Crash Lab — how close is the crash?

CRASH-O-METER

Cracking

how close are we

Fragility88

how much tinder is stacked up — moves slowly

Ignition38

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 28 sources and rewrites this page.

Status: Held at 63 for a fourth run: nothing broke, and the tape is unchanged (VIX 15.46, high yield 270bp, investment grade 81bp, S&P 1.5% off its high, rate volatility down 25% in twenty days). The genuinely new facts — Korea's export bank fleeing to a New Zealand market it has not used in nine years to escape AI issuance, and the FT's reporting that insurers are refusing to wrap the biggest data-centre projects — are sharper measurement of tinder we already count, not fresh tinder.

Reporting from 25 Aug to 26 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.

Bond market dysfunction26 Aughigh

The Kauri market reopens after nine years

AI capex has stopped being an equity story and become a competitor to governments for the world's fixed-income capacity.

The Export-Import Bank of Korea sold NZ$625mn ($372mn) of five-year notes on Tuesday. It was the first Kauri bond — a security issued in New Zealand by a foreign borrower — in nine years. The reason given was not exotic: Sungho Park, a director in the state-owned bank's treasury department, told Bloomberg that issuance costs are rising because of competing sales from "AI-related hyperscalers", and that the bank was "tapping into new markets where the impact is minimal" (Bloomberg). The amount is trivial. The mechanism is not. Bond investors have a finite appetite for duration at any given yield, and that appetite is now being bid for by issuers who do not much care what they pay. Scott Bessent said as much last week: corporate issuance is "almost yield agnostic". Alphabet sold $25bn in a single go. Bloomberg counts more than $410bn borrowed this year for data centres and related AI investment. A yield-agnostic issuer sets the clearing price for everybody else. That is the crowding-out channel, and it runs directly into the story we have been tracking on the other side of the ledger: the Treasury is buying back its own long bonds, at least $4bn per operation from 9 September, precisely to hold that clearing price down. One arm of the system is trying to suppress long yields while another arm of the same system issues into them without a reserve price. A Korean agency detouring to Auckland is a symptom, not an event. But it is the first concrete case we have seen of an official borrower rerouting because of AI capex, and it will not be the last.

The AI capex bubble26 Aughigh

The insurers said no

The insurance market declining to wrap data-centre risk is the clearest external signal yet that this paper is not as diversifiable as its structure implies.

Tech companies are expected to spend $7tn on data centres by 2030. The FT reports this morning that the people who would normally take the tail risk off the lenders' hands are declining to do so: insurance companies "have been reluctant to provide billions of dollars' worth of coverage for the biggest projects, concerned about catastrophic risk as well as concentrated exposure to these giant sites" (FT). This is the part of the machine most readers never see. Project finance works because risk is sliced and sold to whoever is best placed to hold it: construction risk to contractors, demand risk to an offtaker, catastrophe and outage risk to insurers, credit risk to lenders. Each layer that refuses leaves its share sitting with the layer above. So look at what has filled the gap. Nvidia is providing up to $105bn of credit support for OpenAI's Ohio lease, and has organised six asset managers around $500bn of compute financing. The collateral in these structures is GPUs and offtake contracts. Carlos Mendez of Crayhill Capital, who says he has repeatedly turned down chip-collateralised deals lacking structural protections, puts the obsolescence problem plainly: "It's like you're financing a fax machine and then someone invented email." And there is a risk almost nobody has priced: permits. Eric Klar of White & Case notes that local opposition or a lost permit can stop construction midway. "For lenders who have already put money into the project, the question is: what happens then?" When the professional risk-takers step back and the chip vendor steps forward, the risk has not been transferred. It has been concentrated — in the one balance sheet whose revenue depends on the same buildout succeeding.

Crypto and TradFi contagion26 Aughigh

MSCI may evict the index's biggest bitcoin holder

Digital-asset treasury companies are leveraged bets on their own share premium, and index inclusion is the least-discussed pillar holding that premium up.

MSCI is consulting on rules that would exclude from its global indexes companies "primarily engaged in accumulating assets rather than operating businesses". Applied as of May, the methodology would have removed Strategy Inc., Japan's Metaplanet and the uranium vehicle Yellow Cake (Bloomberg). Why this is a funding question rather than a share-price question. Strategy has bought more than $60bn of bitcoin, financed by issuing equity and a stack of preferred and convertible securities. That machine only runs while the shares trade above the value of the coins: sell stock at a premium, buy bitcoin, repeat. Index membership underwrites the premium, because passive funds must own the shares in proportion to their weight regardless of what they think of the strategy. Take the shares out of the index and you get mechanical selling from those funds, a narrower premium, and worse terms on the next security issued to buy the next tranche of coins. The market is not pricing this. Strategy is up 19.6% in five days and 33.5% in twenty, sitting 1.7% below its recent high, with bitcoin at $78,547 and up 22% over the same twenty days. Nothing in the tape suggests a financing threat. Two caveats. MSCI abandoned a narrower crypto-exclusion plan in January before opening this broader review, so consultations of this kind do not always conclude. And the consultation targets non-operating companies generically, not digital assets, which means the outcome turns on a definitional judgement rather than a view on bitcoin. But the pressure point is precisely where we have said the fragility in this structure sits: not the coin, the financing.

Fed, Treasury and policy26 Aughigh

5.8%, and the Treasury is easing

A fiscal authority suppressing long yields while inflation expectations rise is the textbook opening move of fiscal dominance, and the bond market is currently rewarding it.

The Conference Board's August survey put consumer confidence at 89.4, a seven-month low, down from a revised 90.2 in July. In the same release, consumers' expected inflation over the next twelve months rose to 5.8% from 5.6%. Both figures have been checked and confirmed against the Conference Board release and Reuters. Set that beside what the Treasury is doing. Bessent has at least doubled buybacks of 10- to 30-year debt to $4bn per operation from 9 September, and CNBC reports the department is weighing funding those purchases from the Treasury General Account — roughly $950bn parked at the Fed. Big investors have told the FT this pulls directly against the Fed's inflation fight; several Fed officials have said the case runs the other way, towards higher rates (FT). The plumbing: buybacks lower long yields, long yields set mortgage and business borrowing costs, and lower borrowing costs ease financial conditions. That is monetary stimulus delivered through debt management, by an agency with no inflation mandate, while the household inflation expectation it feeds into is rising. Krishna Guha of Evercore ISI says the move may have unsettled "folks on the FOMC as well". For now the market has taken the side of the Treasury. Since the announcement, Treasuries have outperformed swaps, taking the 30-year spread to its tightest since February — Citi's Jason Williams calls it a "Treasury put". The 10-year sits at 4.70% and our rate-volatility gauge is down 25% in twenty days. Kevin Warsh speaks at Jackson Hole on Friday. It is the first time anyone will hear him address this directly.

Household credit26 Augmedium

12.92% delinquent, 3.34% charged off

Aggregate household credit numbers are being held up by prime borrowers while the subprime tail deteriorates, which is how consumer credit cycles always look shortly before the tail stops being a tail.

Two household credit numbers from the second quarter point in opposite directions, and the gap between them is the story. New York Fed data has 12.92% of credit card balances 90 or more days past due — down slightly from 13.1% in the first quarter, which was the highest reading since 2011. JPMorgan's card business, meanwhile, reported a net charge-off rate of 3.34%, down from 3.47% in the first quarter and 3.40% a year earlier. Total household debt was $18.8tn, actually $13bn lower quarter on quarter, with 4.7% of balances in some stage of delinquency. These are not contradictory, and understanding why matters. The 90-plus share is a stock: the proportion of outstanding balances that are deeply impaired. It rises when the impaired pile stops shrinking or when the healthy pile stops growing — and household debt is now flat to falling, so the denominator is no longer bailing out the ratio. A charge-off rate is a flow, and it is specific to one lender's book. JPMorgan's card portfolio skews prime. What the two numbers together describe is a distribution pulling apart, which is the same shape we flagged yesterday in autos: subprime 60-plus delinquencies at 6.13% in July against 0.49% for prime, a ratio of more than twelve to one. Auto debt 90-plus days behind stands at 5.5%, the second-highest in the series. The aggregate looks stable and slightly improving. It is stable because a large, healthy majority is averaging out a tail that keeps getting worse. Lenders exposed to the tail — Oportun, Credit Acceptance — are not trading like it: both are up more than 7% in twenty days.

Hidden leverage and shadow banking25 Augmedium

Borrowed money, on deposit, in Doha

Leverage that is created inside private wealth management is invisible to every measure regulators publish, which is the whole point of it.

UBS is pitching its wealth clients on leveraged deposits at Qatar National Bank, having previously offered the same product for Saudi Arabia's Al Rajhi Bank (Bloomberg). Al Rajhi has advertised around 5.1% on savings; QNB has indicated similar rates this year on deposits of at least £100,000 through its Doha and London branches. The trade is a carry trade with a bank deposit at one end. A client borrows dollars against their portfolio, places the proceeds on deposit at a Gulf bank paying more than the client's funding cost, and keeps the difference. Currency risk is limited — but only limited — because the Saudi and Qatari riyals are pegged to the dollar. Leverage is what turns a spread of a point or two into a return worth pitching. What has to stay true: the peg holds, the client's funding rate does not rise faster than the deposit rate, and the deposit is repaid. The failure mode is not gradual. A peg is either intact or it is not, and the position is concentrated in one bank in one country, which is exactly the risk a diversified deposit book is meant to avoid. This is small beer next to the AI financings. We flag it because of where it sits. Leverage extended to wealthy individuals against securities portfolios is not in the hedge fund leverage statistics the OFR publishes, not in bank credit aggregates as lending to funds, and not in anyone's systemic map. Meanwhile banks continue moving risk the other way: AIB is preparing a significant risk transfer on about €2.5bn of project finance loans, in a market heading for a record $45bn year.

Signs of the times

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

Revenue doubles, shares up 12%

Nvidia reports quarterly results on Wednesday with analysts expecting revenue above $92bn, almost twice the same period last year. Its shares are up just 12% since the start of 2026 — about a fifth as much as an index of leading chipmakers. The market is being offered near-doubling revenue and pricing in something else.

ai capexvaluation

26 AugThe Economist

Investment grade, junk yield

QTS Realty sold $3.9bn of bonds this week to fund a Georgia data centre tied to Microsoft, yielding about 7.23% despite investment-grade ratings; BlackRock paid 7.53% in July on blue-chip paper for a Texas project. When high-grade debt has to be priced for junk buyers, the rating and the risk have parted company.

creditdata centres

26 AugBloomberg (via web sweep)

Everyone launches a stablecoin; volumes cool

Revolut launched EURR, a euro-pegged token, joining Visa, Klarna and several global banks. Bloomberg notes the combined market value of stablecoins has plateaued in 2026 after years of growth, with Visa's on-chain transaction data also cooling. The supply of issuers is growing considerably faster than the demand for the product.

cryptoadoption

26 AugBloomberg

Twenty-two bank charters in nineteen months

Il Messaggero reports the OCC gave preliminary approval on 14 August to World Liberty Trust Company — a national trust bank that would custody the reserves behind the USD1 stablecoin, with a Trump-family entity holding a stake — and that the OCC has approved 22 bank charters in nineteen months, more than in the previous five years combined. Single-source, and we have not seen the OCC document; the charter-count figure is the striking one.

policycryptounconfirmed

25 AugIl Messaggero

The rumour mill

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

Confirmed

“The Treasury basis trade is now larger than it was at its 2019/early-2020 peak.”

Fed-linked estimates put cash-futures basis positions at roughly $830bn as of September 2025, close to double the early-2020 peak, with hedge fund positions at about 3.5% of privately held Treasuries against 2.5% at the last peak. The OFR separately put hedge funds' cash Treasury holdings near $2tn.

Claimed by Thoughtful Money (Adam Taggart)

Unsupported

“The US is spending $700bn this year on AI data centres.”

The $700bn figure is JPMorgan's estimate of 2026 capex for the five largest hyperscalers globally and firm-wide, not US data-centre spending. Actual US data centre project starts were $81.5bn in the first half of 2026 on CoStar figures. The real number is enormous; this is not it.

Claimed by Kitco NEWS

Partly true

“University of Michigan consumer sentiment is at the lowest level in its history.”

Preliminary August 2026 sentiment was 51.0 — below the 1st percentile of the series — but the record low was 44.8 in May 2026. Historically bad, not historically worst, and the claim that every prior approach to this level was followed by a recession is not supported by any source we can find.

Claimed by Palisades Gold Radio

Partly true

“A third of everything American households own is now in stocks, the highest share ever recorded.”

Federal Reserve data has equities at about 32% of household assets, roughly the highest in quarterly records going back to the early 1950s and well above the 27% dot-com peak. "About the highest" rather than a confirmed record, but the direction is right and it matters: household balance sheets are now levered to the same concentration we track in the index.

Claimed by Kitco NEWS

Earlier developments

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

The AI capex bubble26 Aughigh

The vendor is now the credit

When the supplier underwrites the customer, chip demand and chip credit risk stop being two separate things.

Nvidia reports results for the three months to 26 July today. Analysts expect revenue above $92bn, almost double the same quarter last year. The shares are up 12% since the start of 2026 — about a fifth as much as an index of leading chipmakers (The Economist). That gap is the story. On 10 August Nvidia announced a partnership with six large Wall Street firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — aiming to raise more than $500bn for AI compute infrastructure. A week later it agreed to provide up to $105bn of credit support to help OpenAI lease a data centre in Ohio. The plumbing matters more than the headline number. A vendor guarantee on a lease is not a sale. It is a contingent liability whose value depends on whether the tenant can pay rent for the life of the building. If OpenAI pays, Nvidia has bought demand for its chips with a signature rather than cash. If OpenAI cannot, Nvidia owes the landlord — and it owes it at exactly the moment its own order book is falling apart, because the two are the same event. The filings in front of us do not say how the guarantee is accounted for, whether it sits on balance sheet, or what triggers a call. That is the single most important disclosure in tonight's release and we would not assume it appears. The market's verdict so far is a shrug: revenue nearly doubling, the stock up 12% and 5.5% below its high. Either investors have stopped paying for revenue bought with credit, or they have not read the terms.

Hidden leverage and shadow banking26 Auglow

Chips as collateral, amortising to 2049

The AI buildout is being funded with paper that converts a tech-demand bet into a rated bond held by insurers.

Inside the wave of AI project finance is a structure worth pausing on. One vehicle described in this week's sweeps carries $27bn of debt rated A+, priced at 6.58%, amortising out to 2049, secured on GPUs and offtake contracts — with PIMCO reported to have taken around $18bn of it (Bloomberg / Carson Group summary). This is a single trade-press summary and we have not seen the offering documents, so treat the precise figures as indicative. The mechanism is what matters. Debt amortising to 2049 is being secured on hardware whose economic life is measured in single-digit years. That mismatch is not necessarily fatal — the collateral that actually supports the rating is the offtake contract, the promise by a creditworthy tenant to keep paying for compute. Strip away the securitisation language and this is corporate credit exposure to a handful of AI labs and hyperscalers, wearing an asset-backed costume and an investment-grade badge. Who holds it is the point. Not banks. Bond funds, insurers and private credit vehicles — investors who mark to model, face no capital charge for concentration, and do not have to disclose the position by tenant. Carson's estimate is that incremental debt funded about 32% of hyperscaler capex on a trailing basis by mid-2026, against 9% in fiscal 2024. If an offtake contract is renegotiated in 2029, nobody has to write anything down that week. That is precisely the property that made this class of risk attractive to build in the first place.

Household credit25 Augmedium

6.13% against 0.49%

The bottom of the consumer market is at crisis-era delinquency rates while the firms lending to it trade near highs — the divergence resolves when ABS funding, not the borrower, turns.

Fitch's July data put 60+ day delinquencies on subprime auto loans at 6.13%, up from 5.80% at midyear and described as the highest July reading in a series running back to 1994. Prime borrowers, in the same month, were at 0.49% — more than twelve times lower (Fitch data, via web sweep). A separate 6.9% figure is circulating from a different presentation of the same series; we use the dated July number. The rest of the week's household prints point the same way. New York Fed data has 5.5% of auto debt 90+ days behind in Q2, the second-highest in the series. Conference Board confidence fell to 89.4 in August from 90.2, a seven-month low, and one-year inflation expectations rose to 5.8% from 5.6%. New single-family home sales fell 10.5% in July to a 607,000 annual rate, 6.3% below a year ago. Now look at the lenders. Credit Acceptance, which lends at the deepest end of subprime, is up 5.9% in five days and sits 5.3% off its high. Affirm is 0.75% off its high; Synchrony 0.25%. Only Ally and Capital One are soft, both down about 2.4% on the week. That is not irrational. A subprime auto lender earns a spread wide enough to absorb a lot of default, and the securitisation market funding it is still open at tight spreads. Equity holders are paid for the spread until charge-offs exceed it. The delinquency series tells you where the borrower is. The share price tells you the funding is still there. Both are currently true.

Private credit and BDCs25 Auglow

Somebody bid 35% below the mark

A hostile bid is the closest thing to a market price a non-traded BDC ever gets, and it came in a third below the mark.

Saba is reported to have offered to buy into Blue Owl's non-traded BDC at roughly a 35% discount to marked value (web sweep). This is single-sourced trade commentary, we have not seen the tender terms, and the discount a hostile buyer offers is not a valuation — it is an opening bid with a liquidity premium baked in. It still matters, because non-traded BDC net asset value is otherwise an untested number. The manager marks the loans, an independent valuation agent reviews them, and investors redeem at that mark subject to a quarterly cap. There is no market price to argue with. When the top ten non-traded BDCs received redemption requests averaging 13% of assets in Q1 and 14% in Q2 — enough to trigger gates — the queue itself became a price signal, but a silent one. A named outside bidder putting a number on it is louder. The fundamentals underneath are drifting the same way. Non-accruals at the top ten listed BDCs reached 3.95% of debt at cost in Q2, up 20bp on the quarter, on a $83.6bn book. BCRED's non-accruals went from 0.6% to 2.4%, driven substantially by Medallia — marked at 60 cents on a $1.1bn cost basis — and Affordable Care International at 70. Payment-in-kind loans have risen from about 6% of BDC lending in early 2022 to roughly 10%, per the Boston Fed. Meanwhile the listed BDCs sit near their highs: Ares Capital 1.3% off, Blackstone Secured Lending 0.8% off. The traded marks and the contested marks are not telling the same story.

Fed, Treasury and policy26 Augmedium

Jackson Hole, and the accord nobody has signed

If the Fed formalises coordination with Treasury on the long end, the inflation anchor becomes a political variable rather than a monetary one.

Kevin Warsh speaks at Jackson Hole on Friday, his first public remarks since July's contested rate decision. The FT frames the week as a chance to formalise an emerging Fed–Treasury accord, and reports that the rise in term premium is being blamed partly on Warsh's own communication style (FT). The two arms of policy are now saying different things. Warsh has abolished forward guidance and has previously said rising yields were a healthy sign of the market working things out. Bessent, having doubled long-end buybacks to at least $4bn per operation, says "we believe that the yields don't reflect the underlying fundamentals". Mark Cabana of Bank of America puts the market's complaint plainly: "Warsh's 'resolute commitment' to lower inflation is not good enough for the market." What an accord would actually mean: the Fed accepting debt-management as a constraint on its own balance sheet. Warsh wants to shrink that balance sheet. Doing so while Treasury is buying back long bonds would have the two institutions pulling against each other in the same maturity bucket. Coordination avoids that — at the cost of making it explicit that the price of long money is now set jointly by the fiscal authority. The market has largely stopped fighting. The rate-volatility gauge we track is down 25% in twenty days and 15% in five. But the Conference Board's one-year inflation expectation went to 5.8% from 5.6% this week, and gold is at a record, up 14% in twenty days. Somebody is still pricing the other outcome. The New York Post reports, single-sourced and tier-three, that Bessent may escalate as far as halting 20-year issuance. Unconfirmed; worth watching for a denial.

Crypto and TradFi contagion25 Aughigh

Twelve people priced a war over a weekend

Price discovery is migrating to venues with no clearing house, and traditional markets are increasingly opening to prices set there.

When the US and Israel struck Iran on the last Saturday of February, the oil futures markets were shut. Prices moved anyway. On Trade.xyz, a venue built on top of the crypto exchange Hyperliquid, traders bought and sold a blockchain-based oil derivative through the weekend, so that by the time conventional futures reopened, hours of price discovery had already happened somewhere else (Bloomberg). Trade.xyz is run by about a dozen people. It launched in October. It has since done roughly $500bn of trading volume and accounts for more than 99% of activity on Hyperliquid's third-party market system. It offers perpetual futures on crude, precious metals, stock indexes and companies that have not yet listed, including SpaceX, and is lobbying US regulators to allow pre-IPO perps onshore. The structural point is not that this is unregulated — it is offshore and not officially open to Americans. It is that a perpetual future has no expiry and no central counterparty. Margin sits on-chain, positions are liquidated by code, and the funding rate does the work that an expiry date does elsewhere. There is no clearing house to mutualise a loss and no exchange to halt trading. In quiet weekends that produces useful prices. In a violent one it produces a cascade of automated liquidations against a thin book, and a reference price that regulated venues then have to open into on Monday. Meanwhile the tape is euphoric: bitcoin up 21.9% in twenty days, Coinbase up 28% in five and at its period high, Strategy up 37%.

Private credit and BDCs25 Aughigh

The only price on the empire, bid by the empire

The one publicly traded price on a large private insurance-and-credit group may now be set partly by that group.

Guggenheim Investments told lenders on Tuesday that its affiliates "may purchase term loans from time to time in the open market" — the loans in question being the $1.18bn borrowing of GIH Borrower LLC, due 2031, which fell as low as 73 cents on the dollar this week and was indicated at about 77 on Tuesday (Bloomberg). The disclosure calls the loans "an attractive investment opportunity". Why this matters more than an ordinary buyback: Mark Walter's Guggenheim Partners is a private financial empire, and this loan is close to the only instrument attached to it that trades continuously in front of strangers. When investigators are looking at loans made by his insurers that were channelled to other parts of the group, the loan price is the one outside opinion available on the whole structure. If the marginal buyer is an affiliate of the borrower's manager, that opinion stops being independent. The structural detail is worth reading twice. The disclosure says the loans are expected to be bought by an affiliate and held as an investment — not purchased by GIH Borrower and cancelled. So the debt stays outstanding. No leverage comes out of the system. What changes is that a friendly holder sits on the register, at a discount, with whatever consent rights come with the paper. None of this is improper on its face; open-market repurchases are routine. But the filing does not say how much, or whether any has been bought, and Guggenheim did not respond to Bloomberg. We had been treating the bounce from 73 to 77 as a distressed mark being tested. We no longer are.

Crypto and TradFi contagion24 Aughigh

Strategy builds a second cash pile it is allowed to spend

The treasury-company model converts a volatile asset into fixed dollar obligations, and the fix for that is always more equity.

Strategy has created a new balance-sheet bucket called USD Cash, currently holding $1.59bn, which can be used to buy bitcoin, fund preferred dividends and debt payments, repurchase securities, or anything else corporate. It sits beside the original reserve — now $5.1bn — which may only support preferred dividends and interest unless the board says otherwise (Bloomberg). The mechanism to hold in your head: a digital asset treasury company works when it can issue equity above the value of the coins it holds, buy more coins, and thereby raise coins-per-share. The premium is the engine. When the premium goes, issuance stops adding and starts diluting. Meanwhile the preferred stock Strategy sold to fund earlier purchases pays dividends in dollars, on a schedule, and the asset backing them pays nothing at all. That is the squeeze: a non-yielding asset funding a yielding liability. Hence two reserves. One ring-fences dollars for the obligations. The new one restores discretion. "The new USD Cash pool gives Strategy more time and optionality, but it does not remove those underlying obligations," Nansen's Nicolai Sondergaard told Bloomberg. In the preceding week the company sold about $2bn of common shares and bought back $136.4m of its Stretch preferred, which had been trading below par. The tape disagrees with all of this. Strategy is up 37% in five days and sits at its period high; BitMine is up 36%, SharpLink 33%, Coinbase 28%, with bitcoin at $79,144, up 23% in twenty days. Bloomberg's reporting says the financing flywheel "remains impaired". Equity holders are pricing it as repaired. Only one of those can be right.

Fed, Treasury and policy25 Augmedium

The Fed stopped talking. The Treasury started buying.

When the central bank stops signalling and the finance ministry starts backstopping prices, the long end trades on political will rather than expected policy.

Two things have happened to the American policy reaction function in the space of a few months, and they point in opposite directions. First, Fed chair Kevin Warsh has abolished forward guidance. From the June meeting the statement carried no forward-looking language; he has said publicly that "forward guidance is not the business we should be in", withheld his own dot from the projections, and shrunk the statement. We have fact-checked this and it holds up. He delivers a keynote at Jackson Hole on Friday. The Fed has left rates at 3.5–3.75%, most recently on a 9–3 vote in July. Second, the Treasury has taken up the slack. Buybacks of 10- to 30-year paper at least double, from $2bn to $4bn per operation, running 9 September to 4 November. Two senior Treasury officials told CNBC that the near-$1tn General Account is available to fund them — cash already collected in taxes, so the purchases need not be matched by extra bill issuance. And it is working, at least mechanically: Treasuries have outperformed swaps since the announcement, taking the 30-year swap spread to its tightest since February (Bloomberg). So the option the market can actually price is now fiscal, not monetary. That is a real change in plumbing. Two constraints on it: the TGA is a one-off buffer, not an income stream, and it is also the government's working cash. And the buyer is an elected administration with a midterm in front of it, which the New York Post — single outlet, treat carefully — reports is considering going further by suspending 20-year issuance.

Household credit25 Aughigh

Bad borrowers, good lenders — how that holds

Consumer credit stress becomes a financial event only when the securitisation market that funds it closes, not when borrowers stop paying.

Three household numbers landed since yesterday, all confirmed. Conference Board consumer confidence fell to 89.4 in August from 90.2, a seven-month low. Consumers' one-year inflation expectation rose to 5.8% from 5.6%. And new single-family home sales fell 10.5% in July to a 607,000 annualised pace, 6.3% below a year ago, per Census and HUD. Set those beside the credit data. Subprime auto borrowers 60 days or more past due: 6.13% in July, against 0.49% for prime — more than twelve times the rate. New York Fed figures have 5.5% of auto debt 90 days delinquent, the second-highest reading in the series, and 12.8% of credit card balances 90 days or more past due. Now look at the lenders. Synchrony is a quarter of a percent off its period high. Affirm is three quarters of a percent off. Credit Acceptance, which lends at the deepest end of subprime, is up 5.9% in five days. This is not the market being stupid. A subprime lender does not die of delinquency; it prices for delinquency at the outset, at rates north of 20%, and books the spread. It dies of funding — when the securitisation market that buys its loan pools reprices or shuts, and the lender has to hold paper it planned to sell. That has not happened. As long as the ABS bid is there, high loss rates are the business model rather than a threat to it. Which makes the thing to watch not the delinquency print but the subordinate tranche spread on the next subprime auto deal, and whether a shelf comes to market and fails to place.

The AI capex bubble26 Augmedium

Thirty-five billion of first-loss on custom chips

Chip financing is migrating into private-credit junior tranches whose real collateral is an unpublished company's promise to buy compute.

Broadcom is reported to be in talks for $70bn to $80bn of debt, split roughly $45bn senior and $35bn junior, to finance an AI chip deal that benefits Anthropic and others (Bloomberg, reporting CNBC and Reuters). It extends a June arrangement in which Broadcom, Apollo and Blackstone agreed to finance a $35bn expansion of Anthropic's computing capacity using Broadcom's custom chips and networking. The senior tranche is unremarkable. The junior one is the story. Someone is being asked to take roughly $35bn of first-loss risk, and it is worth being precise about what secures it. General-purpose GPUs at least have a market. If a neocloud fails, its H-series racks are worth something to the next buyer, and lenders can argue about the haircut. Custom accelerators designed for one customer's model architecture do not have that. Their resale value outside the intended tenant is closer to scrap than to book. So the collateral is not really the silicon; it is the offtake contract — Anthropic's promise to pay for the compute — and Anthropic is a private company whose financials are not published. That is a credit underwritten on a counterparty you cannot look up, secured by hardware you cannot resell, financed at a size comparable to a mid-cap bank's entire loan book. The junior tranche is where that judgement is expressed. If it prices and places, someone believes the offtake. If the split shifts toward senior, or the junior tranche shrinks, that is the market telling you what it thinks of the contract. Single-chain sourcing here: this is a sweep of CNBC and Reuters reporting, not a filing. Terms may move.

Hidden leverage and shadow banking25 Augmedium

The repo fix has holes, and the leveraged leg sits in them

The post-2020 structural repair to the Treasury basis trade has stopped expanding, and the exemptions sit exactly where the borrowing happens.

The Office of Financial Research published two posts last week. The headline one — hedge funds' cash Treasury holdings reaching about $2tn, scaled by repo rehypothecation and futures margin — we have already counted. The second is the more useful one and has had almost no attention: central clearing in Treasury repo plateaued in the first quarter of 2026 (OFR). Recall what central clearing was for. After March 2020, when the basis trade unwound and dealers could not absorb the flow, the fix was to push Treasury repo through a clearing house so that positions are novated, margined to a single standard, and cannot run bilaterally counterparty by counterparty. It was the main structural answer to the largest leveraged position in the financial system. The OFR names the gaps. As of May 2026 the SEC will not require affiliate repo transactions to be cleared if the affiliated counterparty clears its other eligible repos. FICC does not offer clearing for open or term repos with no fixed end date, so those fall outside the rule entirely. And — the sharpest detail — hedge funds' reverse repo, where they lend cash, is centrally cleared at significantly higher rates than hedge funds' repo, where they borrow it. The leg that creates the leverage is the less-cleared leg. To be clear about what this is: better measurement of risk that has been standing for years, not new risk arriving this morning. For scale, Form PF analysis puts large hedge funds at roughly $2.4tn long and $1.6tn short Treasuries as of September 2025, with about $830bn in the cash-futures basis trade and $305bn in swap-spread arbitrage.

Fed, Treasury and policy25 Aughigh

A put you can see in the swap spread

A backstop in the long end suppresses the one price that was disciplining the fiscal and AI-borrowing complex, and moves the adjustment into the currency instead.

Since the Treasury said on 19 August that it would at least double its buybacks of 10- to 30-year bonds — from $2bn to at least $4bn per operation, running 9 September to 4 November — Treasuries have outperformed equivalent-maturity swaps, narrowing the 30-year spread between the two to the tightest since February (Bloomberg). That is a more informative number than the yield itself. A swap spread measures Treasuries against the pure interest-rate contract that carries no balance sheet, no collateral and no supply. When Treasuries cheapen against swaps, it means nobody has room to hold the bonds. When they richen, as now, it means someone has arrived who will hold them regardless. Traders who were short the long end against swaps have been squeezed out of the position. The mechanism matters. On Monday, Treasury officials told CNBC the purchases could be funded from the Treasury General Account — the government's checking account at the Fed, which stood at $935bn on 20 August (CNBC). Funded with bills, a buyback is a duration swap. Funded from the TGA, it is cash leaving the Fed and entering the private sector — closer to reserve creation than to debt management, done by the Treasury rather than the central bank. Citadel Securities' Nohshad Shah called it "financial repression at the margin": preventing Treasuries from clearing at lower prices "does not eliminate that pressure. It merely shifts it elsewhere." Elsewhere is visible. The dollar hit a three-month low after the announcement and gold is at a record $4,714, up 15% in twenty days. Rate volatility falling while gold rips is not calm. It is the same worry, relocated.

The AI capex bubble25 Augmedium

Investment grade at a junk-bond price

The gap between the rating and the yield is the market's estimate of how much of the AI buildout's risk the rating agencies are not capturing.

QTS Realty Trust sold $3.9bn of bonds this week to fund a Georgia data centre leased to Microsoft. The bonds are rated investment grade. They yield about 7.23%. BlackRock paid 7.53% in July on blue-chip-rated securities for a Texas data-centre project (Bloomberg). Hold those next to the index. Investment-grade option-adjusted spread is 81bp and has not moved in twenty days. The 10-year Treasury is 4.70%, the 30-year around 5.3%. We do not have the maturity or the quoted spread on the QTS deal, so we cannot compute its concession precisely — but a 7.23% coupon on paper the agencies call investment grade is not an investment-grade clearing level in this rate environment. The rating and the price are describing different securities. The reason is what the bond is actually secured on. This is project debt: a single building, a single tenant, a single lease, usually inside a special purpose vehicle. Meta's Hyperion facility carries $27bn of vehicle debt rated A+, priced at 6.58% and amortising to 2049, with PIMCO reported to have taken about $18bn. Meta keeps 20% of the joint venture and the debt stays off its balance sheet. The rating captures the credit of the tenant over the lease term. The price captures what a buyer thinks the concrete is worth in 2049 if the tenant leaves in 2034. Investors are being paid a junk yield to hold that gap, and the buyers, per the reporting, are increasingly junk-bond funds reaching up rather than IG funds reaching down. This is the migration in its purest form: an obligation of the largest companies on earth, structured so that it is nobody's obligation at all.

Private credit and BDCs25 Augmedium

Two names took BCRED's non-accruals from 0.6% to 2.4%

Rising non-accruals plus rising PIK is how a credit book deteriorates without ever printing a loss, and the listed proxies are not pricing it.

Blackstone's flagship non-traded private credit fund, BCRED, saw loans on non-accrual rise from 0.6% of cost to 2.4% between the first and second quarters of 2026. The jump was driven substantially by two borrowers: Medallia, a customer-analytics company, marked at 60 cents on the dollar against a $1.1bn cost basis, and Affordable Care International, a dental network, marked at 70 (via our sweep of Reuters-syndicated and Cliffwater-cited reporting). Across the sector, the twenty largest listed BDCs had a median 2.8% of cost on non-accrual in Q2, up from 2.0% at end-March and the highest since 2017. The ten largest were at 3.95%, an aggregate $3.3bn against an $83.6bn book. Loans paying in kind rather than cash rose from about 6% of BDC lending in early 2022 to roughly 10% by early 2026 (Morningstar). None of this is in the share prices. Blackstone Secured Lending is 0.8% off its period high and up 6.5% in twenty days. Ares Capital is 1.3% off its high. Prospect Capital, the sector's perennial problem child, is sitting exactly at its period high. The Blackstone and Ares management companies are up 8% and 12% over the month. One reading is that a 3.95% non-accrual rate on senior secured loans yielding double digits is simply survivable, which is Blackstone's position and may well be right. The other is that PIK at 10% means a growing share of reported income is an accounting entry rather than cash through the door, and non-accrual is the point at which the fiction is abandoned rather than the point at which the loan goes bad. The equity is voting for the first reading.

Household credit25 Augmedium

6.13%, and the lenders' shares are at their highs

Record subprime delinquency with lender equities at their highs means the losses are landing somewhere other than the originator — and that somewhere is not disclosed weekly.

Fitch's July data put subprime auto loans 60 or more days past due at 6.13%, up from 5.80% at midyear and described as the highest July reading in a series running back to 1994. Prime borrowers were at 0.49% — more than twelve times lower (via our sweep). A separate figure of 6.9% has circulated attributed to Fitch and Equifax; the underlying date on that one is not pinned down and we would not use it. New York Fed data for the second quarter shows 3.00% of auto debt transitioning into serious delinquency, the highest quarterly rate since 2010, and 12.8% of credit-card balances 90 or more days past due. Now the shares. Credit Acceptance, which lends at the deepest end of the subprime auto market, is up 5.9% in five days and 8.4% in twenty. Affirm is 0.75% off its period high. Synchrony is 0.25% off its high. Capital One and Ally are the only names in the complex that are down over five days, and only by about 2%. There is a coherent story in which both are true. A subprime auto lender's economics are a spread: loss rates against a coupon that can run above 20%. Rising delinquency is a cost input, and if the lender has already priced it, tightened underwriting, and can fund itself, higher losses can arrive alongside higher profits. The delinquency rate is a statement about the borrower, not automatically about the lender. What breaks that is the funding leg. Most of this credit is securitised and sold; the equity holders of those deals are the ones absorbing the loss. Which is the point — the household stress is real, it is at a multi-decade high, and it is not sitting anywhere you can see it on a bank balance sheet.

Crypto and TradFi contagion25 Auglow

The buyback showed up in bitcoin first

The digital-asset treasury companies are the most direct read on financial conditions available, and they just went vertical on a Treasury debt-management announcement.

Bitcoin is at $78,752, up 22.5% in twenty days. Coinbase is up 28% in five days and sitting at its period high. Strategy is up 37%, Bitmine 36%, SharpLink 33% — all three digital-asset treasury companies at their period highs, all in the same week. The proximate cause, per crypto-market reporting, was a short squeeze of roughly $2.7bn in liquidated positions, attributed to the loosening in liquidity conditions after the Treasury doubled its long-bond buybacks (Crypto Briefing). This is tier-two attribution and the causal chain is the sort of thing that gets asserted more often than demonstrated. Treat it as a hypothesis. But it is a coherent one, and it is the reason we track this beat. Digital-asset treasury companies are pure leveraged bets on liquidity conditions with a coin attached. Their business model requires their shares to trade above the value of the coins they hold, so that issuing stock to buy more coin is accretive. When that premium compresses, the flywheel stops — which is precisely why Strategy set up a $1.59bn cash pool this week that can be spent on bitcoin, dividends or debt service, alongside a $5.1bn reserve ringfenced for preferred dividends and interest (Bloomberg). A company confident in its financing does not build two separate cash reserves. So the sequence to hold in mind: Treasury intervenes in the long end, financial conditions ease, the most liquidity-sensitive instrument in the world rallies 22% and the leveraged wrappers on it rally 37%. If that channel is real, it works in reverse too.

Hidden leverage and shadow banking25 Aughigh

Project finance joins the risk-transfer queue

Risk transfer is only risk reduction if the buyer is unlevered and unconnected, and on both points the disclosure is thin.

AIB, the Irish bank, is working with Santander and the broker Howden on a significant risk transfer covering about €2.5bn of project finance loans, due to complete later this year (Bloomberg). We flagged the record SRT year on Tuesday — $45bn expected in 2026 against $41bn in 2025, per Crescent Capital — and we are not counting it twice. The new information is the collateral. An SRT works like this: a bank keeps a portfolio of loans on its books but buys protection on the first-loss slice from an investor, who takes a double-digit coupon in exchange. The bank's regulatory capital requirement falls; AIB expects 25 to 30 basis points of core tier 1 benefit. The loans do not move. The risk, notionally, does. Until recently these deals were overwhelmingly corporate loans and mortgages — long, well-behaved data series that investors can price. Project finance is different. It is single-asset, single-cashflow lending, with no diversification within a name and a recovery value that depends entirely on whether the thing gets built and used. AIB's own book includes infrastructure and energy. This is the same category of credit that is currently funding data centres at 7.23%. Two things to watch that this deal does not tell us. First, who the protection sellers are: if they are the same private credit funds the banks lend to elsewhere, the risk has moved sideways, not out. Second, whether any of the financing for the protection comes from banks. The filings do not say.

What would change our mind

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

A data-centre SPV failing to place its junior tranche — the Broadcom structure's reported ~$35bn junior piece is the obvious test — or a deal pulled outright. That would move ignition several points.

Would move the number

MSCI adopting the non-operating-company exclusion, which would create a dated, forced-seller event in digital-asset treasury companies rather than a hypothetical one.

Would move the number

High-yield spreads breaking above roughly 350bp while BDC discounts to NAV widen at the same time — credit and the private-credit equity complex agreeing for once.

Would move the number

Warsh and Bessent visibly at odds on Friday, with the 30-year giving back the swap-spread tightening since the buyback announcement. That would tell us the 'Treasury put' has a price the market can find.

Would move the number

Hit a word you don’t know? Underlined terms above open an explanation without leaving the page. Browse all of Crashopedia →