The most informed lender in the market is quietly walking away.

In February 2026, JPMorgan told investors it was becoming more conservative than its peers on private credit risk. In March, it marked down the software loans that private credit funds had pledged as collateral, and tightened what it would lend against them. In May, it began looking to sell roughly $4bn of private-equity-linked exposure outright.

Not hedge. Not reprice. Sell.

This is the most informed counterparty in the private credit market - the bank that sees the collateral, the fund leverage, and the borrower cash flows - stepping back from the trade. It is not the only such signal. In April the market launched its first standardised instrument for shorting private credit. In May the Financial Stability Board published its first dedicated warning on the asset class. Through 2026 the US insurance regulator has been assembling the authority to re-rate exactly these assets on insurers' balance sheets.

None of this is a forecast. It is a sequence of the best-informed participants in the system repositioning at the same time. The question this paper addresses is not whether private credit and private equity marks are wrong. They are - we will show by how much. The question is what forces everyone else to admit it, and when.

"When the music stops, in terms of liquidity, things will be complicated. But as long as the music is playing, you've got to get up and dance. We're still dancing."

Chuck Prince, CEO of Citigroup, July 2007 - four months before he was forced out. Asked about it later, George Soros observed: the music had already stopped when he said it.

The industry is still dancing. SpaceX is marked on the promise of Starlink and Mars; the value sits in a future that has not yet arrived. OpenAI and Anthropic are valued on the conviction that they will remake the economy the way Amazon remade retail - and they may well be right. But being right about the technology is not the same as being right about the price. Amazon did revolutionise commerce, exactly as the bulls of 1999 said it would. Its shares still fell more than 90% when the dot-com bubble burst, and took until 2009 - the better part of a decade - to reclaim their prior peak. The thesis was correct. The entry price was not. An investor who was right about the company and wrong about the valuation still lost almost everything and waited ten years to break even.

That is the gap this paper is about. Not whether AI matters - it does - but whether the marks assigned to the companies riding it, and to the credit extended against them, bear any relation to what those assets would fetch if anyone had to sell. For two decades, institutional capital never had to find out: rates near zero, multiples expanding, the macro bailing out almost every decision. An entire generation of private equity and private credit professionals has never worked through a genuine credit cycle. That era ended in 2022. What has not yet ended - because, uniquely among asset classes, almost no one is forced to admit it - is the fiction of the marks.

Public and private markets alike are priced for perfection, in equities and in credit. The marks assume the good outcome arrives: that AI delivers on the timetable the valuations require, that the growth underwritten in 2021 materialises. But cracks are appearing. If perfection does not arrive - if AI underdelivers, or physics reasserts itself on the dream of a Martian colony - the fall could be severe, because no margin of safety is priced in anywhere. The history of these moments, 1999 and 2007, is that recognition comes late, then all at once.

Our argument is that the first seeds of that reckoning have already begun to unravel, in a narrow niche of private credit we happen to operate in: lending to software companies. This paper traces how they spread through the private markets - from credit, to the private equity beneath it, to the insurers that hold both. It seems unlikely the public markets would be left untouched. But that is the subject of another piece.

The epicentre: private credit and the mismarked book

Private credit is the 2006 CDO market in different clothes. The assets are different. The opacity is identical. The marks are self-certified. The conflict of interest is structural.

The market, broken by underlying exposure

The standard analysis treats software as the stress cohort - 19–25% of the book. But once software-adjacent industries are included, AI-disrupted exposure rises to 35.6% of BDC portfolios on a median basis.¹ Nearly half the market is in sectors facing structural disruption of the revenue models that justified the original underwriting.

The coverage ratio collapse

Nearly all private credit loans carry floating rates. As the Fed raised rates 525bps between 2022 and 2023, interest expense doubled on outstanding loans while EBITDA grew modestly. 47% of borrowers now have an interest coverage ratio below 1.5×, up from just 7% in Q4 2020.² The floating rate structure that made private credit attractive to investors has become the mechanism destroying the borrowers.

PIK: the smoking gun

Payment-in-kind interest means the borrower cannot pay cash. The lender records revenue it has not received. The NAV is marked on income that does not exist in cash form.

The true default rate

The same disease, the larger patient: private equity

Everything above concerns private credit - the loans. But the loans are the senior claim. They sit above the equity, and get paid first; the PE or VC stake beneath them absorbs losses before the lender takes a cent. This is the point the headline numbers miss. As detailed below, we estimate roughly $123bn of unrecognised loss on the $1.3tn direct-lending book - call it a 10% hit to the senior claim. But the senior claim takes the last loss, not the first. Where a specific loan is genuinely impaired, the equity beneath it has already been wiped before the lender loses a cent. Across a whole book the average is gentler - on the simple arithmetic of a leveraged structure, a blended 10% loss to senior debt implies something like two to three times that on the equity beneath it - but the direction is not in doubt, and that is before recognising that the most distressed equity is not marked at all. The debt marks are a direct read on the equity below, and the equity is the larger pool: a $1.8tn credit book sitting on top of $3.6tn of unsold private equity. We will return to size that loss conservatively; for now the direction is the point.

The mechanism that converts this from logic into recognised loss is refinancing. While the debt performs, everyone holds their marks and the equity loss stays theoretical. But these loans mature, and a company that cannot refinance at the old terms faces one of two outcomes, both of which crystallise the equity. Either the sponsor hands the business to its lenders - the equity is wiped in a single step, and the debt holders become the new owners - or the lenders agree to extend only if fresh equity is injected, and that capital comes in at a drastically lower valuation: a debt-for-equity swap or a down round that forces every existing holder to remark to the new, far lower price. Either path turns "carried at cost" into "marked to reality." This is why the maturity wall matters so much to the timing, and why the recognition does not arrive smoothly: it arrives one refinancing at a time, each one converting a model mark into a transacted price.

Where that equity is forced to trade, the marks do not survive contact. In the venture capital (equity) secondary market, for example, stakes still carried at 2021 valuations change hands at roughly a 68% discount to those marks. Not a discount to some abstract fair value - a discount to the number the fund is reporting to its investors today. The mark is off by two-thirds. And that is the price of the assets liquid enough to trade: the better names, the ones with a bid. The positions that do not trade - the genuine zeros, in companies whose revenue model the SaaSpocalypse has already broken - are not discounted at all. They are still held at cost. How many zeroes are hidden in a book where even the saleable assets clear at a third of their carrying value?

Venture capital is the epicentre of the current crisis but not the only one affected. Both venture capital and private equity value their holdings on the same IPEV framework, and private equity has been buying technology - the sector most exposed to AI disruption - in steadily increasing size. The same method, the same sector risk, the same reluctance to mark. The only question is how far the gap extends from venture into the larger buyout book. The disease is identical. The patient is larger.

The selection bias that hides the rest

The 68% venture capital discount above carries a second, worse implication. What trades in the secondary market is overwhelmingly the higher-quality, more diversified, established-manager material - the assets with a bid. The 2021–2022 vintage software-heavy books that most need price discovery largely do not trade. They sit on GP books at carrying value, extended into continuation vehicles, or held at stale marks while fund lives are quietly prolonged.

"Lower-quality assets often fail to find bids, while portfolios of top-quartile managers continue to clear the market."

Northern Trust, Private Equity Secondaries Analysis, 2025

So the observed discount is a floor on the true markdown, not an estimate of it. The worse the asset, the less likely it appears in the data at all. Whatever the traded discount, the real number across the whole book is worse - because the worst of the book never prints a price.

The machine that keeps the fiction alive

Continuation vehicles - GP-led transactions in which a manager sells assets out of a legacy fund into a new vehicle it also manages - have become the primary mechanism for sustaining PE marks in the absence of genuine exits. GP-led volume reached $115bn in 2025, with continuation vehicles accounting for 89% of that total and roughly 43% of the entire secondary market. The structural conflict is categorical: the GP sets the price, buying from itself, while single-asset vehicles - now 57% of volume - let it cherry-pick the best assets out of the legacy fund and leave remaining LPs holding the residual. Academic work finds these transactions rarely deliver a premium to selling investors. In November 2025, Abu Dhabi Investment Council sued Energy & Minerals Group in the Delaware Court of Chancery, alleging self-dealing and breach of fiduciary duty in exactly this structure - and secured a delay to the deal. It is the first major legal challenge to the mechanism that has kept PE marks intact. When LP confidence in continuation-vehicle pricing breaks, the exit route that has deferred recognition closes, and the marks must follow.¹¹

The unrecognised loss - our estimate

Note on the PE figure. The subordination logic above implies equity losses of two to three times the ~10% taken by the senior debt - i.e. 20–30%. We deliberately do not use that here. PE buyout portfolios are likely better selected and less concentrated in the worst of the SaaSpocalypse than the venture cohort, and we would rather understate than overreach. We therefore apply a 15% discount - below what the subordination maths and the 68% venture print would justify. If anything, the true PE markdown is higher; $540bn is the conservative floor, not the expected case.

For context: the subprime RMBS mark-to-market loss in 2008 was approximately $550bn on a $1.1tn base. The total projected losses including the Alt-A and prime cascade were $1.8 trillion. The unrecognised loss we identify - $663bn - is directly comparable in scale. One important caveat: several industry reports argue private credit is not systemically relevant, citing a market size of $1.8tn against ~$500tn of global financial assets. The denominator is wrong twice over. Total private credit AUM including ABF and structured products is $3.5–4.1tn (Proskauer 2025), not $1.8tn - and even that understates the exposure, because private credit is the trigger, not the mass. The larger charge sits in private equity: $3.6tn of unsold assets carried at marks the secondary market already disputes. The two are joined by the transmission mechanism set out below - stress in a "too small to matter" credit market is precisely what forces recognition in the far larger equity book. Size is not the question. Transmission is.

The transmission: how it spreads

In 2008 the transmission ran through bank balance sheets. Overnight funding, mark-to-market accounting, and interbank contagion made the crisis fast and violent. This time the transmission is slower, broader, and runs through a different set of institutions - but it has an identifiable structure and a specific weak link.

The chain

The first stage is not abstract to us. The SaaSpocalypse has already hit several venture debt lenders in the US, where software exposure surfaces earliest in the loan book. We have deliberately stayed out of generalist software lending, restricting our venture debt exposure to highly defensible verticals - cybersecurity and medical - where the revenue base is not the thing AI disintermediates. That vantage point is why we see the stress building before it reaches public marks, and why we are not carrying it.

The weak link: PE-owned insurers

Pension funds have no forced mark-to-market. They can extend and pretend indefinitely within their investment mandates. But insurance is different - and within insurance, one structure is the precise pressure point.

~$105 billion of insurance assets are in Level 3 positions marked by the same PE firms that originated them, held against policyholder obligations, under regulatory oversight that is actively tightening. The PE firm provides the valuation. The annuity holder bears the risk. This is not a conflict of interest. It is the business model.

The NAIC capital charge - the margin call equivalent

The trigger mechanism is not a discretionary market event. It is a regulatory ratchet that is already in motion.

Applied to Athene's $60bn of self-marked Level 3 positions, currently capitalised at approximately 2.17% (implying ~$1.3bn required capital): a single NAIC-initiated notch downgrade raises required capital to ~$1.9bn. Two notches raises it to ~$4.4bn - exceeding Athene's entire ~$3bn excess capital buffer. Athene does not need to be insolvent for the trigger to fire. It needs the NAIC to re-designate its structured holdings by two notches. That authority is live from 2026.⁵

The 2007 parallel: S&P placed 612 subprime RMBS securities on credit watch on July 11, 2007. That single action triggered margin calls across every repo-financed CDO position in the system simultaneously. The NAIC's 2026 CLO loss model is the equivalent action - it re-designates structured holdings across every PE-linked insurer simultaneously, imposing capital charges that cannot be met without either capital raises or asset sales into an illiquid market.

"The concerns raised a few years ago that annuity assets of PE-owned insurance companies would be used to boost the performance of affiliated private equity funds are no longer theoretical."

AM Best, cited by CEPR, March 2026

The 2007 parallel - exact

At UBS, the decision to internalise Dillon Read's positions was a discrete management decision made to protect relationships. Apollo/Athene have already made that decision - it is baked into the corporate structure. There is no discrete moment. The structure is the problem.

The trigger: why 2027–2028 is the crystallisation point

The thesis does not require a single Lehman moment. It requires the convergence of three structural forces - a regulatory ratchet, a liquidity event, and a maturity wall - all pointing to the same 18-month window.

JPMorgan's three-move exit - the institutional tell

The Rothschild confirmation: AI debt capacity peaks in 2028

Independent of the credit cycle analysis, Rothschild & Co Redburn IDEAS identified 2028 as the year AI infrastructure financing reaches maximum allocation - $2 trillion of debt capacity absorbed, ROI on datacentres already fallen from 16% to 8.5%, hyperscaler land banks complete.⁷

Two independent analytical frameworks - credit cycle mechanics and AI capex economics - converge on the same 12-month window. The capital that should be available to refinance the PE/private credit maturity wall is instead fully committed to AI infrastructure. There is no spare capacity in the system to absorb a credit dislocation, because the system is already at maximum allocation doing something else.

This is not 2008. It is different in ways that matter - and similar in ways that matter more.

The leverage comparison is more nuanced than it first appears. CDO tranching at 30–40× and bank balance sheet leverage at 25–30× were the 2008 amplification mechanisms. Insurance capital charge leverage - 1–2% capital against investment-grade assets - is structurally comparable on a per-dollar basis, with the NAIC override authority as the margin call equivalent. The Fed put is a genuine offset: rate cuts will improve IC ratios for performing borrowers and reduce the cost of refinancing. They will not rescue the cohort whose underlying business model has been structurally disrupted by AI, regardless of the rate environment. Private credit does not blow up overnight. It deflates. But the direction is determined, the regulatory ratchet is live, and the endpoint is the same: $663bn of unrecognised loss finds its way onto institutional balance sheets. The timeline is 2027–2028. The question is whether capital is positioned ahead of it.

The opportunity

Dislocations of this scale create distinct return opportunities at different points on the timeline. The managers positioning for this now - Oaktree ($16bn distressed fund closed February 2025), Ares Special Opportunities ($8.3bn, April 2026), Strategic Value Partners, Marblegate - are already raising capital. The window is open. It will not stay open indefinitely.

"We are talking about a 5+ year problem, as the GFC was, to process all of this liquidity. This is not going to go away in 2025 or 2026."

Chair, Global Private Equity Practice, Bain & Company, January 2026

The opportunity is not a prediction of imminent collapse. It is a structural observation that $663bn of unrecognised institutional losses must eventually be recognised; that the maturity wall makes 2027–2028 the arithmetic deadline for the most stressed cohort; and that the capital positioned ahead of that deadline will be the capital that defines the decade.

Others will spend 2027 and 2028 marking down what they should never have underwritten. That is precisely when we intend to increase ours.

In 2027 or 2028, the myth and the market will collide. Our money is on the market.