Debt-funded AI is selling off. Cash-funded AI is not.

Archived reading, published Mon, 24 Aug 2026 10:21:00 UTC (2 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility89

how much tinder is stacked up — moves slowly

Ignition39

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 21 sources and rewrites this page.

Status: Held at 64 for a sixth run. The only prices that have refreshed since the last reading are VIX (15.13 to 15.89), gold (+0.47%) and bitcoin (-0.11%); equities, credit spreads and yields are still Friday's marks, with high yield at 275bp and investment grade at 82bp. Everything structural in this window — the Treasury twist, the OCC charter count, Eagle Point's datacentre loan — is either a sharper description of risk already standing or too small to move a system-level number.

Reporting from 22 Aug to 24 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 dysfunction24 Aughigh

A twist of $4bn against a debt of $40tn

The Treasury can change who holds the duration, but not how much borrowing there is, and the market spent three sessions demonstrating the difference.

Scott Bessent described last week's intervention as "what I would call a Treasury twist": buy back long-dated bonds, and fund the purchases by issuing more short-dated bills. The Treasury will at least double liquidity-support buybacks in the 10-to-20 and 20-to-30 year buckets from $2bn to at least $4bn per operation, running 9 September to 4 November. Two things about that. First, the size. An eight-week programme of $4bn operations is tens of billions of dollars. One gauge of the federal debt passed $40tn this month. The 30-year auction that preceded all this was $25bn on its own, and cleared at 5.216%, the highest for that tenor since 2001. As a demand offset, this is not large; as a signal, it was large enough to move gold about 5% in a week. Second, the mechanism. Buying long bonds with bill proceeds shortens the average maturity of the federal debt. It reduces the amount of duration the market has to hold, which is the point, and increases the frequency with which the government has to refinance, which is the cost. The buyer of the extra bills is money funds and — via repo — the leveraged Treasury complex the OFR sized at roughly $2tn of hedge-fund cash Treasuries funded overnight. Bessent's own preferred benchmark closed the week at 4.73%, near the highest since he took office. "Every route to lasting relief for the long end runs through something the administration doesn't want," Satori Insights' Matt King told Bloomberg: a smaller deficit, a falling stock market, or less AI investment.

The AI capex bubble23 Augmedium

The AI trade has split along the financing line

Equity in the debt-funded end of AI infrastructure is the first place a rise in AI financing costs shows up, and it is moving while credit spreads and the index are not.

Look at the AI complex over five sessions and the pattern is not risk-off. It is discrimination. Nebius is down 21.1% in five days, having announced on 20 August that it had upsized a convertible bond to $5bn — one of the largest on record — to fund data centres and its AI platform. CoreWeave is down 16.6%. Talen Energy, which sells power to data centres, is down 13.4%; Applied Digital, which builds them on project debt, is down 12.8%; Core Scientific is down 11.6% and 21.8% over twenty days. Vistra is down 8.1%. Over the same twenty days Microsoft is up 26.8%, Oracle 27.4%, Amazon 11.4%. The S&P is 1.6% off its record. The distinction is not what these companies do. It is how they pay for it. A hyperscaler funds capex substantially out of operating cash flow and issues bonds because it is cheap to; Alphabet placed $25bn on 10 August into a capex plan reported at $205bn. A neocloud or a merchant power developer funds capex with project debt, term loans and converts, and its equity is the residual claim on a spread between contracted revenue and the cost of that debt. When the cost of the debt moves, the residual moves several times as much. Demand for the paper itself has not broken: QTS sold $3.9bn of bonds for a Microsoft-linked Georgia data centre with peak orders of about $23bn, roughly six times covered, though at what Bloomberg called junk-like yields on a high-grade deal. So the market is still lending. It is repricing who keeps the difference.

Private credit and BDCs23 Augmedium

A CLO manager is now a data-centre landlord's lender

AI infrastructure risk is being funded by lenders whose losses, if any come, will not appear in a bank's quarterly disclosure.

Eagle Point Credit Management is reported to be providing roughly $1.3bn of private credit for an AI data centre at Hubbard, Texas, tied to Anthropic. The loan is one piece of a $16bn project-finance package for Nexus Data Centers. Eagle Point is known as a manager of CLO equity and structured credit. It is now, in effect, writing single-asset real-estate project debt whose repayment depends on a lease from a company founded in 2021 that does not publish audited accounts. This is the migration Crash Lab keeps pointing at, in one transaction. A bank writing this loan would hold capital against it, mark it against observable credit, and disclose the concentration. A private credit fund holds it at a valuation it determines with an adviser, in a vehicle whose investors cannot easily leave, and against a project whose value if the tenant walks is a building full of specialised equipment in central Texas. The terms are not public. What the reporting does not tell us is the seniority, the tenor, whether the lease is guaranteed by anyone with a balance sheet, or what happens to the collateral if the tenant's compute needs change. Those are the four things that determine whether this is a bond-like asset or a venture bet with a coupon. It fits the wider pattern: Nvidia's platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, described as mobilising more than $500bn of third-party capital with up to $125bn of Nvidia backstops. The capital is arriving. Where it is arriving is balance sheets that do not have to mark it.

Crypto and TradFi contagion23 Augmedium

The treasury companies are outrunning the coin again

DAT equity trading up 28% while the underlying coin rises 12% is leverage rebuilding in the crypto-equity channel, and the premium is the thing that has to hold.

Bitcoin is up 12.1% in five days, to $77,672. Strategy is up 28.1%. BitMine is up 26.3%, SharpLink 26.4%, Coinbase 25.7%. All four closed Friday at their period highs; bitcoin did not. That gap is the whole business model of a digital asset treasury company. A DAT holds coins and issues equity and converts to buy more. It only works while the shares trade above the value of the coins per share, because that is what makes each new issue accretive to existing holders. When the shares trade at a premium, the flywheel spins: issue, buy, report higher coin-per-share, issue again. When the premium closes, the machine stops, and the company becomes a leveraged fund with a stock listing and financing to service. We do not have current NAV multiples for these four; the price data tells us the direction of the premium, not its level. The spark, per Bloomberg, came partly from the bond market: Bessent's buyback plan pulled yields down and weakened the dollar, reviving the debasement trade, with Ray Dalio adding his endorsement. The same trade lifted gold 7.7% in five days. Bloomberg also notes the initial surge came from forced short covering, and that bitcoin remains roughly 43% below its October record. So: a rally in the most leveraged expression of an asset, driven by a fiscal argument and a squeeze, in instruments that need the premium to persist. Nothing here is stress. It is the sort of thing that becomes stress later.

Fed, Treasury and policy23 Aughigh

Twenty-two bank charters in nineteen months

A cohort of deposit-substitute issuers is being chartered quickly, without the backstop that makes deposits safe in a run.

The OCC approved 22 national bank charter applications in the first 19 months of Trump's second term — more than it approved in the previous five years combined. About 40 companies have applied since the start of 2025, roughly equal to the total for the preceding 13 years. Decisions that used to take months or years are now targeted at 120 days. "The OCC is open for business again," the agency's head Jonathan Gould told Congress in June. Many of the new charters are going to fintech and digital-asset firms. World Liberty Financial, tied to the president's family, has conditional approval for a national trust company that would let it issue its USD1 stablecoin directly and custody the dollar assets backing it, rather than using BitGo. The mechanism worth understanding: a trust charter is not a deposit-taking bank licence, but it does bring stablecoin issuance and reserve custody inside the federal perimeter, with the supervisory legitimacy that carries. What it does not bring is deposit insurance or access to the discount window — Custodia's fight for a Fed master account is now at the Supreme Court. So you get instruments that look like deposits to their holders, are supervised like a bank in one respect, and have no lender of last resort behind them. Bank executives have raised the obvious concern: stablecoins compete with deposits, particularly if they start paying something that resembles interest. Bloomberg notes the window may not stay open — regulators taking these applications would face different scrutiny if the November midterms go against the administration.

Household credit23 Augmedium

Two credit card delinquency rates, both from the NY Fed

The gap between the flow into delinquency and the stock of delinquent balances is where next year's charge-offs are already sitting.

This week's household coverage carries two numbers for the same thing. One set of reports says 12.8% of credit card balances were 90 or more days delinquent in Q2 2026, up from 7.6% in Q3 2022. Another says card debt entering serious delinquency was 6.97%, up from 6.93% a year earlier. Auto has the same problem: 5.49% of balances 90+ days delinquent in one telling, 3.00% in serious delinquency in another, the latter described as the highest since 2010. These are not contradictory. They are different statistics being quoted as if they were the same one, and it matters which you are reading. Our reading — inference, not something the reports state — is that the higher figures are stocks: the share of outstanding balances currently sitting 90 days or more past due. The lower figures are flows: the share of balances that transitioned into serious delinquency over the preceding year. A stock accumulates while a flow is elevated and only falls when lenders charge off. A flow tells you what is happening now. So the honest summary is that the flow into serious delinquency is roughly flat year on year, while the stock of already-bad paper keeps building because it has not yet been written off. That is a slower, less dramatic story than "12.8% delinquent" and a worse one for lenders' 2027 provisions. The lenders are unbothered. Synchrony rose 3.6% Friday, Capital One 2.6%, Credit Acceptance 2.4% and within 6% of its high.

Signs of the times

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

More new banks than in five years

The OCC approved 22 national bank charters in 19 months, more than in the previous five years combined, with about 40 applications filed since January 2025 — roughly the total for the preceding 13 years. Among the approvals in progress: a trust company owned by the president's family, which would issue its own stablecoin and hold the dollars backing it.

deregulationstablecoins

23 AugBloomberg

Record convertible, then minus 21%

Nebius announced on 20 August that it had upsized its convertible bond to $5bn, one of the largest on record, to fund data centres. The stock is down 21.1% over the following five sessions — the buyers of the equity and the buyers of the bond appear to have read the same deal very differently.

ai-debtconvertibles

23 AugBloomberg (via web sweep)

$23bn of orders for $3.9bn of bonds

QTS Realty sold $3.9bn of bonds for a Microsoft-linked data centre in Georgia, drawing peak orders of about $23bn — roughly six times covered — at what Bloomberg described as junk-like yields on an ostensibly high-grade deal. Investors want the paper; they are no longer pretending it is safe.

data-centrescredit

23 AugBloomberg (via web sweep)

Subprime auto worse than 2008

Sixty-day delinquencies on US subprime auto loans reached 6.9% on Fitch and Equifax data, against a 2008 peak of 5.0%. Auto lending originations in Q2 were a nominal record $211bn.

consumersubprime

23 AugGlitchwire (via web sweep)

The rumour mill

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

Partly true

“The US dollar has lost 95% of its purchasing power over the past 50 years.”

The 95% figure is real but measured from around 1930, not 1976. Over the last 50 years the loss is closer to 83% — $1 in 1976 buys what about $5.88 buys now. The direction is right; the number has been moved to a different starting line.

Claimed by Wealthion

Partly true

“Central banks bought a record 289 tonnes of gold in Q2, led by Poland and China, but net demand was the weakest in four years.”

The 289t figure checks out on World Gold Council data — up 62% year on year, Poland +51t, China +33t, Russia the largest seller at -22t. But Q2 was a record second quarter, not a four-year low; it is first-half 2026 demand of about 345t that is the weakest since 2022.

Claimed by Kitco NEWS

Confirmed

“US margin debt is at an all-time high.”

FINRA recorded $1.502tn in June 2026, a nominal record and up about 77% from April 2025. Worth pairing with the fact we flagged on Friday: July fell $85bn, the largest monthly drop on record. The peak is real and it is behind us by one month.

Claimed by Thoughtful Money (Adam Taggart)

Earlier developments

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

Private credit and BDCs23 Augmedium

The non-accrual number with a bigger number behind it

The headline non-accrual rate in private credit is a reported figure, and the measure that counts all debt of an already-impaired borrower is half again as large.

Across the ten largest listed BDCs, non-accruals were 3.95% of amortised cost in Q2 2026, up 20 basis points on the quarter — $3.3bn against an $83.6bn debt portfolio. On an adjusted basis, they were 5.95%. The wider FT/Solve cut of the twenty largest BDCs shows the median rising from 2.0% to 2.8% of cost, the highest since around 2017 (Cliffwater/BDC data summary). The gap between 3.95% and 5.95% is the whole story. A loan goes on non-accrual when the lender stops booking interest it no longer expects to collect. That is a decision the lender makes, loan by loan. The adjusted measure asks a different question: if any tranche of a borrower's debt is impaired, treat all of that borrower's debt as suspect. Applied sector-wide, one analysis puts adjusted non-accruals at $17.3bn against $10.0bn reported — a $7.3bn difference that is not concealment, just a different question being answered. Alongside it: PIK is 11% of BDC portfolios, $15.7bn at cost, and $772m of interest income at the largest ten vehicles is attached to loans already impaired. Cliffwater's Direct Lending Index, built from manager-reported marks, says the opposite — non-accruals and PIK stable and below long-run averages, with software markdowns falling from about 3% in Q1 to under 1% in Q2. Both cannot be describing the same portfolios. The filings-based series is the harder one to flatter. The market has picked the soft version. Ares Capital and Blackstone Secured Lending closed Friday 1.3% and 1.0% off their recent highs.

Hidden leverage and shadow banking23 Augmedium

The clearing mandate has doors in it

The reform meant to make the basis trade survivable does not reach the transactions most likely to be pulled first.

The Office of Financial Research put hedge funds' cash Treasury holdings at about $2tn and named the cash-futures basis trade as the main driver, sized by various regulators at somewhere between $350bn and $1.5tn. The spread earned is tiny, so the position is scaled with repo — collateral rehypothecated down a chain of borrow-and-buy transactions, financed overnight because term is more expensive (OFR). The post-2020 fix for this was central clearing. Push Treasury repo into a CCP and the funding stops depending on one dealer's willingness to keep a client's book open on a bad morning; margin is set by a rulebook rather than a credit officer. A companion OFR note describes the holes in it. Under the SEC clearing rule that took effect in May 2026, affiliate repo transactions are not required to be centrally cleared if the affiliated counterparty clears its other eligible repos. And FICC does not offer clearing for open-term repo at all, so open-term trades sit outside the mandate by construction. Separately, hedge funds' cash lending is centrally cleared at materially higher rates than their cash borrowing — the leg that matters for a forced unwind is the less-cleared one. The OFR does not say how much of the $2tn sits in the exempt buckets, and we have not seen a number anywhere that does. The July FOMC minutes, released on 19 August, called leverage vulnerabilities in the financial sector "notable" and said hedge fund leverage remained near all-time highs across all strategies, concentrated in the largest funds. The Europeans have the same shape of problem: non-banks are now over 40% of non-centrally-cleared euro cash-borrowing repo, up from under 30% at end-2020.

The AI capex bubble23 Auglow

The chip vendors are credit-enhancing their own sales

The two largest AI chip suppliers are now underwriting their customers' ability to pay, which moves credit risk into footnotes rather than out of the system.

Broadcom was reported on 20 August to be in talks to raise more than $60bn of debt for an AI chip financing deal benefiting Anthropic and others. The described structure: roughly $30bn of junior debt and a senior secured tranche of $60bn–$70bn, a total that one report puts as high as $100bn, with Broadcom guaranteeing part of it. Ten days earlier, Nvidia announced financing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500bn of third-party capital, saying it could provide backstops of up to $125bn (sweep). The mechanism is old and the scale is not. A vendor guarantee converts a sale that a customer cannot finance on its own credit into a sale that it can. The revenue lands at delivery. The guarantee is a contingent liability, disclosed in a footnote, and it does not become debt on the vendor's balance sheet unless it is called. In between, the vendor books the margin and carries the credit risk of its own customer without the customer appearing as an exposure anywhere a bank supervisor is looking. What has to stay true: the AI labs generate enough cash to service leases and loans on chips that depreciate fast. If they do, the guarantees expire unused and this looks like clever working-capital engineering. We would treat the Broadcom terms carefully. It is reported talks, no filing, and the two reported tranche numbers do not obviously add up to the reported total. What is confirmed is the direction: more than $410bn has been borrowed for data centres this year.

Fed, Treasury and policy23 Aughigh

A one-in-three chance of a rate rise

A central bank that publishes less forces the bond market to price the reaction function itself, which shows up as term premium at the long end.

Futures are pricing roughly a one-in-three chance that the Fed raises rates next month, according to the FT. That is the number to hold in mind this week, because it sits underneath everything else: a 30-year yield at a 19-year high, $40tn of federal debt, and a Treasury buying back its own long bonds. The July minutes — the committee held at 3.5–3.75% on a 9–3 vote — show officials worried that high inflation becomes self-fulfilling once expectations get into wage and price setting. July headline PCE is expected at 0.1% month-on-month and 3.6% year-on-year, with core steady at 3.3% (FT). Two complications, both about measurement rather than the economy. Chair Kevin Warsh has called core PCE "sort of a rough swag as to what's going on". And the measure is due an overhaul that could lower recent readings. So the number the Fed targets is being questioned by the person who targets it, and separately revised, in the same quarter that the long end is repricing on inflation risk. Warsh speaks at Jackson Hole on Friday, his first time as chair, and is expected to explain why he has stripped back the Fed's guidance. In an FT–Chicago Booth poll, nearly 60% of academic economists said they now expect the return to 2% to take longer than they did when he was confirmed in May. Less guidance does not remove uncertainty; it relocates it into term premium. That is one plausible reason the long end has been doing so much of the work this month.

The dollar, gold and reserve status23 Aughigh

Gold's rally is still a recovery

The debasement trade is real, but it is currently recovering ground lost earlier this year rather than pricing something new.

Gold closed Friday at a three-month high, up around 5% on the week and about 14% in August, after the Treasury's surprise ramp-up of long-bond buybacks. The dollar index hit a three-month low the same day. Our series has gold at $4,697, up 7.6% in five days and 16.4% in twenty; the DXY at 98.83, down 2.5% in twenty. One number keeps the story honest. Gold is still roughly $1,000 an ounce below its peak in late January (Business Times). It ground lower through the first half of this year as Warsh asserted the Fed's independence and the Iran war raised the prospect of hikes. What has happened this month is a violent recovery of part of that loss, not a breakout into new territory. Anyone reading the August move as a verdict on the dollar should note the market delivered the opposite verdict in February. The mechanism people are pricing is straightforward enough. If the Treasury is willing to buy its own long debt to hold yields down, the adjustment has to appear somewhere else, and the currency is the obvious candidate. UBS's Bhanu Baweja called the buyback "a very important signal for gold" and said "the dollar will pay the price". Nomura's Charlie McElligott described gold up, dollar down, bitcoin up as a "pressure-release valve". The tell to watch is the one Saxo's Charu Chanana flagged: gold held up even as long-end yields stayed elevated. Normally higher real yields hurt gold. When they stop doing so, the metal has stopped trading as a rate instrument.

Household credit23 Augmedium

Record auto lending into the worst delinquency since 2010

Lenders are writing record volumes of auto credit while the existing book performs at levels last seen in the aftermath of the financial crisis.

The New York Fed's Q2 2026 household file shows auto loan originations of $211bn — a nominal record — in the same quarter that the share of auto debt newly entering serious delinquency reached 3.00%, the highest since 2010, and 5.49% of outstanding auto balances sat 90 or more days delinquent, close to the highest in a series that runs back to 2003 (sweep). Those two facts are about different things. Delinquency is the performance of loans written in 2023 and 2024. Originations are a decision lenders made last quarter, with the delinquency data in front of them. Record volume into a deteriorating back book means one of three things: the new vintages are being written to much better credit, terms are being stretched to keep monthly payments affordable, or car prices have risen enough that the same number of loans is a bigger dollar figure. Fitch's July index showing prime 60-plus delinquency at 0.49% against subprime at 6.13% is consistent with the first. The public data does not let us separate them. Across all household debt, 4.7% of $18.8tn was in some stage of delinquency. The lenders are not behaving as though anything is wrong. Credit Acceptance, which lends at the deepest end of subprime, rose 2.4% on Friday and is up 4.6% in twenty days, 5.9% off its recent high. Capital One is up 7.8% over twenty days. Nothing in these prices says the market expects the 2010-level delinquency rate to become a 2010-level loss rate.

Private credit and BDCs22 Aughigh

The insurer as a piggy bank

The insurance balance sheet is the single largest pool funding private credit, and this is the first evidence that regulators cannot reliably see what is inside it.

Federal prosecutors are investigating Mark Walter, and the probe has revealed that two insurance companies he controlled provided more than $20bn in previously undisclosed financing for dealmaking by other parts of his business empire, the FT reports. Walter has agreed to sell the Los Angeles Lakers and put his Chelsea stake up for sale to clean up the insurers' balance sheets. Guggenheim Partners, which he built on this model, runs $367bn. The plumbing matters more than the man. An annuity is a long, sticky liability: the policyholder cannot demand the money back tomorrow. That is precisely why private capital firms have spent a decade buying insurers — long money can be put into illiquid loans that pay more than corporate bonds, and the difference is the profit. KKR, Apollo and Brookfield have all bought insurers and used them to lend to companies in their own orbit. That is not, in itself, improper; it is disclosed and it is the industry's stated business model. The hazard is the affiliated transaction. When the insurer's asset is the sponsor's funding, the party choosing the loan and the party being financed are the same. The reserves backing retirees' policies become working capital. "What makes Mark Walter unique is that he failed to disclose a tremendous amount of affiliated transactions," a University of Texas law professor told the FT. "But other private capital firms are engaged in essentially the same business model." No charges have been reported. What this changes for us is not the size of the risk — the insurance channel was already the largest untested thing in private credit — but the discovery that $20bn of it can sit unseen until a prosecutor looks.

Bond market dysfunction23 Aughigh

The twist is funded with bills

Shortening the maturity of $40tn of debt to suppress one yield transfers the fiscal risk from the bond market to the next inflation surprise.

Scott Bessent's plan to hold down long-term yields has a funding leg that got less attention than the buying leg. The Treasury will repurchase 10- to 30-year debt at at least $4bn per operation, up from $2bn, from 9 September to 4 November — and it will pay for that by selling more short-dated securities. "What I would call a Treasury twist," Bessent said, per Bloomberg. Here is what that does. Buying back long bonds takes duration out of private hands, which should push long yields down. Selling bills puts the same borrowing back at the front end. The government's total debt is unchanged; its maturity shortens. A 30-year bond fixes the taxpayer's cost for 30 years. A bill re-prices every few months. At today's front end — SOFR at 3.63%, the two-year at 4.19% — that is cheaper. If inflation forces the Fed to raise, the cost of the shortened stock reprices almost immediately, across the whole amount rolled. So the trade converts a term-premium problem into a rollover problem. It is the same swap a company makes when it refinances a fixed-rate bond with a revolver. It also has not worked. The 10-year, Bessent's preferred benchmark, closed the week at 4.73%, near the highest of his tenure. The 30-year touched 5.34%, a 19-year high, and a $25bn 30-year auction cleared at 5.216%, the worst level for that tenor since 2001. Four billion dollars an operation against a $40tn debt stock is a signal, not a supply shift. "Every route to lasting relief for the long end runs through something the administration doesn't want," Satori Insights' Matt King told Bloomberg: a smaller deficit, a lower stock market, or less AI investment.

The AI capex bubble23 Augmedium

Private credit is funding the datacentres now

The riskiest tranche of AI financing is migrating into vehicles that mark their own loans, which is exactly where losses arrive late.

Eagle Point Credit Management is providing roughly $1.3bn of private-credit debt for an AI datacentre in Hubbard, Texas tied to Anthropic — part of a $16bn project-finance package for Nexus Data Centers, according to reporting compiled in this window. One building, sixteen billion dollars. The structure is worth understanding because it is not how the hyperscalers borrow. Alphabet sold $25bn of bonds in August against a company with enormous cash flow; the bond prints a price every day and anyone can see what the market thinks. Project finance is different. The lender's recourse runs to the asset and the lease, not to a corporate balance sheet. What makes the loan good is a tenant who keeps paying rent for fifteen or twenty years. Anthropic does not generate the cash flow to underwrite a twenty-year lease from operations. It is funded by capital raises. So the credit, stripped down, is a bet that the tenant will continue to be able to raise money — which is a bet on the equity market, dressed as a secured loan against concrete and transformers. Bloomberg counts more than $410bn borrowed for datacentres and AI investment this year, and a separate tally puts AI-related debt at $489bn by mid-2026, with the five hyperscalers accounting for nearly 40%. Our arithmetic on those two figures: roughly three-fifths of the AI debt stack is being raised by borrowers who are not Alphabet, Amazon, Meta, Microsoft or Oracle. The part sitting in private credit funds does not print a price. It is marked quarterly, by the manager, against comparable transactions — and Palmer Square said last week that valuation transparency has diminished because there are fewer of those.

Household credit23 Augmedium

Two different numbers for the same delinquency

The gap between record household delinquency and lenders trading near highs is the clearest live disagreement between reporting and prices we have.

The subprime auto figure everyone is quoting this week is 6.9% — the share of subprime auto loans 60 days past due, from Fitch and Equifax, above the 2008 peak of 5.0%. The same sweep also carries a Fitch index reading of 6.13% for July, described as the highest July on record. We cannot reconcile the two, and we are not going to pretend otherwise. Both are records; they are not the same number. The credit card data has the same problem in a more instructive form. One set of coverage of the New York Fed's Q2 release puts 90-plus-day delinquency at 12.8-12.9% of card balances. Another puts it at 6.97%. Both are citing the same publication. The gap is almost certainly stock versus flow: the share of outstanding balances currently 90 days late is a level, while the transition rate into serious delinquency measures how much fresh debt fell over the edge this quarter. Auto shows the same split — 3.00% flowing into serious delinquency, the worst since 2010, against 5.49% of balances already there. If you are reading these numbers, check which one you have. A level near 13% and a flow near 7% describe a book that has been deteriorating for years, not one that broke last quarter. The lenders disagree with all of it. Credit Acceptance, which lends to the deepest subprime auto borrowers, is up 2.4% on the week and sits 5.9% off its high. Synchrony rose 3.6% on Friday. Capital One is 3.8% off its high. Subprime is supposed to be priced for high losses — but the equity market is currently paid to believe these delinquencies convert into charge-offs already in the model.

Crypto and TradFi contagion22 Augmedium

The treasury companies are bid again

Leveraged crypto equity wrappers regaining their premium means the most reflexive financing structure in the market has restarted, and it restarts fastest right before it stops.

Bitcoin is at $77,115, up 11.3% in five days and 20.4% in twenty. The listed vehicles built on top of it have gone up considerably more: Strategy +28.1% on the week, BitMine +26.3%, SharpLink +26.4%, Coinbase +25.7%. All four closed Friday at their period highs. The mechanism inside a digital-asset treasury company is a premium loop. The firm holds coins; its shares trade above the market value of those coins; that premium lets it issue new shares above net asset value and buy more coins, which raises coins-per-share and justifies the premium. Running the loop is the business. When the coin falls, the premium compresses, issuance stops, and the company becomes a closed-end fund with a management team attached. Strategy beating bitcoin by eight percentage points in a week is what premium re-expansion looks like. We do not have NAV figures in front of us, so that is inference from relative price moves rather than a measured premium — but the direction is unambiguous, and it means the equity-issuance channel for these firms has reopened. What is pushing it is not crypto-native. It is the debasement trade. The Treasury's buyback announcement knocked long yields down for a day; the dollar fell, gold rose to a three-month high — up about 14% in August — and bitcoin rallied with them. Nomura's Charlie McElligott called the gold-up, dollar-down, bitcoin-up combination a "pressure-release valve" as Washington tries to suppress long rates. Ray Dalio told investors on Friday to cut bond exposure and hold gold and some bitcoin. None of this is stress. It is the opposite: the risk-seeking end of the market getting refinanced.

Fed, Treasury and policy23 Aughigh

Twenty-two new banks

Deposit-like liabilities are being chartered at a pace not seen since before 2008, and none of the new issuers has been through a run.

The Office of the Comptroller of the Currency has approved 22 national bank charters in the first 19 months of Trump's second term — more than it granted in the previous five years combined, Bloomberg reports from OCC data. About 40 companies have applied since the start of 2025, roughly equal to the total for the preceding 13 years. Decisions that used to take years are now targeted at 120 days. "The OCC is open for business again," the agency's head Jonathan Gould told Congress in June. Many of the new charters are going to fintech and digital-asset firms. Among them is World Liberty Trust Company, tied to the president's family, which has conditional approval to become a national trust bank — a charter that would let it issue its USD1 stablecoin directly and custody the dollars backing it, rather than using BitGo for both. This cuts across our running thesis in two directions at once. A trust charter brings a stablecoin issuer inside the examined perimeter, which is where we generally argue risk should be. But a trust bank is not a full bank: no deposit insurance, no lender of last resort in the ordinary sense. What it holds is a demand liability — a token redeemable at par, on request, backed by a portfolio of short instruments. Traditional bankers quoted by Bloomberg worry that stablecoins are becoming a substitute for deposits, particularly if they start paying something that looks like interest. The thing to watch is not the charter. It is the first redemption day when the portfolio backing a token cannot be liquidated at par as fast as holders want out. Custodia's fight for a Fed master account, now at the Supreme Court, is the same question from the other end.

Private credit and BDCs23 Augmedium

Two names and a fourfold jump

Non-accruals cut reported income immediately while PIK hides the same stress in accrued paper — and PIK is now a ninth of BDC portfolios.

A Blackstone private credit vehicle has reported non-accruals at 2.4% of the portfolio at cost, up from 0.6%, with the increase driven largely by two named borrowers: Medallia and Affordable Care. Payment-in-kind income in the same fund fell slightly, to 7.0% from 7.8% (sweep of last week's filings and coverage). The reporting does not identify which Blackstone vehicle, which matters, because the listed one — Blackstone Secured Lending — closed Friday 1.0% off its recent high. Two mechanisms are worth separating here, because they pull in opposite directions on a BDC's income statement. When a loan goes on non-accrual, the lender stops booking interest it is not receiving. Net investment income falls immediately and visibly. When a borrower is switched to PIK instead, the lender keeps booking interest — it just receives more paper rather than cash. Income looks fine; the cash to pay the dividend has to come from somewhere else. Across the sector, PIK now accounts for 11% of BDC portfolios, or $15.7bn at cost, and at least 24 BDCs filing in the week to 14 August carried $2.26bn of non-accruals (BDC filing data). The median non-accrual rate at the twenty largest listed BDCs went from 2.0% to 2.8% of cost in one quarter. A fourfold jump concentrated in two borrowers is idiosyncratic. A sector-wide move from 2.0% to 2.8% while PIK sits at a ninth of assets is not. Ares Capital trades 1.3% off its high, Blackstone Secured Lending 1.0%. The prices say this is idiosyncratic. The filings say it is broad.

Hidden leverage and shadow banking23 Aughigh

Forty per cent coupons, sold after the crash

A 40% coupon is not yield, it is the price of a put the buyer has written — and retail is now the marginal seller of AI-chip downside protection on two continents.

South Korean retail investors are buying equity-linked securities offering annualised coupons of 40% to 50%, with sales in July reaching a more-than-three-year high, led by notes tied to Samsung Electronics and SK Hynix. This follows the Kospi's 22% plunge last month. The Financial Supervisory Service will tighten oversight from next month, requiring brokerages to warn holders when a product approaches its knock-in level (Bloomberg). An ELS is not a bond. The buyer is, in substance, selling a deep out-of-the-money put on the underlying and receiving the premium as a "coupon". You are paid as long as the stock stays inside a preset range. If it breaches the knock-in barrier, you own the downside, in full. The coupon is 40% because implied volatility on Korean chip stocks is enormous after a 22% index fall — and because the barrier, set as a percentage of a price that has already dropped, is closer than it looks. Korean households learned this in 2023 on China-linked notes, which produced heavy losses. The same instinct is visible in the US. Investors put almost $7bn into the Direxion Daily Semiconductor Bull 3X fund during July and the first two weeks of August, a fund that fell 70% from its June peak to its low; a 2x SK Hynix fund took in over $1bn while dropping 86% peak to trough (FT). Neither flow is large enough to matter systemically. What they tell you is who is now supplying the downside insurance on the AI-chip complex: households, at retail prices, after the first drawdown rather than before it.

Bond market dysfunction23 Aughigh

Half the deficit is now interest

Once interest is the majority of the deficit, the borrowing path stops being a policy choice and becomes arithmetic.

More than half of America's budget deficit now consists of interest payments on past borrowing (The Economist). The deficit is around 6% of GDP; federal debt passed $40tn, roughly 130% of GDP, on 19 August. That single statistic is the whole debt-spiral argument in one line. A deficit made of programme spending responds to policy: you can tax more or spend less. A deficit made of interest responds to the yield curve, and the yield curve responds to the deficit. If the average rate on debt being rolled exceeds nominal growth, the interest bill compounds faster than the economy that services it, and each year's borrowing is larger for reasons no committee voted on. Government yields have risen worldwide this year: 0.6 percentage points on ten-year Treasuries, 0.4 on bunds, 0.8 on JGBs. The Economist attributes part of this to private borrowing for AI data centres making capital costlier for everyone — hyperscaler issuance reached $220bn in 2026 by 10 August, and AI-related debt $489bn by mid-year. Governments and Nvidia's customers are bidding for the same pool of savings. Scott Bessent's answer on 19 August was to double long-end buybacks to at least $4bn per operation. Yields fell for about a day. And yet: the rates market is not behaving like a market in distress. The MOVE index of Treasury volatility is at 8.99, down 28% in twenty days and 42% off its recent high. The two-year sits at 4.19%, the ten-year at 4.69%, the curve at 50bp. Something is being repriced slowly and in an orderly way. That is a different thing from a dislocation, and worth saying so.

The AI capex bubble22 Augmedium

The power leg is repricing first

The AI selloff so far is discriminating precisely by funding structure, which is what a credit repricing looks like before it becomes an equity event.

Look at where the AI complex actually fell last week. Nebius −21% over five days. CoreWeave −17%. Talen Energy −13%, now 21% off its high. Applied Digital −13%. Core Scientific −12%, and −22% over twenty days. Vistra −8%, and −17% over twenty days. Now the other side of the same trade. Microsoft −2.2% over five days but +27% over twenty. Oracle −2.6% over five, +27% over twenty. Alphabet flat. Amazon −1.6%. The split is not about enthusiasm for AI. It is about who can fund. A hyperscaler issues investment-grade bonds against operating cash flow — Alphabet placed $25bn on 10 August against a $205bn capex plan. A neocloud or a merchant power company funds with project debt and convertibles: Nebius upsized a convertible to $5bn on 20 August, one of the largest on record. When the cost of that funding rises, the marginal builder reprices before the incumbent. And it has risen: QTS sold $3.9bn of investment-grade data-centre bonds this month at a 7.23% yield, and BlackRock is paying 7.53% on high-grade paper for a Texas project — levels a middle-tier junk borrower would recognise (Bloomberg). The power names are the sharpest expression of this, because their equity is essentially an option on contracted future capacity that has to be financed before it earns. That is our inference from the price action, not something any filing states. But if the AI trade unwinds in an orderly sequence rather than all at once, this is what the first stage looks like — and Oracle up 3.1% on the day cuts against anything more dramatic.

Crypto and TradFi contagion23 Augmedium

The banks are building the pipes

Crypto losses only threaten the financial system once a regulated balance sheet sits in the path, and this month several volunteered.

Four things happened in the same short window. The OCC conditionally approved World Liberty Financial's application for a national trust bank, which would let it issue its USD1 stablecoin directly and custody the dollar reserves backing it, rather than using BitGo. Citi plans to offer bitcoin custody to institutional clients later this year through its Custody+ platform. Seventeen of the largest US banks, including JPMorgan, Citi, Bank of America, Wells Fargo and HSBC, announced a shared network for tokenised deposits, expected in the first half of 2027. And Mastercard closed its $1.8bn acquisition of BVNK (sweep). These are tier-two reports, not filings, and we would treat any individual item carefully. Together they describe something specific: contagion needs a channel, and the channel is being widened. Until now the crypto-to-banking connection was thin — a few lenders, some ETF plumbing, and Coinbase. Custody at a globally systemic bank is different: it puts operational and legal exposure to digital-asset prices inside a regulated balance sheet. A stablecoin issued inside a national trust charter puts the reserve-asset run risk there too, which is the exact structure that broke in March 2023 in a different guise. Meanwhile the prices have gone vertical. Bitcoin is at $77,729, up 20% in five days. Strategy is up 28% over the same stretch, Coinbase 26%, BitMine 26%, SharpLink 26% — every one of them at its recent high. The treasury companies are outrunning the coin again, which is what leverage does when it works. The rails being laid this year will be tested by a price move at some point. They have not been yet.

The dollar, gold and reserve status23 Augmedium

Gold's rally is a recovery, not a record

Gold rising while long yields also rise is the specific signature of a fiscal-credibility trade rather than a rates trade.

Gold closed Friday at a three-month high, up around 14% in August and about 5% on the week, after the Treasury's surprise ramp-up in long-bond buybacks (Business Times). It sits at $4,680 as we write, with silver up 19% in twenty days and the dollar index down 2.6%. One detail changes the reading. Gold is still about $1,000 an ounce below its peak in late January. The metal ground lower through most of this year as the new Fed chair asserted his independence and the Iran war raised the prospect of rate hikes rather than cuts. What we are watching is a violent recovery from a nine-month decline, not a fresh record — and "gold at a three-month high" and "gold 18% below its high" are both true statements about the same price. That matters for how much you should read into it. The debasement story — fiscal dominance, a Treasury buying its own bonds, a dollar paying the price — is a real and coherent explanation, and UBS's Bhanu Baweja put it plainly: bullion is the beneficiary of an effort to suppress borrowing costs, and "the dollar will pay the price." But part of this move is simply positioning being rebuilt from a low base after a bad year. The cleanest evidence for the fiscal-dominance reading is what gold did not do. Bullion kept climbing even after 30-year Treasuries erased most of their post-buyback gains. Normally gold tracks real yields down. Rising gold with rising long yields is the pattern that says the concern is about the sovereign, not about rates.

What would change our mind

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

High-yield OAS moving above 350bp, or investment grade above 100bp, from today's 275bp and 82bp — that would mean the AI financing repricing has left the equity market and entered credit.

Would move the number

A failed or pulled data-centre bond or convertible: any deal in the Nebius/QTS/Broadcom pipeline that cannot be placed at any price would be the funding channel shutting rather than tightening.

Would move the number

A digital asset treasury company disclosing a discount to the value of its coins, or halting an at-the-market equity programme — that is when the DAT flywheel becomes leverage rather than a funding advantage.

Would move the number

SOFR printing materially above the Fed's target range for more than a day or two, or repo haircuts widening, which is how the $2tn basis-trade position would start to unwind.

Would move the number