How it actually works
Every asset on a balance sheet has to be carried at some number. Accountants sort those numbers into rough tiers. If the asset trades on a screen, the mark is the screen price and there is nothing to argue about. If it does not trade but similar things do, you take the observable price of the similar thing and adjust. If nothing comparable trades at all, you build a model: expected cash flows, a discount rate, an assumption about recovery if it goes wrong, and out comes a number.
That last case is called marking to model, and the honest description is that the owner of the asset tells you what the asset is worth. Valuation agents and auditors sit on top of the process, but they are checking whether the assumptions are defensible, not whether a buyer exists at that price. Those are different questions, and in a stress they give very different answers.
An illustration, with invented round numbers. A fund holds a $100m senior loan to a private company. Nobody has traded that loan since it was written. The borrower is still paying, so the model says 98 cents on and the fund carries it at $98m. A buyer approached today, knowing the fund needs cash, might bid 80. Neither number is a lie. The first is what the loan is worth if held to maturity and everything works; the second is what it is worth this afternoon. The mark is the first number, and the fund's reported performance, its fees, and the price at which new investors buy in and old ones redeem are all calculated from it.
The crucial property of a mark to model is that it does not on its own. A market price falls whether or not anyone wants it to. A model price falls only when something forces the modeller to change an input — a missed payment, a , a loan going on , a rating action, or an actual trade somewhere that establishes an inconvenient comparable. Until one of those arrives, the number sits there being calm.
Why it matters to this crash
Our running thesis is that credit risk has migrated out of banks, which must mark positions and hold capital against being wrong, into vehicles where neither requirement applies. The mark is the specific hinge that migration turns on.
You can see it in almost everything we have covered. When we wrote about consumer legal funding on 19 August, the point was not that the advances are expensive, though at 35 to 45 per cent a year they are. It was that there is no observable secondary market for a plaintiff's settlement expectation, so the marks come from a model nobody outside the deal can check. When we looked at the queue at the private credit exit, the mechanism that made a record 12.4% of NAV in Q2 2026 redemption requests survivable was that funds ration rather than sell — nobody defaults, and no mark has to move. Those figures come from a secondary compilation rather than a filing we have read, so treat the decimals lightly; the structure is not in doubt.
The same pattern runs through the . The $27bn of debt in Meta's Hyperion vehicle, priced at 6.58% and amortising to 2049, is being placed with insurance and pension buyers who do not mark it daily, which is precisely why the maturity can stretch to 2049 against hardware that depreciates in five years. On the insurance side, the TWG-controlled insurers left holding more than $10bn of affiliated investments after their $6.5bn asset exchange are the clearest case there is: the assets are loans to the , there is no market in them, and the value is whatever the related parties agree it is.
The counter-example is instructive. When JPMorgan lends against bitcoin at 30–50% , the collateral has a live price every second, so the transmission runs through margin calls rather than mark-to-market losses. Visible marks do not remove risk. They change its speed, and they make it somebody's problem in public.
What would make this dangerous
The observable tell is a print: an actual transaction at a price well below where the same asset is carried elsewhere. A secondary sale of a data-centre bond materially inside par, a BDC portfolio sold in a wind-down, an insurer forced to liquidate affiliated paper. One real trade re-prices every model that used it as a comparable.
The second is the redemption queue converting from rationing into selling. As long as funds meet 38% of requests and pro-rate the rest, no asset changes hands. The moment a sponsor decides that queueing investors are worse than a discount, the discount becomes public data.
The third is the asset side forcing the issue regardless. Median at the twenty largest listed went from 2.0% to 2.8% of cost between Q1 and Q2 2026, the worst since 2017. Non-accrual is one of the few events that compels a mark to move. A further leg up in that series is the mechanical route from a slow deterioration to a visible one.
The fourth is disclosure moving the wrong way, which it already is. The SEC's clarification that data centres are not "financial assets" removes the risk-retention and asset-level disclosure that would let a buyer build an independent mark rather than accepting the sponsor's. Fewer inputs, more model.