Japan's 10-year hits 3%, first time since 1996

Archived reading, published Tue, 01 Sep 2026 18:22:21 UTC (19 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

Fragility91

how much tinder is stacked up — moves slowly

Ignition41

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 34 sources and rewrites this page.

Why it moved: Ignition up two to 41: the bond rout crossed a genuinely new line — Japan's 10-year yield touched 3% for the first time since 1996 and five-year JGBs hit a record 2.26% — while gold fell 7% in five days into an oil shock and the whole liquid risk complex sold off (VIX +7% to 16.02, Blackstone -4.4%, Oracle -5.4%, Coinbase -6.2%). Fragility holds at 91: the FT's finding that roughly $150bn of data-centre projects are stalled or cancelled on local opposition removes committed capex, offsetting an upward re-measurement of off-balance-sheet AI exposure (BofA's $370bn-by-2029 estimate for the Broadcom structure).

Reporting from 31 Aug to 1 Sep

Crash points

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

What’s moving the needle

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

Bond market dysfunction1 Sephigh

Tokyo joins the rout at three per cent

The last thirty years of cheap global duration rested on Japanese savers having nowhere better to go; at 3% they do.

Japan's 10-year government bond yield reached 3% on Tuesday, the first time since 1996. Five-year JGBs hit a record 2.26%. In the same session UK 10-year gilts touched their highest since June 2008 and the 30-year gilt rose nine basis points to 5.88%, the highest since March 1998. US 10-year Treasuries closed at 4.79%, the 30-year at 5.27%. Brent went through $92 with just five ships transiting the Strait of Hormuz. The Japanese number is the one to hold on to. For thirty years, Japanese institutions — life insurers, banks, pension funds, the largest pool of foreign savings in the world — could not earn a return at home and so bought foreign bonds, much of it US and European paper, often hedged. A domestic long bond yielding 3% changes that arithmetic without anyone announcing anything. It does not require a policy decision, a headline or a crisis; it requires an actuary to run the numbers again at the next asset-allocation meeting. What we cannot see yet is whether that is actually happening. Repatriation flows show up with a long lag in the balance-of-payments data. What is visible is the coincidence: yields rising together across the US, Germany, the UK and Japan, which Reuters and Bloomberg both attribute to energy-driven inflation and fiscal supply rather than any single country's politics. Big Tech's bond issuance is being named in the same sentence — the AI buildout is now large enough to be a supply factor in the global government bond market, which is a sentence that would have read as nonsense two years ago. Minneapolis Fed president Kashkari says the Treasury market is "functioning as it should". On the evidence — no failed auctions, no dealer seizure — he is right. This is a price problem, not yet a plumbing problem.

Fed, Treasury and policy1 Sephigh

$150bn of data centres that voters said no to

For once a piece of the AI pipeline is getting smaller rather than more leveraged, and that genuinely lowers fragility.

Dozens of US data-centre projects worth roughly $150bn have been stalled or cancelled because of local opposition, the FT reports. President Trump responded on Monday by telling Americans opposing them that they wanted to end up "backwards and poor", and urging the country to "let Data Reign". Vice-president Vance put "probably 99 per cent" of the backlash down to higher energy bills in surrounding areas. We have been arguing for a fortnight that the binding constraint on the AI buildout was shifting from capital to power. This is the next step: the constraint is becoming permission. Texas has frozen new grid connections while it works out which of 474 gigawatts of requests are real. Pennsylvania says only 20 of more than 100 proposed projects have actually applied for permits. And now a specific dollar figure has been attached to projects that are simply not going to be built. For the fragility argument this cuts in an unfamiliar direction. Capex that does not happen is debt that does not get raised and leases that do not get signed. $150bn of cancelled projects is $150bn of risk that will not be stacked. Set against that, every cancelled site makes the revenue assumptions underneath the sites that do get financed harder to verify — if a third of the pipeline is political vapour, the demand forecasts justifying the rest were built on the same spreadsheet. The money has noticed. An advocacy group tied to Leading the Future, a super PAC that has raised more than $75mn from Marc Andreessen, Ben Horowitz and OpenAI's Greg Brockman, is running an advertising campaign for data centres in Kansas, Ohio and Wisconsin. When an industry has to buy public consent by the state, its costs have changed.

The dollar, gold and reserve status1 Sepmedium

Gold fell 7% during a war and an oil shock

If gold is falling because people need cash rather than because real yields rose, someone is being margined and we cannot yet see who.

Gold is down 7.07% over five sessions and 2.63% on Tuesday alone, at $4,391. Silver is down 6.59% on the week. This happened while Brent went above $92, while the US and Iran exchanged direct attacks, and while long-dated government bonds sold off across four continents. Gold is still up 3.4% over twenty days, so this is a sharp move inside an uptrend rather than a break — but it is the wrong direction for the week's news. Two explanations fit. The first is mechanical: gold pays no coupon. With the 30-year Treasury at 5.27% and the dollar index up 0.79% on the week, the cost of holding a metal that yields nothing has risen, and gold sells off when real yields rise regardless of what is happening geopolitically. That is the boring answer and it is probably most of it. The second is that gold is what you sell when you need cash. In a liquidation, the asset that goes first is the one with a deep market and an unrealised gain — and after a 20-day run, gold qualifies on both counts. Tuesday's tape is consistent with that: equities down, credit slightly weaker, gold down, dollar up. That is the signature of positions being reduced rather than risk being reallocated. We cannot distinguish the two from price alone, and we are not going to pretend otherwise. What we would watch is the central-bank bid underneath, which runs on a much slower clock: India's reserves hit a record $729.3bn in the week to 21 August, with gold holdings up $2.8bn, and the Reserve Bank of Australia is reported to have cut its target dollar share from 55% to 45%. Those buyers do not care about a bad week.

Household credit1 Sephigh

Cards are getting better. Cars are not.

A car loan is the debt people default on last, so 5.49% of balances 90 days late is a statement about the bottom third of the income distribution, not about credit conditions.

New York Fed data for the second quarter show 5.49% of auto loan balances 90 or more days delinquent, with 3.00% of balances transitioning into that state over the previous four quarters. Equifax, summarising Moody's, says subprime auto borrowers are at or near a 15-year high. Over the same period credit cards improved: bank-issued card 30-plus-day delinquency fell to 2.85%, and 60-plus-day delinquency across all cards fell to 2.69% from 2.87% a year earlier. FICO reports 30-day auto delinquency actually easing to 2.6%. These are not contradictory, and the divergence is the interesting part. The early-stage numbers are stabilising while the late-stage numbers are not. That is what it looks like when a cohort that fell behind two years ago is not recovering — the flow into distress slows, the stock of distress keeps compounding. FICO's own breakdown supports this: 90-day-plus delinquency rose only in the lowest score bands and was flat everywhere else. Add 3.2m student-loan borrowers reported at least 30 days late in the previous six months, carrying an average 38-point score decline, and the same households appear in both series. The market is pricing it selectively rather than systemically. Affirm fell 6.08% on Tuesday and 10.23% on the week; Capital One is off 4.26% over twenty days and Ally 6.81%. But Credit Acceptance, the deep subprime auto lender, is up 1.25% over the same twenty days and sits 1.6% from its high. Deep subprime funds itself through the securitisation market, and while high yield spreads sit at 263bp that funding stays open. The equity is telling you that the lenders' cost of money matters more, for now, than their borrowers' ability to pay.

The AI capex bubble31 Augmedium

A vehicle that buys chips and leases them out

The chips are collateral only if there is a second-hand market for three-year-old accelerators, and nobody has tested that.

Broadcom is reported to be in talks with lenders to raise more than $60bn, possibly $100bn, through a special-purpose vehicle that would buy Broadcom's own custom AI chips and lease them to Anthropic. Bank of America is quoted estimating that total exposure across the structure could reach $370bn by 2029. A previous version of the arrangement, set up in June with Apollo and Blackstone, is said to have raised $35bn. This is tier-2 reporting and the numbers move between accounts, so treat the range as a range. The structure, though, is worth understanding because it is becoming the standard form. A vendor wants to sell chips. The customer cannot fund them out of cash flow. So a separate company is created, funded with borrowed money from private credit and insurance balance sheets, which buys the chips and rents them to the customer. The vendor books a sale. The customer books a lease. The debt sits in the vehicle, which appears on nobody's consolidated balance sheet in a form the reader of an annual report would recognise. Reporting cited by CNBC puts roughly $3tn of off-balance-sheet AI commitments across nine large technology companies; the BIS calls it "shadow borrowing". The question that decides whether this is clever or dangerous is: who owns the residual? If Anthropic stops paying in year four, the lender is left holding used accelerators two generations behind the frontier, in a building that may not have power. Nothing in the reporting tells us how the residual value is set, who guarantees it, or what the chips are assumed to be worth at the end. Until that is disclosed, the loss-bearer is unidentified — which historically has meant it is everybody.

Crypto and TradFi contagion31 Aughigh

Bitcoin fell 1.8%. Its holders fell 6%.

Digital asset treasury companies are a leveraged claim on a volatile asset funded by continuous equity issuance, and the leverage lives in the share price, not the balance sheet.

Bitcoin closed Tuesday down 1.82% at $77,121. The companies whose business is holding it fell three to four times as far: Strategy -5.78%, BitMine -8.00%, SharpLink -5.15%, Coinbase -6.21%, Galaxy Digital -5.17%. Over twenty days the same names are up 28-30% against bitcoin's 21.6%. The wrapper has a beta to the coin, in both directions, and it is running near three. The mechanism is in Strategy's own filing. It bought $369.7m of bitcoin in the week to 30 August, resuming purchases after a ten-week pause. To fund it, the company raised $602.8m by selling common stock and used $151.8m to buy back preferred shares. Read that again: it issued substantially more equity than it spent on the asset it exists to hold, and spent part of the difference retiring a fixed claim ahead of the common. This is the machine working as designed. If the shares trade above the value of the coins per share, issuing stock to buy more coins raises coins-per-share and is accretive; the premium funds the purchases, and the purchases justify the premium. The reflexivity is the product. It runs in reverse just as smoothly — if the premium goes, new issuance dilutes rather than accretes, and the buying stops, which was presumably what the ten-week pause was about. The filing tells us what was bought and how it was funded. It does not tell us where the multiple stands today, and we cannot compute it from the data in front of us. What Tuesday's tape shows is that the equity is repricing faster than the asset — which is what you would expect if investors are marking the premium, not the coin.

Signs of the times

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

Super PAC advertises for data centres

An advocacy group linked to Leading the Future — a super PAC that has raised over $75mn from Marc Andreessen, Ben Horowitz and OpenAI co-founder Greg Brockman — has launched a multimillion-dollar advertising blitz in Kansas, Ohio and Wisconsin to build public support for data centres. The industry that can raise $100bn from lenders on a phone call now has to buy consent one state at a time.

aipolitics

1 SepFinancial Times

Twenty-one banks, one stablecoin, no customers

Goldman Sachs, Bank of America, Citi and Deutsche Bank are among 21 institutions that will form a company this year to issue a dollar stablecoin in the first half of 2027; a rival consortium of 37 banks is launching a euro one sooner. Reuters notes, in the same article, that there are "few signs of demand for stablecoins issued by banks".

cryptobanks

1 SepReuters

Abu Dhabi's spy chief and the Trump coin

Sheikh Tahnoon bin Zayed, deputy ruler of Abu Dhabi and UAE national security adviser, and co-investors own 49% of the holding company of a US bank that has just received preliminary approval to issue USD1, the stablecoin launched by the Trump-family-linked World Liberty Financial. His conglomerate separately moved $30m in a dirham-backed stablecoin, said to be one of the largest single stablecoin transactions in the Middle East.

cryptogeopolitics

31 AugSemafor

Only twenty of a hundred bothered with permits

Pennsylvania says that of more than 100 proposed data centres in the state, only 20 have actually sought permits. Across the Midwest, Mid-Atlantic and South, requests from very large power users exceed 700 gigawatts — more than ten times current estimated US data-centre consumption, and roughly what it takes to power every home in the country.

aipower

1 SepReuters

The rumour mill

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

Confirmed

“The US Strategic Petroleum Reserve is currently being depleted.”

DOE data show the SPR at 289.7m barrels as of 21 August, down 3.7m in a week and the lowest since 1982 — roughly 41% of its 714m-barrel capacity. With Brent above $92 and five ships transiting Hormuz, the buffer is thinner than in any previous oil shock this century.

Claimed by Gregory Mannarino

Partly true

“Musk says roughly 15GW of AI compute built in 2027 cannot be switched on that year, because chip production is growing 40-50% annually while power outside China grows 10-20%.”

The quotes are real and the growth-rate figures are reported accurately, but Musk's own wording ties the 15GW gap to transformers, wiring, cooling and networking, not electricity generation alone. The underlying "consensus estimate" is not independently documented anywhere.

Claimed by Meet Kevin

Partly true

“Q2 corporate profits were among the best in five years with margins at historically elevated levels.”

BEA data put after-tax corporate profits at 19.4% of gross value added, the highest in records going back to the 1940s, and FactSet's blended S&P 500 net margin at 17.0% against a 12.4% five-year average. The additional claim about the best four-quarter change in over twenty years is unsourced. Worth holding alongside everything else here: the earnings are real.

Claimed by Meet Kevin

Earlier developments

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

Bond market dysfunction1 Sephigh

An oil shock arrives in the middle of a bond rout

The long end is now absorbing an inflation shock while the only official tool pointed at it is a liquidity backstop that cannot change duration supply.

Brent rose 1.7% to $92.1 a barrel on Tuesday and only five ships transited the Strait of Hormuz. Wheat hit a three-and-a-half-year high on attacks in the Black Sea. Long bonds sold off worldwide: the 30-year gilt jumped 9bp to 5.88%, its highest since March 1998, and the 10-year gilt reached its highest since June 2008 (Guardian). The US 10-year broke above 4.75% and the 5-year above 4.50% on Monday, both for the first time since January 2025. The 30-year Treasury sits at 5.26%, having closed above 5% on 55 days this year — the most since 2006 (Bloomberg). The plumbing matters here. Treasury's response to the earlier long-end move was to double liquidity-support buybacks from $2bn to at least $4bn per operation, starting 9 September and running to 4 November, funded out of a Treasury General Account near $1tn. That is a market-functioning tool: it buys illiquid off-the-run bonds to keep dealers able to make prices. It does not change the supply of duration, and it does nothing about an inflation impulse. The first announcement bought a day of relief before yields round-tripped. What makes this configuration different from the last fortnight is the direction of the Fed. Kevin Warsh told Jackson Hole that 2% is a "firm and fixed target" and that "we still have work to do." An energy shock into a central bank that has already priced out easing, with a hike live for 16 September, is the one combination in which the long end and the front end can rise together. Both did. Against that: no failed auctions, no dealer balance-sheet seizure, no forced unwinds. This is repricing, not breakage.

The AI capex bubble1 Sephigh

Texas asked for proof and 474 gigawatts appeared

The demand figures underwriting both utility capex and data-centre project debt are partly duplicative filings that cost nothing to make.

Requests for grid connections from very large power users — mostly data centres — now exceed 700 gigawatts across parts of the Midwest, Mid-Atlantic and South, more than ten times industry estimates of what US data centres actually consume today and roughly as much electricity as every home in the country uses. In Texas alone, requests went from about 48GW in 2023 to more than 474GW now. Texas has become the first major hub to freeze new grid connections while it investigates what is real. Pennsylvania says only 20 of more than 100 proposed data centres actually applied for permits (Reuters). The mechanism is worth being precise about. An interconnection request is free or nearly free to file, and developers file with multiple utilities for the same project to preserve optionality. Those queues then become evidence: utilities cite them in capex plans and rate cases, power producers cite them when talking about contracted demand, and lenders to neoclouds and project SPVs treat a signed power agreement as the thing that makes a GPU-stuffed shed financeable. Reuters reports that utilities have cut their data-centre demand figures once upfront payments were required — which is the queue telling on itself. Two consequences pull in opposite directions. Fake demand means the buildout, and the debt attached to it, is smaller than the headline pipeline implies. That lowers systemic risk. But it also means the scarcity story underwriting power valuations is thinner than the queue suggested, and the binding constraint on real projects is now permission rather than capital — which is not something a lender can solve. The prices have noticed. Talen is down 13.6% over twenty days and 22.5% off its high; Vistra is 19% off. Applied Digital is down 22% over twenty days, Core Scientific 31%.

Hidden leverage and shadow banking31 Augmedium

The trade that cares about volatility, not yields

The largest leveraged position in the Treasury market is indifferent to yield levels and highly sensitive to exactly the volatility now arriving.

Analyst estimates put the cash-futures Treasury basis trade at about $830bn, close to double its early-2020 peak, with a further $305bn in swap-spread arbitrage, financed by roughly $3tn of hedge fund repo borrowing — more than double early 2023. Hedge funds' gross Treasury exposure is estimated to have doubled to about $4tn between 2023 and September 2025, $2.4tn long against $1.6tn short, or around 8.5% of outstanding Treasuries. The top 50 funds account for about 90% of the repo borrowing (matterfact). These are commentary estimates, not official data; the BIS, IMF and Fed published nothing new on this in the window. The structure: buy the cash bond, short the future, capture the small spread, and finance the bond in repo at a haircut of a fraction of a percent. Leverage of fifty or a hundred to one is normal because the spread is tiny. Crucially, the position is close to neutral on the direction of yields. What it is not neutral on is volatility. A jump in realised volatility raises futures initial margin and repo haircuts at the same time, and the only way to meet both calls quickly is to sell the cash leg — into a market where dealers are also getting hit. That is why the relevant number this week is not the level of the 30-year but the MOVE index, up 25.5% over five sessions from a low base. Commerzbank called the basis trade "shaky" on 26 August. The RBA's Michele Bullock put it more plainly: "People don't know where the leverage is." The SEC, separately, has subpoenaed banks about one hedge fund's communications with its lenders over the leverage it was taking on. Nothing has broken. But a global long-end selloff is the precise input this position dislikes.

Private credit and BDCs1 Sepmedium

The managers fell. Their loan books rose.

The listed BDCs are trading at or near NAV while their own filings show the worst credit metrics since at least 2021 — one of those two numbers is wrong.

On Tuesday the private credit managers were sold and the funds they run were not. Blackstone -2.4%, Ares Management -2.1%, Apollo -1.9%, KKR -1.5%, Blue Owl -1.6%. Meanwhile Blackstone Secured Lending rose 0.5% and closed at its period high, Ares Capital rose 0.2% and sits 0.5% off its own high, Blue Owl's BDC was flat and FS KKR is up 2.9% on the week. That split is a specific claim about where the damage is. Manager equity is a levered bet on future fundraising, AUM growth and performance fees. BDC equity is a claim on a loan book carried at net asset value. Pricing the first down while marking the second up says: fundraising slows, the book is fine. The reported book is not obviously fine. Non-accruals across the ten largest BDCs reached 3.95% of cost in Q2, up from 3.75% in Q1; the median across the twenty largest rose 80bp in a quarter to 2.8%, the highest since around 2017. Blackstone Secured Lending's own figure went from 0.3% in Q1 2025 to 3.6%. Across 83 BDCs, investments still held were written down by $2.04bn, the largest quarterly write-down in a thirteen-quarter dataset, and positions marked below 90 cents or on non-accrual went from 0.9% to 3.2% (JD Supra, Accelerate). The PIK share of BDC loans has risen from roughly 6% in early 2022 to about 10%. The reconciliation available to us is unattractive: a diligenced, first-lien-heavy book marketed by BlackRock reportedly cleared at 88–92 cents on the manager's own marks. If that is the clearing price, then the discount the market is applying to the fee stream belongs on the loan book as well.

The dollar, gold and reserve status31 Aughigh

De-dollarisation is real and it is not today's trade

Reserve diversification raises the long-run cost of US duration; it does not explain a week in which gold and Treasuries fell together and the dollar rose.

The Reserve Bank of Australia cut the target share of US dollars in its reserve portfolio from 55% to 45% in 2024/25, raising the euro from 20% to 30% — a ten-point shift Deutsche Bank described as a quiet reduction. A Kansas City Fed working paper puts the dollar's share of global FX reserves at 57% in Q1 2026, against 72% in 2000. India's reserves hit a record $729.33bn in the week to 21 August, up $12.42bn, with gold holdings up $2.80bn to $114.22bn (Bloomberg). Every one of those facts is real. None of them explains this week's prices, and it is worth being clear about that rather than pretending the narrative and the tape agree. The dollar index is up 0.64% on the week at 99.55 and the broad index up 0.58%. Gold is down 4.8% over five sessions to $4,414 and silver down 5.9%. Over twenty days both are still up — 4% and 8.9% — so the trend has not broken. But over five days, gold, silver, equities and bonds have fallen together while the dollar rose. That pattern has a boring name: liquidation. When everything falls at once and the funding currency rises, the marginal seller is meeting a margin call, not expressing a view on reserve composition. Reserve diversification is a decade-scale flow of a few percentage points a year against a $100tn-plus asset base; it does not set the weekly price of gold. The useful reading is narrower. Reserve managers are gradually reducing their structural bid for Treasuries at the same time as the US is issuing record duration. That shows up in term premium over years, and in the 30-year's 55 days above 5% this year.

Crypto and TradFi contagion31 Augmedium

Bitcoin is now mortgage collateral

Crypto risk is being wired into mortgage credit and collateral chains before anyone has disclosed the loan-to-value or the top-up terms.

Coinbase and Better have launched bitcoin-backed mortgages, after a waitlist that the companies projected at more than $260m of demand. In the same week Circle minted roughly 5bn USDC, BlackRock's IBIT processed more than $5bn of direct bitcoin-to-ETF conversions after cutting the minimum ticket from $25m to $1m, Franklin Templeton received clearance to use its $2.6bn tokenised BENJI money-market fund inside ETFs and mutual funds and as collateral, and Schwab — a $13tn platform — added SOL, AVAX and LINK (Bitcoin.com weekly recap). The mortgage product is the one that changes a plumbing diagram. A crypto-collateralised mortgage takes an asset that can fall 30% in a fortnight and puts it behind a thirty-year liability on a house. The reporting we have does not say what the loan-to-value is, what the top-up mechanic looks like, who holds the loan after origination, or whether it is intended for securitisation. Those four facts determine whether a bitcoin drawdown produces a margin call on a homeowner, a writedown at a lender, or a downgrade in a securitisation trust. Until somebody publishes them, this is an unmeasured channel between a volatile asset and housing credit. The tokenised money-fund clearance is the quieter version of the same thing. A money fund share used as collateral is fine until the fund gates; then the collateral is illiquid precisely when it is needed. Bitcoin is at $78,300, down 2.4% on the week but up 23.5% over twenty days. Coinbase fell 3.3% on Tuesday, Galaxy 3.7%, the ether treasury companies 2–4% — the equity claims on crypto sold harder than the coin did, which is the ordinary order of things when leverage is somewhere in the stack.

Bond market dysfunction1 Sephigh

Fifty-five days above five percent

A long end that stays above 5% quietly re-underwrites every asset priced off it, whether or not anyone remarks the book.

The yield on the 30-year Treasury has now settled above 5% on 55 days since the start of January — the most in any year since 2006, on Bloomberg's data. It touched 5.34% in mid-August, the highest since 2007 and within 10 basis points of a 22-year high, and sat at 5.27% on Tuesday (Bloomberg). The important word in that sentence is days. A yield spike is a headline; a yield that stays there is a repricing. Every private-credit loan, every data-centre lease, every leveraged buyout underwritten when the long end was at 4% is now being discounted against a curve that has held near 5.3% for most of a year. Treasury's response was to double its long-end liquidity-support buybacks, from $2bn to at least $4bn per operation in the 10–20 and 20–30 year sectors, running 9 September to 4 November. Worth being precise about what that is: the Treasury buys old, illiquid off-the-run bonds and funds the purchases with new issuance, generally bills. It does not reduce the debt. It swaps duration the market does not want for paper it does, and shifts supply toward the front end. Announced on 19 August, it moved yields for minutes; they had round-tripped by the following afternoon. What is not happening matters as much. The WSJ's own summary of the episode: no failed auctions, no dealer balance-sheet seizure, no forced unwinds, nothing resembling March 2020 or the gilt market in 2022. This is price pressure from deficits and supply, met with an official patch. It is a fragility story, not yet an ignition one.

The AI capex bubble31 Augmedium

Two incompatible descriptions of Nvidia's $500bn

Whether the AI build-out is financed on Nvidia's balance sheet or on everyone else's determines who is forced to sell when the returns disappoint.

Andrew Bailey, writing to G20 finance ministers as chair of the Financial Stability Board, warned that debt used to fund the AI build-out "could amplify a future market correction". His concern was not leverage alone: "the issue is not simply that investors are borrowing more but that leverage is interacting with high valuations and market concentration." In setting that up, the Telegraph writes that Nvidia, worth more than $5.2tn, "recently raised $500bn from a consortium of US banks and investors to fund its AI investment spree" (Telegraph). CNBC's account of the same $500bn, carried in our sweep, describes something structurally different: Nvidia partnering with Wall Street firms to establish financing platforms to mobilise $500bn of third-party capital for AI infrastructure. Those are not the same claim. One is half a trillion dollars of Nvidia's own funding. The other is half a trillion dollars of other people's money, arranged by Nvidia, to buy Nvidia's products — which would sit on the balance sheets of neoclouds, SPVs and private credit funds, not Nvidia's. The first is a company levering up. The second is a vendor organising the credit that clears its inventory, with the losses landing somewhere nobody has to mark daily. We cannot reconcile the two figures from what has been published. The distinction is the whole Crash Lab thesis in one number. It is also not academic: Nvidia disclosed in an 8-K on 17 August residual value guarantees with SB Energy covering roughly 4.25GW of IT load, with the aggregate payment obligation capped at $105bn. That is a contingent liability, not a borrowing, and it is exactly the kind of exposure the headline number obscures. The market is untroubled. Nvidia rose 5.9% on the week and sits 3.2% below its high.

Hidden leverage and shadow banking31 Auglow

Treasury repo is drifting back out of clearing

The largest leveraged position in the financial system is financed in a market that is getting less transparent, not more.

One number in this window's leverage sweep deserves more attention than it got: the share of Treasury repo that is centrally cleared fell from 71% at the end of 2025 to 62% in the first quarter of 2026. Why that matters takes a paragraph of plumbing. In a repo, a hedge fund borrows cash against Treasuries. Centrally cleared, both legs face a clearing house, which nets exposures across all its members, demands margin to a published schedule, and knows the size of the position. Bilaterally, the trade sits between the fund and a single dealer, on terms nobody else sees — and a meaningful share of bilateral Treasury repo has historically been done at zero haircut, meaning the borrower posts no cushion at all. The SEC's clearing mandate exists precisely to move this activity into the light. The share moving the other way, if the number holds, is a step backwards. It matters because of what the repo is funding. Analysts recirculating Federal Reserve estimates put large hedge funds' gross Treasury exposure at about $4tn — $2.4tn long, $1.6tn short — financed by roughly $3tn of repo borrowing, with 90% of that borrowing concentrated in the 50 largest funds. Within it, the cash-futures basis trade is sized at about $830bn, close to double its early-2020 peak, plus $305bn of swap-spread arbitrage. The mechanism to fear is not credit loss. It is that a move in yields triggers margin calls, funds sell Treasuries to meet them, yields move further, and the loop closes. That is roughly what happened in March 2020, at a fraction of today's size. Treat the clearing figure carefully: it comes from market commentary in a single sweep, not from an official release, and we have not seen it corroborated.

Household credit31 Aughigh

Subprime auto at a record, credit cards getting better

Losses concentrated in the bottom score bands do not threaten a bank, but they are the collateral behind a large stock of auto ABS.

Household credit is not deteriorating. Parts of it are. Fitch has subprime auto delinquencies at 6.13% in July, up from 5.80% at midyear. Prime borrowers, in the same series, are at 0.49%, up one basis point. New York Fed data for the second quarter shows 5.49% of all auto loan balances at least 90 days delinquent — the second-highest reading in more than twenty years — with 3.0% of balances migrating into that bucket over the past four quarters. Equifax, summarising Moody's, describes subprime auto borrowers as at an all-time high for delinquency. At the same time, cards improved. Commercial bank card delinquencies fell to 2.85% in Q2 from 3.04% a year earlier; the broader measure including private-label and subprime cards fell to 2.69% from 2.87%. FICO's 30-day auto measure ticked down to 2.6%. The average US FICO score held at 714. The reconciliation is in the distribution, not the average. FICO says subsequent 90-plus-day delinquency rates on mortgages and autos rose only in the lowest score bands and were flat everywhere else. Layer on student loans: about 3.2 million borrowers with a payment due were reported delinquent, and they lost an average 38 FICO points year-on-year, while borrowers with no recent delinquency gained 16. So one cohort is being pushed down the score ladder — which raises the price of everything they borrow next — while the aggregate looks fine. That is a slow, compounding problem rather than a spark. The equity market has picked its side. Credit Acceptance, the deep-subprime auto lender, sits 0.36% off its recent high and is up 5.1% over twenty days. Ally, the prime-to-subprime auto lender, is down 4.2%.

Crypto and TradFi contagion31 Aughigh

Strategy raised more equity than it spent on bitcoin

The premium to net asset value is the entire financing model of the treasury companies, and it is the fastest-moving number in this beat.

Strategy bought $369.7m of bitcoin in the week to 30 August, ending a ten-week pause. The interesting part is the funding. It raised $602.8m through common stock sales and bought back $151.8m of preferred shares (Bloomberg). So the company issued roughly $233m more equity than it put into coin, and used part of the difference to retire preferred stock — the instrument that carries a fixed claim ahead of common holders. That is not accumulation for its own sake. It is a balance-sheet simplification funded by selling shares, and it only works while the shares fetch more than the bitcoin behind them. That is the mechanism the whole digital-asset-treasury sector runs on. If the equity trades above the value of the coins it holds, every share issued buys more coin per existing share, and the ratio improves. The premium justifies the issuance and the issuance justifies the premium. When the premium goes, the flywheel does not slow down — it reverses, because issuing shares below asset value destroys the ratio and the only remaining lever is selling coin. Right now the equities are outrunning the asset again. Bitcoin is up 22.8% over twenty days but down 3.0% over five and 0.9% on the day. Strategy is up 40.2% over twenty days and 8.4% over five. BitMine is up 45.6% over twenty days and sits 1.0% below its high; SharpLink is up 39.1%. Coinbase is up 28.4% and within 1.4% of its high. The treasury companies are being valued as leveraged bets on a coin that has stopped rising this week. That gap is the thing to watch, not the coin.

Private credit and BDCs1 Sephigh

The gates now have names

Redemption limits spreading to funds without exposure is how a credit problem becomes a liquidity problem.

Australia's private credit gating has moved from anecdote to named funds. MA Financial has imposed a temporary redemption limit of up to 1% of funds under management per month. CVS Lane has reportedly suspended processing of both application and redemption requests across its funds. The trigger was the A$2.3bn collapse of Sydney builder Bathla Group — and, as Bloomberg reported, one of the funds restricting withdrawals has no exposure to Bathla at all. A gate is not a default. It is the opposite of one: it stops a fund having to sell illiquid loans into a bid that does not exist, which protects the remaining investors and the marks. What it does to the investor who wanted their money is a different question, and what it does to the next investor deciding whether to commit is a different question again. Gates travel by reputation, not by exposure. That is why a fund with no Bathla loans restricted withdrawals anyway. The US numbers behind the same asset class keep hardening. Jefferies has Q2 non-accruals at Blackstone Secured Lending at 3.6%, against 0.3% in Q1 2025; Golub at 2.9%, Blue Owl at 2.8%, Ares Capital at 2.4% — the highest at these funds since at least 2021. Across the ten largest BDCs, non-accruals reached 3.95% of cost. Across 83 funds, investments still held were written down by $2.04bn in the quarter, the largest quarterly write-down in a thirteen-quarter dataset. The listed vehicles disagree, loudly. Blackstone Secured Lending is 0.1% off its recent high. FS KKR is up 11.8% over twenty days, Blue Owl's manager up 11.7%, its BDC up 4.9%. Either the write-downs are already in the price, or the price has not been asked the question.

Private credit and BDCs31 Augmedium

Marked at 64 cents, still called performing

The income these listed private-lending funds report is being sustained by loans their own managers value in the sixties. That means the dividend is a lagging indicator of the portfolio's actual health.

Two loans in the listed private-lending universe are worth looking at closely. Cornerstone OnDemand and Symplr are carried in the mid-60s (63 to 65 cents on the dollar) and neither one has been flagged as a loan where the lender has stopped counting the interest, because it stopped arriving (Accelerate Shares, via web sweep). That combination is the mechanism worth understanding. A mark is a valuation: what the manager thinks the loan is worth today. Deciding to stop booking interest is a separate accounting decision, a judgment that the money is no longer expected to show up. The two can diverge, and right now they are diverging. A loan marked at 64 is a loan the manager believes will lose a third of its principal, and it can still be feeding full contractual interest into net investment income, the number from which the dividend is paid. If that interest is being paid by adding to the debt rather than handing over cash, no money arrives at all; the borrower just hands over more paper. The share of loans paid that way has gone from roughly 6% in early 2022 to about 10% in early 2026 (Blue Vault, same sweep). The aggregate picture is consistent. Across 83 funds, $2.04 billion of investments were written down last quarter, the largest quarterly write-down in the thirteen quarters that dataset covers. Loans where the lender has stopped counting interest at the ten largest funds reached 3.95% of debt at cost, up from 3.75%. The published figures do not agree with each other, and you should know why: some are struck at cost, some at the manager's current valuation, some are medians of a small machine-readable subset. Morningstar DBRS has the average at 3.4%; another panel puts the median at 0.8% of current valuation. The direction is the same everywhere. Meanwhile Ares Capital is up 4.6% over twenty days, FS KKR 11.8%, Blackstone Secured Lending 6.1%. The market is reading these disclosures and buying.

The AI capex bubble31 Augmedium

Broadcom wants up to $100 billion to finance its own buyers

The circular financing of the AI buildout is migrating from chip sellers' own books to private lending funds, where the loan is nobody's disclosed exposure.

A Broadcom-backed financing vehicle was in talks on August 20 to raise $70 billion to $80 billion of debt, with the full structure potentially reaching $100 billion, to fund chip purchases by AI companies including Anthropic (web sweep). In June, Broadcom partnered with Apollo and Blackstone on an AI-infrastructure funding platform, and CNBC reports it is going to the bond market for more money to finance chip purchases. This is the second such structure we have written about in a week. Nvidia has partnered with Wall Street firms to mobilize $500 billion of third-party capital for AI infrastructure. The pattern is the same and it is old: vendor financing. The seller arranges the money the buyer uses to buy the seller's product. Revenue is booked immediately and in full; the credit risk sits somewhere else. The interesting part is where else. In the classic version (Lucent and Nortel in 1999) the chip seller lent the money itself, so the amount owed sat on the seller's own books where an auditor and a stock analyst could see it. Here the lender of record is a private-credit manager and, behind it, an insurance company's investment portfolio. Broadcom's accounts will show a chip sale. Apollo's and Blackstone's funds will show a loan against equipment and contracts. Nowhere in either set of statements does the phrase "we financed our own customer" appear. What we do not know: the structure. This is a talks-stage figure reaching us at second hand, and the terms (who has recourse, what secures the loan, who eats the first loss) are not in anything we have seen. That matters more than the headline number. Apollo trades up 6% over twenty days; Blackstone 6.4%.

Hidden leverage and shadow banking31 Augmedium

The SEC is asking banks what they knew about 4x borrowed money

When a hedge fund borrows from several banks at once, no single bank sees the total. The only time anyone measures the aggregate is after a fund has almost failed.

On August 24, Reuters reported that the SEC, the agency that polices American markets, has sent subpoenas to Wall Street banks that dealt with the hedge fund Situational Awareness, seeking information about the fund's trades and its use of borrowed money after its near-collapse last month. Other coverage puts the fund's total exposure at roughly four times the money it actually had (web sweep). Read the specific ask. The subpoenas seek the fund's communications with lenders about the leverage it was taking on. That is not a question about whether the trades were legal. It is a question about whether each bank knew what the others were doing. This is the structural hole in how hedge funds borrow, and it has not been fixed. A fund borrows against its positions from several banks at once. Each bank sees its own book, prices its own requirements for collateral, and is comfortable. None of them sees the total. The fund's total gross exposure is a number that exists only inside the fund. When the position turns, every lender discovers the same thing at the same moment, and they all try to sell the same collateral. The Reserve Bank of Australia's governor put the general version of this plainly on August 27, talking about private lending funds: "People don't know where the leverage is." Goldman Sachs is down 1% over five days, Morgan Stanley 0.3%. The banks that lend to hedge funds are not being priced as if anyone lost money here, and on the evidence available nobody did. The fund nearly collapsed and did not. That is the honest read: a near miss, followed by a regulator trying to find out how near.

The dollar, gold and reserve status31 Augmedium

Australia's central bank cut its dollars to 45%

When a central bank shifts its reserves away from the dollar, it arrives as a press release rather than a price move, which is why it is easy to underweight.

The Reserve Bank of Australia, the country's central bank, has disclosed that it reduced the US dollar share of its foreign reserve portfolio to 45% from 55%, raising the euro to 30% from 20% (web sweep). Deutsche Bank called it a quiet reduction. Ten percentage points is a large move by the standards of an institution that publishes its allocation once a year and changes it roughly never. Central bank reserve managers are not traders. They set a target, they are slow, and the decision is made by a committee that writes it down. So this is not a market signal in the way a currency move is. It is a stated intention, which is in some respects more durable. Set against the trend: a Kansas City Fed working paper puts the dollar's share of global foreign-exchange reserves at 57% in the first quarter of 2026, down from 72% in 2000. That is about half a point a year. The same body of research makes a point worth holding on to: the euro has captured essentially none of the dollar's lost share globally. Roughly a quarter went to major Asian currencies and three-quarters to non-traditional reserve currencies. Australia is doing something unusual by going to euros. Meanwhile India's reserves hit a record $729.33 billion in the week to August 21, with gold up $2.801 billion to $114.218 billion. The price tape does not corroborate any of this as an emergency. The dollar index, which tracks the dollar against a basket of other currencies, is at 99.5, down 0.2% over twenty days. Gold is up 5.7% over twenty days to $4,488 and down 3.2% over five. What is happening is a slow, announced reallocation, not a run.

Crypto and TradFi contagion31 Auglow

A bitcoin-backed mortgage and a government money-market token as collateral

Crypto is being plumbed into mortgage collateral and fund margin at the same moment its listed wrappers are trading at a widening premium to the assets they hold.

Two things happened in the same week and they run in opposite directions. Coinbase and Better launched bitcoin-backed mortgages, with a waitlist projecting more than $260 million of demand. And Franklin Templeton got SEC clearance to use its $2.6 billion tokenized money-market fund, BENJI, inside ETFs and mutual funds and as collateral pledged to lenders (web sweep). The first is crypto entering the real-economy collateral chain. The mismatch to think about: a 30-year fixed obligation on a house, secured by an asset that trades 24 hours a day and has moved 24.6% in twenty days. Every crypto-backed loan needs a maintenance margin and a trigger at which the lender sells. Nobody has yet run that machinery through a weekend crash at scale with houses on the other side. The filings we have seen do not say what the loan-to-value ratio or the cure period is. The second is the reverse: government money-market exposure being wrapped so it can move instantly and be pledged as collateral. Circle, the company behind the USDC stablecoin (a digital token meant to be worth exactly one dollar), minted about 5 billion USDC in a week. JPMorgan's Kinexys, with Ripple, Mastercard and Ondo, redeemed a tokenized US government bond fund cross-border in under five seconds. Speed is the feature and also the risk. The friction being removed is the same friction that gives a fund manager time to sell assets in an orderly way when investors all want their money back at once. The stock market's verdict is unambiguous. Over twenty days Strategy is up 40.2%, Bitmine 45.6%, SharpLink 39.1%, Coinbase 28.4%, against bitcoin's 24.6%. The companies that hold bitcoin using borrowed money are re-rating faster than the coin they hold, which is what happens when the premium to the value of the underlying assets expands.

Bond market dysfunction31 Augmedium

Washington bought yen in August. It is not holding.

The largest foreign owner of US government bonds has a domestic alternative paying 2.95% for the first time in thirty years, and a currency its own government cannot hold up.

The interest rate on the benchmark Japanese government bond touched 2.95% on Monday, a 30-year high, and the yen weakened again after markets read Kevin Warsh's Jackson Hole remarks as leaving the door open to further rate hikes (Nikkei). The context is what makes this worth a paragraph. In early August the US Treasury Department, which borrows the money the government spends, joined Japan's Ministry of Finance in buying yen, the first joint intervention since 2011, and the first time since 1998 that Washington had bought yen alongside Tokyo. It worked, briefly: the yen went from around 164 to the dollar in late July to around 158. Bessent framed the action as driven by financial stability concerns rather than trade (Reuters column, via Astro Awani). The plumbing is this. Japanese institutions own an enormous stock of foreign bonds, much of it American. Two things determine whether they keep it: the gap between what they earn abroad and what they earn at home, and the currency. A US 10-year bond at 4.73% against a Japanese government bond at 2.95% still favors going abroad, until you pay to insure against the currency moving, at which point the gap can vanish. If the yen falls further, unhedged holders face losses and hedging costs rise; if Japanese bond yields keep climbing, staying home starts to pay. Either route ends in Japanese money selling foreign bonds. Intervention cannot change the yield gap. It can only buy time for policy to. On the tape, the yen ETF is down 1.9% over twenty days: the market has taken back most of what the intervention bought. This is the slow-moving version of a funding shock, and it is not close to one yet.

What would change our mind

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

Japanese life insurers or banks disclosing an actual shift out of foreign bonds into domestic JGBs — that would turn the 3% milestone from a price signal into a flow.

Would move the number

High yield spreads breaking above 350bp, or the HY index moving more than 20bp in a single session; at 263bp credit is still refusing to confirm anything the bond market is saying.

Would move the number

A BDC or private credit fund in the US — not Australia — restricting redemptions, which would take the gating story from a regional property problem to a structural one.

Would move the number

Any disclosure of the residual value assumptions or loss-sharing terms inside the Broadcom-Anthropic chip vehicle; it would let us size a $370bn exposure that is currently a range in tier-2 reporting.

Would move the number