Hook
A bank's private credit book is a ledger with no block explorer. No state root you can query, no transaction receipts, no event logs you can replay against an independent node. When JPMorgan marked down a tranche of its private credit portfolio in March, the only durable record of that decision was a line in a filing and a sentence handed to Semafor. The New York Fed responded by sending examiners into four of the largest banks โ JPMorgan, Wells Fargo, Morgan Stanley, and the U.S. arm of Barclays โ to interrogate how much of their balance sheet is quietly wired to the same asset class they are lending against. The consensus read this as a routine supervisory check. It is not. It is the first time a major central bank has moved, in public view, to audit a market whose defining architectural feature is that it cannot be audited from the outside. That contradiction is the story. Not the markdown. The markdown is just the first crack that let the light through.
Context
Private credit began as a reasonable engineering solution to a real constraint. After 2008, capital rules made it expensive for banks to hold leveraged corporate loans on their books. The lending did not disappear; it migrated. Direct-lending funds, business development companies, and closed-end vehicles stepped in, raised locked-up capital, and originated the loans banks no longer wanted to warehouse. The market grew from a few hundred billion dollars to a figure that plausibly clears two trillion depending on whose definition you accept. For a decade it delivered a seductive return profile: equity-like yields with bond-like volatility. That combination is the tell. Any instrument that reports high returns with low measured variance is either genuinely uncorrelated or simply not being marked to a real price. Private credit is the second kind, and the reason is structural, not fraudulent.
There is no public order book for a private loan. Its value is whatever the manager's model says it is, validated by an auditor once a year and by a valuation committee the manager appoints. This is mark-to-model, the same methodology that turned subprime mortgage tranches into AAA paper in 2007 and that Enron used to turn physical assets into imaginary earnings. I have spent most of my career reconstructing what happened on-chain, where every state transition is permanent and every lie is eventually reconciled against the ledger. The forensic lesson from that work is blunt: an asset whose price is set by the party that profits from the price is not marked, it is narrated.
The AI angle is what converted a slow-burn structural concern into an acute one. For two years the market treated artificial intelligence as a pure growth story โ compute, models, infrastructure, the winners. What the JPMorgan markdown revealed is the other column of the ledger: the losers. A large share of private credit lending over the past five years financed leveraged buyouts of software companies. Those companies were underwritten on a business model โ per-seat SaaS pricing, high gross margins, sticky recurring revenue, low churn โ that generative AI is now dismantling in real time. When a single model can replicate a workflow that used to require twelve seats, the recurring revenue is not recurring. It is decaying. And decaying revenue on a floating-rate loan in a higher-for-longer rate environment is a credit event waiting for a calendar date.
Core
Let me trace the mechanism the way I would trace an exploit, because the structure is identical. An exploit is not a single bug. It is a chain of individually defensible assumptions that, composed, produce a failure. The private credit chain has four links, and each one looks safe in isolation.
Link one is the loan itself. A direct lender extends credit to a software company at a floating rate โ typically SOFR plus a spread. The underwriting assumes the borrower's cash flow is stable and that the loan can be refinanced or repaid at maturity. Both assumptions were imported from a world where software revenue was annuity-like. AI broke the first assumption quietly, through churn, discounting, and seat consolidation, none of which shows up as a headline until it shows up as a missed covenant.
Link two is the fund's leverage. The direct-lending vehicle does not only deploy investor capital. It borrows. Subscription lines let it draw against capital commitments before those commitments are called โ effectively a short-term bridge that flatters the internal rate of return by delaying when investor money is actually deployed. NAV loans let the fund borrow against the net asset value of its own portfolio. Back-leverage lets it borrow against the fund's positions themselves. Each of these instruments is a multiplier on the fund's stated performance. Each is also a place where a valuation error compounds rather than cancels.
Link three is the bank. This is the link the New York Fed went to inspect, and it is the one most people misunderstand. Banks are not primarily exposed to private credit as equity investors. They are exposed as creditors to the creditors. They provide the subscription lines, the NAV loans, the back-leverage, and the revolving facilities that keep the funds liquid. From the bank's perspective this looks like safe, fee-generating, senior-secured lending. The collateral is a diversified pool of loans. The borrower is a sophisticated institutional manager. What could go wrong.
Link four is the circularity. The collateral is valued by the same manager who borrowed against it. The NAV that secures the loan is produced by the model the manager controls. This is the structural flaw, and it is not a bug in one fund โ it is the architecture of the entire market. Tracing the ghost in the smart contract state is trivial compared to tracing the ghost in a NAV loan, because in DeFi the state is public and in private credit the state is a spreadsheet.
Now compose the links. AI erodes software revenue. Software loans in the fund's portfolio lose value. The manager, incentivized to avoid a forced markdown that would trigger margin calls on its own leverage, reports a stable NAV. The bank, relying on that NAV as collateral, keeps lending. Then a borrower defaults, or a redemption is requested, or an auditor pushes back, and the NAV is revised downward. The revision cascades: the NAV loan becomes undercollateralized, the bank issues a margin call, the fund must sell assets into an illiquid market to meet it, and the sale price โ the first real price the asset has ever seen โ is far below the modeled value. That gap is the loss, and it lands on the bank's balance sheet, not the fund's.
JPMorgan's markdown is the moment link one touched link three without passing through a market. There was no trade, no price discovery, no liquidation. A bank looked at its software exposure and decided the modeled value was fiction. That decision is the entire event. It is the equivalent of a node rejecting a block that every other node had already accepted โ the moment the consensus assumption visibly fails.
What makes this worse than a normal credit cycle is the opacity. In a public market, a deteriorating credit shows up as a widening spread, and the spread is information. In private credit, there is no spread, because there is no market. Silence in the logs is louder than the error โ and this market has no logs. The absence of volatility is not evidence of stability; it is evidence of a measurement system that cannot register the thing it is measuring. I watched the same pattern in the Lendf.me exploit in 2020, where the contract reported healthy balances right up until the 3Commas vault drained twenty million dollars through a missing zero-value check. The dashboard said everything was fine. The dashboard was wrong, because the dashboard was built by the people who were about to be robbed.
The regulatory math is straightforward once you see the architecture. The Financial Stability Oversight Council, the IMF, and the ECB have all flagged bank-to-nonbank interconnection as the next systemic channel for years. The New York Fed walking into four banks is the supervisory version of a pre-trade risk check: they want to know the notional before the position blows through the limit. If the examiners find that several banks are lending against the same NAV-flagged collateral through overlapping funds, the correlation is not diversified โ it is concentrated. Dissecting the code reveals the true owner, and in this case the true owner of the risk is the banking system, wearing the costume of a fee-earning intermediary.
Here is the part that should genuinely alarm anyone holding bank equity. Banks have spent a decade telling investors that their exposure to shadow banking is safe because it is senior and secured. Seniority means nothing if the collateral cannot be priced and the borrower cannot be margin-called without triggering a fire sale. Secured means nothing if the security is a loan to a company whose revenue is being eaten by a model it does not own. The risk is not subordinated in the legal stack. It is subordinated in the informational stack, which is worse, because informational subordination cannot be cured by a covenant.
Contrarian
Now the part the bears will hate, and I owe it to the ledger to say it. The bulls are not entirely wrong, and the panic trade is not the clean short it appears to be.
First, private credit has survived two genuine stress tests โ March 2020 and the 2022 rate shock โ without a systemic blowup. That is a real data point, not a talking point. The funds are closed-end, which means they do not face daily redemptions the way a money market fund does. Locked-up capital is a genuine buffer, and it is the single most important structural difference between this market and the 2007 shadow banking complex. A closed-end vehicle can absorb a markdown by simply refusing to sell. It can wait. Illiquidity is a liability when you are forced to transact and an asset when you are not.
Second, floating-rate lending is not purely a headwind. Higher rates mean higher coupon income for the lender, and for a fund that can hold to maturity, rising rates improve returns as long as defaults stay contained. The bull case that private credit is a rate-hedged income machine has real mechanics behind it. The markdown is a credit-quality problem, not a duration problem, and the two should not be conflated.
Third โ and this is the blind spot almost everyone misses โ the AI disruption is a redistribution, not a destruction. Total software spend is not collapsing. It is migrating from seat-based incumbents to infrastructure, models, and platforms. The companies being marked down are the ones whose moat was a pricing model, not a technology. The ones being marked up are the ones selling the picks and shovels. Arbitrage is just theft with better mathematics, and what the AI transition is doing is arbitraging away the margin of every business whose value was a contractual lock-in rather than a capability. The credit losses are real, but they are concentrated, identifiable, and โ for a lender who underwrote cash flow rather than narrative โ survivable.
The genuine error is not that banks lent to software. It is that they lent to software using a valuation framework built for a world where software revenue was predictable. The framework is the bug. The asset is mostly fine, or mostly fixable, depending on the specific borrower. What the New York Fed is really inspecting is not the loans. It is the framework.
Takeaway
So watch the framework, not the headlines. The signal to track over the next two quarters is not whether JPMorgan marks down more loans โ it will โ but whether other banks begin to. A single markdown is a credit decision. Four banks marking down the same category in the same quarter is a revaluation, and a revaluation is when a narrated price becomes a real one. The New York Fed already knows this. That is why it knocked on four doors instead of one.
The forward-looking question is not whether private credit survives โ it will, in some form, because the lending need is real. The question is who is holding the loss when the model finally meets a market. Logic is immutable; intent is often malicious, and the intent baked into mark-to-model is always the same: delay the truth until someone else is holding the bag. The banks think they are the auditor. They are the collateral. When the first NAV loan gets called and the fire sale prints the first honest price this market has ever seen, we will find out which of those four doors was hiding the same loan.
