Hedge funds hold $2tn of Treasuries, financed overnight

Archived reading, published Sun, 23 Aug 2026 14:21:54 UTC (3 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility89

how much tinder is stacked up — moves slowly

Ignition39

how close a spark is — moves fast

Crash Lab watches the machinery under the current boom: the debt paying for AI data centres, private credit, the leverage that has moved out of banks into places nobody has to mark, the bond market and the dollar.

Every four hours it reads the day’s reporting from 18 sources and rewrites this page.

Status: Held at 64. The market data has not changed for a sixth straight window — still Friday's closes, VIX 15.1, high-yield 275bp, investment-grade 82bp, the S&P 1.6% off its record — and bitcoin's 0.4% tick is the only fresh price, so ignition cannot honestly move. Everything genuinely new (the OFR's $2tn basis-trade figure, German pension writedowns booked over four years, Broadcom's re-reported financing) is better measurement of risk already standing; the one real addition, $7bn of retail money into a 3x chip fund, is far too small to move a system-level number.

Reporting from 20 Aug to 23 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 banking23 Augmedium

Two trillion dollars, rolled every morning

The long end has been selling off for a week; the largest single holder of the bonds being sold is a set of funds that finance them overnight and are sized by leverage rather than conviction.

The Office of Financial Research published a blog post on 19 August putting hedge funds' cash Treasury holdings at roughly $2 trillion, against about $1.4 trillion of short Treasury futures. Its own explanation of how the position is funded is worth reading slowly: because the profit on each unit of the cash-futures basis trade is tiny, "hedge funds scale positions by relying heavily on leverage through repo financing and futures margining", and they "typically rely on overnight repo because it is cheaper than term financing". So the mechanism is this. A fund buys a Treasury bond, sells the matching future, and pledges the bond in repo to borrow almost its full value back. The borrowed cash buys more bonds. The collateral is rehypothecated down a chain. The whole structure earns a few basis points and is rolled every single morning. It works until either repo rates spike or the futures exchange raises initial margin — at which point the fund does not choose whether to sell. Two supporting numbers matter. The July FOMC minutes, released the same day, described leverage vulnerabilities as "notable" and said hedge fund leverage was near all-time highs across all strategies and concentrated in the largest funds. And Fed work on Form PF data puts the basis trade itself at roughly $830bn and swap-spread arbitrage at about $305bn as of September 2025. The SEC's Treasury clearing rule took effect in May 2026 and was meant to put more of this through a central counterparty. Per a second OFR post, affiliate repo transactions are exempt where the affiliate clears its other eligible repos, and FICC does not clear open-term repo at all, so that stays outside too. The rule captures less than the headline suggests.

The AI capex bubble23 Aughigh

Seven billion dollars into a falling knife

Retail leverage is now a visible marginal buyer in the AI trade, and it is the kind of buyer that is mechanically forced to sell more as prices fall.

Investors put almost $7bn of net new money into the Direxion Daily Semiconductor Bull 3X Shares during July and the first two weeks of August, according to Morningstar data reported by the FT. Over roughly the same stretch, the NYSE Semiconductor Index it tracks fell more than 20% from its peak — and the fund fell 70% from its June high to its subsequent low. This is not a rounding error of retail enthusiasm. A leveraged ETF promises three times the daily return of an index, which is not the same as three times the return over any longer period. It achieves the daily target by rebalancing its derivative exposure every afternoon: buying more exposure after an up day, selling after a down day. In a market that chops, the product bleeds even if the index ends flat. In a market that falls 20%, it can lose 70%. The pattern repeated in single stocks. A CSOP fund offering double the daily return of SK Hynix took in more than $1bn over six weeks while falling 86% peak to trough. A leveraged Sandisk fund took $350mn while dropping 85%; the underlying stock fell 46% in July alone. Goldman's Ben Snider described the behaviour as "buying the dip and selling the rip". Worth noting where the pain is and is not. Nvidia is 4.7% off its recent high and up 3.7% over twenty days. The 20%-plus drawdown is in the broader chip complex — memory, Korea, the second tier. The index and the megacap have separated, and the leveraged money went into the index.

Crypto and TradFi contagion23 Augmedium

The treasury companies outran the coin again

When a bitcoin treasury company raises equity at a premium and keeps the cash, it is telling you something about its own leverage that the share price is not.

Bitcoin is at $77,408, up 19.7% in five days. The equities built on top of it moved further: Strategy +28.1%, Bitmine +26.3%, SharpLink +26.4%, Coinbase +25.7%, all four sitting exactly at their recent highs. Gold is at a record and the dollar index is down 2.6% over twenty days. Bloomberg's framing of the week — Bessent's buyback lifting long yields for barely a day while gold and bitcoin rallied and the dollar fell — has Nomura's Charlie McElligott calling it a "pressure-release valve". Two things cut against reading this as fresh money arriving. Spot bitcoin ETFs saw $389.7mn of net outflows in the week of 10 August after $853.5mn of inflows the week before, and global crypto ETPs shed about $382.5mn. The regulated wrapper was being sold while the coin rose. And Strategy, the original digital-asset treasury company, disclosed that it sold about $333.7mn of its own stock last week, bought and sold no bitcoin at all, and ended the week with a $4.8bn dollar reserve. That is worth sitting with. The mechanism of a treasury company is that a premium to the value of its coins lets it issue equity and buy more coins, raising coins-per-share. Issuing equity and holding the proceeds in dollars does not do that. It either means management is building a buffer for debt service and preferred dividends, or it means they think their own shares are dearer than the asset. Both readings are informative.

Household credit23 Augmedium

Two different numbers for the same delinquency

The deep-subprime tail is at or near record stress while the prime books of the listed lenders are merely soft — and the two are being reported as one story.

The subprime auto figure circulating this week is 6.9%, sourced to Fitch and Equifax, described as 60-day delinquency and above the 2008 peak. A second write-up of what appears to be the same Fitch index gives 6.13% for July 2026, described as the highest July reading in a series going back to 1994. Both cannot be the same measurement. One is likely 60-day, the other 60-plus days past due, or one is seasonally adjusted and the other not. The reporting in front of us does not resolve it, and readers should hold the alarming version loosely until the primary release is visible. The New York Fed numbers are firmer, because they come from one file. In Q2 2026, 5.49% of auto-loan balances were 90 or more days delinquent, the second-highest reading in a series back to 2003, and the flow of auto debt into 90-plus delinquency ran at 3.00%. Total household debt is $18.8tn with 4.7% in some stage of delinquency. The same caution applies to credit cards, where 12.8% and 6.97% are both being quoted for Q2. They measure different things: the share of balances sitting 90-plus days delinquent, and the flow of balances newly entering that bucket each quarter. Quoting them interchangeably makes the situation look twice as bad as it is. Against all this, Capital One reported a July net charge-off rate of 4.12% on domestic cards with 3.48% of balances 30-plus days delinquent — elevated, not distressed. The lenders came off last week: Sallie Mae -5.7%, Ally -4.4%, Capital One -3.8% over five days, though all remain up over twenty.

Private credit and BDCs22 Aughigh

German dentists, shrimp farms, half of €2.2bn

It shows what the reporting lag on unmarked private assets actually looks like: four years, and a number the fund managers themselves do not trust.

German professional pension schemes — the Versorgungswerke that doctors, lawyers, architects and dentists are required to pay into — are facing more than €2bn of losses and writedowns on property and private-market investments, on the FT's calculations from annual reports and disclosures. Funds in Hesse, Schleswig-Holstein, North Rhine-Westphalia and Lower Saxony have booked well over €800mn of impairments in four years. The individual cases are where the texture is. A pension fund for Berlin's dentists warned in December that investments in hotels, a US plastics recycling start-up and a shrimp-farming venture may have wiped out more than half of its €2.2bn in assets. Bayerische Versorgungskammer, the largest manager of professional pension money in Germany, said late last year it could lose half of a €1.6bn US property portfolio run through Deutsche Finance. The sector as a whole manages over €300bn. The plumbing point is not that private assets fell — property bought before 2022 and marked against rates that then rose is an old story. It is when the losses appeared. These are annual-report disclosures, arriving three and four years after the event, from institutions with no daily marks, no redeeming investors and, one fund manager told the FT, no consistent disclosure of unrealised positions at all: "We should be happy if the losses are only €2bn." That is the migration thesis in one paragraph. Risk left the banks, where it is marked and supervised, and settled in vehicles where the loss becomes visible when someone gets round to writing it down. The German funds say the losses are manageable and will not affect benefits. That may well be true. It is also, definitionally, unverifiable from outside.

The AI capex bubble22 Augmedium

Broadcom guarantees the debt that buys its chips

Vendor guarantees convert customer default risk into supplier contingent liabilities that sit outside the reported revenue and the reported debt.

Broadcom is reported to be seeking more than $60bn of debt — some accounts put the full package as high as $100bn — to finance AI chip purchases benefiting Anthropic and others. The structure described is roughly $60bn–$70bn of senior secured notes plus about $30bn of junior debt, with Broadcom guaranteeing part of it. This is not new risk this morning; it has been re-reported across several days, and we count it once. What makes it worth a dispatch is the pattern completing. Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500bn of third-party capital, saying it could provide backstops of up to $125bn. A vehicle called the AI XPV Platform, launched in June, has reportedly already provided Anthropic $35bn for data-centre construction. Eagle Point is putting roughly $1.3bn of private credit into an Anthropic-linked Texas site inside a $16bn project package. The mechanism is the same each time. The chip vendor books revenue now. The credit risk of the customer is transferred to third-party lenders. And the vendor writes a guarantee or backstop that brings a slice of that risk back — but as a contingent liability, not as a receivable, and generally not as a line on the income statement. What we cannot establish from the reporting: how much of the Broadcom guarantee is first-loss versus senior, whether it is disclosed as an off-balance-sheet commitment, and what triggers it. Those three answers determine whether this is ordinary trade credit or a leveraged bet by a chipmaker on the solvency of one lab.

Signs of the times

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

Berlin dentists, hotels and shrimp

A pension fund for Berlin's dentists warned in December that investments in hotels, a US plastics recycling start-up and a shrimp-farming venture may have wiped out more than half of its €2.2bn in assets. German professional pension schemes collectively manage over €300bn and disclose their private-market marks once a year.

private marketspensionsmarks

22 AugFinancial Times

$7bn into a fund down 70%

Investors put nearly $7bn of net new money into the Direxion Daily Semiconductor Bull 3X ETF during July and the first half of August, a fund that had fallen 70% from its June peak. A leveraged SK Hynix fund took in over $1bn while dropping 86%.

retail leveragechipsetfs

23 AugFinancial Times

Bitcoin company accumulates dollars

Strategy sold $333.7mn of its own stock last week, bought no bitcoin, and finished with a $4.8bn dollar reserve — a company whose entire premise is converting equity into coins, briefly running as a dollar fund.

treasury companiesequity issuance

22 AugBlockhead

A vehicle from June has lent $35bn

The AI XPV Platform, launched in June 2026, has reportedly already provided Anthropic with $35bn for data-centre construction. Two months old, and lending at a pace that would rank it among the largest project-finance operations in the country.

vendor financingspvsspeed

22 AugBloomberg

The rumour mill

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

Partly true

“Microsoft has changed the depreciation schedule on its offices and data centres from 15 years to 25 years.”

CFO Amy Hood said on 29 July that the change takes effect at the start of FY27 — it is prospective, not yet in the filed accounts. Extending useful life spreads the same capex over ten more years of reported earnings, which matters a great deal for how AI infrastructure spending shows up in profit.

Claimed by Meet Kevin

Confirmed

“PIMCO estimates capex could absorb about 94% of hyperscaler operating cash flow this year and next.”

PIMCO's May 2026 report puts it at roughly 93–94% for 2026–27, against 33–40% in 2022–23. That is the arithmetic behind the bond issuance: at 94%, the buyback and dividend money has to come from somewhere else.

Claimed by Mark Moss

Partly true

“The Treasury is bailing out the long end because buyers of US debt are vanishing.”

The buyback doubling to at least $4bn per operation is real and Treasury called it liquidity support. But Reuters reported demand was intact with no buyers' strike apparent, and the 30-year auction cleared — expensively, at 5.216% — rather than failing. Expensive is not the same as absent.

Claimed by Gregory Mannarino

Confirmed

“Data-centre securitisation issuance grew from about $2.4bn in 2020 to $15.5bn last year.”

Multiple outlets carry the same Bloomberg-compiled figures. A sixfold rise in five years in a market that packages data-centre lease cash flows into tradeable bonds — and whose collateral is buildings whose tenants' business models are four years old.

Claimed by Mark Moss

Earlier developments

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

Bond market dysfunction23 Aughigh

A 19-year high, reached without a panic

The long end has settled at a level that reprices every leveraged balance sheet in the system, and it got there calmly enough that nobody has been forced to do anything about it yet.

The $25bn 30-year Treasury auction last week cleared at 5.216% — the highest yield for that auction since 2001. The previous day's 10-year drew the highest financing cost since 2007. On Friday the 30-year traded around 5.34%, its highest since 2007, having briefly touched 5.1765% after the Treasury announced it would at least double its buybacks of 10- to 30-year paper to $4bn per operation from 9 September (Reuters, NPR). The striking part is the volatility. The MOVE index — the options market's price for rate turbulence — is down 18% in five days and 28% in twenty. TLT, the long-bond ETF, is unchanged over five days and down 1.1% over twenty. The long end has repriced to a two-decade high in yield without anything that looks like a disorderly market. That cuts two ways, and it is worth being clear about both. A quiet repricing means no forced sellers, no margin spiral, no dealer balance sheet buckling. It also means the move is not a spasm that reverses. If 5.3% is where the market now clears US duration, every borrower priced off it — the hyperscalers, the leveraged loan market, the mortgage — pays that indefinitely. The Washington Post's framing is the one to keep: debt past $40tn, and the Treasury competing for the same capital as the companies building data centres. Bessent's buybacks change the composition of what is outstanding, not the quantity of duration the market must hold. Four sessions on, the market has said so.

The dollar, gold and reserve status22 Augmedium

The adjustment went to the currency instead

Capping long yields does not remove fiscal pressure, it relocates it — and three separate hard-asset markets are now pricing the relocation at once.

Gold closed at $4,680, up 3.6% in a day and 6.9% in five, sitting at its high for the period. Silver is up 7.3% in five days, also at its high. Bitcoin is up 18.5% in five days to about $76,700. The dollar index fell to a three-month low against the euro, which touched $1.1711, its strongest since 14 May (Livemint/Reuters). The mechanism is simple enough to state. A government facing a bond market that demands 5.3% for thirty-year money has three ways to respond: pay it, borrow less, or lean on the price. Buybacks are leaning on the price. If the leaning works even partially, the pressure does not vanish — it moves to whatever is not being managed. That is the exchange rate. "Bessent's efforts to suppress U.S. yields haven't done much for U.S. yields, but it's undermined the dollar," Marc Chandler of Bannockburn told Reuters. "The market is pushing back." Nomura's Charlie McElligott called the gold-up, dollar-down, bitcoin-up combination a "pressure-release valve" (Business Times). Ray Dalio spent Friday telling investors to cut bonds and hold gold and some bitcoin. Proportion matters. The broad dollar index is down 1.35% over twenty days; the narrow DXY 2.6%. This is a drift, not a run, and a three-month low is not a crisis low. What makes it worth watching is the company it keeps: gold, silver and bitcoin all at period highs simultaneously is a portfolio decision about the unit of account, not a trade in any one of them.

Crypto and TradFi contagion23 Augmedium

A stablecoin issuer becomes its own custodian

A stablecoin whose issuer holds its own reserves has the same maturity and liquidity mismatch a bank has, minus the backstops a bank gets.

World Liberty Financial has received conditional OCC approval to form World Liberty Trust Company, a national trust bank. The charter would let it issue its USD1 stablecoin directly and hold the dollar assets backing it itself, rather than using BitGo for those functions (CoinDesk). Worth being precise about what a national trust bank is and is not. It is federally chartered and federally supervised, which is a real upgrade in oversight. It does not take deposits, so it has no FDIC insurance. And whether it can hold its reserves at the Federal Reserve itself — a master account — is not settled: that is exactly the question in Custodia Bank's long fight for an account, which has now reached the Supreme Court, with briefs from Senator Cynthia Lummis, former Senator Pat Toomey, the Blockchain Association and the Digital Chamber. So the plumbing looks like this. Reserves that were held by a third-party custodian move in-house. Unless a master account is granted, those reserves sit where every other private dollar sits: in Treasury bills, in repo, and in commercial bank accounts. A redemption wave against the token is therefore a forced sale of bills and a withdrawal from a bank, on the same day, at whatever price the market offers. That is the 2023 regional-bank problem with the token wrapper on top. The reporting does not give the composition or the maturity of USD1's reserves, and until it does, the size of that transmission is unknown. What is now clear is the direction of travel: stablecoin issuance is being pulled into the chartered banking perimeter without the deposit insurance or central bank access that makes the perimeter safe.

The AI capex bubble22 Augmedium

What exactly secures a GPU loan

The AI buildout is being financed against collateral that is worth a lot precisely when nobody needs to seize it, and guaranteed by the vendor selling the collateral.

Nvidia's financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are meant to mobilise more than $500bn of third-party capital, with Nvidia providing backstops of up to $125bn. We have flagged the headline number before. The detail that has since emerged is the structure: the debt sits in special purpose vehicles and is collateralised by Nvidia GPUs and offtake contracts (Bloomberg, MarketScale). Think about what that collateral package is. In a normal secured loan you want collateral whose value is uncorrelated with the borrower's fortunes — property, receivables from someone else. Here both legs are the same bet. If AI compute demand softens enough that the borrower cannot pay, the second-hand price of the chips falls and the offtake contract is worth whatever the counterparty is then good for. The security and the credit fail together. The backstop closes the loop. Nvidia is guaranteeing, up to $125bn, debt raised to buy Nvidia hardware. That is not risk transferred to third-party capital; it is Nvidia's own revenue risk taken back in a different legal form, sitting off the operating balance sheet in vehicles it does not consolidate. The filings available do not say how the backstop is accounted for, when it would be recognised, or what triggers it. The same shape appears at Broadcom, reported to be seeking over $60bn — possibly up to $100bn including roughly $30bn of junior debt — for chips destined for Anthropic, with Broadcom guaranteeing part of it. A separate vehicle, the AI XPV Platform launched in June, has already supplied Anthropic $35bn.

Household credit23 Augmedium

The delinquency pipeline is filling faster than it drains

The consumer credit book is deteriorating in level but not in rate of change, which is why lenders' shares are near highs and why the reckoning is being pushed into later quarters.

The New York Fed's Q2 2026 file shows 12.8% of credit card balances are 90 or more days delinquent, against 7.6% in Q3 2022, on $1.26tn of card debt (reported summary). Auto is worse than the headline suggests: 5.49% of auto balances are 90+ days late, just below the 5.60% record set in Q1, and Fitch/Equifax data put 60-day subprime auto delinquency at 6.9%, above the 2008 peak of around 5% (glitchwire summary of Fitch data). The interesting number is the one that is not moving. The flow into serious card delinquency — the share of balances newly transitioning past 90 days — was about 6.97%, against 6.93% a year earlier. Essentially flat. Meanwhile the stock of balances sitting in serious delinquency rose from 12.2% to about 12.9%. A flat inflow with a rising standing balance means the pipe is draining more slowly: bad loans are being resolved, cured or charged off at a lower rate than they arrive. That can be forbearance, longer collections timelines, or simply borrowers who go delinquent and stay there. Whatever the cause, it defers the moment of recognition rather than avoiding it. Capital One's July master-trust data is the freshest read: a 4.12% annualised net charge-off rate on domestic cards and 3.48% of balances 30+ days delinquent. That is elevated but not accelerating. The stock is at a level that would once have been called a consumer recession; the flow is not. Both can be true, and the second is what lenders trade on — Capital One is up 7.8% in twenty days.

Hidden leverage and shadow banking23 Augmedium

Who is buying the bonds nobody wants

The long end is being cleared by leverage that funds itself one day at a time, which is fine until the day it isn't.

The Office of Financial Research puts hedge funds' cash Treasury holdings at $2.0tn at the end of 2025, against $1.4tn of short Treasury futures — the two legs of the cash-futures basis trade. Because the spread captured is tiny, the OFR notes, "hedge funds scale positions by relying heavily on leverage through repo financing and futures margining," and they typically fund in overnight repo because term financing costs more, rehypothecating the same collateral through a chain of borrow-and-buy transactions (summary of OFR blog). Put that next to the auctions. The $25bn 30-year that cleared at a 25-year-high yield was absorbed, and part of what absorbs supply at the margin is exactly this: leveraged relative-value funds buying the cash bond and selling the future, financed overnight. The market clearing calmly and the market being financed with overnight money are the same fact seen from two sides. The OFR is explicit about the failure mode: "if repo rates spike or futures margin requirements increase abruptly, hedge funds may be forced to unwind positions rapidly," and "a disorderly unwinding of highly leveraged basis trades could amplify volatility in the cash and futures markets." This is March 2020's mechanism, at larger size. The same structural shift is visible in Europe. A BIS review published on 17 August notes that non-bank financial institutions now account for more than 40% of non-centrally-cleared euro-denominated cash-borrowing repo, up from less than 30% at the end of 2020. Nothing in funding markets is tight today. SOFR is 3.63% and has barely moved. But the buyer of last resort in the world's most important market is a borrower whose lender can change its mind overnight.

Hidden leverage and shadow banking21 Aughigh

The fire sale that never reached the tape

The largest AI-fund blow-up of this cycle was resolved by one buyer's balance sheet rather than by the market, which means the price at which those positions would actually clear under stress is still unknown.

Citadel has told investors it has sold more than 80% of the "aggregate risk" it took on from Situational Awareness, completing close to 100 block trades worth more than $4bn in three weeks (FT). The sequence is worth setting out, because it is the clearest picture we have of how this cycle absorbs a shock. Situational Awareness, run by 24-year-old OpenAI alumnus Leopold Aschenbrenner, gathered more than $20bn of assets in two years. July's tech sell-off left it unable to meet losses through individual asset sales. It sold the bulk of its stock book to Citadel in a deal assembled in 24 hours, at roughly a 10% discount. Its biggest holdings, including Sandisk and Micron, then recovered, and Citadel has been distributing them into a rising market. Two things follow, and they point in opposite directions. The good one: a concentrated, distressed $4bn-plus position has been broken up and passed to a wide set of buyers. That is fragility genuinely coming out of the system, not merely being re-measured. Listed prices were never forced to clear a liquidation, and now they do not have to. The uncomfortable one: the reason the position never hit the tape is that a single private balance sheet with extraordinary risk appetite was willing to warehouse it overnight at a 10% haircut. That is not a market mechanism. It is a bilateral trade, priced by whoever picks up the phone, and there is no obligation on anyone to do it next time. The system did not clear the risk. One firm agreed to hold it until the price came back.

Private credit and BDCs23 Augmedium

Software loans, marked down where nobody trades

This is the first sign in a listed private-credit portfolio that the AI boom is producing credit losers as well as winners, and it arrived as a valuation markdown rather than a default.

Palmer Square Capital BDC's portfolio fell about 3.6% in fair value over the second quarter, from $1.15bn to $1.11bn, even though its non-accruals stayed very low at 0.29% of fair value and 149bp at cost. Management attributed the drop to markdowns in software loans, as software and AI-related borrowers weakened towards quarter-end and, in their words, valuation transparency diminished because there were fewer transactions. That last clause is the whole private-credit question in one sentence. A private loan has no screen price. It is valued by reference to comparable trades, comparable spreads, and the manager's own model. When the comparables dry up, the input that moves the mark is judgement. Here a manager has told investors, on the record, that the observable inputs got thinner and the marks came down anyway. Most managers in that position mark down less, or not at all, and nothing in the accounting compels them to. The second point is the identity of the borrowers. Software companies are the classic private-credit asset: recurring revenue, high margins, no hard collateral, lent against a multiple of EBITDA. They are also the most exposed to AI displacing what they sell. Crash Lab's running argument is that risk has moved into places where losses do not have to be marked; this is a small, specific instance of that risk being marked, in the sector where the AI boom's losers are most likely to appear first. Context: Morningstar puts non-accruals at the top ten BDCs at 3.95% of debt at cost in Q2, up 20bp on the quarter, against the 2.8% median across the twenty largest listed BDCs we covered on Saturday.

Fed, Treasury and policy23 Aughigh

What a buyback actually does

Buybacks shift duration risk from investors to the government's own refinancing schedule, and the fact that they are being tried at all is the clearest measure of Washington's pain threshold on long yields.

The Treasury's surprise announcement on 19 August was that it will at least double the size of its liquidity-support buybacks of 10-to-30-year debt, from $2bn to at least $4bn per operation, running 9 September to 4 November. The 30-year fell to 5.1765%, then closed at 5.247% the next day and traded around 5.27% on Friday (Reuters, Washington Post). It is worth being precise about what a buyback is, because it is often described as if it were quantitative easing. It is not. The Treasury has no printing press. Money spent buying back long bonds has to be raised by issuing other debt, and in practice that means bills. So the operation does not reduce the government's borrowing; it shortens the maturity of it. Duration risk comes off private balance sheets today and reappears as rollover risk on the government's own, to be refinanced at whatever rates prevail. That is our reading of the mechanism, not something the Treasury has stated. The market has responded by selling the bonds anyway and selling the currency as well — the dollar hit a three-month low. Three times in three weeks, the FT notes, Bessent has effectively asked investors to stop selling his bonds (FT). The part to watch is institutional. St Louis Fed president Alberto Musalem said on 20 August that the Fed sets policy independently of debt management. Bessent said any rate decision is "completely separate" — while adding that the two institutions would work together if there were any change in the Fed's balance sheet. Nobody had asked about the balance sheet.

The AI capex bubble22 Augmedium

The Treasury and the hyperscalers are in the same queue

AI capex and the federal deficit are now competing for the same duration buyers, which makes the long end of the Treasury curve the binding constraint on the buildout.

Companies have borrowed more than $410bn for data centres and other AI investment this year, according to Bloomberg's own tally. Hyperscaler issuance alone was $220bn by 10 August on BNP Paribas data. Net Treasury issuance is running at roughly $2tn across 2026 and 2027, with the national debt past $40tn. These are not separate stories. They are bids for the same pool of long-duration money. When the Washington Post says the Treasury "competes with" the hyperscalers for investors, that is the mechanism by which an AI capex cycle becomes a bond market problem: every incremental campus financed with 20-year paper raises the yield at which the government can sell 20-year paper, and vice versa. Barclays strategists make the same point about corporate issuance adding to pressure on long rates (Business Times). BofA has drawn the loop closed. Its note argues the administration is now focused on capping long-term borrowing costs partly because higher costs would undermine the AI investment boom — and warns that if the 30-year cannot be pulled below 5%, the result is a weaker dollar and deleveraging in exactly the leveraged assets that boom rests on: AI stocks, private credit, cyclical financials (ANI). That is a single-source, tier-three summary of a sell-side note, and should be read as one house's framework rather than fact. The framework is still useful, because it names the transmission channel: the AI trade's biggest external dependency is not chip supply or power. It is the price of thirty-year money.

Bond market dysfunction23 Augmedium

The clearing mandate has two doors left open

The main post-2020 regulatory repair to the Treasury market leaves affiliate and open-term repo outside central clearing, so the most leveraged corner of the world's most important bond market remains partly unmeasured.

The SEC's Treasury central-clearing rule took effect in May 2026. It was the regulatory answer to the basis trade — the hedge fund strategy of owning cash Treasuries against short futures, financed in repo, which the Office of Financial Research now sizes at $2.0tn of cash holdings against $1.4tn of short futures as of end-2025. Two exemptions are worth knowing about, both flagged in an OFR post this week. First, affiliate repo transactions do not have to be centrally cleared, provided the affiliated counterparty centrally clears its other eligible Treasury repos. Second, FICC does not offer central clearing for open-term repo — repo with no fixed maturity, rolled until someone stops it — so those trades remain outside the mandate entirely. Why this matters mechanically: central clearing does two things. It standardises margin, so a leveraged position cannot be financed at a haircut some dealer decided was fine, and it puts a clearing house between counterparties so that one failure does not cascade. Trades routed through affiliates, or booked open-term, get neither. Whatever share of the $2tn sits in those channels is financed on bilateral terms nobody publishes. We want to be careful here. These are not loopholes discovered this week; they are design features written into the rule, and the OFR is describing its own regulatory architecture, not sounding an alarm. Nothing has broken. But the reason the mandate exists is that in March 2020 the basis trade unwound violently and the Treasury market stopped functioning. A fix with two exempt channels is a partial fix, and the size of what flows through them is not disclosed.

Crypto and TradFi contagion22 Augmedium

The coin rose 17%. Its holding company rose 28%.

Treasury companies rising faster than the coin they hold means premiums to net asset value are re-expanding, which is what funds their buying and what disappears first when sentiment turns.

Bitcoin is at $75,906, up 17.4% in five days after touching $79,455 on Friday. Strategy — the original digital asset treasury company, whose business is holding bitcoin — is up 28.1% over the same five days and closed at its period high. SharpLink is up 26.4%, BitMine up 26.3%, Coinbase up 25.7%, all three also at their period highs. The gap between 17% and 28% is the part that carries information. A company that holds a fixed quantity of an asset should, mechanically, move with that asset, geared by whatever leverage sits on its balance sheet. Strategy disclosed in an SEC filing that last week it sold about $333.7m of its own stock, bought and sold no bitcoin, and finished with a $4.8bn dollar reserve (Blockhead). So bitcoin per share did not rise. It fell fractionally, because shares were issued and no coins were added. Which means the outperformance is not asset appreciation. It is the premium to net asset value expanding again — investors paying more per underlying coin than they did a week ago. That premium is not decoration. It is the machine: issuing stock above NAV is how these companies buy more coin without borrowing, and it works in exactly one direction. When the premium compresses, issuance stops, and the marginal buyer of the asset disappears. The wider context is the debasement trade — Nomura's Charlie McElligott called the gold-up, dollar-down, bitcoin-up combination a "pressure-release valve" as Washington tries to hold long rates down. Ray Dalio told investors on Friday to cut bonds and hold gold and some bitcoin.

Household credit22 Augmedium

The delinquency data and the share prices disagree

Consumer credit stress is now at levels the lenders' share prices are explicitly not discounting, which means the loss is being absorbed somewhere less visible than a bank income statement.

The New York Fed's Q2 numbers are not ambiguous. Credit-card balances rose $21bn to $1.26tn, and the share of card balances 90 or more days delinquent reached 12.8%, against 7.6% in Q2 2022. The flow of auto-loan balances into 90-day delinquency hit 3.00%, the highest since 2010 and above the 2.34% recorded in Q3 2007. Mortgage 30-day delinquency is the highest since 2015. Capital One's own July disclosure put domestic card charge-offs at a 4.12% annualised rate with 3.48% of balances 30 days past due (sweep of NY Fed and issuer data). On subprime auto there are two figures in circulation from the same Fitch data and we cannot reconcile them: one report gives a 60-day delinquency rate of 6.9%, above the 2008 peak; another gives 6.13% for July, described as a record for any July since 1994 (glitchwire). Both are bad. Only one is historic. Now the prices. Capital One is up 7.8% in twenty days and sits 3.8% off its high. Synchrony is up 9.5%, Credit Acceptance — deep subprime auto — up 4.6% and 2.4% in the last five days alone. Oportun is up 33% in twenty days. There is a coherent story on the equity side: charge-offs are already reserved for, lenders reprice new business upward, and the losses on the worst paper have been sold into securitisations held by somebody else. That is the mechanism worth holding onto. The delinquency is real; the question is only who holds the tranche where it lands, and the answer is increasingly not the listed lender whose shares you are watching.

Hidden leverage and shadow banking22 Augmedium

Two trillion dollars of Treasuries, bought on repo

The long end is where policy is currently straining, and the long end is financed by an unregulated $2tn position that unwinds by selling exactly what is already falling.

The Office of Financial Research puts hedge funds' cash Treasury holdings at $2.0tn at the end of 2025, against $1.4tn of short Treasury futures positions (via market summary). That pairing is the basis trade, and it is worth explaining because it is the largest leveraged position in the financial system that nobody votes on. A Treasury future trades slightly rich to the bond it settles against, because asset managers who want duration find it cheaper to buy the future than the bond. A hedge fund harvests that gap: buy the actual bond, sell the future, wait for the two prices to converge. The gap is a few basis points, so the only way to make it a business is leverage. The fund pledges the bond in the repo market, borrows almost its full value, buys another bond, repeats. Fifty to one is not unusual. Nothing about that is fraudulent and it genuinely tightens the market most of the time. The fragility is in the two things that can move against it at once: a repo lender raising haircuts, and the futures exchange raising margin. Both tend to happen on the same day, and the only way to meet both is to sell the cash Treasury — into a market that is already why the margin went up. The BIS adds a second data point: non-banks now account for more than 40% of non-centrally-cleared euro repo borrowing, up from under 30% at the end of 2020. This is our thesis in one sentence. The financing that supports the world's safest asset has moved to institutions with no lender of last resort.

The dollar, gold and reserve status22 Aughigh

The yield they suppress comes out of the currency

Yield suppression does not remove fiscal risk from the system; it relocates it to the currency, where the US has fewer tools and more foreign holders watching.

Three interventions in three weeks. The Treasury bought yen to support the Japanese currency. It said future yen operations would use an unusual facility chosen specifically to avoid selling Treasury bonds. Then, unscheduled, it announced it would double the pace of long-dated buybacks to at least $4bn per operation from 9 September (FT). The FT's reading of the yen purchase is worth stating plainly: bond market participants do not think it was solidarity. They think it was an attempt to stop Japan solving its own currency problem by selling its very large stock of US debt. Whether or not that is right, the choice of facility tells you what the Treasury is optimising for. The result so far: the 30-year yield fell for about a day and was back at 5.27% on Friday, having touched 5.33% earlier in the week — the highest since 2007. The dollar, meanwhile, fell to a three-month low against the euro, which briefly reached $1.1711 (Reuters via Livemint). "Bessent's efforts to suppress U.S. yields haven't done much for U.S. yields, but it's undermined the dollar," said Marc Chandler of Bannockburn. That is the mechanism. Fiscal stress has to be expressed somewhere. If the authorities lean against the bond price, it moves into the exchange rate and into the assets people buy when they do not want the exchange rate. The dollar index is down 2.6% in twenty days, gold is up 15% to a three-month high, and bitcoin is up 19% in five days — while still down about 17% on the year. This is a hedge against the currency, not a bid for growth.

The AI capex bubble22 Augmedium

What a GPU is worth when nobody wants it

The AI buildout is increasingly financed against collateral whose value is set by the same company whose products the debt is buying.

The structure of Nvidia's $500bn financing effort has become clearer, and the collateral is the interesting part. The platforms Nvidia announced with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR place the borrowing in special purpose vehicles, secured on Nvidia GPUs and offtake contracts, with Nvidia itself offering backstops of up to $125bn (sweep of Bloomberg and Reuters reporting). We have not seen the documents; this description comes from secondary reporting and the specifics of any backstop matter enormously. Take it at face value and the chain has three links. The lender is secured on chips. The chips are worth what a functioning secondary market says they are worth, which nobody has tested at scale in a downturn — and the depreciation schedule is set by Nvidia's own product cycle. The second link is the offtake contract, which is only as good as the AI company signing it, several of which are not yet cash-generative. The third link is Nvidia's backstop, which is the vendor guaranteeing the debt raised to buy the vendor's product. Each link is fine on its own. What we cannot see is whether they are independent. In a scenario where GPU resale values fall, it is likely because demand has fallen, which is also when offtake contracts get renegotiated, which is also when the backstop is called. Correlated collateral is how structured credit usually breaks. More than $410bn has now been borrowed for data centres and AI this year, on Bloomberg's count. The money is real. The recovery value in default is a modelling assumption.

Private credit and BDCs22 Aughigh

The insurer's balance sheet is selling a basketball team

A forced seller has appeared, and the assets being sold are sports teams rather than loans — which tells you which side of the balance sheet the marks are still holding.

Mark Walter has agreed to sell the Los Angeles Lakers and put his stake in Chelsea Football Club up for sale, in order to clean up the balance sheets of the insurance arms of his group. Federal prosecutors are investigating, and the probe has established that two insurers he controlled provided more than $20bn of previously undisclosed financing to other parts of his empire (FT). The reason to care is not the trophy assets. It is the model. An insurance company holds reserves against policies it has written — retirement annuities, mostly. Those reserves have to be invested, and for a decade the higher-return answer has been private credit. Guggenheim, with $367bn under management, was an early and large practitioner. KKR, Apollo and Brookfield all own insurers and all lend into their own orbit. Affiliated lending is where the wiring gets uncomfortable, because the loan's value is a mark, and the mark is set by a party with an interest in it. "What makes Mark Walter unique is that he failed to disclose a tremendous amount of affiliated transactions, which have the potential to loot the insurer," the University of Texas law professor Andrew Granato told the FT. "But other private capital firms are engaged in essentially the same business model." The market is distinguishing carefully. The listed BDCs are steady — Ares Capital up 0.4% on the week, Blackstone Secured Lending up 0.3%. The managers with insurance balance sheets are the ones that fell: Apollo down 5.3% in five days, KKR down 4.9%. Investors are marking the structure, not the loans.

Crypto and TradFi contagion22 Augmedium

Strategy sold stock and bought no bitcoin

When a treasury company issues stock to hold cash rather than coins, the flywheel that justified its premium is running in reverse even as the premium expands.

In its latest filing week Strategy sold about $333.7m of its own shares, purchased and sold no bitcoin, and finished with a dollar reserve of $4.8bn (Blockhead). The original digital-asset treasury machine worked like this: the shares trade above the value of the coins the company holds, so it issues shares, buys coins with the proceeds, and the coins-per-share number goes up. Shareholders get more bitcoin exposure than they paid for; the premium justifies itself. The flywheel only turns while the premium exists. Selling equity to accumulate dollars is a different machine. It is prudent — preferred dividends and interest are payable in currency, not in bitcoin, and a $4.8bn buffer means no forced coin sale in a drawdown. But it converts the vehicle from an accumulation story into a financing entity managing its liabilities, and it does so at exactly the moment the shares have re-rated hard. The price action is the part that does not fit. In five days MSTR is up 28%, ether treasury companies BitMine and SharpLink are up 26% each, and Coinbase is up 26% to a period high — against bitcoin up 19% and still down roughly 17% on the year, by BofA's count. The treasury companies are outrunning the coin again, which is how the premium rebuilds. Meanwhile the regulated wrapper went the other way: US spot bitcoin ETFs saw $5.4bn of net outflows in the first half, the worst half-year on record, with $4.5bn of that in June alone. Leverage-adjacent equity is bid; the plain fund is being redeemed.

What would change our mind

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

SOFR printing materially above the interest on reserves for more than a day, or CME raising initial margin on Treasury futures — either would force the $2tn basis book to shrink into a falling market.

Would move the number

High-yield OAS above roughly 350bp from today's 275bp, or a data-centre SPV bond deal being pulled or repriced sharply wide of guidance.

Would move the number

Q3 BDC filings showing median non-accruals falling back toward 2.0%, or a large non-traded BDC meeting a full redemption quarter without gating — both would be genuine evidence the private-credit mark is holding.

Would move the number

A named vendor guarantee — Nvidia's $125bn backstop or Broadcom's — appearing as a quantified contingent liability in a 10-Q, which would turn an estimate into a disclosure and could move fragility either way.

Would move the number