Private credit is now lending against Nvidia chips

Archived reading, published Tue, 01 Sep 2026 02:19:47 UTC (8 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

Ignition35

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 13 sources and rewrites this page.

Why it moved: Ignition down two to 35: markets have not moved since our last run (same August 31 close, the VIX, the market's gauge of how much turbulence traders expect over the next month, at 14.92, and the extra interest risky borrowers pay over the government sitting at 260 basis points and 8.8% tighter over twenty days) and this window is largely the same AI-debt and fund-lending numbers reaching us from new sources, with no funding channel closing. Fragility holds at 92: the one genuinely new commitment we can see (a $2.4bn loan to Iren to buy Nvidia processors) is real but small against a stack we already carry, and nothing was unwound.

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.

Private credit and BDCs31 Augmedium

A $2.4bn loan secured on chips

Private credit is moving from lending against a company's earnings to lending against hardware that loses value every year, and nobody has a market price for the collateral.

Iren, a bitcoin miner that rebuilt itself as an AI data-center operator, has raised $2.4bn of debt to buy Nvidia's latest processors for a campus in Canada. The package splits evenly: a $1.2bn senior loan and $1.2bn of senior notes, both secured against the company's assets. One account of the deal says Blue Owl, one of the largest private lenders, led it. This is the merger of our two most-watched stories. Private credit's pitch has always been that it lends against businesses with cash flows, at interest rates that float with the market, on promises it negotiates itself. Lending against a fleet of processors is a different exercise. The collateral is hardware whose resale value depends entirely on how quickly the next generation arrives and on whether the customer paying for the computing power is still there in three years. There is no observable secondhand price for a used rack of these chips. The lender's valuation will be a model, not a real market price. The reporting does not give the percentage of the chips' value the lender is willing to advance, the length of the loan, or what exactly sits inside the security package: chips, the campus, the customer contracts, or all three. Nebius has been doing the adjacent thing openly, pledging infrastructure and customer contracts as security, and in July borrowed $775m against existing processors and the cash flows from one highly rated customer. Meanwhile the same week produced Jefferies' count of loans where the lender has stopped counting the interest because it stopped arriving, at the big listed funds that borrow money, lend it to mid-sized private companies, and pass the interest to shareholders. Those non-paying loans are at their highest since at least 2021: Blackstone Secured Lending at 3.6% against 0.3% a year and a bit ago, Golub 2.9%, Blue Owl 2.8%, Ares 2.4%. The stock market does not care. Blue Owl's management company is up 11.7% over twenty days, its lending fund up 4.9%, FS KKR up 11.8%. Rising bad loans and a rising share price can both be right, but only if you think the new loans are better than the old ones.

Hidden leverage and shadow banking31 Augmedium

The basis trade is $830bn and less of it runs through a central clearinghouse

The largest bet built on borrowed money in the US government bond market is growing while the share of it financed through a central clearinghouse is shrinking.

Analyst estimates circulating this week put the cash-futures basis trade in US government bonds at about $830bn, close to double its early-2020 peak, with a further $305bn in related bets, financed by roughly $3tn of cash borrowed overnight against bonds put up as security, more than twice the level of early 2023. Hedge funds' total government-bond exposure is put at $4tn ($2.4tn of bets that prices will rise, $1.6tn that they will fall), around 8.5% of all outstanding US government bonds, with about 90% of the borrowing sitting in the fifty largest funds. The trade itself is simple. Government bond futures usually trade a little above the price of the underlying bond because leveraged buyers want exposure without funding it. A hedge fund buys the bond, sells the future, and collects the difference. The gap is a few hundredths of a percentage point, so the trade only pays at enormous size, which means enormous overnight borrowing against the bonds. It works until the lender demands more security for that borrowing, at which point the fund must post cash it does not have and sells the bond, into a market that is already falling. The number to watch is not the size but the plumbing. One market note says overnight borrowing that runs through a central clearinghouse (where margin is set by rule) fell from 71% of the market at end-2025 to 62% in the first quarter of 2026. The rest is arranged privately between two firms, with the lender setting the terms by judgment, at the exact moment judgment is worst. A rising private share means more of this borrowed money is financed on terms nobody outside the two firms can see. All of this is analyst commentary, not official data: no release from the Bank for International Settlements (the central bankers' own bank), the IMF (which lends to countries in trouble), or the Federal Reserve (America's central bank) in the window. What is official: the SEC, the agency that polices American markets, has subpoenaed banks over the near-collapsed fund Situational Awareness, seeking its communications with lenders about the borrowed money it was taking on.

The AI capex bubble31 Aughigh

The FSB chair writes to the G20 about Nvidia

The body that coordinates global financial regulation has now named AI-related borrowing as a risk to the whole system, which tells you it is large and tells you nobody has a tool for it.

Andrew Bailey, in his twice-yearly letter to G20 finance ministers written as chair of the Financial Stability Board (the body that coordinates global financial regulation), says the debt used to fund the AI build-out "could amplify a future market correction." The framing is the important part: "The issue is not simply that investors are borrowing more but that leverage is interacting with high valuations and market concentration." The Telegraph's accompanying numbers give it scale. Nvidia is worth more than $5.2tn, its shares are up 850% in five years, and it recently raised $500bn from a group of US banks and investors to fund its AI investment program. Set that beside CNBC's reporting that some of the largest cloud companies have pushed capital spending above 100% of operating cash flow, meaning the next data center is not funded by the business, it is funded by somebody's balance sheet (the list of what they own and owe), and CoreWeave finishing the second quarter with around $35bn of debt. What has changed here is what officials are willing to say, not what is standing in the system. A letter to the G20 does not alter a single capital requirement, a single lease, or a single promise in a loan contract. The Financial Stability Board has no ability to make a bank hold more capital against a loan secured on processors. Bailey is describing risk that was already built, in the same week that JPMorgan was shopping $5bn of debt for a company valued at $2.4bn three weeks earlier and Nebius closed $5.75bn of convertible bonds. One item we cannot confirm and flag as such: a single report says a Broadcom-backed vehicle was in talks on August 20 to raise $70bn to $80bn for a chip-financing arrangement supporting AI companies including Anthropic, with a total possibly near $100bn. One outlet. Treat it as unconfirmed until it is not.

Fed, Treasury and policy31 Aughigh

Treasury may buy its own long bonds with its own cash

The government is becoming a buyer of its own long-dated debt, funded by its cash balance, which changes both the structure of the debt and how honestly the interest rate reflects what investors think.

From September 9 to November 4, the Treasury Department (which borrows the money the government spends) will double the maximum size of its buybacks of government IOUs that do not come due for ten to thirty years, from $2bn to at least $4bn per operation. Scott Bessent said the market for thirty-year bonds was "especially poor" and called the program a kind of "Treasury twist." The announcement followed the interest rate the government pays to borrow for thirty years touching about 5.34%, a nineteen-year high; that rate fell within minutes and had round-tripped by the following afternoon. The mechanism deserves attention. A buyback is the Treasury purchasing its own outstanding long-dated bonds in the secondary market. Someone has to pay for them. Officials have signaled that the Treasury General Account, the government's cash balance at the Federal Reserve (America's central bank, which sets interest rates), about $940bn on August 19 and near $1tn, could be used. Run down the cash pile, or replace the retired long bonds with short-term bills, and you have shortened the length of debt the public holds. That is precisely what Operation Twist did in 2011. The difference is that this is the borrower doing it, not the central bank, and the borrower has an obvious interest in the price of its own debt. Two consequences. First, financing with short-term bills makes the $40tn debt stock more sensitive to short-term interest rates, awkward when futures are pricing better-than-even odds of a rate increase on September 16. Second, once there is a discretionary official buyer in the long end of the market, the thirty-year interest rate becomes a slightly less honest reading of what investors think of the country's fiscal path. What this is not: dysfunction. No failed auctions, no dealer balance-sheet seizure, no forced selling. This is price stress and a policy response to it, which is a different, and earlier, stage.

Crypto and TradFi contagion31 Aughigh

Strategy sold $603m of stock to buy $370m of bitcoin

The treasury-company model feeds on itself in both directions, and it has just restarted on the upswing.

Strategy has resumed buying bitcoin after a ten-week pause, purchasing $369.7m of the tokens in the week ended August 30. To fund it, the company raised $602.8m through the sale of common stock, and used part of the difference to buy back $151.8m of its own preferred shares. The second half of that sentence is the interesting one. The digital-asset treasury model only works while the shares trade above the value of the coins the company holds. If they do, issuing stock and buying coin adds coin per share, and the loop feeds itself. If they do not, issuing stock destroys value for existing holders and the loop stops, which is roughly what the ten-week pause looked like, after what Bloomberg describes as a crypto downturn that pressured the shares, weakened confidence in the financing model and prompted a balance-sheet overhaul. Buying back preferred stock with common-equity proceeds is a move to reduce fixed obligations in the middle of a move to add borrowed exposure. Preferred shares sit ahead of common and carry a fixed cost that has to be paid whatever bitcoin does; buying them back shrinks that fixed claim. That the company chose to spend a quarter of its raise doing that, rather than buying more coin, is a statement about what that preferred was costing. The prices say the loop is running again. Over twenty days Strategy is up 40%, BitMine up 46%, SharpLink up 39%, Coinbase up 28%, against bitcoin up 23% and, at $78,289, down slightly on the day. The company wrappers are outrunning the asset itself by fifteen to twenty points. That is what borrowed exposure to a rising coin looks like. It is also what it looks like on the way back down.

Household credit31 Aughigh

The average credit score is 714. The bottom band is falling

Consumer stress is concentrating in the lowest credit-score band while the average looks stable, which is exactly how the average stops looking stable.

The average American FICO score (the number lenders use to decide whether to lend to you and at what rate) is unchanged at 714. Underneath it, about 3.2 million consumers with a student loan payment due were reported behind on payments, and that group lost an average of 38 points year on year. Consistent payers gained 6 points. Borrowers with no recent missed payments gained 16. A 38-point drop is not a statistic, it is a repricing. It moves a borrower across the boundaries that lenders use mechanically: the car-loan rate, the credit-card limit, the mortgage overlay, the insurance quote in states that use scores. Someone who missed student loan payments in 2026 pays more for everything else they borrow in 2027, which makes the next missed payment more likely. FICO's own follow-up data shows the pattern: subsequent severe delinquency on mortgages and car loans rose only in the lowest score bands and was flat across higher ranges. The rest of the household picture is a clean split. Credit cards are healing: bank delinquency of thirty or more days at 2.85%, down from 3.04% a year earlier. Car loans are not. Fitch (one of the three firms that grade how likely a borrower is to pay you back) has subprime car-loan delinquency at 6.13% in July against 5.80% at midyear, prime at 0.49%; Equifax, summarizing Moody's, calls subprime car-loan delinquency an all-time high; one in eighteen car-loan balances was three or more months late in the second quarter (5.49%, per the New York Fed). The market disagrees, and specifically. Credit Acceptance, the deep-subprime car lender, is up 5.1% over twenty days and sits 0.4% below its recent high. Its borrowers are at a record delinquency rate. Both can be true: deep-subprime lenders price for losses and often earn more when competitors retreat. But the stock is not carrying any doubt at all.

Signs of the times

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

The liquidator invested in the lender

Hall Chadwick managing partner Richard Albarran, appointed over the assets of dozens of defaulted borrowers, disclosed to ASIC (Australia's corporate regulator) that he is a personal investor in Blackbird Capital Group, the lender that appointed him, and that his firm vets Blackbird's loans. Blackbird has referred him 50 matters in 24 months, ten times what he had previously disclosed.

private creditconflictsaustralia

30 AugAustralian Financial Review

A $5.2tn company raised $500bn

Nvidia, worth more than $5.2tn and up 850% in five years, recently raised $500bn from a group of US banks and investors to fund its AI investment program. That is a single financing round larger than the annual GDP of Belgium.

ai capexvendor financing

31 AugThe Telegraph

You can now mortgage a house with bitcoin

Coinbase and Better launched bitcoin-backed mortgages after a waitlist projected more than $260m of demand. A volatile asset is now being used as security for the least liquid asset most households will ever own.

cryptohousingcollateral

31 AugBitcoin.com weekly recap

Five billion dollars, one week, one issuer

Circle minted roughly 5 billion USDC (a digital token meant to be worth exactly one dollar) in a single week, one of its largest pushes. That supply is now expanding at a rate that would make it a mid-sized money market fund complex every seven days.

stablecoinsliquidity

31 AugBitcoin.com weekly recap

The rumour mill

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

Partly true

“The government just quietly revised away nearly a million jobs that never existed.”

Two different revisions are being mixed up. The roughly 900,000 downward revision was for the year to March 2025, published in September 2025 and finalized at -862,000 in February 2026. The revision released on August 28, 2026, for the year to March 2026, was -79,000 total: private payrolls down 178,000, government up 99,000. Both were announced on a published schedule, not quietly.

Claimed by Gregory Mannarino

Confirmed

“The KOSPI went parabolic and then fell more than 20% in roughly a month.”

The index more than doubled in 2026, peaked at 9,114.55 on June 22 and fell around 22% in July alone, its steepest monthly decline since 2008, for a peak-to-trough drop of roughly 38 to 44%. Worth keeping in view as the nearest live example of what a concentrated, retail-driven run-up does on the way down.

Claimed by George Gammon

Confirmed

“227,548 US properties had foreclosure filings in the first half of 2026, up 21%, with activity rising in 77% of metros.”

ATTOM's mid-year report gives exactly those figures, and 28% above the first half of 2024. Still low by historical standards, but it is the direction that matters against a household picture where stress is concentrating in the lowest score bands.

Claimed by GoldSilver (Mike Maloney)

Partly true

“Futures now price roughly two-in-three odds of a quarter-point hike on 16 September.”

The direction is right, the number is high. After the Jackson Hole speech, Reuters had about 60%, CNBC 56%, MarketWatch 57.4% and Forbes 58%, against roughly 36 to 40% before the speech. Better than even, not two in three.

Claimed by Kitco NEWS

Earlier developments

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

Fed, Treasury and policy31 Aughigh

The two ends of the curve now have different owners

A rate rise now transmits fastest to exactly the floating-rate private-credit borrowers whose non-accruals are already at post-2021 highs.

Two weeks before the 16 September FOMC, the market has flipped. Before Kevin Warsh spoke at Jackson Hole on 28 August, futures priced roughly a 36% chance of a quarter-point hike; afterwards, between 56% and 60% depending on which contract you read. MarketWatch put it at 57.4% against 35.9%; Forbes at 58% against 36%. Warsh's line, as reported, was that "two percent is a firm and fixed target" and that "short-term interest rates are the primary tool". This was already the internal direction of travel. Three FOMC members dissented in favour of a hike in July, when the rate was held at 3.50–3.75% for a fifth straight meeting, and four of the twelve Reserve Banks voted to raise the discount rate. Now put that next to the other arm of the state. From 9 September through 4 November, the Treasury will buy back at least $4bn of 10–30 year debt per operation, double the previous cap. Bessent called long-bond liquidity "especially poor" and the programme a "Treasury twist" (Guardian). So the Fed may push the front end up while the Treasury leans on the long end from below. The 2-year is already at 4.34%, up 7bp on the week; 2s10s is 39bp. A flattening curve driven by policy in both directions is not an academic problem for anyone funding floating-rate assets. Every direct loan in a BDC portfolio is priced off SOFR, currently 3.65%. Borrowers already generating 3.95% non-accruals at the ten largest BDCs would pay 25bp more, immediately.

The AI capex bubble30 Augmedium

The chips arrive. The transformers do not.

Power assets falling while everyone insists power is the bottleneck is the clearest price-versus-narrative disagreement in the AI complex right now.

Elon Musk posted on 29 August that the "consensus estimate is that ~15GW of AI compute produced in 2027 cannot be turned on in 2027", and that the binding constraint is not electricity as such but "transformers, wiring, liquid-cooling, (massive) chillers & complex networking". The Information separately reported that developers working with Google, Microsoft and Oracle fear chips with nowhere to plug in. Both are claims rather than audited facts, and we have marked them as such. The scale is not in dispute. At Piketon, Ohio, OpenAI intends to lease eight gigawatts from SoftBank, which plans to spend up to $500bn and build more than nine gigawatts of new gas generation in the region (n-tv). Here is the part that does not fit. If power and power equipment are the scarce input, the companies that own generation should be repricing upward. They are doing the opposite. Talen Energy is down 14.1% over twenty days and sits 22% below its recent high. Vistra is down 11.9% and 18.7% off. Applied Digital, which builds on project debt, is down 13.8%; Core Scientific is down 28.5% and a third off its high. Two readings are available. Either the equity market has decided the scarcity rent is already in the PPAs and the upside accrues to the tenant, not the landlord — or it has quietly stopped believing that all the announced load will actually show up. The second reading is the one that matters for anyone lending against a twenty-year lease.

Private credit and BDCs31 Augmedium

A loan marked at 63 cents that is still performing

A loan that pays interest in more loan can stay 'performing' well past the point where the mark says the lender is impaired.

Two names to hold on to: Cornerstone OnDemand and Symplr. Both are marked in the 63–65 cent range in BDC portfolios. Neither is on non-accrual. That is not necessarily an abuse. A loan goes on non-accrual when the lender stops expecting to collect interest. A loan can be marked at 63 because the market for that credit is thin and the exit is uncertain, while the borrower still pays. But there is a mechanism that makes the distinction slippery, and it is growing. Payment-in-kind lets a borrower settle interest by issuing more debt rather than sending cash. The share of BDC loans using PIK has risen from about 6% in early 2022 to roughly 10% by early 2026, according to Blue Vault; a separate count puts PIK loans at about 16% of total investments at listed BDCs, with five funds holding three-quarters of it. The two figures do not reconcile and we cannot reconcile them — they are almost certainly measuring different things (loans with any PIK feature versus PIK-only), and nobody publishes a clean series. Apollo says PIK yields in listed BDCs have started to rise, driven by software-sector stress. That is the same sector as Cornerstone and Symplr. Meanwhile the arithmetic underneath keeps moving: non-accruals at the ten largest BDCs reached 3.95% of debt at cost in Q2, and across 83 funds $2.04bn was written down — the largest quarterly write-down in the thirteen quarters of data available. The shares do not care. FS KKR is up 11.8% over twenty days, Blue Owl's manager up 11.7%, Blackstone Secured Lending sits 0.1% off its high.

Household credit31 Aughigh

Cards are getting better. Cars are at a record.

Aggregate household credit numbers are being held up by prime borrowers while the bottom band records all-time-high auto delinquency — averages will keep looking calm until they suddenly don't.

The household data this week splits cleanly in two, and the split is the story. Getting worse: Fitch reports subprime auto delinquencies at 6.13% in July, up from 5.80% at midyear. The New York Fed's Q2 data show 5.5% of all US auto debt at least 90 days past due — the second-highest reading in more than twenty years. Equifax, summarising Moody's, says subprime borrowers are at an all-time high for auto delinquency. On student loans, FICO counts about 3.2 million consumers with a recent delinquency — 15% of those with a payment due — carrying an average 38-point year-on-year score decline. ATTOM's mid-year report, which we have fact-checked as confirmed, counted 227,548 properties with foreclosure filings in the first half of 2026, up 21% year on year, with activity rising in 77% of metro areas. Getting better: 30-day-plus credit card delinquency fell to 2.85% in Q2 from 3.04% a year earlier, and Equifax's 60-day measure fell to 2.69% from 2.87%. Prime auto delinquency is 0.49%. Early-stage mortgage delinquency eased to 1.35%. The average FICO score is unchanged at 714. FICO's own cut of the data is the sharpest version: 90-day-plus delinquency on mortgages and autos rose only in the lowest score bands, and was flat everywhere else. This is not a deteriorating consumer. It is a consumer splitting into two populations, one of which is fine. Credit Acceptance, which lends at the deepest subprime end, trades 0.72% off its high and up 5.1% in twenty days. Equity holders have concluded that lending to distressed borrowers at those rates is a good business, and so far this cycle they have been right.

Crypto and TradFi contagion31 Augmedium

Bitcoin is being moved indoors, one wrapper at a time

Crypto only becomes a systemic problem when it is collateral inside regulated balance sheets, and that is precisely what was built this week.

BlackRock processed more than $5bn of direct bitcoin-into-ETF conversions after cutting the minimum ticket from $25m to $1m. Circle minted roughly 5bn USDC in a single week. Franklin Templeton received SEC clearance to use its $2.6bn tokenised BENJI money-market fund inside ETFs and mutual funds — and as collateral. JPMorgan's Kinexys, with Ripple, Mastercard and Ondo, settled a cross-border redemption of a tokenised Treasury fund in under five seconds. Each of these is presented as plumbing, and each is. But taken together they describe a specific migration: coins that used to sit in self-custody, outside the regulated system, are being converted into share classes held in brokerage accounts, and tokenised fund units are being made eligible as collateral. That is how an asset with 40% annual drawdowns acquires the ability to trigger a margin call somewhere that matters. The direction of travel is visible in price. Bitcoin is at $78,952, up 24.2% in twenty days. Strategy is up 40.2%, BitMine up 45.6%, SharpLink up 39.1%, Coinbase up 28.4% — all within about 3% of their recent highs. Three months ago the digital-asset treasury model was described as broken; Strategy had stopped buying for ten weeks. It has now bought $369.7m of bitcoin funded by $602.8m of common stock sales, and used $151.8m to buy back preferred (Bloomberg). The equity issuance machine works again. That is the whole model: it functions when the shares trade above the coins, and stops when they don't.

The dollar, gold and reserve status30 Aughigh

Australia's central bank cut its dollars from 55% to 45%

Reserve diversification is slow enough to ignore for years and then determines who has to absorb the long end when it matters.

The Reserve Bank of Australia disclosed that during 2025 it reduced the target US dollar share of its foreign reserves from 55% to 45%, raising the euro from 20% to 30%. Deutsche Bank called it a quiet reduction. It is ten percentage points from one mid-sized reserve manager, which is not a systemic event on its own — but it is a decision taken deliberately, disclosed after the fact, and of a kind that other reserve managers make by watching each other. The broader series is undramatic and relentless: the dollar's share of global FX reserves has slipped from a little over 70% at the turn of the century to a little under 60%, roughly half a point a year. Notably, the euro has captured none of that — Asian major currencies took about a quarter and non-traditional reserve currencies about three-quarters. Gold is the other side of the same trade. India's reserves hit a record $729.3bn, with gold holdings up $2.801bn to $114.218bn in the week to 21 August. Spot gold rose 10% in August, its biggest monthly gain since January, before giving back 3.1% over the past five sessions once Warsh committed to fighting inflation (Bloomberg). Connect this to the bond desk. Half a point a year of reserve share is a slow withdrawal of the least price-sensitive buyer of long-dated Treasuries. When that buyer steps back and the deficit runs near 6% of GDP against $40tn of debt, the marginal purchaser of the 30-year becomes someone who demands compensation. Which is roughly why the Treasury is now buying its own bonds back.

The AI capex bubble31 Augmedium

Volta wants twice its own valuation in debt

The debt-funded tier of the AI buildout is now borrowing multiples of its own equity value, and the lenders' recovery depends entirely on tenant credit nobody has disclosed.

On August 27 JPMorgan began early outreach to lenders on a $5bn debt package for Volta Infra Holdings, to fund an AI data-center buildout. Three weeks earlier, in early August, Volta had raised $300m of equity at a $2.4bn valuation, backed by Andreessen Horowitz, Altimeter, Nvidia and Michael Dell (pulse24media). So the debt being arranged is roughly twice what the entire company was worth on the equity market a month ago. That is not automatically absurd. It is how project finance works. Lenders to a data center are not really lending to the developer. They are lending against contracted lease payments from whoever occupies the building, secured on the concrete, the power connection and the machines. The developer's equity value is close to irrelevant if the tenant paying the lease is a safe borrower. Which makes the identity of the tenant the whole trade, and the reporting does not say who it is. If the leases are with Microsoft or Amazon, this is a utility bond wearing a venture-capital costume. If they are with a smaller cloud company or an AI lab, the lender is underwriting a startup's promise to pay for fifteen years, dressed as infrastructure. There is also the familiar circularity: Nvidia is an equity backer of a company borrowing money that will largely be spent on Nvidia chips. This is one week's haul. Nebius closed a $5.75bn convertible, IREN arranged $2.4bn in term loans and secured notes for Blackwell Ultra GPUs. Over $13bn of new debt from smaller cloud and AI infrastructure companies in seven days. The equity market is not buying the whole complex. Applied Digital is down 14.7% over twenty days, Core Scientific down 28.7%, IREN down 8%. Nvidia, which supplies all of them, is up 6.3%.

Crypto and TradFi contagion31 Aughigh

Saylor's machine restarts, and buys back its own preferred

The digital-asset treasury model is a leverage machine that only runs when the stock trades above the value of the bitcoin behind it, and it just demonstrated both that it can stop and that it can restart.

Strategy bought $369.7m of bitcoin in the week ended August 30, its first purchase after a ten-week pause, according to a filing reported by Bloomberg. It funded this by selling $602.8m of common stock, and separately bought back $151.8m of preferred shares. The pause is the informative part. Strategy's model only works in one direction: issue shares at a premium to the value of the bitcoin behind each share, buy more bitcoin, and every existing holder ends up with more coin behind their stock. Issue below that value and the arithmetic reverses. Each sale hands existing holders less bitcoin than they had. A treasury company that stops buying for ten weeks is telling you its premium had gone. It is back. Over twenty days bitcoin is up 24.1%; Strategy's shares are up 38.8%. Over five days bitcoin is flat (down 0.2%) while the stock added 7.3%. The premium is expanding faster than the asset behind it, which is exactly the condition that makes the flywheel spin again. The preferred buyback is the quieter signal. Strategy's preferred stock carries fixed dividends that must be paid in dollars, out of a company whose assets pay nothing. Retiring $151.8m of it reduces the hard claim on a balance sheet that has no income. Of the $602.8m raised, $369.7m went into bitcoin and $151.8m into shrinking the fixed obligations, leaving roughly $81m unaccounted for in the reported figures. On our reckoning this reduces fragility at the margin: equity replacing preferred stock is less borrowed money, not more. The vulnerability is unchanged in kind. It is entirely a function of a premium that has now proven it can vanish for ten weeks.

Private credit and BDCs31 Augmedium

One fund stopped taking money in, as well as out

A fund refusing new deposits is a manager conceding its own published price is unreliable, which is the closest thing to a real market test this asset class produces.

The collapse of Sydney builder Bathla Group, with A$2.3bn of debt, has Australian private credit funds limiting withdrawals, including, as we noted yesterday, one with no exposure to the developer at all (Bloomberg). Two specifics have now surfaced. MA Financial has imposed a temporary cap on withdrawals of up to 1% of funds under management per month. CVS Lane has reportedly suspended processing of application and redemption requests across its funds (Accelerate). Suspending withdrawals is defensive: it stops a run on assets that cannot be sold quickly. Suspending deposits is a different statement. It means the manager does not want new money at the current published value, because if the prices in the books are too high, every new investor buys assets worth less than they pay, and the existing holders quietly receive the difference. A manager who closes the door in both directions is telling you they do not trust their own price. That is the mechanism this beat exists to watch. Private credit's prices are the manager's opinion until a transaction tests them. When a transaction does test them (a BlackRock-marketed book of senior loans reportedly clearing at 88–92 cents on the manager's own valuations, Ares cutting a European continuation vehicle from €1bn to €400m rather than accept the buyers' price) the opinion has been consistently too high by something like a tenth. The listed market disagrees, loudly. Blue Owl's lending fund rose 2.9% today, Blackstone Secured Lending 1.7%, FS KKR 3.1% on the week, Apollo 3.6%. The extra interest shaky borrowers pay over the government fell to 2.6 percentage points. Either Australia is a local property accident, or the listed vehicles are being priced by people who have not read the late-payment rates.

Fed, Treasury and policy31 Aughigh

The FSB writes a letter. Nothing happens.

The most senior financial-stability official in the world has named the mechanism, publicly, and the price of risk did not move.

Andrew Bailey, as chair of the Financial Stability Board (the body that coordinates financial regulation across the G20), has written to G20 finance ministers warning that debt used to fund the AI boom "could amplify a future market correction", and that "the issue is not simply that investors are borrowing more but that leverage is interacting with high valuations and market concentration" (Telegraph, carried by four outlets). That sentence is the right one. Any of the three ingredients alone is survivable. Borrowed money against a diversified book is fine. High valuations funded by equity are fine, because the loss stops at the shareholder. Concentration in a well-capitalized sector is fine. Combined, they produce a market where a repricing forces sales of unrelated assets by borrowers meeting demands from their lenders for more cash, which is the transmission channel Bailey names. The letter also states that Nvidia, worth over $5.2tn, "recently raised $500bn from a consortium of US banks and investors to fund its AI investment spree." We cannot reconcile that with CNBC's version, which describes Nvidia as having helped establish financing platforms to mobilize $500bn of other people's capital for AI infrastructure, money raised by others, deployed around Nvidia, not borrowed onto its own books. The difference matters enormously: one is the world's most valuable company loading up on debt, the other is it organizing other people's debt. The public reporting does not settle it. What the FSB can do about any of this is: write letters. It has no rulebook and no money of its own. Its chair is telling the G20 that the risk sits outside the banks that the G20 spent 2009 to 2019 fortifying. Market response: the S&P 500 fell 0.36%, the VIX (the market's gauge of how much turbulence traders expect over the next month; under 20 is calm) sits at 15.05, and the extra interest shaky borrowers pay over the government tightened.

Hidden leverage and shadow banking31 Augmedium

The Treasury is now the backstop for a hedge fund trade

The largest position built on borrowed money in the financial system is being quietly supported by the Treasury Department's own buying, which is not a formal program, has no conditions, and nobody voted on it.

Analyst estimates circulating this week put the cash-futures basis trade (where hedge funds borrow heavily to profit from a small gap between two nearly identical prices for US government bonds) at $830bn, close to double its early-2020 peak, with a further $305bn in a related trade exploiting gaps in interest-rate swaps, financed by roughly $3tn of cash borrowed overnight against bonds put up as security, and about 90% of that borrowing concentrated in the fifty largest hedge funds (matterfact). Hedge funds are estimated to hold 8.5% of all outstanding US government bonds. The trade is simple and, individually, sensible. A fund buys a government bond, sells the matching futures contract, and pockets the few hundredths of a percentage point between them. The gap is tiny, so you need thirty or fifty times borrowed money to make it a business, which you get by financing the bond with overnight borrowing. It works until lenders demand more security or the futures exchange raises its cash requirements, at which point every fund holding the trade must sell the same bonds on the same afternoon. That is the March 2020 mechanism, at roughly twice the size. Now set that next to policy. From September 9 the Treasury Department, which borrows the money the government spends, doubles its buybacks of government bonds that do not come due for ten to thirty years to at least $4bn per operation, after the interest rate the government pays to borrow for thirty years touched 5.33%, a 19-year high, then gave back the relief within a day. Treasury Secretary Bessent called liquidity in those long-dated bonds "especially poor" and described the program as a kind of "Treasury twist." Buying back exactly the bonds where the leveraged trade lives is, in effect, a public bid supporting a private bet. It may well be prudent plumbing. It is also a standing subsidy to a position the Federal Reserve (America's central bank, which sets interest rates) does not formally backstop, delivered by the debt manager rather than the central bank, with, as Tom Hoenig noted at Jackson Hole, the Fed conspicuously silent about it.

Private credit and BDCs31 Aughigh

The gate that had nothing to gate against

The first gate to be imposed by a fund with no exposure to the underlying failure is the moment credit stress becomes liquidity stress, and that is how the migration of risk out of banks actually bites.

Bathla Group, a Sydney builder, has collapsed owing A$2.3bn. Australian private credit funds that lent to it have limited investor withdrawals. That is ordinary. What is not ordinary is that a fund with no exposure to Bathla at all has also put new limits on how much investors can pull out (Bloomberg). The mechanism is worth being precise about, because it is the one that turns a credit event into a liquidity event. An open-ended private credit fund offers investors monthly or quarterly redemption against a book of three-to-five year construction loans that cannot be sold in a month. As long as redemptions are small, the cash buffer covers them. Once redemptions rise, the investor who leaves first is paid at the manager's own mark, in cash, out of the buffer — and everyone left behind holds a more concentrated, less liquid book. That is a first-mover advantage, and it is why a fund with no bad loans still has a very good reason to gate before it has a problem. A gate protects the investors who stay. It also tells every other investor in every other fund exactly what to think about. This matters beyond Australia because the same structure is now standard in US non-traded vehicles, where aggregate redemption requests hit a record 12.4% of NAV in Q2, up from 10.4% in Q1. ASIC warned on private credit on 27 August, and RBA Governor Michele Bullock's line was the plainest thing any official has said this month: "People don't know where the leverage is."

Private credit and BDCs31 Augmedium

Non-accruals tripled. The shares are at their highs.

Every mark in private credit is a manager's opinion until someone sells, and the three sales we can observe this month all cleared about ten percent below the opinion.

Hold two numbers next to each other. Jefferies puts Q2 2026 non-accruals at Blackstone Secured Lending at 3.6% of the portfolio, against 0.3% in Q1 2025 — a twelvefold rise. Golub is at 2.9% from 1.2%, Blue Owl 2.8% from 1.4%, Ares Capital 2.4% from 1.5%. Across the twenty largest BDCs the median is 2.8%, up 80bp in a quarter. Fitch's trailing-twelve-month private credit default rate through July is 6.1%. Now the tape. BXSL closed at $24.93, 0.34% off its period high and up 5.8% in twenty days. Ares Capital is 0.12% off its high. FS KKR is up 11.6% in twenty days, Blue Owl's BDC up 4.1%, and the managers — KKR, Apollo, Blue Owl — rose again today while the broad index fell half a percent. These cannot both be casual readings. A BDC share is a claim on net asset value, and net asset value is the manager's own mark on loans that do not trade. The August evidence on what those marks are worth is not encouraging: a diligenced, first-lien-heavy US direct lending book marketed by BlackRock reportedly cleared at 88–92 cents on the manager's marks, and LP-led secondaries have priced near 90% of NAV all year. BlackRock TCP Capital took a 10.4% NAV hit selling $523m of loans. Ares cut a European continuation vehicle from €1bn to €400m rather than accept the discount buyers wanted. So the observable clearing price for this asset class is roughly ninety cents, and the equity that owns it trades at its high. One of those is wrong. We do not think it is the secondary bid.

The AI capex bubble31 Augmedium

Nvidia's $105bn promise on an Ohio uranium site

The most creditworthy company in the AI complex is now writing insurance on the least creditworthy part of it, and the exposure lives in a footnote rather than a capital ratio.

At Piketon, Ohio, on the grounds of a former uranium enrichment plant, SoftBank is building an eight-gigawatt server farm that OpenAI will rent. SoftBank has said it will put up to $500bn into it. To make the financing work, Nvidia has provided a guarantee: if OpenAI fails as tenant and the facility has to be sold at a discount, Nvidia absorbs up to $105bn of the loss. In exchange, Nvidia is the exclusive chip supplier. With that backstop behind it, SoftBank borrows more cheaply. This is reported by N-TV, citing the Wall Street Journal and documents filed with the SEC. We have not seen the filing. The structure is a residual value guarantee, and it does a specific thing: it lends Nvidia's balance sheet strength to a counterparty without moving any asset onto Nvidia's balance sheet. The lender to the project is no longer underwriting OpenAI's ability to pay rent for a decade. It is underwriting Nvidia's ability to cover a shortfall. What has to stay true for it to work is that GPUs and the buildings around them hold their value — and the resale market that would set that value barely exists, because almost nobody has yet tried to sell a used eight-gigawatt AI campus. The accounting is the part nobody outside the companies can see. A guarantee of this kind is a contingent liability whose fair value depends on a probability of tenant default and a recovery assumption, both of which are the issuer's own estimate. It is exactly the sort of exposure that is measured and capitalised inside a bank and merely disclosed outside one. That is the whole Crash Lab thesis in one footnote.

The AI capex bubble30 Augmedium

The power leg is down a fifth. Nvidia isn't.

The AI trade is being decomposed in public into the parts paid in cash and the parts paid in promises, and power is being sorted into the second pile.

Talen Energy closed at $294.33, down 14.5% in twenty days and 22.4% below its recent high. Vistra is down 12.5% over the same twenty days and 19.3% off its high. These are the companies that sell electricity to data centres — the demand story everyone agrees is real. Over the same period Nvidia is up 6.1% and Supermicro up 30.3%. That gap is the market drawing a line, and it is not the line between AI winners and losers. It is the line between who gets paid now and who gets paid later. Nvidia sells a chip and books cash. A merchant power producer signs a long-dated contract and its equity value is a function of what price it gets, for how long, from a customer whose own capex plans are financed. Marking the power names down a fifth while marking the chip vendor up is a statement that the market believes the near-term orders and doubts the duration of the demand behind them, or the price at which it will be served. That is our inference from the prices; no reporting in this window explains the move. It fits the rest of the tape. Today the buyers fell and the vendor rose: Alphabet −2.3%, Amazon −2.2%, Meta −1.5%, Microsoft −1.1%, Oracle −1.4%, Nvidia +0.9%. Alphabet is down 9.3% in twenty days, Amazon 8.3%. CNBC reports capital spending at some hyperscalers has now passed 100% of operating cash flow, which is the point at which every additional gigawatt must be financed rather than funded. The market is pricing the financed part more harshly than the funded part, and it is right to.

The dollar, gold and reserve status30 Augmedium

Australia's central bank sold a tenth of its dollars

Official-sector dollar demand is being managed by facility rather than by market, which works until the facility is the only bid.

The Reserve Bank of Australia's annual report discloses that during 2024/25 it cut the US dollar share of its foreign reserve portfolio from 55% to 45%, and raised the euro share from 20% to 30%. Deutsche Bank called it a quiet reduction. It is a ten-point reallocation by a G20 central bank, executed without a speech. The context is a slow, consistent drift: IMF COFER puts the dollar at 57.13% of global reserves in Q1 2026, against 71% in 2000, sliding about half a point a year. The interesting detail in that research is where the share went. The euro has captured none of it; roughly a quarter went to major Asian currencies and about three-quarters to non-traditional reserve currencies. Reserve managers are not switching to a rival. They are spreading out. Why this matters operationally rather than symbolically: a reserve manager reducing its dollar share is, mechanically, a seller of Treasuries. Washington appears to understand this. The same reporting notes that the US used euros rather than dollars in a recent yen-support operation, and pushed the Federal Reserve to widen a facility so the Bank of Japan could pledge Treasuries for cash instead of selling them. That is a plumbing fix designed to stop official-sector dollar management from showing up as price in the long end — the same long end where the 30-year touched 5.33%, a 19-year high, and where the Treasury has doubled buybacks to at least $4bn per operation from 9 September. Gold, meanwhile, is up 9.1% in twenty days despite giving back 3.7% this week after Warsh's hawkish Jackson Hole remarks. India's reserves hit a record $729.3bn, of which $114.2bn is gold.

Household credit31 Aughigh

Subprime auto went back up in July

The bottom score band is deteriorating on its own while the average holds, which is what a credit cycle looks like before it is visible in an average.

Fitch's auto ABS data had been telling a recovery story: subprime delinquencies at 6.50% at the end of 2025, improving to 5.80% by mid-2026. July reversed a large part of that, back to 6.13%. Prime moved from 0.48% to 0.49%, which is nothing. The New York Fed's Q2 figures put 5.49% of all auto loan balances at 90 days or more past due — the second-highest reading in more than twenty years — with 3.00% of balances rolling into that bucket over the past four quarters. We lowered the household leg of our fragility estimate two days ago on improving card data, and that data still holds: bank card 30-day delinquency at 2.85% against 3.04% a year earlier, the broader 60-day measure at 2.69% from 2.87%. The July auto print takes part of that back. Cards and cars are diverging, and the reason is that a car loan is secured on a depreciating asset with an insurance and fuel bill attached, so it is the payment a stretched household abandons second, after nothing. The distribution matters more than the average. FICO's average score is unchanged at 714, and 90-plus-day delinquency on mortgages and autos rose only in the lowest score bands — flat everywhere else. Separately, about 3.2 million consumers with a student loan payment due went delinquent and lost an average 38 FICO points year on year, while consistent payers gained six. A 38-point drop moves a borrower down a pricing tier on every product they touch. The equities are unimpressed but not asleep: Affirm fell 5.2% today, Ally is down 5.2% in twenty days, Capital One 1.4%.

Crypto and TradFi contagion31 Augmedium

A tokenised money fund becomes collateral

Crypto's contagion channel to traditional finance has never been price; it is the moment its instruments become collateral in somebody else's funding.

Franklin Templeton has cleared the way, via an SEC no-action position, to use its $2.6bn tokenised money market fund BENJI inside its ETFs and mutual funds — for cash management and as collateral. Separately, Circle minted roughly 5bn USDC in a single week, one of its largest issuance runs, and Standard Chartered and HSBC completed what is described as the first bank-to-bank transfer of tokenised deposits over SWIFT's blockchain ledger, with 17 banks on six continents preparing live transactions. The collateral point is the one to hold onto. Collateral is the load-bearing wall of the financial system: it is what lets one institution lend to another against something it can seize instantly. For decades that has meant Treasuries and money fund shares whose legal and operational characteristics are exhaustively tested. A tokenised money fund share is a claim on the same underlying bills, but the settlement layer, the redemption mechanics under stress and the behaviour of the register when a chain is congested are not. $2.6bn is small enough that this is a pilot, not an exposure. The question is what it looks like at $260bn. On the other side of the beat, the digital asset treasury companies have re-inflated with the coin rather than ahead of it: bitcoin is up 22.7% in twenty days to $77,955, Strategy up 33.5%, BitMine up 36.7%, SharpLink up 31.2%, Coinbase up 21.5%. Two days ago these names fell while bitcoin did not. The premium is back, and it is still the same premium: paying more than a dollar for a dollar of coin, on the theory that the vehicle can keep issuing shares above net asset value.

What would change our mind

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

A US lending fund or interval fund limiting withdrawals, as MA Financial and CVS Lane have done in Australia. That would mean the liquidity problem has traveled to the market that matters.

Would move the number

The extra interest risky borrowers pay over the government moving above 350 basis points from 260, or a secondhand price for a loan secured on processors coming in below 90 cents on the dollar. Either would test valuations that have so far only been modeled, not traded.

Would move the number

The Federal Reserve raising rates on September 16, or an auction of twenty- or thirty-year government bonds attracting far fewer buyers than expected while the expanded buybacks are running.

Would move the number

Genuine relief: a large private-credit portfolio selling at or above the value the lender had it on the books for, or the largest cloud companies guiding capital spending lower without a corresponding sell-off in the lenders financing them.

Would move the number