Investment-grade AI bonds are now priced like junk

Archived reading, published Sat, 22 Aug 2026 22:19:50 UTC (4 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility90

how much tinder is stacked up — moves slowly

Ignition41

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 22 sources and rewrites this page.

Status: Held at 66. The market data is again Friday's closes — VIX 15.1, high-yield spreads 275bp, investment-grade 82bp, the S&P 1.6% off its record — so ignition has not moved, though its composition has: rate volatility is 28% lower than a month ago while the funding-dependent AI names fell hard in five days (Nebius -21%, CoreWeave -17%, Talen -13%). Fragility stays at 90 because Bloomberg's data-centre bond pricing is a sharper measurement of what AI debt costs, not a new stack of it.

Reporting from 20 Aug to 22 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.

The AI capex bubble22 Aughigh

Rated investment grade, sold to junk buyers

The cost of AI infrastructure debt is now set by junk buyers, whose participation is the least reliable part of the capital structure.

QTS Realty Trust sold $3.9bn of bonds this week to fund a Georgia data centre tied to Microsoft. The bonds carry investment-grade ratings. They yielded about 7.23%. BlackRock is paying 7.53% on blue-chip paper sold in July for a Meta-linked Texas project. In both cases, underwriters placed the debt with high-yield investors as well as high-grade ones (Bloomberg). On our arithmetic, that is roughly the yield of the average junk bond. The high-yield index trades at 275bp over Treasuries and the ten-year is at 4.69%. So a bond the agencies call safe is clearing at the price of one they do not. There are two readings and they are not mutually exclusive. One: the natural investment-grade buyer base is full of this sector, so the deals need money from somewhere else. Two: investors do not believe the ratings, which rest on long leases from creditworthy tenants for buildings whose economics depend on those tenants still wanting them in 2040. Either way the identity of the marginal buyer matters. "We're seeing high-yield investors become tourists in investment-grade technology debt," Advent Capital's Steven Schweitzer told Bloomberg. Tourists go home. A high-yield fund buying IG paper for the carry is the first seller when its own market widens, and it has no mandate obligation to show up for the next deal. Companies have borrowed more than $410bn for data centres and AI this year on Bloomberg's count. That pipeline is now partly dependent on a buyer who is there opportunistically. What would make this benign: the deals keep clearing at these yields with genuine IG demand. What would not: the next one prices wider, or gets pulled.

The AI capex bubble22 Augmedium

The complex splits into funders and funded

The market is now discriminating between hyperscalers and the leveraged periphery that builds for them, which is where any AI credit event starts.

Over five days the Nasdaq fell 2.1%. Over the same five days Nebius fell 21.1%, CoreWeave 16.6%, Talen Energy 13.4%, Applied Digital 12.8%, Core Scientific 11.6% and Vistra 8.1%. Microsoft was down 2.2%, Alphabet 0.3%. That is not a sector selling off. That is the market sorting the AI trade into companies that pay for capex out of operating cash flow and companies that pay for it out of the capital markets. The second group is being repriced; the first is not. Nebius is the sharpest case. It upsized a convertible bond to $5bn on 20 August, described as one of the largest on record, and the stock fell a fifth in the same week. Part of that is mechanical and worth naming: convertible buyers are typically arbitrageurs who hedge by shorting the underlying stock, so a large convert creates a large short. That is our inference from the structure, not something the filings state, and the size of the fall looks larger than hedging alone would explain. The power names are a different mechanism. Talen and Vistra are the pure expression of the assumption that data centres get built and plug in. Texas has halted approvals of roughly 1,800 of them and politicians in both parties have discovered that voters dislike them (WSJ). Note the ordering. Equity in the funding-dependent names is being marked down while the bonds financing the same buildings clear at 7.2%. That is the correct sequence — equity takes first loss. But equity is also the collateral for the next raise, and several of these companies need one.

Fed, Treasury and policy22 Aughigh

Treasury steps into the long end, the Fed wants out

Fiscal and monetary duration policy are pulling opposite ways in a market already clearing 30-year paper at the worst yields since 2001.

Two arms of the state are moving in opposite directions in the same market. The Treasury announced on 19 August, unscheduled, that it would at least double buybacks of longer-dated debt to $4bn per operation between 9 September and 4 November, and Scott Bessent then said it could be more. The relief lasted about a day: the 30-year fell to 5.18%, then went back to 5.25% and touched 5.27% on Friday. The dollar fell to a three-month low and has not recovered (FT). Meanwhile Fed chair Kevin Warsh has named a balance-sheet task force — Jeremy Stein, Karen Dynan, Raghuram Rajan — to propose reforms by year-end, with an explicit brief to shrink the Fed's footprint in markets. Darrell Duffie's view, given to FTAV, is that shrinking the balance sheet is possible but not currently necessary, would take years, and that the asset mix will drift towards Treasury bills (FT). Put those together. The Fed is planning to hold less duration. The Treasury is buying duration back from the market. Buybacks have to be funded by issuing something else, and the Treasury has not said what — the assumption in circulation, that it is bills dollar-for-dollar, is not established by anything we can see. If it is bills, the government's average maturity shortens, which lowers today's coupon cost and raises the amount that has to be rolled every quarter. Warsh speaks at Jackson Hole on Friday. Three of twelve FOMC voters wanted a hike in July. This is the single scheduled event most likely to move the long end in the next week.

Private credit and BDCs22 Augmedium

The market is marking the managers, not the loans

Insurance balance sheets are the largest untested holder of private credit, and the first public unwind of one is happening now.

Over five days: Apollo -5.3%, KKR -4.9%, Blue Owl -4.4%. Over the same five days the funds those firms run barely moved — Ares Capital +0.4%, Blackstone Secured Lending +0.3%, Golub -0.5%. Investors are selling the fee streams while leaving the loan books alone. The context is the insurance channel. Federal prosecutors are investigating Mark Walter's businesses, and the probe has revealed that two insurers he controlled provided more than $20bn of previously undisclosed financing to other parts of his empire. Walter has agreed to sell the Lakers and put his Chelsea stake up for sale to clean up the insurers' balance sheets (FT). Guggenheim runs $367bn. We counted the $20bn earlier this week. What is new is the remedy: a forced seller has appeared, and what is being sold is sports teams, not credit. That is the good version of a deleveraging — the Lakers do not need a levered buyer and their sale does not mark anybody else's book. The mechanism worth understanding is why insurers hold this paper at all. Annuity liabilities are long and sticky, so an insurer can genuinely afford illiquid loans that a bank cannot. That is the legitimate trade, and KKR, Apollo and Brookfield all run versions of it. The dangerous variant is affiliate lending, where the asset is a loan to a related party and the price of that asset is set inside the same house. "Other private capital firms are engaged in essentially the same business model," the University of Texas's Andrew Granato told the FT — his distinction is disclosure, not structure. The equity market appears to agree that the risk sits at the manager. We would watch whether it stays there.

Crypto and TradFi contagion22 Augmedium

A mutual fund's cash can now live on a chain

Regulated fund cash management is being wired into blockchain settlement at the same time as the levered crypto equity complex is at its highs.

The SEC has issued a no-action letter allowing Franklin Templeton's traditional registered funds to hold shares of its onchain BENJI money market fund for cash management. Mastercard has closed its $1.8bn acquisition of BVNK, a stablecoin infrastructure company. Ripple Prime closed an upsized $275m private placement of senior unsecured notes (CoinDesk). None of these is a leverage story. All three are plumbing. The Franklin Templeton letter is the most consequential: the cash sleeve of an ordinary registered fund can now sit in a tokenised money market fund, which means the fund's ability to settle depends on a blockchain being available and on the tokenised vehicle's redemption mechanics working under stress. Those mechanics have never been tested in a redemption run. Neither had the ones in prime money funds in 2008, and those were older. Separately, and unrelated in mechanism, the digital-asset treasury complex has gone vertical. Bitcoin is up 19.5% in five days. Strategy is up 28.1% and at its high for the period; BitMine 26.3%, SharpLink 26.4%, Coinbase 25.7%, all at their period highs. Gold is at a record and the dollar at a three-month low in the same week. That pairing is the tell. These are not the prices of a funding squeeze; they are the prices of a bid for anything that is not a dollar-denominated government promise, expressed with leverage. Which means it works in both directions. The treasury companies are levered beta on a coin, and their premium to the coins they hold is the first thing to compress when the coin stops going up.

Signs of the times

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

Tether's $120m Uruguayan ghost mines

Tether announced two bitcoin mining sites in Uruguay in 2023 and quietly abandoned both after a dispute over how much electricity it would be supplied; one person with direct knowledge estimated the project likely cost around $120m, and little was left to show for it. A researcher described the model as "plug-and-play" — literally pulling the plug and moving.

cryptoenergyabandoned capex

21 AugReuters (via Channel NewsAsia)

A $2.25bn bond for Jane Street's servers

Zenith Arc completed a $2.25bn data-centre bond deal to fund a facility leased to Jane Street, according to Bloomberg. The proprietary trading firm is now a data-centre credit story, and its lease is the collateral.

securitisationai capex

21 AugBloomberg (via web sweep)

"People really effing hate data centers"

Abdul El-Sayed, running for Senate in Michigan, said that at a July campaign event; his Republican opponent Mike Rogers now wants a "pause" in the sector. Texas Governor Greg Abbott declared his state the "epicenter of AI development" in November alongside a $40bn Google investment, and this month halted approvals of some 1,800 data centres.

politicspermittingai capex

21 AugThe Wall Street Journal

The rumour mill

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

Confirmed

“Central banks bought a record 289 tonnes of gold in Q2 2026, more than five times the revised Q1 figure of 57 tonnes, with Poland adding 51 tonnes and China 33 tonnes.”

The World Gold Council's Gold Demand Trends report carries all these figures, including the 62% year-on-year rise from 177.9 tonnes. With gold at a record and the dollar at a three-month low, this is the official-sector leg of the same trade.

Claimed by Coin Bureau

Confirmed

“US spot bitcoin ETFs had their worst half-year on record in H1 2026, with $5.4bn of net outflows including roughly $4.5bn in June across a 13-day outflow streak.”

Confirmed by multiple data trackers. Worth holding next to the price: bitcoin is up 19.5% in five days regardless, which means the marginal buyer this month is not the ETF channel.

Claimed by Coin Bureau

Partly true

“Michael Saylor sold bitcoin during the recent run of negative crypto news.”

Strategy the company sold 1,690 BTC between 3 and 9 August for about $108.6m, and last week sold $333.7m of its own stock while buying no bitcoin and ending with a $4.8bn dollar reserve. Saylor personally says he has never sold a satoshi. The company and the man are not the same balance sheet.

Claimed by Anthony Pompliano

Partly true

“Bessent's buyback intervention failed — long yields reversed almost entirely, and rose again after his CNBC interview floating bigger buybacks.”

The direction is right and well documented: the 30-year fell to about 5.19% on the announcement, was back at 5.25% the next day and touched 5.27% on Friday. The specific attribution to a Deutsche Bank note is not established.

Claimed by Meet Kevin

Earlier developments

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

Private credit and BDCs21 Augmedium

Non-accruals jumped 40% in a quarter. Nobody sold.

Non-accruals are the earliest hard number in private credit that the manager cannot mark to its own opinion, and they are now rising faster than the equity of the lenders implies.

The median share of loans on non-accrual at the twenty largest listed US business development companies rose to 2.8% of cost in the second quarter, up from 2.0% at the end of March — the highest level since roughly 2017, on Solve data cited by the FT. Fitch's US private credit default rate reached 6.1% for the twelve months to July, up from 6.0% in June, a record for the series. Non-accrual is a specific thing. It means the lender has stopped booking interest income on a loan, either because the borrower has stopped paying or because the lender no longer expects it to. It is not a write-off; it is the step before one. And it is measured here against cost, not fair value, so a loan already marked down still counts at what was originally lent. A 0.8 percentage-point move in three months is fast for that measure. The dispersion matters as much as the median. FS KKR reported 7.1% of its credit portfolio as problematic. Investcorp's small BDC saw NAV per share fall from $3.65 to $3.44, down 5.7% in a quarter, with five investments on non-accrual. Goldman Sachs BDC went the other way: non-accruals down from 3.2% to 2.9% of fair value, NAV off less than 1%. The shares do not care yet. Ares Capital closed the week 1.3% off its recent high and up 6% in twenty days; Blackstone Secured Lending is 1.0% off its high; Prospect Capital, the perennial stress case, is up 11% in twenty days and sits 0.6% below its high. Only Blue Owl's BDC (-3.7% in five days) and FS KKR (-3.5%) sold off. That is defensible arithmetic rather than complacency: on portfolios yielding around 10%, 2.8% of loans paying nothing still leaves the dividend covered. The loss arrives later, in NAV.

The AI capex bubble21 Augmedium

The AI build is now bidding against the Treasury

When a data-centre bond prices at junk yields and a 23-year amortisation lands in a bond fund, the duration risk of the AI build has been moved to savers who will not mark it for years.

Two numbers from the same week. The US sold $25bn of 30-year bonds at 5.216%, the highest auction yield since 2001. And QTS Realty sold a $3.9bn investment-grade bond to fund a Microsoft-linked data centre in Georgia at what Bloomberg described as junk-like yields. Those facts are connected, and the Treasury Secretary said so himself: Bessent told CNBC his buyback expansion was partly about supporting a thinly traded market that has to "compete with a lot of corporate issuance at higher yields, including for artificial intelligence infrastructure." Hyperscaler debt issuance reached $220bn in 2026 by 10 August on BNP Paribas data; one estimate puts global AI-related issuance at $489bn by mid-year, with five companies accounting for nearly 40%. The structures are where the interesting risk sits. Meta's Hyperion vehicle carries $27bn of debt rated A+, priced at 6.58%, amortising to 2049, with PIMCO reported to have taken about $18bn of it. That is a bond fund lending for twenty-three years against a building whose economic value depends on chips that depreciate in a fraction of that time and on a lease that is the entire credit. Meta's follow-up El Paso deal, around $12.5bn, priced wider than the comparable 2025 transaction. The cost of AI money is rising even as the volume does. The equity of the same complex is not going along with it. Nebius upsized a convertible to $5bn, one of the largest on record, and its shares fell 21% in five days. CoreWeave is down 16.6% over the week, Applied Digital 12.8%, Core Scientific 11.6%, Talen 13.4%. Debt markets are still clearing these deals. Equity markets are marking down the people who have to keep doing them.

Crypto and TradFi contagion22 Augmedium

Strategy sold stock, bought no bitcoin, kept the cash

The digital-asset treasury trade only works while these companies are net buyers; the biggest one just spent a week not buying while its stock rose 28%.

In the week to 16 August, Strategy — the original digital-asset treasury company — sold about $333.7m of its own shares, made no bitcoin purchases and no sales, and ended the week with a $4.8bn dollar reserve. The mechanism these companies run on is simple. Issue equity at a premium to the value of the coins you hold, use the proceeds to buy more coins, and the coins per share rises even though the share count did. It works only while the premium exists. When a treasury company sells stock and does not buy the asset, one of two things is happening: either the premium no longer justifies buying, or the cash is needed for something else — servicing the preferred and convertible obligations that sit above the common shareholders in the capital structure. The filing does not say which. What makes it odd is the price action. MSTR rose 28% in five days and finished at its recent high. Bitcoin is up 19.7% over the same stretch to around $77,200. The ether treasury companies did the same thing: BitMine up 26% in five days and 44% in twenty, SharpLink up 26% and 36%, both at their period highs. Coinbase is up 25.7% in a week. So the whole complex re-rated hard while the largest and longest-running member of it chose to hold dollars. We do not know Strategy's premium to net asset value from the material we have, and without it we cannot say whether this is discipline or necessity. But a treasury company accumulating fiat is, on its face, the flywheel running in reverse.

Household credit22 Aughigh

Two consumers, one dataset

Prime card losses look priced; subprime auto is at a record and is where the collateral fraud of the last two years also showed up.

The New York Fed's second-quarter household debt report is a study in averages hiding a split. Total household debt fell $13bn to $18.8tn. Credit card balances rose $21bn to $1.26tn, and the share of card balances 90 or more days delinquent reached 12.8%, against 7.6% in mid-2022. Auto loans 90+ days late hit 5.49% of balances, just below the 5.60% Q1 peak and the highest since 2010. Two measures are being quoted interchangeably and should not be. The 12.8% is a stock — the proportion of all card balances currently sitting in serious delinquency. The 6.97% also reported is a flow — the annualised rate at which balances are newly transitioning into 90+ days late. The stock is high because bad balances sit there for quarters before being charged off; the flow is what tells you about the next few months. On the flow, the monthly issuer data is calmer than the headline. Capital One reported a 4.12% annualised net charge-off rate on domestic cards for July and 3.48% of balances in 30+ day delinquency. Charge-offs are realised losses; 30-day delinquency is the pipeline feeding them. A pipeline running below the current loss rate means losses are more likely to flatten than accelerate. Which is why the lenders' shares look the way they do: Capital One up 7.8% in twenty days, Synchrony up 9.5%, Credit Acceptance up 4.6%, Oportun up 33%. The exception is subprime auto. Fitch's 60+ day subprime delinquency index printed 6.13% for July, the highest July reading in a series going back to 1994. That is not the same borrower as the Capital One card book, and it is not stabilising.

Bond market dysfunction22 Aughigh

The yields didn't fall. The dollar did.

Fiscal stress that cannot express itself in yields expresses itself in the currency, and that is a slower, wider channel than a bond selloff.

The Treasury's unscheduled announcement on 19 August that it would at least double buybacks of long-dated debt, from $2bn to $4bn per operation between 9 September and 4 November, bought about a day of relief. The 30-year fell to 5.1765% and then rose 5.4bp to 5.247% the next session. By Friday it was around 5.24–5.27%, roughly where it started. Bessent then said the size could go above $4bn per issue. What did move was the currency. The dollar hit a three-month low against the euro, which traded as high as $1.1711. "Bessent's efforts to suppress US yields haven't done much for US yields, but it's undermined the dollar," Marc Chandler of Bannockburn told Reuters. "The market is pushing back." That is the mechanism worth understanding. A buyback does not retire debt; it swaps long bonds for short bills. The government's interest bill becomes more sensitive to the front end, and the duration the market has to absorb falls temporarily. If investors believe the motive is price management rather than housekeeping, they do not stop demanding compensation — they demand it somewhere else. The somewhere else is the exchange rate. The FT's markets column makes the second point: predictability is the entire product in a $32tn market, and support measures get read as an admission that something has broken. This is the third unscheduled Treasury intervention in three weeks, following yen purchases and a facility explicitly designed to avoid selling Treasuries. Gross federal debt crossed $40tn on 18 August. Rate volatility, meanwhile, sits near a twenty-day low. The long end is repricing in an orderly way. It is repricing all the same.

The AI capex bubble21 Aughigh

The index is fine. Its financiers are not.

The AI buildout depends on continuous refinancing by companies whose equity is the junior tranche of every future deal, and that equity is being marked down while the index is not.

Over the last five sessions, Nebius fell 21.1%, CoreWeave 16.6%, Talen Energy 13.4%, Applied Digital 12.8%, Core Scientific 11.6% and Vistra 8.1%. Over the same five sessions the S&P 500 fell 1.4% and now sits 1.6% below its record. High-yield spreads are 275bp. Investment grade is 82bp. Nothing in the credit market registers what is happening to this specific group of companies. The grouping is not random. These are the firms that build and power AI capacity without a hyperscaler balance sheet behind them, and they fund themselves with project debt, converts and offtake contracts. Nebius upsized a convertible to $5bn on 20 August — $3bn due 2030, $2bn due 2034, and one of the largest on record (Reuters, via sweep). A convert is equity sold at a distance. When the stock drops a fifth in a week, the distance gets longer, and the next tranche costs more. The power names are a different mechanism with the same shape. Talen and Vistra are priced off the assumption that data-centre demand arrives on schedule at contracted prices. Texas has halted approvals of roughly 1,800 data centres over power and water strain, and a Republican governor is being attacked from the left on electricity bills (WSJ). The reporting says demand is overwhelming: $220bn of hyperscaler bond issuance in 2026, QTS placing $3.9bn at junk-like yields for a Microsoft-linked site, Meta's El Paso paper printing wider than last year's comparable. The prices say the equity funding channel for everyone who is not a hyperscaler is closing first. Where reporting and prices disagree, we would watch the prices.

Bond market dysfunction22 Aughigh

Three surprises in three weeks

Three unscheduled interventions in three weeks have moved the currency instead of the yield, which is the expensive way for a fiscal problem to resolve.

On Wednesday, without warning, the US Treasury announced it would at least double the size of its buybacks of longer-dated debt, from $2bn to $4bn per operation, running from 9 September to 4 November. On Thursday Scott Bessent said the number could go higher still. On Friday the 30-year yield closed at 5.27%, the highest in nearly two decades, and the dollar hit a three-month low against the euro. Buybacks are ordinary housekeeping. Doing them unscheduled is not. As the FT puts it, predictability is the whole game in a $32tn market, and support measures get read as an admission that something has broken (FT). The mechanism matters. A buyback does not reduce the debt; it swaps the maturity profile, retiring long bonds and funding that with bills. It removes duration from the market without removing a dollar of borrowing. If the fiscal path is unchanged — and the debt crossed $40tn this week, with the $25bn 30-year auction clearing at 5.216%, the highest for that tenor since 2001 — then capping the yield does not remove the adjustment. It relocates it. "Bessent's efforts to suppress US yields haven't done much for US yields, but it's undermined the dollar," Marc Chandler of Bannockburn told Reuters. "The market is pushing back." The third surprise is the one fewest people noticed: three weeks ago Bessent bought yen, and said future interventions would use an unusual facility designed specifically to avoid selling Treasuries. Bond investors read that as an attempt to stop Japan funding yen support out of its stock of US debt. Warsh speaks at Jackson Hole on Friday.

Private credit and BDCs22 Aughigh

Selling the Lakers to recapitalise an insurer

The retirement savings funding private credit sit behind a disclosure regime, not a capital regime, and one participant has just been shown to have driven straight through it.

Federal prosecutors are investigating Mark Walter, and the investigation has established that two insurance companies he controlled provided more than $20bn in previously undisclosed financing to other parts of his business empire. Walter has agreed to sell the Los Angeles Lakers and put his Chelsea stake up for sale, in a rush to clean up those insurers' balance sheets (FT). The structure is worth understanding because it is not unusual. Annuity liabilities are long-dated and cannot run, which makes them the ideal funding for illiquid private credit — that is the whole insight behind private capital's move into insurance, and Guggenheim reached $367bn under management on the back of it. The check on the arrangement is not capital; it is disclosure. Affiliated transactions have to be visible, and the assets have to be valued by somebody with no stake in the answer. Which is why prosecutors also subpoenaed the ratings firm that graded some of the bonds. The FT's sourcing is blunt about the read-across. "What makes Mark Walter unique is that he failed to disclose a tremendous amount of affiliated transactions," says Andrew Granato of the University of Texas. "But other private capital firms are engaged in essentially the same business model." KKR, Apollo and Brookfield all own insurers that lend into their own orbits. The distinction between fraud and business model is the whole question, and this window does not resolve it. What the window does show is a forced seller: trophy assets going to market on a schedule set by regulators rather than by the owner. Apollo is off 5.3% in five days and 7.2% from its recent high; the S&P is off 1.6%.

Hidden leverage and shadow banking22 Auglow

Banks sold €905bn of risk. Some came back.

This is the risk-migration thesis in its purest form: a market that grew eighteenfold in a decade, moving credit from the regulated balance sheet to the unregulated one — sometimes on the regulated balance sheet's money.

European and global banks sold exposure to €378bn of credit through significant risk transfer deals in 2025, according to figures attributed to the International Association of Credit Portfolio Managers. There are now more than 800 SRT transactions outstanding across 62 entities, with a live credit balance of €905bn. A decade ago the figure was a little over €50bn (IACPM, via web sweep). An SRT works like this. A bank holds a portfolio of loans and buys credit protection on the first-loss piece from a fund — usually a private credit or hedge fund — which posts cash against it. The bank keeps the loans and the client relationship, hands over the first losses, and gets regulatory capital relief for the reduction in risk. Everyone gains something real. The loans stay on the balance sheet where a supervisor can see them. The complication is the one the ECB has flagged: banks sometimes finance the funds that buy the protection. The sweep quotes ECB supervisory communications to the effect that where a bank lends to the investor taking the transferred risk, the risk that appears to have left the balance sheet may in substance be partly retained, with lower overall capital coverage against it. The circle is not perfectly closed — the fund's own equity absorbs losses first — but the bank's exposure is no longer zero, and the capital held against it is. What we do not have is a number for how much of the €905bn is bank-financed, or a US figure at all. This is a single secondary source citing a trade body, not a primary release, and we would treat the €905bn as an order of magnitude rather than a measurement.

Crypto and TradFi contagion21 Augmedium

A hundred times leverage on a stock you cannot buy

$1.4tn of leveraged exposure to real-world assets now sits in a market that produces no position data anyone can supervise.

Global trading volume in perpetual futures hit a record $86.2tn in 2025, up 47% on the year, according to CoinGecko. Perps tied to real-world assets — stocks, commodities, anything not a token — went from $23.6bn in the first half of 2025 to $1.4tn in the first half of 2026. That is roughly sixtyfold in twelve months (Nikkei). A perpetual future has no expiry date. A periodic funding payment between longs and shorts keeps it tethered to the spot price, and the only constraint on size is the margin the venue demands. Kalshi, which is US-regulated, caps bitcoin perps at six times. Bybit allows a hundred. What that combination now buys you is exposure to assets you have no legal route to. Before humanoid-robot maker Unitree listed in Shanghai this week, its pre-IPO perpetual future traded near $100 on several platforms, more than four times the listing price. One contract, xyz.UNITREE, did over $105m of volume on Hyperliquid within 24 hours of the IPO. The shares then opened up 629%. "The perpetual market called the mania before it happened," a Pepperstone strategist told Nikkei. The point for this site is not the mania. It is that this is leveraged equity exposure with no prime broker, no 13F, no short-interest report and no margin data anybody aggregates. Gross positioning in these instruments is invisible until a liquidation cascade makes it visible. We have spent this month tracing leverage out of banks and into private funds and money-market plumbing. This is the same migration, one step further out: the leverage has left the reporting perimeter entirely.

Household credit22 Augmedium

Household debt shrank. The card book did not.

The aggregate household numbers are being used to argue the consumer is healthy, and they cannot bear that weight — the stress is real but concentrated in the two smallest books.

Two things in the New York Fed's Q2 data point in opposite directions, and the gap between them is the story. Total US household debt fell $13bn on the quarter to $18.8tn. Aggregate delinquency eased to 4.7% of balances from 4.8%. On those numbers the American household is fine. Inside them: credit card balances rose $21bn to $1.26tn, and the share of card balances 90 or more days delinquent reached 12.8%, against 7.6% in the second quarter of 2022 (NY Fed data, via sweep). Auto balances 90+ days late stand at 5.49%, just below the 5.60% Q1 peak and the highest since 2010. Fitch's July index put subprime auto 60+ day delinquency at 6.13%, the highest July reading in a series running back to 1994. The reason the aggregate can improve while the card book deteriorates is composition. Mortgages are most of household debt, and most of them were written by high-score borrowers at coupons that will never be refinanced. That block is inert and it dominates the average. The revolving and auto books are where the marginal borrower actually lives, and they are 8% and 6% of the total. A household crisis confined to those lines will not show up in the headline number until it is large. The lenders are not behaving as if it is large. Capital One is up 7.8% over twenty days, Synchrony 9.5%, Affirm 9.6%. The market's view is that a 12.8% serious delinquency rate is fully priced into the APR, and on the evidence of the last two years it has been right. One caution: the 12.8% figure reaches us through commentary on the NY Fed release rather than the release itself, and it is a share of balances, not of accounts.

Hidden leverage and shadow banking22 Augmedium

The 2007 pipe has been repointed at equities

If money funds stop rolling conduit paper, equity leverage disappears in days and the selling lands on the most crowded stocks.

A JPMorgan note picked up by the FT this morning makes an observation that ought to be better known: the growth in asset-backed commercial paper issuance coincides with, and may be directly connected to, growth in equity repo volumes. The plumbing. ABCP is very short-dated debt, mostly issued daily, bought overwhelmingly by money-market funds looking to earn a few extra basis points on cash. It is issued by conduits — vehicles sponsored by, but not consolidated onto, a bank. JPMorgan's Teresa Ho, Pankaj Vohra and Molly Herckis write that certain ABCP structures, including independent sponsor programmes, "can provide banks with off-balance-sheet solutions/optimization of funding and potentially favorable accounting treatment by using a conduit to intermediate transactions with a counterparty" (FT). Strip out the language and the chain is: your cash sits in a money fund; the fund buys overnight paper from a conduit; the conduit finances equity collateral; a hedge fund — or a retail brokerage client — holds a leveraged stock position at the far end. This is the structure that funded subprime mortgages into 2007 and seized in the 2020 dash for cash. What is different is the collateral. Equities can be sold in a morning, which sounds reassuring until you notice that everyone deleveraging at once is what makes a momentum unwind. Goldman's global momentum index had its worst systematic long-short day in more than two years last Wednesday; Morgan Stanley said it was the first time in five years its pure momentum index fell more than 4% while the S&P rose. The note does not say how much of the ABCP growth is equity-linked. Without that number we cannot size the exposure, and we have not raised fragility for it. But this is exactly the migration Crash Lab tracks: leverage financed off balance sheet, by the most run-prone liability in the system, and counted nowhere in bank capital.

The dollar, gold and reserve status22 Augmedium

Reserve managers say they are leaving. Gold agrees.

The marginal foreign official buyer of Treasuries is the thing that has quietly absorbed US deficits for thirty years, and it says it is stepping back.

An OMFIF survey published this week found that, for the first time, more central banks plan to reduce dollar exposure over the next decade than to increase it. A separate 2026 World Gold Council survey found 74% of reserve managers expect the dollar's share of global reserves to fall over five years, 89% expect central-bank gold holdings to rise over the next year, and 45% intend to add gold themselves. The crossover has already happened on one measure. Per the European Central Bank's June review, gold made up 27% of total official reserves at end-2025 against 22% for US Treasuries — the first time gold has led since 1996. Why this is a plumbing story and not a talking point. A reserve manager holding Treasuries is lending to the US government; the coupon comes back, the position rolls, and the bid is automatic. A reserve manager holding gold is holding a bar that pays nothing and finances nobody. Every percentage point that shifts is a percentage point of deficit that has to find a private buyer at a price. That is the same buyer Scott Bessent is now competing with himself, having doubled long-end buybacks to at least $4bn per operation from 9 September. Prices are consistent. Gold closed at $4,680, up 3.6% in a day and 15% in twenty, at a record. Silver is up 19% in twenty days. The dollar index is 98.84, down 2.6% over the month, with the euro at its highest since 14 May. The caution: these are surveys of intent, and intent is cheap. What is not a survey is the 30-year Treasury at 5.33% last week, its highest since 2007, and the record price of the asset that competes with it.

Private credit and BDCs22 Aughigh

Non-accruals at a decade high, BDCs at their highs

The market is pricing BDC credit as a rounding error while the one unmanipulable credit metric has gone up 40% in a quarter.

Two numbers from the past week. The median non-accrual ratio across the twenty largest listed BDCs rose from 2.0% of cost at the end of March to 2.8% in the second quarter — the highest since the aftermath of the 2017 oil bust, on FT analysis of Solve data. And Fitch's US private credit default rate reached 6.1% for the twelve months to July, up from 6.0% in June, a record for the series. A non-accrual is a loan where the borrower has stopped paying interest, or the lender no longer expects them to. It is the least gameable number a BDC publishes, because unlike a mark it does not depend on a valuation committee. Now the prices. Ares Capital, the largest listed BDC, sits 1.3% below its recent high and is up 6% in twenty days. Blackstone Secured Lending is 1.0% off. Prospect Capital — the perennial stress case — is 0.6% off its high and up 11% in a month. Blue Owl's BDC and FS KKR are the softer ones, down 3.7% and 3.5% over five days. The two claims can both be true. A 2.8% non-accrual rate with normal senior-secured recoveries is a low single-digit hit to net asset value, and the dividend still covers. Equity holders are being paid roughly 10% to wait, and they are waiting. What would break the reconciliation is the accrual side. Payment-in-kind income across BDC portfolios rose from about 6% in 2022 to about 10% by early 2026, per Boston Fed work cited in our sweep. PIK is interest recognised as income and added to the loan balance rather than received in cash. A rising PIK share and a rising non-accrual share are the same borrower, six months apart.

The AI capex bubble21 Augmedium

AI debt is getting more expensive one deal at a time

Credit indices will be the last place the AI financing cycle turns up, because the risk is being written in vehicles no index tracks.

The spread indices say nothing is happening: investment grade at 82bp, high yield at 275bp, both roughly where they were a month ago. The individual deals say something else. Meta's $12.5bn El Paso data-centre bond priced at higher yields than the comparable Hyperion deal, which came at 6.58% on an A+ rating with PIMCO taking about $18bn of the $27bn. QTS Realty sold a $3.9bn high-grade bond to fund a Microsoft-linked centre in Georgia at what Bloomberg called junk-like yields. Nebius upsized a convertible to $5bn — $3bn due 2030, $2bn due 2034 — one of the largest converts on record. Hyperscaler bond issuance hit $220bn in 2026 as of 10 August, on BNP Paribas data. And Broadcom is in talks for $70–80bn, roughly $45bn senior and $35bn junior, with Blackstone and Apollo, issued through a special-purpose vehicle, to buy chips for Anthropic and others. The structure is now standard: an SPV, an offtake contract with a creditworthy tenant, sometimes a vendor guarantee, and a rating that reflects the contract rather than the technology. It is a good structure if the tenant pays for twenty years. It is a real-estate loan if the tenant does not. The equity market is discriminating in the same direction. Over five days: Nebius −21%, CoreWeave −17%, Applied Digital −13%, Talen −13%, Vistra −8%, while Microsoft, Alphabet and Oracle are within a few percent of their highs. Money is being priced project by project, not sector-wide. That is healthier than the alternative. It also means the sector's cost of capital is rising well before any credit index shows it.

Household credit22 Augmedium

Subprime auto is at records. Subprime lenders are near highs.

Household credit stress is real but concentrated at the bottom, and the market is right that lenders can price it — until the funding side balks.

The New York Fed's Q2 household debt report puts 90-plus-day delinquency on auto balances at 5.49%, just below the 5.60% first-quarter peak and the highest since 2010. Fitch's July index shows subprime 60-plus-day delinquencies at 6.13%, the highest July reading in a series running back to 1994. Credit card balances rose $21bn to $1.26tn. One caveat we cannot resolve: summaries of the same Fed release put card 90-plus-day delinquency at 12.8% of balances, while others report 6.97%. Those are not the same measure and the sweeps do not say which. We are not going to build an argument on a number that appears twice at half the size. What is clear is the divergence. Aggregate household delinquency actually fell, to 4.7% from 4.8%. Deep subprime auto is at a record. Those two facts describe two different populations, and the second one is small enough not to move the average. Which is roughly what the equities say. Over twenty days: Capital One +7.8%, Synchrony +9.5%, Credit Acceptance +4.6%, Oportun +33%. Consumer lenders are not bets on borrowers being fine. They are bets on the spread between what they charge and what they pay, minus losses they have already provisioned for. Charge-offs at record levels are survivable if funding is cheap and APRs are high, and both currently hold. The channel that matters is not the delinquency rate but the securitisation bid, which is where these loans go. Tricolor showed how that fails — not gradually, through defaults, but suddenly, when the collateral turns out not to exist.

Crypto and TradFi contagion22 Augmedium

The bitcoin company is accumulating dollars

When the flagship treasury company chooses cash over the coin, the mechanism that makes these vehicles work is being paused by the people who run it.

Between 10 and 16 August, Strategy sold about $333.7m of its own stock, bought no bitcoin, sold no bitcoin, and lifted its US dollar reserve to $4.8bn. The digital asset treasury model has one engine. Issue equity at a premium to the value of the coins you already hold, use the cash to buy more coins, and the coins-per-share number goes up. It works precisely as long as the stock trades above the net asset value of the holdings. Selling stock and holding the proceeds in dollars is the same engine with the second stroke missing. There are innocent readings. A $4.8bn dollar reserve services preferred dividends and convertible obligations without touching the bitcoin, which is prudent liability management and something the older version of this company was criticised for not doing. There is a less innocent reading, which is that management does not think the coin is the best use of the dollar right now. The disclosure does not say which. What makes it interesting is the price action around it. MSTR is up 28% in five days and sits at its recent high. Bitcoin is at $76,774, up 19% over five days and down 2% today. Coinbase is up 26% and at its high; BitMine and SharpLink, both ether treasury companies, are up 26% each. The whole complex ripped in the same week the dollar hit a three-month low and gold hit a record. That is the tell. This rally is not a bet on blockchain adoption. It is the same trade as the gold bid, in a different costume — and it is being expressed through vehicles that carry leverage gold does not.

Fed, Treasury and policy22 Aughigh

The two arms of policy are pulling opposite ways

When the Treasury tries to hold down a yield the Fed may be about to push up, the adjustment happens in the exchange rate — which is exactly what gold at a record and the dollar at a three-month low are telling you.

On 19 August the Treasury said it would at least double its buybacks of long-dated Treasuries, from $2bn to at least $4bn per operation, running 9 September to 4 November and targeting the 10–20 and 20–30 year sectors. The next day Scott Bessent told CNBC it could be more than $4bn per issue (Gulf Times). This followed a $25bn 30-year auction that cleared at 5.216%, the highest stop for that tenor since 2001, and a 30-year yield that touched 5.33% on 18 August, the highest since 2007. It did not work for long. The 30-year rallied on the announcement and then rose again, to 5.247%. Charlie McElligott at Nomura called the plan a "band-aid on a bullet hole"; Jim Caron at Morgan Stanley Investment Management put it more flatly: "The Treasury simply can't control long-term yields" (FT). The plumbing matters. A buyback does not retire debt. It swaps an old, illiquid long bond for cash raised elsewhere — in practice, bills. So the operation shortens the average maturity of a $40tn debt stock and moves duration risk from private balance sheets onto the rollover calendar. It also announces that the issuer cares about a price it cannot set. Meanwhile the July FOMC minutes, released the same day, showed three dissents in favour of a quarter-point hike and "many" participants saying one would be needed if inflation did not fall. So the fiscal authority is buying duration while the monetary authority discusses tightening. The pressure has to come out somewhere, and this week it came out in the currency: the dollar hit a three-month low against the euro. "Bessent's efforts to suppress U.S. yields haven't done much for U.S. yields, but it's undermined the dollar," said Marc Chandler of Bannockburn (Livemint/Reuters). Kevin Warsh speaks at Jackson Hole on Friday.

What would change our mind

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

A data-centre bond deal that gets pulled, or one that needs a sponsor guarantee to clear — the QTS and BlackRock deals worked at 7.2-7.5%, and the first that doesn't is the signal.

Would move the number

High-yield OAS above 350bp, or investment-grade above 110bp, which would mean the credit market has stopped treating the AI build as a rates story and started treating it as a credit story.

Would move the number

A listed BDC cutting its dividend or a large non-traded BDC formally gating redemptions, which would turn the non-accrual data into a liquidity event.

Would move the number

The 30-year above 5.50% with the dollar falling at the same time — the combination that says the buyer strike is about credibility rather than supply.

Would move the number