The AI Bet Hiding in Private Credit

One of the biggest bets that AI wouldn’t change anything wasn’t made by a skeptic. It was made by lenders, and most of them didn’t know they were making it.

Between 2015 and the end of 2025, private credit loans to software-as-a-service companies grew from almost $8 billion to more than $500 billion, according to the Bank for International Settlements. That’s 19 percent of all direct lending — nearly one dollar in five. Every one of those loans was underwritten on some version of the same premise: subscription software revenue is sticky. Customers don’t rip out the system their business runs on. The revenue you see this year will be there next year, and the year after, and that’s what makes it safe to lend against.

That premise is precisely the thing AI attacks.

In March I wrote that DeFi is rebuilding the shadow banking system on-chain — banking functions performed outside banking regulation, with the same structural fragilities that broke in 2008. I still think that’s right. But it was the small version of the argument. The large version is sitting in private credit, and this year it got its first real test.

The safest loan in the world

It’s worth remembering why lenders loved software in the first place, because the logic was good.

A mature SaaS company has high gross margins, low capital needs, and revenue that renews automatically. Its customers face real switching costs — data migrations, retraining, integrations that took a year to build. Net revenue retention above 100 percent meant the existing customer base grew on its own. For a lender, that profile is close to ideal: predictable cash flow with a moat around it. It became normal to lend against recurring revenue to software companies that weren’t profitable yet, on the theory that the revenue was the collateral and the revenue wasn’t going anywhere.

Private equity sponsors bought software companies by the hundreds on the same thesis, and private credit funds financed the buyouts. The trade was so consistent it stopped looking like a trade. It looked like a sector.

The unstated assumption underneath all of it was that the software would still be needed in five years, at roughly the same price, by roughly the same number of seats. Nobody wrote that assumption into a credit memo, because nobody thought it needed defending.

What AI does to a moat

I’ve written about what AI does to legal work: it collapses the cost of producing the work while leaving judgment intact. The same collapse is happening to software itself, and it hits the subscription model in three places at once.

It makes software cheaper to build. A function that once justified a vendor relationship can increasingly be built in-house, or assembled from general-purpose models, at a fraction of the cost. I’ve watched this from the inside — a four-lawyer firm built for itself the kind of system legal-tech companies sell as a platform. Every customer who can do that is a customer some vendor no longer has.

It makes switching cheaper. The moat around a lot of enterprise software was migration pain: the data, the workflows, the integrations. Those are exactly the tasks AI is getting good at. A switching cost that took a team six months now takes far less, and a moat that’s cheaper to cross isn’t much of a moat.

And it shrinks the seat count. Much of SaaS is priced per user. When agents do work that used to require a person logged into a tool, the number of people logged in goes down — even if the amount of work getting done goes up. The software can be more useful than ever and still bill for less of it.

None of this means every software company is in trouble. Plenty will adapt, and some will come out of it with much better businesses. But a credit portfolio doesn’t care whether the average company adapts. It cares whether these borrowers can pay, on schedule, out of the revenue that was underwritten. Public markets understood that quickly: software stocks fell almost 30 percent between October 2025 and February 2026, per the BIS. Private loans don’t reprice like that. That’s the problem.

The marks move slower than the world

A public bond tells you what the market thinks of it every second. A private loan tells you what its holder thinks of it once a quarter, through a valuation process that is part model, part judgment, and structurally slow. In a stable world that’s a reasonable trade-off. In a world where one technology is repricing a whole sector at once, it means the numbers you can see lag the risk you’re actually holding.

You can watch the lag in the data. While software stocks were falling 30 percent, business development companies — the vehicles that hold about a fifth of U.S. direct loans — saw their share prices fall about 10 percent on average, with the most software-heavy ones trailing by around five more points. Then the reported numbers started catching up. Nonaccrual loans at BDCs jumped 40 percent in a single quarter in early 2026, and information technology was the sector driving the most deterioration. Fitch’s private credit default rate hit a record 6 percent for the twelve months through April.

The most telling detail is how those defaults happen. According to Fitch, more than half were driven by interest deferrals and the introduction of payment-in-kind — borrowers paying interest with more debt instead of cash. That isn’t a missed payment in the way most people picture one. It’s a loan that keeps growing while the cash stops coming, and it can look current on a quarterly statement long after it has stopped performing in any sense a borrower would recognize. A loan that pays its interest in more loan isn’t healthy. It’s postponed.

This is the shadow banking signature. Not hidden fraud — honest valuation processes, followed in good faith, that are built to absorb small shocks slowly and are poorly suited to one large shock arriving all at once.

The run is quarterly

The second piece of the parallel is liquidity, and it’s where private credit has changed the most in the last few years.

Private credit used to be funded mostly by institutions with long horizons — pensions, endowments, insurers — through closed-end funds that didn’t promise anyone their money back early. Increasingly, it’s funded by wealthy individuals through semi-liquid vehicles: non-traded BDCs and interval funds that hold illiquid loans but offer periodic redemptions, typically capped at 5 percent of the fund per quarter.

That’s a maturity and liquidity mismatch, the same one I flagged in DeFi lending protocols. The assets are multi-year loans that can’t be sold quickly without a discount. The liabilities are investors who were told they could get out on a schedule. Most of the time those two things coexist peacefully, because most investors don’t want out at the same time.

This year, a lot of them did. In the first quarter of 2026, investors asked to pull $20.8 billion from private credit funds managing roughly $300 billion, and only a little over half of it was paid out, per the Financial Times. Some managers paid redemptions above their caps. Others held the line at 5 percent to avoid selling loans into a weak market. In the second quarter, requests at some funds ran far past the cap — one large fund received requests for 17 percent of its shares.

I want to be fair to the gates. A redemption cap is a legitimate design choice, disclosed up front, and holding it protects the investors who stay from being diluted by fire sales. Managers who enforced their caps were doing what their documents said. But it’s worth being precise about what a gate is. In DeFi, the run happens in minutes. In a semi-liquid credit fund, it happens in quarterly installments of 5 percent. The cap doesn’t prevent the run. It slows it to a speed the fund can survive.

Two hundred borrowers, one risk

Here is the part that makes this an AI story rather than an ordinary credit cycle.

Diversification is the core promise of a credit fund. Lend to two hundred companies across dozens of industries, and no single failure can sink you. The mortgage market before 2008 made the same promise — thousands of loans, across every region — and it failed because every one of those loans rested on one shared assumption about house prices. The loans looked independent. The risk wasn’t.

A portfolio of software loans can be diversified by borrower, by vertical, by geography, by sponsor, and still rest on one shared assumption: that subscription revenue holds. AI is a shock to that assumption everywhere at the same time. Two hundred borrowers, one risk factor. That isn’t diversification. It’s concentration in a costume.

I don’t think this is 2008. Most of this risk sits in funds rather than on bank balance sheets, leverage is far lower than it was then, and the funds that gated are working as designed rather than collapsing. A painful credit cycle concentrated in one sector is not a financial crisis, and I’d be suspicious of anyone who tells you otherwise with confidence. But the structural features — marks that lag reality, a liquidity promise layered over illiquid assets, and correlated exposure that looks like diversification — are exactly the ones that make a cycle worse than it needs to be. They were the problem in shadow banking. They’re the problem here.

What this means if you’re building

I’ll be direct about my own position first. I’m Chief Legal Officer of Real Finance, which builds infrastructure for private credit markets, so I have a stake in the view that this market needs better plumbing. Weigh what follows accordingly.

If you run a software company with a private credit facility, your lender is now underwriting AI risk whether or not you’ve discussed it. Expect the question at your next amendment or refinancing, and expect it to show up in pricing and covenants before it shows up in conversation. Know when your facility matures, what a PIK toggle actually costs you, and what your net retention says about your exposure. The time to tell your lender your AI story is before they form their own.

If you’re raising venture debt or an ARR-based loan, the market is repricing the thing you’re borrowing against. Model what happens to your facility if revenue multiples compress or a few large customers decide to build instead of buy. Recurring revenue is still good collateral. It’s just no longer assumed to be permanent.

If you’re investing in a semi-liquid credit fund, read the redemption terms as the actual product. Quarterly liquidity subject to a 5 percent cap is a promise with a ceiling, and this year showed exactly where the ceiling is.

If you’re building in fintech or credit infrastructure, the opportunity is the problem itself. Almost every fragility here is an information problem: valuations that lag reality, underlying assets that are hard to verify, risk that can’t be transferred quickly or priced transparently. Markets with verifiable data on what’s underneath them, and a real way to move risk to whoever wants to hold it, get through shocks like this one better than markets without. That’s the infrastructure worth building.

The frontier doesn’t hold still

I’m a frontier optimist, and I’m optimistic about this too. AI collapsing the cost of software is good news for almost everyone — more people can build, tools get cheaper, and every company that runs on software gets more out of each dollar it spends. It’s bad news mainly for capital that was priced on the assumption that none of that would happen.

That’s the real lesson, and it’s bigger than private credit. Every credit model is a bet that the future will look enough like the past to repay the loan. The models that broke in 2008 bet on house prices. The ones under pressure now bet on software staying the way it was. The technology isn’t the risk. Underwriting as though the frontier would stand still was.