Broadcom will guarantee the debt used to buy Broadcom chips

Archived reading, published Fri, 21 Aug 2026 10:19:20 UTC (5 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

-1 since the last reading

Fragility89

how much tinder is stacked up — moves slowly

Ignition45

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 23 sources and rewrites this page.

Why it moved: Down one, from 68 to 67. Ignition gives back a point because last window's distressed mark has not propagated: high-yield spreads tightened to 273bp, rate volatility is at a 20-day low, the 10-year rallied to 4.65% and bitcoin is up 24% in five days — not the price action of a funding squeeze. Fragility holds at 89: Broadcom's $100bn vendor-guaranteed financing is genuinely new tinder but is still in talks, and Nvidia's $500bn platforms were counted a week ago.

Reporting from 19 Aug to 21 Aug

Crash points

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

What’s moving the needle

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

Hidden leverage and shadow banking21 Aughigh

The 100x contract that trades a stock you cannot buy

A trillion-dollar leveraged market in real-world assets has appeared in a year, and none of it shows up in the leverage measures anyone monitors.

Global trading volume in perpetual futures — futures with no expiry date, born in crypto — hit a record $86.2tn in notional terms in 2025, up 47% on the year, according to CoinGecko data cited by Nikkei. Volume in perps tied to real-world assets rather than coins went from $23.6bn in the first half of 2025 to $1.4tn in the first half of 2026. That is a 59-fold increase in twelve months. The mechanics are worth understanding because they are unlike a normal derivative. A perp never settles. It is kept near the spot price by a funding rate paid between longs and shorts, and platforms extend leverage that ranges from 6x on the US-regulated Kalshi to 100x on Bybit for certain bitcoin contracts. Nobody has to source or borrow the underlying asset. That means a perp can exist on something a foreign investor is legally barred from owning — which is exactly how it is now being used. Before humanoid-robot maker Unitree listed in Shanghai, its pre-IPO perpetual futures traded at nearly $100, more than four times the listing price. One contract, xyz.UNITREE, did over $105m of volume in 24 hours on Hyperliquid (Nikkei). The relevance to this site's argument is direct. This is leveraged equity exposure held outside any prime broker, unrecorded in margin-debt statistics, uncleared, and mostly offshore. When it unwinds, it unwinds against a spot market that is not connected to it. No regulator is counting it.

Private credit and BDCs21 Augmedium

A 6.1% default rate and a 2.8% non-accrual rate

The market is repricing the private-credit fundraising machine well before it reprices the loan books, which is the order these things usually happen in.

Fitch's US private credit default rate reached 6.1% for the trailing twelve months to July, up from 6.0% in June and the highest reading in the published series, according to summaries of its release. Set that against the other number circulating this week: a median non-accrual ratio of 2.8% across the twenty largest listed BDCs in Q2, up from 2.0% in Q1. These measure different things and the gap is instructive. A default rate counts borrowers that missed a payment or did a distressed exchange at any point over twelve months. A non-accrual ratio is a snapshot: loans on which a fund has stopped booking interest, on the date of the balance sheet. And it depends heavily on whether you measure at cost or at fair value. Blue Owl's BDC reported 2.8% non-accruals at cost and 0.8% at fair value — meaning the fund has already written those loans down by roughly seven-tenths of their original value. The write-down is the honest number, and it happened before the label changed. The listed prices tell you where the market thinks the pain is. Over five days the BDCs holding the loans barely moved — Ares Capital -2.0%, Blackstone Secured Lending -1.3%, Golub -1.2% — while the managers who collect the fees fell hard: Blue Owl -9.9%, Apollo -9.1%, KKR -7.2%, Ares Management -6.1%. Equity in a manager is a levered claim on future fundraising, and with $9.6bn of unmet redemption requests queued at non-traded BDCs, fundraising is the thing under threat, not yet the loans.

The AI capex bubble20 Augmedium

The vendor guarantees the loan that buys its product

Vendor guarantees move AI credit risk off the borrower's balance sheet and onto the seller's, where it is disclosed as a contingency rather than as debt.

Broadcom is in talks with lenders to raise more than $60bn of debt to finance AI chips for Anthropic and others, Bloomberg reports. The structure under discussion: a senior-secured tranche of roughly $60bn–$70bn, a junior tranche of about $30bn, issued out of a special-purpose vehicle, total potentially $100bn. Blackstone and Apollo are in talks to participate, following the $35bn compute-financing platform the three firms set up in June. Broadcom would guarantee a portion of the senior tranche (Bloomberg). That last clause is the whole story. In an ordinary sale, the vendor books revenue and walks away with the credit risk sitting on the buyer's balance sheet. Here the vendor books the revenue and then writes a guarantee that hands part of the credit risk back to itself. If Anthropic's compute demand materialises and the SPV services its debt, Broadcom has sold $100bn of hardware it might otherwise not have sold. If it does not, the guarantee converts a booked sale into a cash liability, at a moment when the collateral — chips and a half-built campus — is worth least. This is the same shape as Nvidia's arrangements: MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR to mobilise more than $500bn, with Nvidia indicating backstops of up to $125bn, roughly a quarter. The deal is not signed and the terms are described as subject to change. What matters is the direction: chip vendors are now the credit enhancement in the AI buildout, and the debt sits in vehicles nobody consolidates.

Household credit20 Augmedium

Subprime auto is past its 2008 peak. The lenders noticed.

Subprime auto stress above its 2008 level, in a labour market that is still near full employment, means the credit deterioration is structural rather than cyclical.

Sixty-day delinquency on US subprime auto loans has reached 6.9%, according to Fitch and Equifax data summarised this week — against a 2008 crisis peak of 5.0%. The New York Fed's Q2 2026 household report puts auto balances at $1.71tn with 5.49% of them 90 or more days past due, marginally better than 5.60% in Q1 but near the worst in a series going back to 2003. Credit-card balances are $1.263tn, with 12.92% of balances 90+ days delinquent and the flow into serious delinquency at 6.97%. One caveat we cannot resolve: the same sweep of coverage carries Fitch's subprime 60+ figure at 5.67% for June alongside the 6.9% number, and the two are not reconciled anywhere we can see. Treat 6.9% as the alarming end of a range, not a settled print. Note too that aggregate household delinquency actually improved, to 4.7% of balances. This is a distribution problem, not a population-wide one. The share prices are the interesting part. On Thursday Capital One fell 3.7%, Synchrony 3.7%, Sallie Mae 2.8% and Ally 2.0%, on a day the S&P lost 0.87%. Over five days the consumer-credit complex is down 4–7% while high-yield spreads tightened. Equity holders in the lenders are marking something the credit market is not. The mechanism to watch is not defaults themselves but funding: these lenders finance through securitisation. Deterioration shows up first in subordinate ABS tranches, then in origination volume, and only later in headline delinquency.

The dollar, gold and reserve status20 Aughigh

Gold is the price of Treasury's intervention

When a government intervenes to cap its own borrowing costs, the currency, not the bond, becomes the shock absorber.

Gold is at $4,639, up 2.7% in a day, 5.9% in five days and 14.1% in twenty. Silver is up 18.5% over the same twenty days. The dollar index is down 2.8%. Those three moves together are one trade, and the trade has a specific cause. On Wednesday the Treasury said it would at least double its buybacks of 10- to 30-year debt, from $2bn to at least $4bn per operation, running from 9 September to 4 November. It came after the 30-year auction cleared at 5.216%, the highest since 2001, and after the 30-year yield touched roughly 5.33%, the highest since 2007. It also follows the first US purchases of yen in three decades on 31 July, and an opened door to cutting long-end issuance. Bloomberg's description — the most interventionist Treasury secretary in decades — is not editorialising (Bloomberg). The plumbing: if the government is willing to buy its own long bonds to hold yields down while running a 6% deficit, then the adjustment has to appear somewhere else. Andrew Canobi of Franklin Templeton put it to Bloomberg as Bessent being "prepared to sacrifice a bit of dollar strength in order to keep term yields somewhat in check. Something has to be the relief valve." Gold is what the relief valve looks like on a chart. Separately, and single-sourced to Chinese state media, the LBMA has suspended the Good Delivery status of three Chinese gold companies after they were placed on a US sanctions list. If accurate, it is another reason for reserve managers to hold metal in vaults they control.

Crypto and TradFi contagion20 Augmedium

Bitcoin up 24% in five days, and nobody can say why

Crypto collateral is entering bank lending books just as the collateral demonstrates how fast it moves.

Bitcoin is at $77,773, up 6.5% on the day and 23.8% over five sessions. Essentially the entire twenty-day gain of 23.9% arrived in that one week. Over the same five days the S&P 500 fell 2.0% and the Dow fell 2.0%. The wrappers moved more than the coin, as they always do. Strategy is up 15.7% in five days and sitting at its high for the period. The ether treasury companies BitMine and SharpLink are up 17.8% and 20.4%, both also at period highs. Coinbase is up 11.9%, Galaxy 12.0%. A digital-asset treasury company is a levered claim on a coin plus a claim on the market's willingness to pay a premium for the wrapper; both legs are working at once, which is what a period high in all three at the same time means. We have no sourced explanation for the move in the evidence available, and we are not going to invent one. What we can say is what has changed structurally around it: JPMorgan has opened bitcoin-backed lending to institutional clients, taking bitcoin and ether as collateral against dollar loans at haircuts of 30–50%. On a $1m bitcoin position, that is roughly $500,000–$700,000 of borrowing. That haircut is the interesting number. It is conservative — three times what a bank would take on Treasuries — and it means the bank has priced a very large drawdown. But it also means a real bank balance sheet is now directly exposed to the price of an asset that just moved 24% in a week in the direction nobody had to explain.

Signs of the times

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

The derivative called the IPO first

Before Unitree listed in Shanghai, pre-IPO perpetual futures on the humanoid-robot maker traded near $100, over four times the listing price; the shares then opened up 629%. One contract did $105m of volume in 24 hours on a decentralised exchange. "The perpetual market called the mania before it happened," said Pepperstone's Dilin Wu.

leveragechinaderivatives

21 AugNikkei Asia

Fifty-nine times in twelve months

Trading volume in perpetual futures tied to real-world assets rather than crypto went from $23.6bn in the first half of 2025 to $1.4tn in the first half of 2026, per CoinGecko data. Total perp volume set a record $86.2tn in 2025.

leverageshadow-banking

21 AugNikkei Asia

A data-centre bond amortising to 2049

Meta's Hyperion vehicle issued about $27bn of A+ rated debt at 6.58%, amortising out to 2049, with PIMCO taking roughly $18bn. The loan runs twenty-three years against a building whose most valuable contents — the chips — are typically depreciated over about five.

ai-capexdurationoff-balance-sheet

20 AugCarson Group (web sweep)

$105bn for land, power and shell

Nvidia's regulatory filing caps its credit support for OpenAI's Ohio campus at $105bn — and that figure covers only the property, the energy and the warehouses, not the more than 1mn Nvidia chips that will go inside them. OpenAI takes a 20-year lease on a site that will not be finished until 2032.

ai-capexvendor-financing

21 AugFinancial Times

The rumour mill

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

Partly true

“The dollar's share of global central bank FX reserves has fallen from above 70% in 2000 to roughly 52–53% today.”

The direction is right and the destination is wrong. IMF COFER data put the dollar at about 71–72% in 1999–2001 and 57.13% in Q1 2026 — a large decline, but no reputable series shows 52–53%. Overstating a real trend makes it easier to dismiss.

Claimed by George Gammon

Partly true

“Central banks now hold more gold than US Treasuries, trending back toward 1980.”

The crossover appears real: analyses of ECB and World Gold Council data put gold at roughly 27% of official reserves against 22% for Treasuries at end-2025, the first such gap since about 1996. The comparison to 1980, when gold dominated reserves outright, is not supported.

Claimed by George Gammon

Partly true

“The Fed began 'significant stealth easing' in mid-December, buying Treasury bills, with the balance sheet expanding on net this year.”

The Fed did end runoff and start reserve-management T-bill purchases in December 2025, and Reuters reported on 17 August that the balance sheet is gradually growing again. The specific figures cited — roughly $150bn, plus surges in C&I and non-bank lending — are not established, and the Fed frames it as reserve management, not stimulus.

Claimed by Thoughtful Money (Adam Taggart)

Partly true

“Bessent's buyback programme adds roughly $14bn of long-bond purchases over seven weeks, funded by selling bills.”

The size and timing check out: $2bn to at least $4bn per operation, 9 September to 4 November, at least $14bn of additional purchases, taking maximum repurchases to $83bn. That it is funded by bill issuance is an inference, not a stated policy detail.

Claimed by Thoughtful Money (Adam Taggart)

Earlier developments

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

Private credit and BDCs20 Aughigh

The rater, not just the rated

If the ratings on affiliated private loans held by annuity writers cannot be relied on, the problem is not one insurer but the valuation method the whole insurance-plus-private-credit complex uses.

The US Attorney's Office for the Southern District of New York has gathered records of Egan-Jones Ratings Co.'s work with Delaware Life Insurance and Clear Spring Life and Annuity, two insurers controlled by Mark Walter, including by subpoena. Bloomberg reports that Egan-Jones provided the only known ratings for more than a fifth of the bonds those insurers held (Bloomberg). Prosecutors are trying to work out how more than $20bn of loans on the insurers' books ended up funding entities affiliated with Walter's other businesses — a fact not disclosed until this year. Why the rater matters. An insurer's regulatory capital requirement is a function of what its assets are rated. Public bonds have several ratings and a market price; private loans held by an annuity writer often have neither. What they have instead is a rating from a nationally recognised agency, and if only one agency has looked at the paper, that agency's opinion is effectively the asset's official value for solvency purposes. When more than a fifth of the portfolio has one rating from one small firm, the capital adequacy of the insurer rests on a single supplier's judgement. That is our inference, not Bloomberg's reporting, but it is why the subpoena is the interesting part. Egan-Jones says it "occasionally receive[s] subpoenas in the ordinary course" and is not a subject or target. A subpoena to a third party is routine information-gathering, not an accusation. The market has already voted on the credit. Guggenheim's $1.2bn loan — Walter runs Guggenheim as well as TWG — was quoted at 78 cents on Thursday against 96 the week before (FT). Listed private-credit managers have sold off all week. High-yield spreads have not moved at all.

The AI capex bubble20 Aughigh

China's robots are sold to the people who supply the data

The same accounting question that hangs over Nvidia's vendor financing now applies to an entire Chinese industry with 50-plus companies heading for public markets.

Chinese humanoid robot makers are earning much of their revenue selling machines to government-backed training centres, which then collect training data and sell it back to the robot makers (FT). The centres are often co-funded by local governments and by the robot makers themselves. Run the cash through the circle. The maker books a robot sale as revenue. The centre, part-funded by the maker, employs humans to teleoperate the machine and generate data. The maker then pays for that data as a cost of goods or R&D. Net cash out of the ecosystem can be close to nothing, but one side of the loop lands on a revenue line that investors capitalise at a growth multiple, and the other side lands somewhere less visible. Only a small share of the data is sold on to non-robotics buyers such as carmakers, training-centre staff told the FT. The FT draws the comparison to Nvidia's circular financing of AI data centres directly, and it is the right one, with one difference worth noting: in China the counterparty completing the circle is a local government with a policy target, not a private buyer with a return requirement. That makes the demand more durable in the short run and less informative about real demand in the long run. The money involved is not small. Nearly 370 humanoid startups have been founded in two years and more than 50 are listed or preparing to list. Unitree rose more than 600% after listing on Shanghai's Star Market on Wednesday, at a $50bn valuation. "This model can't last," one investor told the FT. Early-stage holders are reportedly looking for exits into later-stage hands — which is what distribution of a concentrated position looks like when it works, and what the last leg of a bubble looks like when it doesn't.

The AI capex bubble20 Augmedium

Data-centre debt is getting more expensive, deal by deal

Long-dated project debt on single-tenant AI buildings is being absorbed by insurance and annuity balance sheets, where nobody has to mark it until a tenant fails.

The AI buildout's financing costs are visible now in a run of individual prints. Meta's Hyperion vehicle raised roughly $27bn of A+ rated debt at 6.58%, amortising to 2049, with PIMCO taking about $18bn and Blue Owl-led investors holding around 80% of the equity. Meta's $12.5bn El Paso bond priced on 28 July at a higher yield than a structurally comparable 2025 deal. QTS Realty offered $3.9bn for a Microsoft-linked Georgia site at what Bloomberg described as junk-like yields. Nebius upsized a convertible sale to $5bn on 20 August. Broadcom is in talks for $60bn–$100bn, part of it guaranteed by Broadcom itself and issued through an SPV (Bloomberg). A 6.58% coupon on A+ paper amortising for 23 years is the number to hold on to. It says the buyers are pricing a single-tenant, single-purpose building against a 4.65% Treasury and demanding nearly two points for the tenant risk and the residual value, whatever the rating says. And the buyers are insurers and annuity money: PIMCO and Blue Owl are not warehousing this to flip it. It ends up backing retirement liabilities, marked at model prices, for a quarter of a century. Much of the deal detail above reaches us through aggregation of Bloomberg and CNBC reporting rather than the primary articles, so treat individual figures as approximately right rather than exact (CNBC summary). The equity is not buying the story as enthusiastically as the debt. Over five days CoreWeave is down 15.5%, Nebius 13.6%, Oracle 9.1%, Talen 11.6%. Spreads on the bonds financing the same assets have not budged.

Hidden leverage and shadow banking19 Aughigh

Two prop firms, $11.4bn up and $15bn down

The firms that now make more money trading than the big banks do report nothing, and their losses show up only when someone chooses to tell a reporter.

Hudson River Trading posted a record $11.4bn in quarterly trading revenue and a $7.4bn net profit for the three months to June, and projected no losses of the kind that hit Jane Street this quarter (FT). Jane Street reported a $15bn loss in July, after an investment in Leopold Aschenbrenner's hedge fund Situational Awareness and other AI stocks went against it. It has still made more than $40bn of net trading revenue in the past year. These are not marginal firms. In 2025 Jane Street, Hudson River and Citadel Securities made more than $60bn of trading revenue between them, beating the trading desks at JPMorgan and Goldman Sachs. They are, functionally, the market's liquidity providers in equities, ETFs and increasingly credit. They are also private partnerships. They do not publish balance sheets, they do not disclose gross exposure, and nobody outside their prime brokers knows their leverage. That is the part that matters. A $15bn loss in a single month absorbed quietly by a firm with $40bn of annual revenue is a demonstration of capital strength. The same loss at a firm with less capital would be a margin call at a prime broker, and the prime brokers are Morgan Stanley and Goldman Sachs — down 5.0% and 3.8% on the week, more than the index. There is also a tell in the composition of Jane Street's loss: a market-maker took a directional position in AI equities and a hedge fund built around an AI thesis. Liquidity provision and AI beta are not supposed to be the same trade. This quarter they were.

The dollar, gold and reserve status20 Aughigh

Japanese savers are quietly leaving the yen

Household-level currency diversification in the world's largest creditor nation is slow, one-way and much harder for policy to reverse than a speculative position.

Foreign-currency deposits at Japanese banks grew by a record ¥3.98tn ($25.1bn) year on year in the April–June quarter, to an average ¥26.1tn ($165bn), up 18% (Nikkei). It is the largest increase since Japan deregulated foreign-currency deposits in 1998, and it is bigger than the country's entire ¥1.7tn trade deficit for fiscal 2025. Retail non-yen deposits alone rose 8% to ¥6.7tn. The behavioural shift is the point. Japanese retail investors have historically been contrarian — buying foreign currency when the yen was strong. Now they are buying dollars at a 39-year low, past 163, after a joint US-Japan intervention failed to hold it. "Holding just yen is a risk," one Tokyo office worker told Nikkei. That is capital flight in polite clothing, and it is self-reinforcing: every household that hedges adds to the selling pressure that justifies the hedge. This connects to the Treasury market by a short chain. Scott Bessent bought yen on 31 July, the first US purchase in three decades — an action read as reducing Japan's need to sell its Treasury holdings to fund its own yen defence (Bloomberg). The US is now defending a foreign currency to protect the bid for its own bonds. Japanese households are, in aggregate, taking the other side. Meanwhile the dollar index is down 2.7% in twenty days to 98.7 and gold is at $4,609, up 2.1% in a day and 13.3% in a month. Both moves are consistent with the same reading: the world is diversifying out of the yen and the dollar at once, and into metal.

Crypto and TradFi contagion20 Augmedium

JPMorgan will lend against your bitcoin at a 30-50% haircut

Regulated bank balance sheets are now taking crypto price risk through a collateral haircut, which is a much faster transmission channel than an ETF.

JPMorgan is opening bitcoin-backed lending to institutional clients: bitcoin and ether held at third-party custodians can be pledged as collateral for dollar loans, with reported haircuts of 30% to 50% — so $1m of bitcoin supports roughly $500,000 to $700,000 of borrowing (sweep of crypto-industry reporting). This comes to us through aggregation rather than a tier-one story, so treat the exact haircut range as indicative. The haircut is the contagion parameter, and it is worth understanding what it does. At a 40% haircut, the bank is protected against a 40% drawdown before it has to make a margin call. Below that, the borrower must post more collateral or sell — and the thing they will sell is bitcoin, into the drawdown that caused the call. That is the standard mechanism by which a collateral market turns a price move into a liquidation cascade. It works fine until enough borrowers are in the same trade at the same strike. Bitcoin is at $75,418, up 3.8% on the day and 20.1% in five days. That is a large move to be building a bank lending book against. The more interesting number is a comparison. Over the same five days Strategy rose 15.7% — less than the coin it holds. The levered vehicle underperforming its underlying asset in a sharp rally is what a compressing premium to net asset value looks like. Coinbase (+11.9%) and Galaxy (+12.0%) also lagged. The treasury companies are sitting at their recent highs, but the market is paying less for the wrapper than it did.

Private credit and BDCs20 Aughigh

A loan fell 18 points in a week

This is the first time a price, rather than an investigation, has told us what the market thinks the Walter complex is worth.

The $1.2bn loan to Guggenheim Partners' $367bn asset management division was quoted at 78 cents on the dollar on Thursday, down from 96 cents at the end of last week (FT). Guggenheim is privately held. It has no share price. The loan is the only continuous market opinion that exists about it, which is why an 18-point move matters more here than a similar move in a listed name would. The fall followed a hastily arranged creditor call on Wednesday led by Guggenheim Investments president Dina DiLorenzo, convened to review the health of GIH Borrower, the vehicle that took out the loan. Three things came out of it. Executives said they were not aware of any active probe into Guggenheim Investments connected to the investigation into the TWG-backed insurers. Two of the four clients of Guggenheim Private Investments are affiliated with Walter's wider businesses. And GIH Borrower did not include advisory fees accrued by that unit in its second-quarter results, producing a sharp year-on-year drop in revenues and profits. That last point is the mechanism. Fee revenue booked from affiliated clients is revenue whose collectability depends on the affiliate. Stop recognising it and the borrower's earnings shrink; shrink the earnings and the leverage ratio on a $1.2bn loan moves without a single loan asset changing hands. Separately, Bloomberg reports the Justice Department has subpoenaed Egan-Jones, which provided the only known rating on more than a fifth of the bonds held by Delaware Life and Clear Spring Life. The SDNY is examining how more than $20bn of loans on those insurers' books funded affiliated entities — a figure that was $6.5bn in reporting a week ago and above $10bn on Thursday. Insurer capital charges are set off ratings. When one small firm is the sole rater on a fifth of the book, the capital adequacy of the insurer rests on that firm's judgement.

The AI capex bubble20 Aughigh

Broadcom will guarantee the loans that buy its chips

The credit quality of AI infrastructure debt increasingly rests on chipmakers guaranteeing their own customers, which concentrates the risk rather than distributing it.

Broadcom is in talks with lenders to raise more than $60bn of debt to finance AI chips for Anthropic and others, Bloomberg reports. The structure under discussion: a senior secured tranche of roughly $60bn–$70bn, a junior tranche of about $30bn, debt issued by a special-purpose vehicle, and Broadcom guaranteeing a portion of the senior piece. Total potentially $100bn. Blackstone and Apollo are in talks to participate, following the $35bn compute-financing platform the three set up in June. Read that structure carefully. Anthropic does not have a balance sheet that supports $60bn of secured debt. The SPV does not either — it owns chips whose second-hand value nobody has tested at scale. What makes the senior tranche financeable is the guarantee from the company selling the chips. The lender is not really underwriting Anthropic; it is underwriting Broadcom, plus collateral. This is now the industry pattern rather than an exception. Nvidia signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500bn of third-party capital, with Nvidia able to backstop up to a quarter — around $125bn. Nvidia is backstopping part of the value of OpenAI's Ohio campus in return for exclusivity on the first 5GW (WSJ). The correlation is the problem. A vendor guarantee gets called when the customer cannot pay, which is the same state of the world in which the vendor's own order book is collapsing. Bank capital rules exist partly to stop that concentration; an SPV has no capital rules at all. Bloomberg says terms are still being ironed out, and the size, triggers and accounting of Broadcom's guarantee are not public. We would like to see whether it lands on the balance sheet or in a footnote.

Hidden leverage and shadow banking20 Aughigh

Momentum fell 4% on a day the market rose

Leverage inside market-neutral strategies is invisible in index levels until the day a crowded factor breaks and everyone de-grosses at once.

Morgan Stanley told clients it was the first time in at least five years that its "pure momentum" index had fallen more than 4% while the S&P 500 finished higher on the same day. Goldman Sachs said its global momentum index was well outside historical norms and that systematic long-short hedge funds were down 1.4% by 1pm in New York — their worst day in more than two years (FT). Why does a factor move that size matter when the index was up? Because systematic long-short funds are, by construction, close to market-neutral. They do not earn the index return; they earn the spread between what they are long and what they are short, and to turn a small spread into a fund-sized return they run the book at multiples of capital. A 4% adverse move in the factor is therefore not a 4% day. It is a large fraction of the year. Then the second-order effect. Risk managers respond to that by cutting gross exposure, which means selling the longs and buying back the shorts — the same trades, in the same direction, across firms holding the same crowded factor. That is the transmission mechanism, and it is why an unrelated event can matter. Wednesday's session included both the Treasury's surprise buyback expansion and Moderna nearly tripling on skin-cancer trial results. One portfolio manager put it plainly: it "doesn't take a lot of [Modernas] to get risk managers to freak out a bit". Context against the alarm: systematic long-short funds were still up 1.7% for the month as of Goldman's note. This was a factor unwind, not a fund failure. But it is the second such unwind since the AI-stock selloff of July, and crowded-factor deleveraging is one of the few things that can move equity prices with no news attached.

Household credit20 Augmedium

Two subprime auto numbers, both attributed to Fitch

Household stress is now concentrated in cohorts small enough that aggregate delinquency can improve while the loss curve on subprime paper steepens.

One widely circulated write-up this week put 60-day delinquencies on US subprime auto loans at 6.9%, citing Fitch Ratings and Equifax, and noted that the 2008 peak was 5.0%. Another summary of the same week's data, also citing Fitch, put subprime 60-plus-day delinquency at 5.67% in June. We cannot reconcile the two. They are probably different cuts — 60-day versus 60-plus, a single month versus a rolling series — but the sources given do not say, and a reader should treat the 6.9% figure as unconfirmed until a primary Fitch release carries it. The New York Fed's Q2 numbers are firmer and more interesting for being mixed. Auto balances 90-plus days past due: 5.49%, down from 5.60% in Q1 but near the highest in a series going back to 2003. Credit card balances 90-plus days delinquent: 12.92%, with the flow into serious delinquency at 6.97%. Aggregate household delinquency: 4.7%, an improvement. Card balances $1.263tn, auto balances $1.71tn. The shape is a middle that is fine and a tail that is not. Aggregate delinquency improves because most households are current; the subprime auto and revolving-card cohorts deteriorate independently of the average, because those borrowers were already using credit as income substitution before rates rose. The equity market half-agrees. On Thursday, Capital One fell 3.7%, Synchrony 3.7%, Sallie Mae 2.8%, Affirm 3.0% and Ally 2.0%, against an S&P down 0.87%. But all of those except Ally are still up over twenty days, and Oportun is up 38%. That is a market pricing a funding-cost story, not a credit story. If the 6.9% figure is real, it is pricing the wrong one.

The dollar, gold and reserve status20 Aughigh

Everything that isn't a dollar went up

A Treasury that manages the yield curve directly transfers the adjustment to the currency, and the gold price is where you see it first.

Gold is at $4,586, up 4.7% in five days and at its high for the period. Silver is up 6.0% and also at its high. The dollar index is down 0.9% in five days and 2.7% in twenty. Bitcoin is up 19.2% in five days. Over the same week the S&P 500 fell 2.0%. That combination — hard assets up, the currency down, equities down — is a currency story rather than a risk-on story. The proximate cause is policy. Scott Bessent has doubled the size of Treasury's long-end buybacks, said on Thursday they could exceed $4bn per operation, opened the door to cutting long-dated issuance, and in July oversaw the first US purchases of yen in three decades (Bloomberg). The plumbing point is simple and worth stating slowly. If you cap long yields by buying back long bonds and funding the purchases with bills, you have not reduced what the government owes. You have shortened its duration and moved the adjustment somewhere else. The candidates are inflation and the exchange rate. "Something has to be the relief valve," as Franklin Templeton's Andrew Canobi told Bloomberg; Reed Capital's Gerald Gan put it more bluntly — "the dollar certainly is the biggest casualty". One further wrinkle, from a single and interested source. China Daily, a state outlet, reports in an editorial that Washington placed three Chinese gold companies on the UFLPA entity list and that the LBMA subsequently suspended their Good Delivery status. We have no second source. If accurate, it is worth watching: Good Delivery is the passport into London bullion clearing, and every use of it as a sanctions instrument raises the return to Shanghai building a rival standard.

Crypto and TradFi contagion20 Augmedium

The treasury companies are levered beta again

The treasury-company premium is the mechanism that converts crypto price moves into equity issuance, and it now sits alongside bank lending against the same collateral.

Bitcoin rose 3.1% on Thursday and 19.2% over five days, to $74,892. The listed vehicles that hold it rose more: Strategy up 7.8% on the day and 15.7% on the week, sitting exactly at its high for the period; BitMine up 6.5% and 17.8%; SharpLink up 7.5% and 20.4%, both also at period highs. Coinbase gained 11.9% on the week and Galaxy Digital 12.0%. This is the digital asset treasury model working in the direction it is designed to work. A DAT holds coins and funds them with equity and convertible debt. When the share price trades above the value of the coins, the company can issue stock, buy more coins and make each existing share better off — the issuance is accretive, so it issues, so the coin bid strengthens, so the price rises. It is a machine that runs on its own premium. The same machine runs backwards. Below net asset value, issuing shares destroys value per share, the buying stops, and any leverage in the structure has to be serviced out of a static coin pile. This week's reporting sweep looked specifically for a named DAT trading below NAV and found none, which is worth stating: we cannot currently source current mNAV figures for these names, so the prices above are all we have. What has changed structurally is the connection to bank credit. JPMorgan is now lending dollars to institutions against bitcoin and ether held with third-party custodians, at haircuts reported at 30% to 50%. A 40% haircut is a large cushion. It is also a margin call in waiting, and it is the first time a coin drawdown has a direct channel into a G-SIB's loan book rather than into a crypto lender's.

Private credit and BDCs19 Augmedium

The lenders are being marked down. The loans are not.

The listed managers are the only daily-priced window into a $2tn asset class that otherwise reports quarterly at its own valuations.

Five days of price action in the private-credit complex: Blue Owl -9.9%, Apollo -9.1%, KKR -7.2%, Ares -6.1%, Blackstone -5.3%. Over the same week the high-yield spread index barely moved, closing at 273 basis points, and the investment-grade index at 81bp. The listed BDCs — the vehicles that actually hold the loans — mostly fell less: Ares Capital -2.0%, Blackstone Secured Lending -1.3%, Golub -1.2%. Blue Owl's own BDC was the exception at -5.2%. That pattern is worth sitting with. A manager's share price is a claim on two things: the fee stream from assets under management, and the credibility of the marks those assets carry. It is the most junior claim in the structure — junior to the BDC's shareholders, who are junior to the loans themselves. So it reprices first, and it reprices on doubts that credit indices are not built to express. What doubts? The FT reported this week that Mark Walter held talks with Apollo through the summer about a multibillion-dollar loan against his Los Angeles Lakers stake, before selling the team at a $12.5bn valuation instead — cash raised while federal prosecutors examine insurers he controls that made more than $20bn of undisclosed related-party loans (FT). Delaware Life has swapped out $6.5bn of affiliated assets; more than $10bn remains. The insurance-plus-private-credit model that Walter helped popularise is now the industry's core business. Equity holders in the firms that run it are discounting something. The loan indices, which are quoted off dealer marks on liquid instruments, are not. Both cannot be right, and the equity market is usually earlier.

The AI capex bubble20 Augmedium

Investment-grade tenants, junk-grade bonds

The bond market is already saying these leases are not the same credit as the tenants signing them.

Four data-centre financings surfaced in this week's reporting, and the pricing is the story. QTS Realty sold a $3.9bn bond against a Microsoft-linked campus in Georgia at what one account calls "junk-like yields". Eagle Point Credit Management extended roughly $1.3bn of private credit against a Texas AI data centre tied to Anthropic, inside a $16bn project-finance package for Nexus Data Centers. Zenith Arc raised $2.25bn against a facility leased to Jane Street. Nebius upsized its own debt sale — reported at $4.5bn of convertibles split 2030 and 2034 in one account and $5bn in another; the two figures cannot be reconciled from what has been published (CNBC sweep). The mechanism to hold onto: in a single-asset project financing, the bond is a claim on one building and one lease. If the tenant is Microsoft, the credit should price close to Microsoft. It does not. Meta's Hyperion vehicle priced $27bn at 6.58% with an A+ rating; the El Paso deal came wider than a comparable 2025 structure. The spread between the tenant's own cost of funds and the project's is the market's price for everything the lease does not cover: the residual value of a purpose-built shell full of depreciating accelerators, the enforceability of the lease if the tenant restructures, and the fact that nobody has re-let one of these buildings yet. That is a real risk premium, honestly quoted. What is less comfortable is who is absorbing it. Not banks. Pimco took roughly $18bn of Hyperion. Blue Owl-led investors reportedly hold 80% of the equity. Eagle Point is a credit fund. This is the migration thesis in its purest form: a new, untested, long-dated asset class landing almost entirely outside the regulated banking system.

Household credit20 Aughigh

Wall Street has securitised personal injury lawsuits

When collateral this exotic gets securitised, the marginal buyer of credit risk has run out of ordinary things to buy.

The New York Times has documented an asset class we had not seen packaged before: cash advances to personal-injury plaintiffs, bundled into asset-backed securities. The advances carry fees and interest averaging 35 to 45 per cent a year. One New York woman who said she turned her ankle on a sidewalk received $76,500 in advances and owed at least $1.4mn by the time her case settled, according to court records (NYT). The structure is genuinely clever, which is the problem. The advances are non-recourse: if the case fails, the plaintiff owes nothing. That makes the credit uncorrelated to unemployment, house prices or rates — the pitch every structurer wants to make. What the pool is actually exposed to is court calendars and insurers' willingness to settle. Change either and the whole pool's cash flows move at once, in a way no consumer-credit model captures, because there is no cycle of history to model against. The Times places this alongside securitisations of songwriting royalties, fast-food franchises and data centres. That list is the whole argument of this site in one sentence. Bank lending is measured, capitalised against and stress-tested. These are debts owed by people to funders who fund themselves in the capital markets, sold to investors who mark them by model. Nobody in the chain has to hold capital against being wrong. For scale: this is a small market against $1.263tn of credit-card balances, of which 12.92% is 90-plus days delinquent per the New York Fed's Q2 data. It matters less as a size risk than as a signal of how far down the collateral barrel the securitisation machine is now reaching for yield.

The dollar, gold and reserve status20 Aughigh

Japanese savers are now selling the yen too

The largest foreign holder of Treasuries is under domestic currency pressure that Washington is now spending its own reserves to contain.

Foreign-currency deposits at Japanese banks grew by ¥3.98tn ($25.1bn) year on year in the April–June quarter, the largest increase since Japan fully deregulated such accounts in 1998. Total foreign-currency deposits averaged ¥26.1tn ($165bn), up 18%. Retail non-yen deposits rose 8% to ¥6.7tn. Sony Bank's foreign-currency book passed ¥800bn, fifteen months after passing ¥700bn (Nikkei). The behavioural change matters more than the amount. Japanese retail investors were historically contrarian — they bought foreign currency when the yen was strong and it looked cheap. They are now buying dollars with the yen at a 39-year low beyond 163, having watched a joint US-Japan intervention fail to hold it. That is not a valuation trade. That is capital flight by households, and it is self-reinforcing: every yen of deposits shifted is a yen sold. Now trace the loop. Japan intervenes to support the yen. To fund intervention it can sell Treasuries. Bessent, according to Bloomberg, bought yen on 31 July — the first US purchase in three decades — a move read as reducing Japan's need to sell down its Treasury stockpile (Bloomberg). In other words: the US Treasury is buying a foreign currency to protect the bid for its own long bond. And Japanese households are on the other side, selling yen for dollars every month. The official sector and the retail sector are pushing in opposite directions on the same exchange rate. Gold is up 13% in twenty days, the dollar index down 2.5%. Those two prices are the market's summary of the arrangement.

Bond market dysfunction19 Aughigh

The debt ceiling just moved forward a year

Treasury's improvised support for the long end depends on frictionless bill issuance, and the debt ceiling is the one thing that reliably stops it.

Gross federal debt crossed $40tn on Tuesday, according to Treasury data released Wednesday. It has grown $3tn in a year, the fastest pace outside the pandemic. Debt held by the public exceeds $32tn, roughly the size of the economy (FT). The number itself is a milestone, not a mechanism. The mechanism is what it does to the calendar. Congress set the debt limit at $41.1tn last year. Six months ago the CBO expected borrowing to top out at $39.4tn this fiscal year; it is already at $39.9tn and counting, partly because of revenue lost when Trump's tariffs were invalidated. Budget analysts now think the ceiling binds by early next year (Washington Post). Here is why that collides badly with everything else. Treasury is currently managing the long end by hand: doubling liquidity-support buybacks from $2bn to at least $4bn per operation from 9 September, leaning ever harder on bill issuance, and — Bessent said on CNBC on Thursday — possibly going beyond $4bn per issue. Buying back long bonds while funding with bills shortens the average maturity of the debt. That works as long as bills roll effortlessly. A debt-ceiling episode is precisely the event that interrupts bill supply, forces the cash balance down and then requires a violent rebuild afterwards. Doing that on a shortened maturity profile, with the New York Fed's term premium estimate near 80bp and 30-year auctions clearing at the highest yields since 2001, is a different exercise from doing it in 2023. Interest costs this year are projected above $1tn, more than the Pentagon budget.

Crypto and TradFi contagion20 Augmedium

JPMorgan will now lend against your bitcoin

A bank credit channel from the coin price now exists, at the same moment crypto and credit-linked equities are moving in opposite directions.

JPMorgan is opening bitcoin-backed lending to institutional clients: bitcoin and ether held with third-party custodians, pledged as collateral for dollar loans, at reported haircuts of 30 to 50 per cent. A $1mn bitcoin position supports roughly $500,000 to $700,000 of borrowing (sweep). The haircuts are conservative and that is the point worth being fair about: at a 50% advance rate, the bank survives a halving of the collateral before it is impaired. This is not 2021 crypto lending. What is new is that a regulated deposit-taker's balance sheet now has a direct, contractual link to the coin price, and the transmission runs through margin calls rather than through mark-to-market losses. The risk is not that JPMorgan loses money on any single loan. It is that a fast drawdown produces simultaneous calls across a client base that all holds the same two assets and all sells into the same order book. Meanwhile the market is going the other way, hard. Bitcoin is up 15.2% in five days to $72,630. Coinbase +11.9%, Galaxy +12.0%. All three digital-asset treasury companies we track are at their period highs: Strategy +15.7% over five days, SharpLink +20.4%, BitMine +17.8%. And equities did the opposite — the S&P fell 0.87% on the day, the Dow 1.32%. That split is unusual and informative. Risk appetite has not vanished; it has relocated out of things that depend on credit and into things that depend on nothing but flow. Treasury companies trading at highs means the equity-issuance window is open, which means more coins get bought, which is what the recent price is partly made of.

What would change our mind

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

High-yield OAS moving above 350bp from 273bp, or the Guggenheim loan trading below 70 — either would mean last week's distressed mark has become a market-wide repricing rather than one borrower's problem.

Would move the number

The Broadcom financing pricing and clearing at a spread comparable to Meta's Hyperion deal, which would tell us the buyer base for vendor-guaranteed AI paper is deep and real, not thin — and would lower our fragility estimate.

Would move the number

A non-traded BDC suspending repurchases entirely rather than gating at 5%, or a listed manager cutting fundraising guidance, which would confirm the redemption queue is becoming a funding event.

Would move the number

The 30-year yield sustaining above 5.4% despite doubled buybacks running from 9 September, which would mean the intervention has failed and the term premium is being set against the Treasury rather than by it.

Would move the number