AI's hidden $1.1 trillion debt. BDC non-accruals jump.

Archived reading, published Mon, 31 Aug 2026 06:22:27 UTC (21 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

Fragility92

how much tinder is stacked up — moves slowly

Ignition33

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 8 sources and rewrites this page.

Why it moved: Fragility up two points to 92: new reporting quantifies $1.09–1.16 trillion in off-balance-sheet AI lease commitments, and BDC non-accruals jumped significantly (e.g., BXSL from 0.3% to 3.6%). Ignition holds at 33: market indicators remain calm, and no new funding channels closed in this window.

Reporting from 27 Aug to 31 Aug

Crash points

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

What’s moving the needle

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

The AI capex bubble31 Aughigh

AI's Trillion-Dollar Shadow Debt

This quantifies a massive, hidden liability in the AI buildout, shifting risk outside traditional balance sheets and making the true scale of leverage harder to assess.

New reporting reveals that cloud service providers (CSPs) now hold between $1.09 trillion and $1.16 trillion in signed future lease commitments for AI data centers. This figure is nearly four times the $285 billion in lease liabilities currently recognized on their balance sheets, effectively pushing a massive portion of AI infrastructure costs into off-balance-sheet structures. This mechanism allows hyperscalers to finance their enormous capital expenditures—which in some cases now exceed 100% of operating cash flow—without the debt appearing directly on their books. Concrete examples include NVIDIA's $105 billion residual-value guarantee on OpenAI's leased data center in Ohio and Google's $44 billion rent guarantee for Anthropic. These arrangements, alongside new debt deals like JPMorgan's $5 billion package for Volta and Nebius's $5.75 billion in convertible notes, illustrate a growing reliance on complex, opaque financing that shifts risk and leverage outside traditional, regulated banking channels.

Bond market dysfunction31 Aughigh

Treasury's Buyback Fails to Calm Long End

The market's rejection of Treasury's intervention suggests deeper structural issues in the bond market and questions the effectiveness of policy tools in managing fiscal concerns.

The U.S. Treasury's attempt to stabilize the long-dated bond market proved ineffective last week. On August 19, Treasury Secretary Scott Bessent announced that long-end bond buybacks would be doubled from $2 billion to at least $4 billion per operation, targeting 10- to 30-year securities. This intervention followed the 30-year Treasury yield hitting a 19-year high of 5.337%. While yields briefly fell after the announcement, they quickly round-tripped back to pre-announcement levels within a day. By the end of the week, the 10-year Treasury yield was back near 4.73%, and the 30-year yield remained around 5.24%–5.30%, near multi-decade highs. Market commentary explicitly noted an absence of classic dysfunction indicators like failed auctions or dealer balance-sheet seizure, suggesting the intervention was more about price support than addressing a plumbing crisis, yet it failed to provide sustained relief.

Private credit and BDCs30 Aughigh

Private Credit Non-Accruals Soar

A sharp deterioration in private credit asset quality indicates underlying stress in the loans themselves, not just funding, raising concerns about potential losses.

Private credit asset quality deteriorated sharply in Q2 2026, with non-accrual rates at several major Business Development Companies (BDCs) reaching their highest levels since at least 2021. Jefferies reported that Blackstone Secured Lending Fund (BXSL) saw its non-accrual rate jump from 0.3% in Q1 2025 to 3.6% in Q2 2026. Other significant increases include Golub Capital BDC from 1.2% to 2.9%, Blue Owl Capital from 1.4% to 2.8%, and Ares Capital (ARCC) from 1.5% to 2.4%. Across the 20 largest BDCs, aggregate non-accruals rose from 2.0% at the end of March to 2.8% in Q2. A broader analysis of 66 funds showed $3.86 billion of debt moving from performing into stressed marks, and $0.65 billion into deep discount, with only $0.53 billion recovering. This rapid shift indicates a faster-than-expected deterioration in the underlying loan portfolios, highlighting significant stress in private credit assets.

Household credit30 Aughigh

Subprime Auto Delinquencies Hit 6.13%

Rising delinquencies in subprime auto loans and falling retail sales indicate increasing stress for the most vulnerable consumers, which can ripple through the broader economy.

Stress in household credit is rising, particularly in the subprime auto loan sector. Fitch reported that subprime auto delinquencies re-deteriorated to 6.13% in July, up from 5.80% at midyear. Broader data from the New York Fed showed that 5.5% of all U.S. auto debt was at least 90 days past due in Q2 2026, marking the second-highest level in over 20 years. This deterioration is concentrated in the lowest credit score bands, as FICO noted that 3.2 million student-loan delinquents saw an average 38-point drop in their FICO scores. Compounding this, U.S. retail sales fell 0.6% month-over-month in July, the biggest drop in over a year, with core control-group sales also down 0.4%. These figures collectively point to increasing financial strain on vulnerable households, impacting their ability to service debt and spend.

Private credit and BDCs29 Aughigh

Australian Private Credit Funds Gate Investors

These gates are a concrete example of funding channels closing in private credit, indicating liquidity stress and potential contagion risks for investors seeking to withdraw capital.

The Australian Securities and Investments Commission (ASIC) has warned of the "first significant cracks" appearing in the country's private credit market. This follows actions by several major funds to restrict investor redemptions. MA Financial, a prominent firm, announced a temporary redemption limit of up to 1% of funds under management per month across its affected private credit funds. Even more severely, CVS Lane, described as a major private credit fund, fully suspended the processing of both application and redemption requests across multiple funds. These measures are explicitly framed as responses to liquidity constraints and stress within the private credit sector, indicating that some funding channels are beginning to close or tighten.

Signs of the times

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

$1.1 Trillion AI Lease Debt

Cloud providers have signed $1.09–1.16 trillion in future lease commitments for AI data centers, nearly four times their recognized lease liabilities of $285 billion, pushing massive infrastructure costs off-balance-sheet.

aileverageoff-balance-sheet

30 AugUnited Daily News

Liquidator Invests in Lender

Hall Chadwick managing partner Richard Albarran, a prominent liquidator, admitted he is a personal investor in Blackbird Capital, a lender that referred 50 insolvency matters to his firm, ten times more than previously disclosed.

private creditconflictsopacity

30 AugAustralian Financial Review

SEC Proposes Semi-Annual Reporting

The SEC has proposed allowing companies to report earnings only twice a year and exempting most from outside audits of internal controls, a rollback of post-Enron regulations that critics warn could hide fraud or financial stress.

policytransparencyregulation

30 AugThe New York Times

Volta, Just Out of Stealth, Secures $5B

JPMorgan Chase is arranging a $5 billion debt package for Volta Infrastructure Holdings, a company that only recently emerged from stealth mode, to fund its AI data center buildout.

aifundingspeed

31 AugPerplexity (sonar-pro)

The rumour mill

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

Confirmed

“US retail sales fell 0.6% month-over-month in July (reported by Reuters/Census), the biggest drop in over a year, with core control-group sales also down 0.4%.”

U.S. retail and food services sales fell 0.6% month-over-month in July 2026, and the retail “control group” also fell 0.4%, with multiple outlets noting this was the biggest overall retail-sales drop in more than a year.

Claimed by Reventure Consulting

Confirmed

“The INTerpath-001 phase 3 trial, a Moderna-Merck partnership for a personalized mRNA cancer therapy, met its primary endpoint of significantly reducing melanoma recurrence, announced around August 19.”

Merck and Moderna have publicly reported that the Phase 3 INTerpath‑001 trial of their personalized mRNA cancer therapy met its primary endpoint of recurrence‑free survival and a key secondary endpoint of distant metastasis‑free survival, with topline results announced on 19 August 2026.

Claimed by Meet Kevin

Confirmed

“Flock says its network now processes more than 20 billion vehicle detections every month across 49 states.”

Flock Safety itself, as well as multiple recent reports, state that its camera network operates in 49 U.S. states and performs around/more than 20 billion vehicle scans (detections) every month.

Claimed by Michael Bordenaro

Confirmed

“On August 19, the US Treasury overrode the buyback plan it had announced two weeks earlier, doubling the maximum size of each operation from $2 billion to at least $4 billion focused on 10–30 year bonds (running September 9 to November 4), and Bessent told CNBC the next day the Treasury could buy even more than $4 billion if needed.”

On 19 August 2026, the U.S. Treasury formally announced that it was doubling the maximum size of its liquidity-support buyback operations from $2 billion to at least $4 billion per operation, effective for 10–30 year bonds from 9 September to 4 November 2026, and on 20 August 2026 Treasury Secretary Scott Bessent told CNBC that the buybacks “could be more than the $4 billion per issue.”

Claimed by Coin Bureau

Earlier developments

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

Private credit and BDCs30 Augmedium

Seven dollars going bad for every one getting better

When a fund lends to a private company, the price it puts on that loan is its own estimate until somebody actually sells the loan. Those loans are now sliding into trouble seven times faster than they are recovering, and the funds' shares are trading as if nothing has happened.

Jefferies' second-quarter 2026 numbers, circulated on August 28: at Blackstone Secured Lending, the share of loans where the borrower has stopped paying interest is 3.6% of the portfolio, against 0.3% in the first quarter of 2025. That is a twelvefold increase in fifteen months. Golub Capital's fund is at 2.9%, up from 1.2%. Blue Owl is at 2.8%, from 1.4%. Ares Capital is at 2.4%, from 1.5%. Across the twenty largest funds of this kind the figure is 2.8% in the second quarter, up from 2.0% at end-March, so the deterioration is accelerating inside the year, not just against a soft 2025 base. Morningstar DBRS, using a wider universe and calling the trend "manageable," has the average at 3.4% in the second quarter, from 3.1% at year-end. The data on how fast loans are moving between categories matters more. Across 59 funds, loans carried below 90% of what was paid for them, or where the borrower has stopped paying, rose from 0.9% to 3.2% of cost between the fourth quarter of 2025 and the first quarter of 2026. On matched positions across 66 funds, $3.86 billion of loans moved from performing into troubled marks in a single quarter, $0.65 billion moved from troubled into deeply discounted, and only $0.53 billion recovered. This is the better instrument. Whether a lender admits a borrower has stopped paying is a label the lender applies, and it lags. The share of loans the lender itself marks below 90 cents on the dollar is where the lender's own estimate moves first. And the share prices: Ares Capital is up 6.3% over twenty days, Blackstone Secured Lending 7.1%, FS KKR 15.9%, Blue Owl's management company 19.1%. Blackstone Secured Lending sits 1.1% below its recent high with non-paying loans twelve times what they were fifteen months ago. Either the market believes 3.6% is the peak, or it is pricing the dividend checks these funds mail to shareholders and ignoring the loans underneath. Meanwhile, requests to pull money out of private lending funds that do not trade on an exchange hit a record 12.4% of the funds' stated value in the second quarter, from 10.4% in the first. Money is asking to leave a book that is rolling the wrong way.

The AI capex bubble29 Augmedium

The chipmaker guarantees the campus. The tenant is building a better chip.

Guarantees from the chip supplier are what make single-tenant data center campuses possible to finance. Those guarantees are underwritten on the assumption that the tenant keeps buying the guarantor's chips.

TrendForce's account of Nvidia's own disclosure: at the PORTS-Pike site in Ohio, SB Energy builds, owns and operates the facility; OpenAI leases 8 gigawatts of computing capacity; Nvidia takes exclusive rights to supply the computing power and provides a guarantee that the initial 4.25 gigawatts of infrastructure will still be worth something when the lease ends, with total payment obligations capped at $105 billion. That guarantee covers land, power and buildings. It does not cover the chips. A guarantee like this moves the risk that a building becomes worthless from the landlord to the chipmaker, and that transfer is what makes a lender willing to fund a campus with only one tenant at all. Read narrowly, the structure is defensible: Nvidia is standing behind the parts with alternative uses (dirt, substations, a shell) and not behind the silicon, where obsolescence actually lives. Google has done something similar in reverse, guaranteeing $44 billion of Anthropic's data center rent in exchange for roughly 20% ownership of the halls and power projects, with Broadcom pledging to make up the difference on chips that do not sell. What has to stay true is that the anchor tenant keeps wanting the computing power. On August 25, Bloomberg reported that OpenAI's own inference chip, Jalapeño, co-designed with Broadcom, beat Nvidia's GB300 in OpenAI's benchmarking on AI work per unit of power and on response speed. Those are OpenAI's own tests, on OpenAI's own workloads, and volume production is not imminent. But the customer whose lease payments Nvidia has guaranteed to $105 billion is designing itself out of Nvidia's product line. Nvidia fell 4.66% on August 28. The smaller data center companies fell harder: IREN down 12.6% in a day, Core Scientific 31% below its high, Applied Digital down 7.6%. Microsoft closed at its record and Amazon rose 4%. The market is separating the companies whose own cash can fund this from the ones that must borrow.

Hidden leverage and shadow banking30 Augmedium

The regulator is asking the banks, not the fund

The fix that was supposed to make the most crowded trade in the government bond market survivable was central clearing, where a middleman stands between the two sides of every deal. The share of trades going through that middleman is falling while the trade itself grows.

Reuters reported on August 24 that the SEC, the agency that polices American markets, has sent subpoenas to Wall Street banks that dealt with Situational Awareness, seeking information about the fund's trades and its use of borrowed money after its near-collapse last month. Other coverage puts the borrowed money at roughly four times the fund's actual capital. Four times is not an extreme number by the standards of investment firms that manage money for the wealthy. The interesting part is the direction of the subpoena. A fund's total borrowing is visible in full to nobody: each bank that lends to it sees only the slice it finances, the fund sees the whole picture, and the regulator sees a delayed and aggregated form. When the SEC wants to reconstruct what a fund was actually carrying, it has to go to the lenders and add up the pieces. That is the shape of the problem in one procedural detail. Alongside it, a number from this week that got no attention: the share of overnight government bond lending that goes through a central clearing house, where a middleman stands between the two sides and absorbs a failure, fell from 71% at end-2025 to 62% in the first quarter of 2026. Central clearing is the fix built after the 2020 crisis. It puts a clearing house between the cash lender and the borrower and spreads a default across the system instead of leaving it with whichever bank happened to be on the other side. The share moving backward means a larger share of the financing behind the most crowded trade in the government bond market (roughly $830 billion of positions that profit from a small gap between two nearly identical prices, inside about $4 trillion of total hedge fund government bond exposure, 90% of it in fifty funds) is back to being a private handshake between two firms. We would want the official series before leaning on that clearing figure; it comes from market commentary, not a regulator's release. But if it holds, the plumbing built to make this trade survivable is being used less, not more, while the trade gets bigger.

Crypto and TradFi contagionmedium

The coin moved two percent. The proxies moved eight.

Companies that hold bitcoin on their balance sheets transmit crypto swings into the stock market at roughly 1.6 times the move, and their ability to keep buying coins depends entirely on their shares trading above the value of what they already hold.

On August 28, Galaxy Digital fell 8.2%, SharpLink 7.4%, Strategy 7.4%, Bitmine 7.2% and Coinbase 6.4%. Bitcoin, over the five sessions to Sunday, fell 1.66%, to $77,718. Over twenty days the same group runs the other way: bitcoin up 22.3%, Strategy up 36.5%, Bitmine up 37.7%, SharpLink up 33.9%. Call it 1.6 times in both directions. That is what a company built around holding digital coins is: a pile of coins wrapped in a corporate structure, whose share price is the coin multiplied by a premium over the value of what the company actually holds. The premium expands when the company can sell new shares at a high price and buy more coins per existing share, and it compresses when it cannot. The important thing about this kind of amplification is that there is no lender demanding more cash behind it. Nobody lends against these positions overnight; nothing gets forcibly sold when the premium halves. The shares simply fall faster than the asset. That is why a 7% day in five listed companies leaves no fingerprint in overnight lending markets and no fingerprint in the VIX, which closed at 14.43. The transmission channel, if there is one, is the buying cycle. These companies have been a persistent buyer of coins, funded by selling stock at a premium to the coins they already hold. If the premium goes to zero, that buying stops, not violently, just permanently. Nothing in this window suggests any of them is a forced seller, and we should say that plainly: no gates, no demands for cash, no disclosed redemptions. What we are watching is a buyer at the margin whose ability to buy is a function of its own share price.

The dollar, gold and reserve status30 Aughigh

Australia's central bank cut its dollars by ten points

When a central bank changes what it holds in reserve, the decision is slow, deliberate and hard to reverse. That shift is now showing up in gold and in central bank portfolios, while the dollar's exchange rate itself does nothing.

The Reserve Bank of Australia's annual report discloses that it cut the US dollar share of its foreign reserve portfolio from 55% to 45% in 2024/25, while raising the euro from 20% to 30%. Deutsche Bank called it a quiet reduction. It is a ten-point move in one financial year by a major central bank. Set against the global picture, it is striking twice over. IMF data put the dollar at 57.13% of allocated global reserves in the first quarter of 2026, down from roughly 71% in 2000, a drift of about half a point a year. The RBA did twenty years of that drift in one. And the destination is unusual: in the global aggregate the euro has captured essentially none of the dollar's lost share, with about a quarter going to major Asian currencies and three-quarters going to smaller reserve currencies. The RBA went to euros anyway. Central bank reserve managers are not traders. They set target allocations, and those targets are sticky for years. A change of this size is a considered decision about what you want to be holding in a crisis, taken once and then lived with. It is not a market call and it will not reverse on a good inflation report. The price action says the same thing in a different currency. Gold rose 10% in August, its best month since January, which Bloomberg ties to Treasury Secretary Bessent's plan to at least double government purchases of bonds that do not come due for ten to thirty years, and to worries about the dollar's purchasing power. It then fell 3.3% on Friday after Federal Reserve chair Kevin Warsh's pledge to fight inflation; silver fell 4.4%. Gold is still up 9.6% over twenty days at $4,489. Meanwhile the dollar index, which tracks the dollar against a basket of other currencies, is 99.6, down 0.3% over twenty days. The move away from the dollar is being expressed in gold and in reserve portfolios, not in the exchange rate.

Household credit30 Aughigh

One in eighteen auto loan dollars is ninety days late

A flat national average that conceals record stress among the weakest borrowers is two credit cycles running at once, and only one of them is visible in the headline numbers.

The New York Fed's second-quarter data show 5.5% of US auto debt at least 90 days past due, the second-highest reading in more than twenty years. Fitch has July delinquencies on auto loans to the weakest borrowers at 6.13%, up from 5.80% at midyear, though still below the 6.50% at the end of 2025. For the strongest borrowers, it is 0.49%. Moody's, via Equifax, describes delinquency among the weakest auto borrowers as at an all-time high while overall auto delinquency is elevated but stable. Credit cards are going the other way. Card delinquency of 30 days or more fell to 2.85% in the second quarter from 3.04% a year earlier; Equifax's broader 60-day-plus measure fell to 2.69% from 2.87%. Mortgage early-stage delinquency eased to 1.35% from 1.42%. The average credit score is unchanged at 714. Credit scores explain how both can be true. Ninety-day-plus delinquency on mortgages and autos rose exclusively in the lowest score bands and was flat across every higher range. Separately, about 3.2 million borrowers with a student loan payment due went delinquent and lost an average of 38 credit-score points year on year, while consistent payers gained 6 points and those with no recent delinquency gained 16. That is not one household sector with a mild problem. It is two populations moving apart, whose average happens to be steady. Auto is where the weaker one shows up first, because the car loses value faster than the loan gets paid down and a weak borrower cannot borrow again to get out of the problem. We would not call this a threat to the system on its own. Loans to the weakest auto borrowers are small relative to anything that could break a bank. This is a timing signal on the consumer, not a source of large losses. And the lenders' shares agree: Credit Acceptance, the specialist in the weakest borrowers, sits 1.9% off its high; Ally is down 2.6% over twenty days; Affirm is 1% off its high.

Private credit and BDCs30 Augmedium

The price was offered. The seller declined.

The only outside price check private credit has is an actual transaction, and a transaction that gets canceled leaves the marks untested.

Ares Management has scaled back a European private credit continuation vehicle from €1 billion to €400 million after secondary investors demanded deeper discounts on the underlying loans than Ares was prepared to accept (Fund Finance Friday summary). This is worth slowing down on, because it is one of the few moments in private credit where an outside party puts a number on a loan. In a fund that lends to mid-sized private companies, or in a drawdown fund, the manager decides what its own loans are worth, with a valuation agent's blessing. Nobody actually buys or sells. A continuation vehicle is different: real buyers with their own money bid for the same loans, and the bid is a price. When that bid comes in below what the manager says the loans are worth, the manager has two options. Accept it, and take the markdown across the whole fund. That is what BlackRock's TCP Capital did: it sold $523 million of loans, about 48% of its debt assets, cut its borrowed money from 1.38 times its equity to 0.4 times, and ate a 10.4% hit to the value of the fund, $0.68 a share against a June value of $6.58. Or shrink the deal, keep the loans, and keep the mark. Ares took the second route. Neither choice is improper, and the loans may pay in full. But the mechanism matters: the escape hatch from private credit (selling to secondaries) narrows precisely when the price on offer is one you do not want to print. Loans where the borrower has stopped paying interest across the twenty largest funds that lend to private companies went to 2.8% in the second quarter from 2.0% in the first. Meanwhile Ares's stock is up 11% in twenty days, its lending fund up 6%, Blackstone's lending fund up 7%, and the extra interest that shaky borrowers pay over the government is 7% tighter. The market is not pricing this at all.

The AI capex bubble30 Auglow

A trillion in rent, four times what is on the books

The AI buildout's biggest liability is a lease obligation that does not appear on the list of what a company owes until the concrete is poured.

One institutional estimate circulating in Taipei this weekend: the big cloud providers have signed $1.09 trillion to $1.16 trillion of future data-center lease payments that have not yet been executed, against roughly $285 billion of lease liabilities actually recognized on their books. Roughly four to one. The same note puts cloud-provider bond sales this year at about $220 billion as of August 25 (United Daily News). Treat the exact figure with care. It is a tier-three report of an analyst estimate, not a filing. But the accounting mechanism behind it is not in dispute. A lease becomes something a company owes on its books when it commences, meaning when the building is available for use. A twenty-year commitment on a campus that opens in 2028 sits in a footnote until then. So the fastest-growing obligation in corporate America is largely invisible in the line item where you would look for it, and visible only in disclosure that the SEC, the agency that polices American markets, has now proposed companies file half as often. Meta's Hyperion campus in Richland Parish is the template our fact-check could partially verify: a roughly $27 billion joint venture with Blue Owl funds owning 80% and Meta 20%, debt maturing in 2049, the campus leased back to Meta. The claim that it is the largest private credit deal ever is not established by any outlet we can find; the structure is. What has to stay true is simple. AI revenue must grow fast enough to cover fixed payments that begin when the buildings open, whether or not the demand arrives on schedule. The market is split on that: Microsoft sits at its high, while CoreWeave is 22% off, Core Scientific 31% off, IREN 21% off. The tenants are priced for it. The landlords are not.

Private credit and BDCs30 Aughigh

The liquidator was an investor in the lender

Private credit's loss numbers depend on recovery processes that, in at least one case, are run by people with a stake in the lender.

Richard Albarran, managing partner of the accounting firm Hall Chadwick, disclosed to Australia's corporate regulator last week that Blackbird Capital Group has referred 50 insolvency matters to his firm over 24 months, ten times what he had previously disclosed, and that he is a personal investor in Blackbird, whose loans his firm vets (Australian Financial Review). Follow the chain. A private lender writes a loan. An outside firm assesses it. The borrower fails to pay. The lender appoints a liquidator to take the borrower's assets and recover what it can. The recovery number is what ultimately determines the loss on the loan, and therefore what the loan is said to be worth, and therefore the return the lender reports to its investors. On this disclosure, one man's firm sits at the assessment end and the recovery end, and he has money in the lender in between. The filing does not say that any loan was mispriced or any recovery mishandled, and we are not asserting either. What it illustrates is structural. Bank lending is checked by supervisors, capital rules and a published provision for loan losses. Private lending is checked by relationships between a small number of firms who all have reason to want the same answer. When those firms are also each other's clients and each other's investors, the check is not a check. This lands in the one market where the gates are already closing: ASIC, Australia's markets regulator, warned on August 27 of the "first significant cracks" in Australian private credit; MA Financial has capped redemptions at 1% of funds under management a month; CVS Lane suspended applications and redemptions outright. Australia is small. It is also early.

Crypto and TradFi contagionmedium

Fifty thousand ether, twelve seconds, one insurance fund

The most leveraged corner of finance prices itself in seconds; the one we worry about most does not price itself at all.

At 04:51:03 UTC on August 20, a Hyperliquid account known as pension-usdt.eth held a short position of 50,000 ether, roughly $108 million in face value. By 04:51:15 it did not. The forced sale ran across five transactions in twelve seconds, cost the trader somewhere between $24 million and $27 million (around half their reported lifetime shorting profits) and the final 1,417 ether could not be filled in the order book at all, so the exchange's insurance fund absorbed them. Our fact-check rates the account, size, timing and backstop as supported; the claimed 3x borrowed-money ratio and $1,700 average entry are not. That was one position inside the largest short squeeze on record. Across August 19 and 20, roughly $2.7 billion of positions were forcibly closed in 24 hours, about 91% of them bets that prices would fall, hitting some 172,000 accounts. It was the biggest one-sided forced-selling event in CoinGlass data going back to 2021, per Bloomberg. The mechanism is worth understanding because it is the opposite of everything else on this site. On a perpetual futures venue, when a trader's cash cushion runs out the liquidation engine seizes the position and sells it into the order book immediately. If the order book cannot take it, the insurance fund covers the gap. If that empties, profitable traders on the other side get their positions closed against their will. Losses are spread across users within seconds, at a price everyone can see, with no creditor and no negotiation. Which is why it never reaches bank capital, and why bitcoin is up 23% in twenty days to $78,600, partly on covering rather than new money. Compare with the dispatch above, where a price takes a quarter, a valuation agent and a canceled auction. Crypto's plumbing is violent and honest. Private credit's is calm and unverified.

The dollar, gold and reserve status30 Augmedium

Washington bought yen, and paid in euros

Reserve managers are trimming dollars at the margin while the US Treasury spends its non-dollar reserves defending the yen.

The joint US-Japan intervention that halted the yen's slide (from about ¥164 per dollar in late July to around ¥158) was executed on the American side by the New York Fed selling euros out of the Exchange Stabilization Fund, rather than dollars, according to Financial Times reporting relayed by multiple outlets and checked by us as partly true. The framing that this was the first joint yen-buying operation since 1998 is contested: there was coordinated US-Japan action after the 2011 earthquake. The motive commonly attributed (that the Treasury Department, which borrows the money the government spends, sold euros to avoid weighing on the dollar) is inference by commentators, not something either government has stated. The plumbing is the interesting part. The Exchange Stabilization Fund holds foreign currency assets. To buy yen you must sell something, and what you sell determines who bears the price impact. Selling euros keeps the operation from adding to the supply of dollars, and it also spends down a finite stock of non-dollar reserves. This is the United States using its own reserve assets to defend somebody else's currency, on the argument (Bessent's, publicly) that it was a financial stability matter. Set against that: the Reserve Bank of Australia cut the dollar share of its reserve portfolio from 55% to 45% in 2024/25 and raised the euro share from 20% to 30%. IMF data still put the dollar at 57.13% of allocated global reserves in the first quarter of 2026, against 71% in 2000. Gold rose 10% in August, its best month since January, before retreating 3.3% on Friday after Kevin Warsh's Jackson Hole pledge on inflation (Bloomberg). None of this is a dollar crisis. The dollar index, which tracks the dollar against a basket of other currencies, is at 99.68 and up 0.8% on the week. It is a slow reallocation, and the US spending euros to hold a yen line is one more data point in it.

Household credit30 Augmedium

Korea hikes into two thousand trillion won of debt

Korea is the live test of what higher rates do to variable-rate household debt, and Warsh has left a US hike on the table.

The Bank of Korea, the country's central bank, raised its policy rate to 3% on Thursday, a second consecutive hike, into a household sector that has just passed 2,019.8 trillion won ($1.45 trillion) of total credit for the first time. Self-employed borrowers owe 1,095 trillion won, a record; 1.6 million of them hold three or more loans, owing 645 trillion won between them; their late-payment rate has climbed to 2.04%, the highest since 2015 (Korea Times). What makes Korea the cleanest experiment anywhere in the world right now is the rate structure. About 65.5% of self-employed loans are variable, meaning the interest rate moves when the central bank moves. The Bank of Korea's own estimate is that a 0.25 percentage point move adds 1.8 trillion won to their annual interest bill, roughly 560,000 won per borrower. That transmits within a quarter. In the United States, a 30-year fixed mortgage means a hiking cycle takes years to reach the household's cash flow and mostly never reaches existing borrowers at all. In Korea it arrives with the next statement. So if you want to know what tightening does to a maximally indebted household sector, you do not have to model it. You can watch it, and the late-payment series is already at an eleven-year high before this hike lands. The US household picture, for contrast, keeps splitting. Credit card 30-day late payments were 2.85% in the second quarter against 3.04% a year earlier; FICO's average score held at 714. The damage is concentrated: late payments on car loans to the shakiest borrowers rose again to 6.13% in July after improving at midyear, and New York Fed data show 5.5% of all car-loan debt 90 days or more late, the second-highest reading in over twenty years.

Fed, Treasury and policy30 Aughigh

Fewer numbers, and nobody checking the machine

Publicly listed companies were the part of the system with reliable, frequent, independently checked numbers. Both proposals reduce that.

The SEC (the agency that polices American markets) has proposed two changes to corporate reporting. The first would let listed companies file earnings twice a year instead of four times, ending a rule that has run for more than fifty years. The second, which has drawn far less attention, would exempt most companies the SEC regulates from bringing in outside auditors to verify their internal books and controls. That requirement was the one Congress wrote in 2002 after Enron collapsed and took its accounting firm, Arthur Andersen, with it (New York Times). The first is arguable. Britain and much of Europe already report every six months, and the case that quarterly reporting encourages companies to chase short-term results is a real one, made by serious people. The second is different in kind. An internal-controls audit does not check whether the numbers are good news. It checks whether the process that produces the numbers is capable of producing true ones. Take it away and you are left trusting management's own assessment of management's own systems, which is precisely the arrangement that failed in 2001. "If the quality of reporting information from the financial system deteriorates, then that absolutely leads to financial-sector risks of the kind that have bitten us before, as in 2008," Simon Johnson of the Systemic Risk Council told the Times. Our running argument is that risk has migrated to places where nobody has to price things honestly: private lending funds, insurance company books, separate companies set up to hold data-center leases off the parent's balance sheet. This proposal reduces the frequency and verification of those prices in the one part of the system that still had both. It is a proposal, not a rule, and it may be softened. But note the direction: at the point in the cycle when you would most want more measurement, the measurement is being thinned.

The AI capex bubble29 Aughigh

IREN borrowed $2.4bn. Its shares fell 15% that week.

The same company is being valued two ways at once: down 15% in the stock market, unchanged in the debt that funded it.

The loan: $1.2bn in senior secured borrowing plus $1.2bn of senior secured bonds, both at 9% interest, coming due two and a half years after funding, arranged by Blue Owl funds with Pimco. The deal is structured to let IREN draw the money in stages as it buys Nvidia Blackwell Ultra processors for a Canadian campus (Bloomberg). The stock market's verdict, in the same week: IREN down 12.6% in a single session and 15.4% over five, now 21% below its recent high. Core Scientific, a similar company, down 6.2% on the day and 20% over twenty sessions. Applied Digital down 7.6%. The power companies that sell electricity into the same buildout went with them: Talen down 5.7% on the week, Vistra down 7.5% over twenty days. Meanwhile the extra interest that shaky borrowers pay compared with the government sat at 2.63 percentage points, 7% lower over twenty days. The debt market is not repricing this at all. The stock market is. That split is the thing to hold onto. Shareholders take the first loss, so their prices move first, and what they are repricing is a specific piece of arithmetic: thirty-month money at 9% interest, secured on hardware that loses value on an uncertain schedule and is being replaced by newer models on roughly an annual cycle. For the lender to be repaid, either the campus generates contracted revenue covering 9% plus the hardware's loss of value within thirty months, or the processors are worth enough at the end to borrow against again. The filing does not tell us what revenue contracts sit behind the Canadian campus. The lender here is a private lending fund. It will price this loan quarterly, at a value of its own construction. The stock prices itself every second the market is open, and it is marking down.

Private credit and BDCs30 Augmedium

A record 12.4% asked for their money back

Redemption queues, not missed payments, are how open-ended lending funds actually break, and the queue is at a record.

Aggregate requests to withdraw money from private lending funds that are not traded on exchanges reached 12.4% of total fund value in the second quarter, a record, up from 10.4% in the first (Fund Finance Friday). Loans where the borrower has stopped paying interest at the twenty largest funds that lend to mid-sized private companies (known as BDCs) went to 2.8% of portfolios, up from 2.0% at end of March. Jefferies puts Blackstone Secured Lending at 3.6%, against 0.3% in the first quarter of 2025, a twelvefold rise in five quarters (Tribune India). The mechanism worth understanding is the redemption number, not the default number. These funds promise investors periodic windows to take their money out, capped at a share of total fund value. The cap is the product. When requests come in below the cap, everyone gets paid and nothing is visible. When requests exceed it, investors get a fraction of what they asked for, and the rest rolls into next quarter's queue, which raises next quarter's requests, because rational people who got 40% of what they asked for will ask for more next time. The reporting gives us requests, not what was actually paid out. We cannot say from this how much money actually left. Now the disagreement. Over the last twenty sessions Blue Owl's management company stock is up 19.1%, FS KKR up 15.9%, Blackstone Secured Lending up 7.1%, Ares Capital up 6.3%. Every listed name on our private lending board is within about 5% of its high. The stock market is pricing rising non-payments and record withdrawal requests as noise inside a fee stream. One of those two readings is wrong. We do not know which, and neither does the market. But the prices that would settle it are set by the managers whose shares are rallying.

Crypto and TradFi contagion30 Augmedium

The coin was flat. The wrappers fell seven percent.

A bitcoin treasury company's stock is a leveraged bet on other people's continued willingness to buy the leverage, which is a thing that can stop without the underlying asset moving at all.

On Friday, Strategy fell 7.35%, BitMine 7.18%, SharpLink 7.44%, Galaxy Digital 8.22% and Coinbase 6.43%. Over the same five sessions bitcoin moved 0.37%. Over twenty days the ratio runs the other way and tells you the same thing: bitcoin is up 23.4%, Strategy 36.5%, BitMine 37.7%, SharpLink 33.9%. These companies deliver roughly one and a half times the coin's move in both directions. That is not borrowed money in the usual sense. Most of these companies have not pledged their coins as security for loans. It is the premium. A company whose main business is holding bitcoin is worth its coins plus the value of its ability to keep issuing new shares above the value of those coins and buying more coins with the proceeds. The second part is what moves. When investors think the share-issuance machine keeps running, the stock trades at a multiple of the holdings and every new share adds to the coins held per share. When they stop believing it, the premium compresses toward one and the stock falls with the coin standing perfectly still. That is what Friday was. The part that matters for contagion is what does not compress. Convertible bonds issued during the premium years do not shrink when the premium does. They come due on their own schedule, at full face value, and the assets backing them are the coins. We do not have current premium-to-holdings figures for these names in this window, so treat the mechanism as the point, not the magnitude. Bitcoin itself sat at $78,858 on Saturday, described in Bloomberg's gold coverage as the metal's digital proxy, lifted by the same Treasury buyback trade.

The dollar, gold and reserve status30 Aughigh

Gold's best month since January, and a 3% Friday

Gold is now pricing the credibility of the Federal Reserve (America's central bank) against the Treasury Department (which borrows the money the government spends), not the inflation print.

Spot gold is up 10% in August, its biggest monthly climb since January, and buyers have moved into exotic options and complex bets rather than simple upside bets, a rally Bloomberg describes as "orderly" (Bloomberg). The stated driver is not inflation data or a war headline. It is Treasury Secretary Scott Bessent's plan to at least double Treasury purchases of outstanding government bonds that do not come due for ten to thirty years, which pressured the dollar and lifted both gold and bitcoin. Read the causal chain slowly, because it is the interesting part. The Treasury Department announces it will buy back more of its own long-dated debt to hold down the interest rate on long borrowing. Investors conclude that the government's tolerance for a higher long-term cost of borrowing is limited. If the price of long money is going to be managed rather than discovered by the market, the adjustment shows up somewhere else, in the currency. So they buy hard assets. Gold is not trading as a hedge against a bond market crisis here. It is trading as the expression of one. Then Friday: gold fell 3.26%, silver 4.37%, and the dollar index (which tracks the dollar against a basket of other currencies) rose 0.52%, after Federal Reserve chair Kevin Warsh pledged to fight inflation at Jackson Hole. Which is the test. If the Fed is genuinely willing to raise interest rates into a fiscal situation the Treasury is trying to cushion, the debasement trade is wrong and gold gives back a good deal of August. If Warsh backs down in September, it is right. Two institutions are pulling in opposite directions on the same interest rate, and the metal is keeping score.

Hidden leverage and shadow banking29 Aughigh

Principal-protected, ten percent, and frozen

The pattern showing up in Australian private lending funds has an offshore retail cousin, and it is the same promise that breaks: fixed returns from a strategy that cannot easily be cashed out.

Roughly HK$1bn ($128m) belonging to mainland Chinese and Hong Kong investors is frozen in accounts linked to ZD Group, a Hong Kong family office founded in 2018 (Caixin). Investors have formed groups to seek repayment and gone to the police. The SFC, Hong Kong's securities regulator, declined to comment. What was sold is worth spelling out. For a HK$1m minimum, retail clients could act as cornerstone or anchor investors in Hong Kong stock market listings. The product was principal-protected, paid a fixed return above 10%, and handed investors about 30% of any additional gains. Later tiers dropped principal protection, raised the minimum to HK$2m and offered up to 70% of trading gains. Some investors pooled money to reach the thresholds. They were introduced through social media. There is no strategy in public markets that pays a guaranteed 10% and keeps the upside optional. What that promise requires is a reliable supply of underpriced allocations, and the complaint filed in March alleges how it was meant to be produced: "IPO cornering," taking most of an offering to control how many shares are freely available and push the price. In July the SFC ordered Futu Securities to freeze up to HK$125m of assets over suspected false demand for IPO shares. The money itself moved through a British Virgin Islands entity, Renascence Global Capital SPC, having previously gone through a limited partnership whose controlling partner changed in 2025. The freeze has been confirmed to investors only verbally, by a law firm, with no written documentation. HK$1bn is not systemic. The shape is: offshore wrapper, changed sponsor, retail promise of a fixed high return, redemptions stopped, and nobody able to produce a document.

What would change our mind

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

A sustained decline in long-dated Treasury yields, indicating market confidence in fiscal sustainability.

Would move the number

A reversal in BDC non-accrual rates, showing improved asset quality in private credit.

Would move the number

Evidence that AI off-balance-sheet commitments are being brought onto corporate balance sheets or are being financed with equity.

Would move the number

A significant tightening of credit spreads (HYG, LQD) alongside rising equity markets.

Would move the number