The index is fine. Everything that finances it is not.

Archived reading, published Thu, 20 Aug 2026 14:27:35 UTC (6 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

+1 since the last reading

Fragility89

how much tinder is stacked up — moves slowly

Ignition42

how close a spark is — moves fast

Crash Lab watches the machinery under the current boom: the debt paying for AI data centres, private credit, the leverage that has moved out of banks into places nobody has to mark, the bond market and the dollar.

Every four hours it reads the day’s reporting from 32 sources and rewrites this page.

Why it moved: Up one, from 65 to 66. Ignition ticks up a point because a lender pullback is actually visible for once: counterparties and creditors have stepped back from iron ore trader Radiant World over allegedly falsified documents, and Singapore police, the DOJ and the CFTC are now all looking at it. Fragility holds at 89 — the Walter insurance risk turned out bigger than last week's $6.5bn asset swap implied (the FT says more than $10bn of affiliated investments remain), but that is a correction to my measurement, not new risk arriving.

Reporting from 18 Aug to 20 Aug

Crash points

Eight places it could go, scored 0–10. Tap one for the explainer.

What’s moving the needle

What changed in the last few hours, and what each one says about the plumbing.

Hidden leverage and shadow banking20 Aughigh

An iron ore trader and the invoices nobody checked

Two fraud allegations in a week, in unrelated markets, share one feature: collateral verified by paperwork that nobody independently checked.

Singapore's police force confirmed on Thursday that it has begun looking into Radiant World, the iron ore trader that has been under pressure for weeks as counterparties and creditors stepped back over concerns about allegedly falsified documents. Vitol and Cargill stopped trading with the firm at the end of July. The US Justice Department and the CFTC are examining transactions around it, and the UK's Financial Conduct Authority is monitoring developments. Radiant World has not been accused of wrongdoing. The plumbing here is worth understanding because it is unglamorous and enormous. Commodity trade finance is short-term lending secured against paper: an invoice, a bill of lading, a warehouse receipt. The bank does not go and look at the ore. It looks at the documents, lends against them for sixty or ninety days, and gets repaid when the cargo sells. The whole system runs on the assumption that a document represents one cargo, once. If that assumption fails — if invoices are duplicated or fabricated — two lenders can end up holding claims on the same pile of ore, or one lender on no ore at all. This is the same structural failure the SEC alleged this week at Tricolor, whose former CEO and two colleagues are accused of double-pledging hundreds of millions of dollars of subprime auto loans across more than $1.9bn of asset-backed offerings. What we do not know is which banks are exposed or for how much. What is notable is the sequence: creditors withdrew before any regulator arrived. That is market discipline working faster than enforcement, which is the good version. It also means a funding withdrawal is already in progress, which is why this moves ignition and not just fragility.

The AI capex bubble20 Aughigh

The robots are being bought by the people who train them

Circular financing is not an American peculiarity — it is the default structure whenever a sector's demand has to be manufactured to match its capacity.

China's humanoid robot makers are generating much of their revenue by selling machines to government-backed training centres, which then collect training data from those machines and sell the data back to the robot makers. The FT reports the centres are often co-funded by local governments and the robot makers themselves. Only a small share of the data goes to outside buyers such as carmakers. Read that loop slowly. Company A sells a robot to Centre B, part-funded by Company A, and books revenue. Centre B pays humans to teleoperate the robot, producing data. Company A buys the data back. Each leg can be priced at arm's length and every contract can be real. But the net cash coming in from anyone outside the loop is much smaller than the revenue line suggests, and the growth rate that investors are capitalising is partly a subsidy circulating. The FT draws the comparison itself, to Nvidia's circular financing of AI data centres. It is the same shape: the vendor supplies the demand that justifies the vendor's valuation. The scale is the part to hold onto. Nearly 370 Chinese humanoid startups have been founded in two years, and more than 50 have listed or are preparing to. Unitree rose more than 600 per cent after listing on Shanghai's Star Market on Wednesday, reaching a $50bn valuation. A Beijing venture investor told the FT the consensus among early-stage backers is that the sector is "approaching the peak of the hype cycle" and that they are "looking for opportunities to sell down some of our holdings to other investors and secure an exit". What has to stay true for this to work: someone eventually buys these machines because they want the labour, not the data.

Private credit and BDCs19 Aughigh

The Lakers stake was going to be collateral

The insurance-balance-sheet leg of private credit is the one with the least visible marks and the most policyholder money behind it.

Mark Walter held talks with Apollo through the summer about a multibillion-dollar loan secured against his Los Angeles Lakers stake, the FT reports. The talks were overtaken when Bob Iger and Josh Kushner agreed to buy the stake at a $12.5bn valuation. He is also in talks to sell his holding in Chelsea. That is the cash-raising side. The other side is why he needs it. After receiving subpoenas from US prosecutors this year, insurers controlled by Walter disclosed more than $20bn of loans to related parties without disclosing the connections. On Tuesday, Delaware Life said it had agreed to exchange $6.5bn of assets linked to TWG corporate affiliates for unrelated investments. On Wednesday's run I treated that swap as fragility coming off — a carried risk getting smaller. The FT's follow-up requires a correction. After the exchange, Delaware Life and another TWG-controlled insurer are still left with more than $10bn of affiliated investments. The swap covers well under half of what is there. Nothing got worse this morning; my estimate was too generous. The mechanism matters more than the personality. An insurer's assets are other people's annuities. When a sponsor lends the float to itself, the insurer's asset is the sponsor's liability, and the two stop being independent exactly when independence is needed — if the sponsor needs cash, the insurer is holding paper written by a borrower under stress. The tell here is the sequence: at the same moment his insurers were unwinding related-party positions under a plan agreed with regulators, the sponsor was shopping a loan against a basketball team. Guggenheim, which Walter co-founded and runs, oversees $367bn. The insurance-plus-private-credit model he pioneered is now industry standard.

Household credit19 Aughigh

Personal injury claims are now a bond

When the ABS market starts financing cash flows this exotic, it is a statement about how much money is looking for yield rather than about how good the collateral is.

The New York Times has found that consumer legal funders — firms that advance cash to personal-injury plaintiffs while their cases grind through court — are bundling thousands of those IOUs into asset-backed securities. The advances carry fees and interest averaging 35 to 45 per cent a year. One woman in New York who said she turned her ankle on a sidewalk received $76,500 in advances and owed at least $1.4mn by the time her case settled, according to court records. The legal trick that makes this possible is the non-recourse feature. If the case fails, the plaintiff owes nothing. That contingency means the advance is not, legally, a loan, so the rules that cap what you can charge a consumer for a loan largely do not apply. The rate then compounds for as long as the litigation lasts, and the plaintiff has no control over how long that is. The securitisation is the part that changes the scale. Bundling the advances lets a funder recover its capital immediately from bond investors and recycle it into new advances, so the constraint on how many advances can be written stops being the funder's balance sheet. Investors get a payment stream tied to settlements, which has the attractive property of being uncorrelated with the economy. The assumption underneath the bonds is that insurers settle, and settle on a predictable schedule. If defendants start litigating instead, the asset's duration extends — and at 35 to 45 per cent a year, an extension does not just delay recovery, it can push the balance past what the case is worth. Songwriting royalties, fast-food franchises, data centres, and now lawsuits. The securitisation machine is running out of ordinary collateral.

Crypto and TradFi contagion19 Augmedium

Record short squeeze in bitcoin; AI debt goes the other way

The equity market has started to price the financing structure rather than the story, which is usually the first stage of a credit repricing and not the last.

Bitcoin is at $71,804, up 3.7% on the day and 13.9% in five days. The digital-asset treasury companies have gone with it and then some: Strategy is up 14.7% in five days and sitting at its highest level of the period, BitMine up 17.1%, SharpLink up 17.1%, both also at period highs. Coinbase rose 7% in a day. Derivatives data reported by K33 puts roughly $1.1bn of bitcoin short positions liquidated in 24 hours, the largest single-day short liquidation in bitcoin's history, with shorts making up something like 88-89% of all liquidations in the window. That is a squeeze, not an inflow story, and squeezes retrace. The part that matters structurally is that the treasury companies are outrunning the coin again. Over twenty days bitcoin is up 14.3% and Strategy is up 18.9%; BitMine is up 29.1%. These vehicles fund themselves by issuing stock above the value of the coins they hold and buying more coins with the proceeds, so an expanding premium is both the symptom and the fuel. It is worth remembering that Strategy spent last week doing the opposite — repurchasing preferred and building a $4.8bn cash reserve rather than adding leverage. Now the contrast. Over the same five days CoreWeave fell 13.1%, Nebius 15.2%, Talen 12.4%, Oracle 8.7%, Meta 8.3%. Blue Owl fell 8.8%, Apollo 7.9%, KKR 6.3%. The S&P is 1.45% off its record and high-yield spreads have not moved from 275bp. The index is being held up by companies that do not need to borrow. Everything in it that does need to borrow is being marked down, quietly, one sector at a time.

Crypto and TradFi contagion19 Augmedium

A bank designed so the deposits can leave at 3am

The 2023 bank runs were the fastest in history because of technology; this one is being built with the friction deliberately removed.

Erebor Bank — founded by Palmer Luckey, backed by Peter Thiel and Andreessen Horowitz — got its charter from regulators only this year and is now raising capital at an $8bn valuation, nearly double the $4.35bn it reached late last year. It is based in Columbus, Ohio, and its clients are drawn from the technology and defence firms of California. It added hundreds of customers in the second quarter. Its stated mission is to fill the gap left by Silicon Valley Bank. Bloomberg's opinion desk makes the obvious point about concentration: lenders serving one boom-and-bust industry have a bad history, and Erebor is pitching to a homogeneous group of founders, investors and their companies, many of them close to its own backers. The less obvious point is in the product design. Erebor plans to use blockchain rails so clients can move their money at any time. SVB's run in March 2023 was constrained, such as it was, by wire cut-offs, bank operating hours and the mechanics of the payment system; $42bn still left in a day. A bank whose depositors can move balances at three in the morning on a Sunday has removed the last physical brake on a run, and it is doing so for a depositor base that reads the same group chats. What we do not know is the thing that actually killed SVB: the asset side. Duration mismatch plus concentration is fatal; concentration alone is survivable if the assets are Treasury bills. Erebor has not been through a cycle and there is no public balance sheet to judge. But a bank being valued at $8bn before it has funded anything through a downturn is itself a data point about the price of optimism.

Signs of the times

Stuff you wouldn’t have believed was possible until 2026.

Hudson River Trading posted $11.4bn of trading revenue and $7.4bn of net profit in the three months to June, a record. In the same market, Jane Street lost $15bn in July on an investment in a 24-year-old's AI hedge fund. In 2025 Hudson River, Jane Street and Citadel Securities between them took more than $60bn of trading revenue — more than the trading desks of JPMorgan or Goldman.

19 AugFinancial Times

Unitree rose more than 600 per cent on its Shanghai Star Market debut on Wednesday, reaching a $50bn valuation. Nearly 370 Chinese humanoid robot startups have been founded in the past two years and more than 50 have listed or are preparing to.

20 AugFinancial Times

Eric Trump told an earnings call in February that his Miami-based American Bitcoin was "fast becoming the leader in Bitcoin" with "the strongest brand of all", and thanked a short list of executives "and everyone at American Bitcoin". A filing a month later showed the company had two full-time employees, according to Forbes.

19 AugForbes (via Capital.gr)

Meta's Hyperion data centre vehicle carries about $27bn of debt, rated A+, priced at 6.58% and amortising to 2049, with PIMCO reportedly taking around $18bn of it. That is a 23-year repayment schedule secured on a building whose contents are re-rented by the hour.

20 AugCarson Group (via web sweep)

The rumour mill

What the crash-callers are saying, checked against real reporting.

Partly true

“$1.1bn of crypto short positions were liquidated in hours — the largest short liquidation on record, and over 90% of all liquidations in the window.”

K33 and derivatives data support roughly $1.1bn of bitcoin short liquidations in 24 hours and describe it as the largest daily bitcoin short liquidation ever, against prior records of $757mn in May 2021 and $694mn in November 2025. But shorts were about 88-89% of the $1.345bn total, not over 90%, and the record is specific to bitcoin rather than all crypto.

Claimed by Coin Bureau

Partly true

“Nvidia is raising $500 billion to help finance AI infrastructure.”

Nvidia's own release says it has signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500bn of third-party capital over time, with Nvidia backstopping up to about a quarter. It is not raising $500bn onto its own balance sheet — the distinction is the whole point of the structure.

Claimed by Heresy Financial

Partly true

“Oracle carries about $248bn of off-balance-sheet lease commitments on top of $130bn of debt, both agencies have it on negative watch, and its 5-year CDS is at a 16-year high.”

The lease number is real and now larger — Oracle disclosed $248bn of uncommenced lease commitments at 30 November 2025 and about $260bn at 31 May 2026, nearly seven times its recognised lease liabilities, on 15-to-19-year terms starting FY2027-29. The ratings claim overstates it: Moody's has a negative outlook, S&P downgraded rather than placing Oracle on negative watch.

Claimed by Coin Bureau

Partly true

“An Nvidia H100 that rented for $8 an hour in early 2024 was renting for $2-3 by late 2025.”

BCG puts the peak at about $8/hour in early 2024 and the trough at $1.96/hour in late 2025 — a steeper fall than claimed. But prices rebounded to $2.64 by April 2026, and mid-2026 median on-demand rates are quoted around $3.99 at GPU clouds. The collapse happened; the recovery is the part the bears leave out.

Claimed by Coin Bureau

Earlier developments

Dispatches from previous readings. The same argument, no longer the news.

The AI capex bubble19 Augmedium

Data centres are not financial assets, says the SEC

The fastest-growing form of AI infrastructure debt has just been moved further outside the disclosure regime built after the last securitisation blow-up.

The SEC has told data-centre owners that they do not have to comply with certain disclosure and risk-retention rules when they issue asset-backed securities. The reasoning: a data centre is not a "financial asset" like a mortgage or a car loan. The clarification was prompted by a request from Latham & Watkins, a major adviser to the AI sector, according to the Telegraph. Two things are being switched off here. Risk retention is the post-crisis rule that makes the sponsor of a securitisation keep a slice — usually 5% — so that it eats the first losses alongside the buyers. Regulation AB is the disclosure regime that forces the sponsor to publish standardised, asset-level data so buyers can build their own model rather than trusting the ratings. Neither rule changes the economics of a bad deal. Both change how much a buyer can see, and how much the seller cares. The technical argument is genuinely defensible. A data centre is a building with a tenant, not a pool of consumer loans, and the loan-level disclosure template does not fit. But the effect lands precisely where the risk is. The thing that will determine whether a data-centre bond pays is tenant concentration and the residual value of the building if the tenant leaves — which is exactly what Nvidia is now writing backstops against, up to $125bn across its financing platforms. Scale: JP Morgan sees data-centre securitisation reaching $30bn–$40bn a year, up from about $27bn in 2025. That is small against Morgan Stanley's $250bn–$300bn estimate for total AI debt issuance this year. It is the growth rate, not the level, that matters. One source carries this story. Treat the detail carefully. The direction is not in doubt: Professor Raghavendra Rau of Cambridge Judge put it as the SEC "replacing regulatory constraints with a reliance on market discipline." Market discipline requires disclosure. That is the part being relaxed.

Private credit and BDCs19 Aughigh

The same cars, pledged twice

Private credit's entire structure rests on collateral that is described in documents rather than independently inspected, and Tricolor shows how long a false description can survive.

The SEC has sued three former executives of Tricolor Holdings — founder and chief executive Daniel Chu, finance chief Jerome Kollar and senior finance director Ameryn Seibold — over a scheme that raised more than $1.9bn from asset-backed securities investors on collateral that was not what it appeared to be. Kollar and Seibold have pleaded guilty and are cooperating. Chu's lawyer calls the case a "rehash" of prior allegations and says many are inaccurate. The mechanic at the centre of it is double pledging: using the same used cars as collateral for more than one loan. "These defendants defrauded investors based on bogus collateral and violated the integrity of our private credit markets," the SEC's enforcement director David Woodcock said. Why this is possible is worth understanding, because it is structural rather than exotic. When a bank lends against a car, the bank holds the title and its regulator inspects the file. When an originator sells the loans into a securitisation, the pool is described in a prospectus, a servicer collects the payments and a trustee holds the paperwork — but nobody physically walks the lot. The verification is contractual. If the originator lies about which assets back which deal, the lie can run for years, and it did: from at least 2020 until Tricolor filed for bankruptcy in September 2025. This matters beyond subprime auto. The private credit boom rests on assets that are described rather than inspected — NAV loans against fund stakes, loans against sports teams, and now buildings full of chips. Each of those relies on somebody's assertion about what the collateral is and what it is worth. The enforcement itself is mildly reassuring. Somebody is checking. It just takes a bankruptcy to trigger the check.

Household credit19 Augmedium

Worse than 2008, and the lenders are near their highs

Subprime auto stress is past its 2008 peak and the lenders' shares are near record highs — because the loss sits with the securitisation buyer, not the originator.

Sixty-day delinquencies on US subprime auto loans have reached 6.9%, according to Fitch and Equifax data — above the 5.0% peak of the 2008 crisis. Treat the precise number with care: a separate read of Fitch's June data puts the same series at 5.67%, and the two are not obviously measuring the same thing. The New York Fed's Q2 figures are less disputable. Auto balances 90+ days past due sit at 5.49% of the $1.71tn outstanding, near the highest in a series going back to 2003, and the flow of loans newly entering serious delinquency rose to 3.00% annualised from 2.97%. Now look at the share prices. Credit Acceptance, which lends at the deep subprime end, is 8.7% off its recent high and rose 2.1% on the day. Capital One is 2.5% off its high and up 9.9% in a month. Synchrony is 1.7% off its high. Oportun, a near-prime lender, is up 35% in twenty days. Affirm rose 5.1% in a session. That gap is not irrational, and it is the useful thing here. The equity of a subprime lender is not a claim on the loans. Most of the credit risk has been sold into securitisations; what the shareholder owns is the origination and servicing fee stream, plus a retained residual that only takes losses after the ABS buyers' subordination is exhausted. Rising delinquencies eat the residual before they touch the stock. High delinquencies also mean high yields, and a lender that can still originate is earning them. So the equity market is saying: the losses land somewhere else. It is probably right about that. The question Crash Lab keeps returning to is who "somewhere else" is, and whether they have marked it. One genuine improvement: aggregate household delinquency fell to 4.7% of balances in Q2. The stress is concentrated, not general.

Bond market dysfunction20 Aughigh

Forty trillion, and a buyback JPMorgan doesn't believe

The government is managing its borrowing cost by shortening the maturity of its debt, which lowers today's yield and moves the risk to the refinancing.

Gross US federal debt crossed $40tn on Tuesday. It has grown $3tn in a year — the fastest pace outside the pandemic, on the FT's calculation. Debt held by the public now exceeds $32tn, roughly the size of the economy, and annual interest costs have passed $1tn, more than the defence budget. The Treasury's response arrived the same week: buyback operations in the 10-to-30-year sector doubled from $2bn to $4bn apiece, running 9 September to 4 November. The 30-year fell 10bp to 5.18% on the news, off its 5.33% high, the most since 2007. JPMorgan's rates team is not convinced. "Absent real fiscal consolidation, we fear the markets will view this action as lacking credibility," strategists including Jay Barry wrote. Their point is that the US runs a 6% deficit at full employment, and a buyback addresses the symptom. The plumbing is worth being precise about. A buyback is not money creation. Treasury purchases an old, illiquid long bond and must fund that purchase by issuing something else. It has already opened the door to cutting long-dated issuance and deepening its reliance on bills. The net effect, if that is the path, is to shorten the average maturity of the national debt: less duration for the market to absorb today, more of the debt rolling at whatever the short rate is tomorrow. That is a trade, not a solution, and it is the same trade a corporate treasurer makes when long money gets expensive. What we do not know is the funding mix. Treasury has not said. The programme was created in 2024 as market-functioning support and explicitly "not intended to ameliorate periods of acute market stress." It is now being used because yields are high before a midterm. Everyone can see that, including the buyers.

Hidden leverage and shadow banking20 Augmedium

Non-banks now do 40% of Europe's bilateral repo

Funding leverage is moving into bilateral markets where no single party can see how much any borrower owes in total.

The BIS reported this week that non-bank financial institutions now account for more than 40% of non-centrally cleared euro-denominated cash-borrowing repo, up from less than 30% at the end of 2020, and said the share is likely to keep growing. The Basel Committee separately flagged the growth of significant risk transfers — the trades by which banks buy protection on their own loan books — as warranting continued monitoring. Those two sentences describe the same migration from opposite ends. Repo is how leveraged investors fund positions: post a bond, borrow cash against it, buy another bond. When the trade clears through a central counterparty, the haircut is standardised, the exposure is netted and a regulator can see the aggregate. Non-centrally cleared repo is bilateral. The haircut is negotiated, the terms are private, and nobody has a consolidated view of how much leverage one borrower has taken from five different lenders. A hedge fund whose collateral falls in value gets a margin call from each of them separately, on the same day. The lending side of the same shift showed up in a small, concrete deal. Starwood Property Trust and Realterm provided a $672m loan against 78 industrial outdoor storage lots — 830 acres across 33 markets — refinancing an existing $486m loan. That is $186m more debt on the same dirt. Realterm's head of credit noted the firm has been expanding its lending platform "as traditional banks pull back." Meanwhile the banks themselves had a poor session on an up day for the index: Zions fell 3.7%, the regional bank ETF 2.4%, Goldman 1.8% and Morgan Stanley 1.5% while the S&P rose 0.2%. One day is noise. But it is the second week of regionals lagging, and it is the sort of divergence that shows up before anyone explains it.

Crypto and TradFi contagion18 Aughigh

The buyer of last resort left and the price went up

A leverage loop we had flagged as a forced-selling risk is unwinding in an orderly market rather than a disorderly one.

Bitcoin is at $71,784, up 13.9% in five days. Strategy — the company that was the largest marginal buyer of bitcoin for six years — has not bought any since mid-June. Over the past five weeks it has sold about $2.1bn of common stock, repurchased about $347m of its STRC preferred, and sold $213.3m of bitcoin to build a cash reserve now standing at $4.8bn against a roughly $59bn coin position. This is the flywheel in reverse, and it is worth being clear about why the flywheel existed. Strategy could issue equity above the value of the bitcoin behind each share, use the proceeds to buy more bitcoin, and thereby raise bitcoin-per-share for existing holders. That works only while the premium exists. When it collapses, issuing stock destroys value, and the company's fixed obligations — preferred dividends, interest — have to be paid in dollars it no longer earns from issuance. Hence: sell some coins, hold cash, buy back the expensive preferred. The fear was that this becomes a doom loop — a large holder selling into a falling market, forcing more selling. What has happened instead is that the market absorbed the retreat and then rallied without it. Strategy's own shares rose 12.6% on the day, and sit at their high for the period, as do the two ether treasury companies, BitMine and SharpLink. "Strategy's trading is pro-cyclical," Orbit Markets' Caroline Mauron told Bloomberg. "They are able to buy more when it goes up, and they are forced to sell when it goes down." We have been carrying digital-asset treasuries as a source of forced selling. On this evidence the deleveraging is happening in an orderly way in a rising market, which is the best case. That is a small reduction in the risk we were counting, and we are counting it. Strategy's stock is still down about 73% over a year.

Private credit and BDCs19 Augmedium

The queue at the private credit exit

Private credit's promise of quarterly liquidity against illiquid loans is being tested for the first time at scale, and the test is being passed by rationing rather than selling.

Investors in non-traded business development companies asked for 12.4% of net asset value back in the second quarter of 2026 — a record. Sponsors generally cap quarterly repurchases at about 5% of NAV. Only 38% of requests were met, leaving roughly $9.6bn queued. Across the first half, $12.7bn was withdrawn, $5.9bn of it in Q2. These numbers come from a secondary compilation, not from a primary filing we have seen, so treat the decimal places lightly. The structure they describe is not in doubt. Here is the mechanism. A non-traded BDC holds illiquid senior loans to mid-sized private companies. It promises investors quarterly liquidity, but only up to a cap — typically 5% of NAV a quarter — because it cannot sell a third of its loan book in ninety days. In calm conditions the cap is theoretical: nobody asks for more than 2%. Once requests exceed the cap, the fund pro-rates. Every investor gets a fraction of what they asked for, and the rational response is to ask for more next quarter to get the same amount out. The queue becomes self-feeding. Nobody has to default for this to happen, and no mark has to move. This sits directly on top of the asset-side number we wrote about earlier this week: median non-accruals at the twenty largest listed BDCs rose from 2.0% to 2.8% of cost between Q1 and Q2, the worst since 2017. Assets deteriorating and liabilities running are the two halves of the same problem. The listed BDCs, which cannot be redeemed at all — you just sell the share — are calm. ARCC is 1.9% off its high, BXSL up 0.7% on the week. That is the whole point of listing: the stress shows up in the price instead of the queue.

The AI capex bubble19 Aughigh

The market has split the AI trade in two

The AI selloff is concentrated precisely in the companies whose capex is funded by other people's debt, which is where the losses would land first.

Over the last twenty trading days Microsoft is up 24.1% and Amazon up 8.6%. Over the same period Core Scientific is down 20.8%, Vistra down 14.4%, Talen down 14.6% and Meta down 12.9%. In five days: CoreWeave -15.6%, Nebius -13.6% (nearly ten percent of it yesterday), Talen -11.7%, Core Scientific -10.2%, Applied Digital -9.3%. This is not the market losing faith in AI. It is the market separating the companies that pay for compute out of cash flow from the companies that pay for it with project debt, power contracts and leases. Meta is the tell. It is the one hyperscaler in the losing column, and it is the one that has leaned hardest on structures that keep the borrowing off its own balance sheet: a Louisiana campus vehicle reported at more than $27bn with Blue Owl-led investors holding about 80% and Pimco leading roughly $26bn of debt, and a $12.5bn El Paso data-centre bond priced in July at higher yields than a structurally similar 2025 deal (CNBC). That pricing detail is the whole story in one number. The same sponsor, the same asset type, a year apart, and lenders want more. Goldman analysts put hyperscaler lease commitments at about $1.5tn, up from roughly $200bn five years ago. Debt that does not appear on a balance sheet until the lease commences is still debt someone is funding. What has to stay true for this to work: the residual value of a data centre full of three-year-old GPUs, and a tenant who keeps paying for twenty years. Equity investors in the levered names are quietly repricing both. Investment-grade spreads, at 82bp, are not.

Bond market dysfunction20 Aughigh

Nobody fears rates. They want paying to hold them.

A high term premium with low rate volatility means the long end is being repriced structurally, not panicking — and buybacks address panic, not structure.

Two things in the bond market are moving in opposite directions and both are true. The New York Fed's estimate of the term premium is around 80 basis points, close to its highest in twelve years. That is the extra yield investors demand purely for the risk of owning a long bond rather than rolling short paper. It produced a 30-year auction on 11 August that cleared at 5.216%, the highest for that maturity since 2001, and a 10-year auction the next day at the highest financing cost since 2007 (Reuters via web sweep). The same repricing hit Japan, where the 10-year JGB reached 2.93%, its highest since 1996, and France, where the 10-year hit its highest since 2009. Meanwhile the MOVE index — the market's price for rate volatility — sits at 9.83, down 10.7% in five days and 26.9% in a month, and 37% below its recent high. Read together: nobody expects a violent surprise from the Fed. What has changed is the price of duration itself. Investors are not panicking about the next six months; they are declining to lend for thirty years at the old price. That is a structural repricing, and it is much harder for policy to fix than a scare. Which is why the Treasury's response is interesting. Doubling long-end buybacks from $2bn to at least $4bn per operation — roughly $32bn a quarter by analysts' reckoning — knocked 10bp off the 30-year on announcement. Against $32tn of outstanding Treasuries and a debt stock that has grown $3tn in a year, the size is a rounding error. It works as a signal or it does not work.

Hidden leverage and shadow banking19 Aughigh

One fund blew up and made hedging cheap

Cheap volatility after a blow-up is the mechanism by which a system that just took a loss ends up carrying more risk, not less.

Hudson River Trading made $11.4bn in trading revenue in the three months to June and $7.4bn of net profit, a record (FT). Jane Street lost $15bn in July on an investment in Leopold Aschenbrenner's Situational Awareness fund and other AI positions — and still generated more than $40bn of net trading revenue over the year to last Friday. The knock-on is more interesting than either number. Banks had been carrying large exposures to leveraged AI vehicles and to Situational Awareness through the summer, and hedged the jump risk by buying semiconductor options. That hedging demand is what made chip volatility expensive. Now the exposure is gone. "Now they don't need those hedges, and I believe unwinding of those hedges has made volatility in the sector inexpensive," Zed Francis of Convexitas told CNBC. Implied volatility on the VanEck Semiconductor ETF fell from 65% last month to 40%, the lowest since February. The ratio of open puts to calls dropped to 1.89, the most call-heavy since early April. So: a leveraged fund collapses, the banks that were exposed to it stop hedging, protection on the most crowded sector in the market gets cheap, and positioning turns more bullish than it has been in four months. The failure did not make the system more cautious. It made caution cheaper and therefore rarer. One person disagrees. On Monday a single trader bought 20,100 SMH puts for $129m — more than a third of all premium traded in the ETF that day, and 3.5 times the next-largest trade on the tape. Deep in the money, almost certainly a synthetic short.

Household credit19 Aughigh

They securitised the personal injury lawsuits

The securitisation machine has run out of conventional consumer collateral and is now buying claims on litigation outcomes.

A woman in New York who said she turned her ankle on a sidewalk took $76,500 in cash advances from a consumer legal funder while her case ran. By the time it settled she owed at least $1.4m, according to court records reported by the New York Times. Consumer legal funding works like this. A plaintiff waiting years for a personal injury case to resolve needs rent money. A funder advances it at 35–45% a year. If the case fails, nothing is owed — which is what lets the industry argue it is not lending. Most cases do not fail; insurers settle. So the funder collects. The new part is what happens next. The biggest funders are bundling thousands of these IOUs into asset-backed securities. Investors buy a claim on the future settlements; the funder gets its capital back immediately and lends it out again. The collateral is a pool of lawsuits whose value depends on the behaviour of insurance company claims departments, priced by a model nobody outside the deal can check. That is the Crash Lab thesis in one product. Risk that a bank would have to hold against capital, and mark, has been rebuilt in a structure where neither applies — alongside data centres, fast-food franchises and songwriting royalties. Background stress in the same population: auto loan balances hit $1.71tn in Q2 with 5.49% of balances 90 or more days delinquent, the worst since 2010, and credit card debt is at $1.26tn with roughly one dollar in eight seriously past due. One compilation puts subprime auto 60-day delinquency at 6.9%, above the 2008 peak; Fitch's June figure was 5.67%. We would use the lower one. Capital One is 2.5% off its high. Synchrony 1.7%. The lenders are not worried.

Crypto and TradFi contagion18 Aughigh

The treasury companies are pure beta now

Digital asset treasury companies have gone from being the bid under bitcoin to being levered bets on it, which changes what happens on the next drawdown.

Bitcoin rose 10% in five days to $69,323. On Wednesday, Strategy rose 12.6% in a single session to close at its highest in three months. BitMine Immersion rose 10.7%, SharpLink 12.2%, Coinbase 9.7%, Galaxy Digital 7.2%. That is a levered response to a coin move, which is exactly what these instruments are supposed to deliver — except that the mechanism that used to justify the leverage has been switched off. Strategy has not bought bitcoin since mid-June. Over the last five weeks it sold about $2.1bn of common stock, used the proceeds to buy back roughly $347m of STRC preferred shares and build a $4.8bn cash reserve, and sold $213.3m of bitcoin (Bloomberg). It still holds about $59bn of the coin. The original flywheel worked because the shares traded above the value of the bitcoin behind them. Issue stock at a premium, buy coins, coins-per-share rises, premium justified. When the premium goes, issuing stock destroys value for existing holders, and the machine has to run the other way: sell equity to retire preferred, sell coins to pay the preferred dividend. Strategy is now a company that services its own capital structure. Its stock is down about 73% over the past year against bitcoin's 45%. So what moved 12.6% yesterday is not an accumulation engine. It is a leveraged claim on a coin, with a fixed-charge burden on top, responding to a 10% rally. The distinction matters in the other direction: the same structure amplifies the way down, and unlike in previous cycles there is no longer a buyer of last resort standing underneath the market. The biggest marginal buyer of the last six years is now a marginal seller.

Fed, Treasury and policy19 Aughigh

Three dissents, five years above target

The single largest source of ignition risk in the next month is not a lender failing but a hawkish Fed meeting a weakening labour market while the Treasury manipulates the other end of the curve.

The Federal Reserve's July minutes, released Wednesday, show a 9-3 vote to hold rates at 3.5–3.75 per cent. It was the first time since 2016 that three FOMC members dissented in the same direction on a policy decision, and they dissented towards higher rates (New York Times). Inflation has now overshot the 2 per cent target for five consecutive years. Several participants said price pressures "appeared broad based" and that current settings were not taming them. A few argued that hiking in July would "likely help forestall the need for a steeper and potentially more costly sequence of tightening moves at a later stage." Kevin Warsh, in his first months as chair, has declined to say whether price stability requires higher rates. Now set that beside what the Treasury did on the same day. Scott Bessent doubled the size of long-end buyback operations, from $2bn to at least $4bn, explicitly to support the 10-to-30-year sector after the 30-year yield hit 5.34 per cent, its highest since 2007. One arm of the state is buying duration to push the long end down. The other is drifting towards pushing the short end up. If both happen, the curve flattens from an unusual direction: policy-driven at both ends, market-driven at neither. The 2s10s is at 52bp today. The complication is the labour market. Payrolls fell 23,000 in July against expectations of about 80,000 added, and participation dropped to 61.4 per cent as 264,000 people left the workforce. A central bank that hikes into that is not obviously wrong — but it is the textbook shape of a policy error, and the bond market is pricing almost none of it. The MOVE index is down 27 per cent in a month.

The AI capex bubble19 Augmedium

What Nvidia actually promised in Ohio

The credit of the AI buildout is being manufactured by one company's balance sheet, and the price divergence between Nvidia and the firms it finances is the market saying it knows that.

The Ohio deal has a structure now, and it is narrower than the headline number suggested. OpenAI has signed a 20-year lease on a 10-gigawatt campus in southern Ohio with SoftBank's SB Energy, partly sited on a former uranium-enrichment facility owned by the Department of Energy. Nvidia has agreed to backstop a portion of the value of the completed data centre, starting with phase one — around 5GW. In return, Nvidia becomes the exclusive chip supplier for the first half of the site and takes an equity stake in SB Energy. Crucially, the commitment "is structured to require a series of mitigation steps before the chip maker would be obligated to make any payments": if OpenAI walked, SB Energy would first have to try to re-lease the site (Wall Street Journal). That is not a guarantee of $105bn of rent. It is a residual-value floor with a waterfall in front of it. Our estimate of Nvidia's direct exposure here goes down on this reporting, not up. What has not changed is why the structure exists. Lenders will not fund a project of this size against OpenAI's own credit. They will fund it against Nvidia's. The whole edifice rests on one assumption that nobody has tested: that a purpose-built gigawatt-scale GPU hall retains substantial value with a tenant other than the one it was designed for. Every residual-value backstop in the AI complex is a bet on that sentence. The market is voting on it unevenly. Nvidia is 3.5 per cent off its high. The companies that actually carry data-centre debt are not: CoreWeave is down 15.6 per cent in five days, Nebius 13.6 per cent, Applied Digital 9.3 per cent, Core Scientific 20.8 per cent in a month, Talen Energy 11.7 per cent in five days.

Bond market dysfunction19 Aughigh

$40 trillion, and a debt ceiling in the diary

The long end got a policy bid, but the calendar now contains a hard fiscal deadline shortly after that bid expires.

US gross federal debt crossed $40tn on Tuesday. It has grown $3tn in a year — the fastest pace outside the pandemic. Debt held by the public now exceeds $32tn, roughly the size of the economy (Financial Times). Annual interest payments are projected to top $1tn this year, more than the Pentagon budget. The round number is not the news. The date is. Congress set the debt limit at $41.1tn last year. Budget analysts now say borrowing could reach it by early next year — earlier than planned, in part because of billions in revenue lost when Trump's tariffs were invalidated, which forced Treasury to borrow faster to cover the same bills (Washington Post). Six months ago the CBO thought total borrowing would top out at $39.4tn this fiscal year. It was already $39.9tn on Monday. Why this matters mechanically: a debt-limit standoff forces Treasury into extraordinary measures, which means running down its cash balance and skewing issuance to bills. That drains and then floods the front end at exactly the moment the Treasury is trying to manage the long end with buybacks. Note the buyback programme's end date: the day after the November elections. The debt limit arrives after that. For now the intervention worked. The 30-year fell 10bp to 5.18 per cent, TLT rose 1.7 per cent, and the dollar hit a three-month low. Bloomberg's Marcus Ashworth called Bessent's move an "inappropriate co-opting" of a programme created for market functioning, not for acute stress. Both things can be true: it is a distortion, and it worked.

Household credit19 Aughigh

The collateral was pledged more than once

Consumer credit stress is measurable and rising, the securitisation chain has just been shown to have a verification hole in it, and the equities are priced as though neither is happening.

The SEC has sued Tricolor Holdings' founder Daniel Chu, former CFO Jerome Kollar and former senior finance director Ameryn Seibold over the subprime auto lender's collapse. The allegation is specific: from 2020 until the September 2025 bankruptcy, the firm raised more than $1.9bn through asset-backed securities while making false statements about its finances, faking documents, and pledging the same vehicles as collateral multiple times — double pledging (Bloomberg). Kollar and Seibold have pleaded guilty and are cooperating. Chu's lawyer calls the case a "rehash" and says many allegations are inaccurate. Read the SEC enforcement director's sentence carefully: "These defendants defrauded investors based on bogus collateral and violated the integrity of our private credit markets." A subprime auto ABS is now described by the regulator as private credit. That is the correct description, and it tells you where the boundary now sits. The mechanism worth understanding: an auto ABS is a promise that thousands of specific cars exist, are worth something, and are pledged to you and only you. Nobody checks each car. The structure substitutes a servicer's representations and an auditor's sampling for verification. When those fail, the tranche does not degrade gracefully — the collateral simply was not there. The backdrop is not comfortable. New York Fed Q2 data put auto balances at $1.71tn with serious delinquency at 5.49 per cent of balances; Fitch and Equifax put subprime 60-plus-day delinquency at 5.67 per cent in June. Pending home sales fell 2.3 per cent in July to 71.2, matching the second-worst reading since 2001. The shares disagree. Credit Acceptance rose 2.1 per cent on the day, Affirm 5.1 per cent, Oportun 35 per cent in a month. Only Ally and Sallie Mae are soft.

Private credit and BDCs19 Aughigh

Apollo saw it in April 2024 and shorted it instead

Private credit's core weakness is not credit quality but the absence of any process that forces two holders of the same asset to agree on what it is worth.

A 112-page filing in the First Brands bankruptcy, from the emergency chief executive Charles Moore, contains a small and very clarifying story. Katsumi Global, a Japanese trade-finance player, tried to sell part of its exposure to First Brands' factored invoices to Apollo. Apollo declined "after correctly identifying the numerous red flags discussed herein, even without the benefit of the specific disclosure known to Katsumi" (FT Alphaville). The filing quotes an internal Apollo presentation from April 2024 — 17 months before the collapse. Its key issues: "the cash flow of the business has never matched the Company's claims about its margins"; a claimed 20 per cent operating margin at a company that had been acquiring businesses with sub-10 per cent EBITDA margins; and a chief executive who "has controlled dozens of entities over the years, many of which have ended up in litigation or liquidation (or both). Simple Google searches reveal almost nothing about him." This is the private-credit marking problem in a single anecdote. One institution was carrying the exposure at something close to face value and trying to sell it. Another had concluded, from public information and a Google search, that the accounts were not credible. In a public market that disagreement becomes a spread and both sides get marked. In private credit there is no mechanism that forces the two views to meet. Apollo eventually expressed its view the only way available — by shorting First Brands' public debt. Sector-wide, the FT's count of median non-accruals at the 20 largest listed BDCs at 2.8 per cent of cost, up from 2.0 per cent in a quarter, is the closest thing to a market price the asset class produces. The BDC shares still trade within two per cent of their highs.

Crypto and TradFi contagion19 Auglow

Tether got audited. Strategy is selling. Both are up.

Two of the largest untested claims in crypto — Tether's backing and Strategy's premium — are being resolved in an orderly way rather than in a panic, which is what fragility falling actually looks like.

Tether has completed its first full independent financial audit. KPMG issued an unqualified opinion on the statements, which showed reserves exceeding liabilities by $6.8bn at the end of 2025 (CNN Brasil). Treat the sourcing carefully — this reached us through one outlet and we have not seen the statements. But if it stands, it is the clearest example this year of a risk getting smaller rather than larger. For a decade the single largest untested mark in crypto has been the question of whether the biggest dollar stablecoin is actually backed. A Big Four firm has now signed an unqualified opinion on it. That is not a rally; it is a measurement. We are marking that fragility down. The other half of the picture is Strategy, and it is stranger. Michael Saylor has not bought bitcoin since mid-June. Over five weeks the company sold about $2.1bn of common stock, bought back $347m of STRC preferred shares, sold $213.3m of bitcoin, and built its cash reserve to $4.8bn against roughly $59bn of holdings (Bloomberg). The flywheel that made the sector — issue equity above net asset value, buy coins, repeat — is deliberately running backwards. And the stock rose 12.6 per cent on Tuesday, to sit at the top of its recent range, with bitcoin up 10.5 per cent in five days to $69,626. Note what that conceals: Strategy is down about 73 per cent over a year while bitcoin is off 45 per cent. The gap between those two numbers is the premium deflating. The ether treasury companies moved harder still — SharpLink up 12.2 per cent, Bitmine 10.7 per cent. A sector deleveraging into a rally is the least dangerous version of this story.

What would change our mind

The specific, observable things that would move the number - in either direction.

High-yield OAS breaking above 350bp from 275bp, or IG through 100bp — the credit market has been the flat line in every reading for a month and is doing the most work holding the number down.

Would move the number

A named bank disclosing a writedown on Radiant World exposure, which would turn a documents scandal into a measurable loss in bank trade finance.

Would move the number

A non-traded BDC suspending repurchases outright rather than pro-rating them, or a second BDC following BlackRock TCP Capital's reported 19% NAV cut.

Would move the number

The 30-year holding below 5% after Treasury's expanded buybacks actually begin on 9 September — that would be an untested policy mark testing well, and would take ignition down two or three points.

Would move the number