platform,
• competitors dramatically reduce pricing, and
• the company’s product becomes easier to replicate.

The lender may technically remain protected by covenant terms while economic protection is disappearing.

That suggests AI-era lending needs more operational triggers.

Potential AI-era covenant signals

• net revenue retention deterioration,
• seat contraction,
• AI substitution rates,
• gross margin compression,
• model-provider concentration,
• customer migration toward AI-native workflows,
• revenue per human seat,
• percentage of product functionality reproducible through general-purpose models,
• contract duration versus loan duration,
• and the ratio of proprietary-data revenue to commodity-model revenue.

There Is Also a Shared-Borrower Problem

Private credit is often described as diversified because many funds participate across different deals.

That can be misleading.

The BIS notes that several large BDCs are exposed to a shared pool of software borrowers.

A single company can therefore appear across multiple portfolios.

A sector shock may consequently travel through funds that investors assumed were independent.

This is another distinction between manager diversification and economic diversification.

Owning five private-credit funds that lend to the same class of highly leveraged software businesses is not necessarily five different trades.

It can be one software-duration trade expressed through five managers.

The $500 Billion Number May Still Understate the Problem

There are at least three layers of exposure.

First is direct lending to software itself.

Second is lending to AI and technology companies.

Third is lending to businesses outside the technology sector whose revenue models can also be affected by AI.

Call centers.

Professional services.

Marketing agencies.

Business-process outsourcing.

Research providers.

Recruitment businesses.

Certain education companies.

Back-office service providers.

The true private-credit AI exposure is therefore larger than the technology bucket.

AI is a horizontal productivity shock.

Industry classifications are vertical.

That creates another measurement blind spot.

The Real Exposure Is Labor Substitution Beta

Instead of asking whether a borrower is an AI company, lenders may need to ask:

How much of this borrower’s revenue ultimately depends on expensive human work remaining expensive?

That is a very different way to classify credit.

A law-software provider and a legal-services company may sit in different sectors but face exposure to the same AI capability.

A customer-support SaaS platform and an outsourced call center can both be affected by autonomous service agents.

The correct unit of analysis becomes the underlying economic task.

DN calls this:

Labor Substitution Beta.

It measures how sensitive a company’s economics are to AI reducing the cost of the labor or expertise its product monetizes.

Private Credit Is Not Necessarily the Loser

None of this makes private credit structurally bearish.

In some ways, private lenders may be better positioned than bond investors.

Private loans can be renegotiated.

Lenders can demand covenants.

They can secure collateral.

They can receive board information.

They can modify pricing.

They can insist on amortization.

They can inject capital alongside sponsors.

And private-credit firms can move toward financing the winners of the AI transition.

The strongest lenders may therefore turn AI disruption into an advantage.

But only if they recognize that the old software underwriting template has changed.

The Best AI-Era Software Borrower May Look Different

The most resilient borrower may increasingly possess several characteristics:

• mission-critical workflow ownership,
• proprietary or regulated data,
• high integration depth,
• measurable switching costs,
• outcome-based rather than purely per-seat pricing,
• AI products that expand customer value rather than cannibalize seats,
• multiple model providers or model portability,
• low customer concentration,
• strong free-cash-flow conversion,
• and debt that matures well inside the defensible life of the revenue stream.

The New Credit Question

The old question was:

How predictable is this company’s revenue?

The better question is becoming:

How predictable is the economic relevance of the product generating that revenue?

That is considerably harder.

And it is exactly why the opportunity exists.

DN Revenue Half-Life vs Debt Maturity Engine

The proprietary model below introduces a variable traditional credit analysis rarely measures explicitly:

Revenue Half-Life.

For this model, Revenue Half-Life means the user-selected stress period over which half of the borrower’s currently defensible revenue could be repriced, displaced or lost under technological disruption.

It is not a prediction.

Its purpose is to force the analyst to compare technological time with financial time.

How to Read the Tool

The model deliberately does not try to predict which software company will fail.

It asks whether the debt survives significantly longer than the economic assumptions supporting it.

Suppose a borrower has a 4.5-year loan but the analyst believes its currently defensible revenue has a 24-month stress half-life.

The revenue-duration gap is 30 months.

If the debt barely amortizes during those two years, most of the principal remains outstanding precisely when the borrower would need to prove that its new product economics can replace the old ones.

That is the private-credit equivalent of the AI hardware refresh wall discussed in our previous research.

The asset changed.

The financing did not.

A New Private-Credit Metric: Debt Years per Revenue Half-Life

There is a simple ratio DN believes deserves attention:

Debt Years / Revenue Half-Life.

If a four-year loan finances revenue assumed to remain defensible for eight years, the lender has a substantial duration buffer.

If a four-year loan finances a business whose core economic proposition could be repriced in 18 months, the lender effectively owns several technology cycles inside one credit maturity.

The larger this ratio becomes, the more important:

• rapid amortization,
• strong covenants,
• high recovery values,
• proprietary data,
• switching friction,
• and sponsor willingness to inject new equity become.

What Would Prove the Thesis Wrong?

A strong thesis requires a falsification test.

This one weakens if:

• AI consistently expands SaaS revenue rather than compressing it,
• net revenue retention remains strong across highly exposed software categories,
• per-seat pricing proves resilient despite labor automation,
• software borrowers successfully reprice toward outcomes and usage,
• AI-native competition fails to reduce switching costs,
• BDC software marks remain stable through multiple frontier-model cycles,
• PIK usage falls rather than increases,
• defaults remain contained without widespread amendments,
• and private lenders begin earning materially larger spreads for technological-disruption risk.

That outcome is plausible.

Many incumbent software companies possess distribution, data, workflows and customer trust that AI-native challengers cannot recreate overnight.

AI may make strong incumbents stronger.

But that reinforces the main argument.

Lenders need to differentiate between companies.

Broadly similar spreads and maturities are increasingly difficult to justify if the durability of the underlying revenues diverges sharply.

The Bigger Macro Risk

Private credit does not exist in isolation.

The Financial Stability Board estimates the market at roughly $1.5 trillion to $2 trillion at the end of 2024.

It has also documented deepening connections with banks, insurers and private-equity firms.

Available regulatory data capture roughly $220 billion of drawn and undrawn bank credit lines to private-credit funds, while commercial estimates of the exposure can be higher.

This matters because a software shock need not remain inside a software loan.

A deterioration can move through:

borrower,

private-credit fund,

BDC,

private-equity sponsor,

bank funding line,

insurance portfolio,

and eventually the investors seeking liquidity from those vehicles.

The transmission is slower than a crypto liquidation.

It is not necessarily smaller.

The Most Important Insight

For more than a decade, software credit was built around an extraordinarily attractive assumption:

revenues could compound faster than debt while the underlying product remained relevant.

AI does not automatically destroy that model.

It changes the burden of proof.

A lender can no longer look only at whether revenue is recurring.

The lender has to ask:

What exactly makes this revenue difficult to replace?

That question leads away from ARR and toward:

data,

workflow ownership,

authorization,

regulation,

switching cost,

network effects,

outcome economics,

and technological substitutability.

Private credit's AI problem is therefore not simply that some software companies might fail.

It is that a financing market designed around predictable revenue may be entering an era where the predictability itself has become the risk.

The loan may last four years.

The business model no longer has to.