Private credit lends $2.4bn against GPUs at nine percent

Archived reading, published Sun, 30 Aug 2026 10:17:18 UTC (22 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility90

how much tinder is stacked up — moves slowly

Ignition34

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 11 sources and rewrites this page.

Status: Fragility up one to 90: the Blue Owl/Pimco $2.4bn loan to IREN is new borrowed money stacked on this week, not a sharper measurement of old borrowed money. It is thirty-month money at 9%, secured on graphics chips. Ignition holds at 34. This is the seventh consecutive run on the same August 28 tape: the VIX, the market's gauge of how much turbulence traders expect over the next month, sat at 14.43 (under 20 is calm). The extra interest that shaky borrowers pay compared with the government was 2.63 percentage points and had tightened 7% over twenty days. For safer borrowers it was 0.79 percentage points. Government bonds that do not come due for twenty or thirty years cleared without failed auctions, no sign of dealers running out of room on their books, and no forced selling.

Reporting from 28 Aug to 30 Aug

Crash points

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

What’s moving the needle

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

The AI capex bubble30 Aughigh

Nine percent, secured on chips, due in thirty months

Private lending funds are now betting that AI chips will hold their value over a loan term shorter than the chips' assumed useful life.

Blue Owl funds have led a $2.4bn debt package for IREN, the Sydney-listed former bitcoin miner, to buy Nvidia Blackwell Ultra graphics chips for a data-center campus in Canada. The structure, per the filing reported by Bloomberg: $1.2bn in senior secured term loans and $1.2bn of senior secured notes, both at 9%, both coming due two and a half years after funding, drawn in stages as equipment is purchased. Pimco is also in. Put that next to the other AI data-center borrowing priced this month. QTS sold $3.9bn of bonds rated safe enough for conservative investors at about 7.23%, for a Microsoft-tied Georgia site. IREN is paying roughly 1.8 percentage points more, and the difference is what the lenders can seize if things go wrong: QTS lenders have a building with power. IREN's lenders have racks of chips. The loan's due date is the part to watch. The big cloud companies write off AI servers over roughly five to six years. This loan comes due in thirty months. Whoever holds it in early 2029 will either have been paid down by contracted cloud revenue, or will be trying to borrow again against second-hand Blackwell Ultras in a market where nobody has yet observed what a three-year-old graphics chip fetches. The question of what those chips will be worth does not wait for the depreciation schedule; it falls due at the refinancing. This is the migration story in one deal. No bank is lending here. The loans sit in Blue Owl and Pimco funds, valued quarterly by the managers themselves, and Blue Owl is the same platform whose funds that lend to mid-sized private companies had 2.8% of their loans go bad enough that the lender stopped counting the interest (Jefferies' figure for the second quarter, up from 1.4% in the first quarter of 2025). The stock market took the other side in the same week. IREN fell 12.6% on August 28 and 15.4% over five days, and sits 21% below its recent high. Debt ranks ahead of shares in a bankruptcy, so both views can be right. The gap between them is still the most informative thing on the screen.

Bond market dysfunction29 Aughigh

Nobody's auction failed

Government bonds that do not come due for thirty years sold off hard and still found buyers. The risk being created here is a policy precedent, not a funding failure.

It is worth writing down the things that did not happen in the US government bond market this month, because the absence is doing the work. The interest rate the government pays to borrow for thirty years reached about 5.33%, a 19-year high. On August 19 the Treasury Department, which borrows the money the government spends, announced it would at least double the size of its program to buy back its own long-dated bonds and support trading conditions, from $2bn to at least $4bn per operation, running September 9 to November 4. Interest rates fell on the news, then reversed within a day: the ten-year rate finished the week near 4.73% and the thirty-year near 5.27%, both above where they started. What the reporting does not show, on repeated checks, is any of the actual markers of a market breaking down. No auctions that failed to attract enough buyers. No evidence that the big banks that stand between the government and investors ran out of room on their books. No forced selling. Trading was orderly. Nobody has produced a poorly received auction or a breakdown in the plumbing for this window. This was a price move driven by investors demanding more compensation for tying their money up for a long time with a government running deficits this large, not a mechanical failure. That distinction matters for how we score this. A market that repriced and kept clearing is a market that tested a level and held. On the plumbing, things are genuinely fine, and we are not going to pretend otherwise because the interest-rate number is dramatic. The fragility created here is different and slower: precedent. The buyback tool has now been scaled up in response to a level of interest rates that was uncomfortable rather than broken. Once an intervention has been used at 5.3% in an orderly market, the bar for withholding it at 5.6% in a disorderly one is lower, and everyone pricing long-term government debt knows it.

Fed, Treasury and policy29 Augmedium

Two arms of the state pulling on the same long-term debt

If the Treasury Department is managing the interest rate on long-term bonds while the Federal Reserve manages the overnight rate, the compensation investors demand for tying money up becomes a political price.

Tom Hoenig, the former Kansas City Fed president, told Semafor at Jackson Hole that the Federal Reserve, America's central bank, "has been a little too quiet" about the Treasury's campaign to bring long-term interest rates down by buying back its own debt. "What they're trying to do is not identify with Treasury, and so they're not going to mention it." The mechanical problem is easy to state. The Treasury Department is buying government bonds that do not come due for decades (at least $4bn per operation from September 9) to pull long-term debt out of the market. Fed Chair Kevin Warsh is shrinking the Fed's holdings of the same bonds, which puts that debt back into the market. Both are arms of the same government acting on the same instrument in opposite directions, and only one of them is supposed to be setting the price of money. It gets sharper if the Fed raises interest rates. Futures markets raised the odds of a near-term move after Warsh's August 28 Jackson Hole remarks. Hoenig's read: Warsh has "a very difficult needle to thread", and if he moves he will need a great many explanations beforehand — "and I don't mean to the market, I mean to the Treasury." That is a former central banker saying out loud that a decision on interest rates now requires managing the government's borrowing arm. The technical name for this is fiscal dominance, and it is not established by one interview. But the observable test is simple and near: if buyback sizes rise above $4bn while the Fed continues shrinking its holdings and says nothing about it, the Treasury Department is setting long-term rates and the Fed is setting the overnight rate, which is a different institutional arrangement from the one described in the textbooks. Neel Kashkari's position, stated August 23, is that the market is "functioning as it should" and the Fed can simply focus on its own rate. Both things cannot stay true indefinitely.

Hidden leverage and shadow banking29 Augmedium

Principal-protected, over ten percent, now frozen

Retail money reaching institutional strategies through offshore vehicles is the same wiring as frozen private lending funds, with less documentation.

About HK$1bn ($128m) belonging to mainland Chinese and Hong Kong investors is frozen in accounts linked to ZD Group, a Hong Kong "joint family office" founded in 2018, Caixin reports. Investors have formed groups to seek repayment and gone to the police. The Securities and Futures Commission, the agency that polices Hong Kong's markets, declined to comment. The product is the interesting part. ZD sold small investors the chance to act as cornerstone or anchor investors in Hong Kong stock-market listings, a role normally reserved for institutions. The first tier required HK$1m, offered a guarantee of getting your money back plus a fixed return above 10%, plus about 30% of additional gains. A later tier required HK$2m, dropped the guarantee, and offered up to 70% of trading gains. Some investors pooled money to reach the minimums. A fixed return above 10% with a guarantee on your principal, from a strategy that buys newly listed shares, is not a feature of the strategy. It is a promise, and someone's balance sheet has to fund it when the strategy underperforms. Nobody in the chain here appears to have had one. The chain itself is the second lesson. The problem sits not at ZD Group but at Renascence Global Capital SPC, a company set up in the British Virgin Islands to hold the assets off the parent's books, with accounts at an overseas branch of a Chinese bank. Money previously flowed through a different vehicle, MS Global Investment LPF, whose general partner changed in 2025; investors were then told to send funds to RGC accounts. The only confirmation of the freeze came verbally, from a law firm, with no written documentation. Separately, a March complaint alleged ZD engaged in "IPO cornering," taking most of an offering to control how many shares were available to trade and inflate the price. On July 30 the SFC ordered Futu Securities to freeze up to HK$125m over suspected false demand for IPO shares. $128m is small. The structure (retail money, a promised return, an offshore shell company, verbal confirmations) is not.

Household credit29 Aughigh

Brazilians now spend 26.6% of income servicing debt

A record share of income going to debt payments, combined with a regulator moving on lenders rather than borrowers, is what the late stage of a consumer lending cycle looks like from the inside.

Brazil's household debt payments as a share of income, excluding mortgages, hit a record 26.6% in June, up from 25.7% at the end of last year, central bank data released on Friday show. That means the average Brazilian household is spending more than a quarter of its income just keeping up with non-housing debt. More than half of the country's credit card users, 52.8 million people, carry revolving balances or installment loans. The central bank is preparing to act, and the shape of the action is the informative part. The IMF, which lends to countries in trouble, recommended in July a cap on how much of an individual's income can go to debt payments. Brazilian policymakers are instead going first at the lenders: limits on institutions rather than eligibility rules on borrowers. Citi read the Financial Stability Committee's statement as a shift of focus, away from the pace of lending growth and toward the cost and structure of the products themselves. Two people familiar with the discussions told Reuters the concern is that expensive credit products, weak transparency, and thin financial education are combining to push more lending onto already-indebted borrowers. Why it belongs in this publication rather than an emerging-markets note: a cap on borrowers bites immediately and visibly and hurts the borrower who is refused. A curb on lenders bites slowly and hurts the lender's growth rate, and it is what regulators reach for when they want to slow a lending cycle without being blamed for stopping it. Watching which tool a central bank picks tells you how far along the cycle it thinks it is. The US comparison is not reassuring in one direction and is reassuring in another. American household stress is real but narrow. FICO puts the rise in 90-day mortgage and auto late payments exclusively in the lowest credit-score bands, with higher bands flat, and late payments on car loans to the riskiest borrowers at 6.13% in July. Brazil's number is not narrow. It is the average.

The dollar, gold and reserve status28 Aughigh

Treasury bought yen with euros and will not say how much

The US is intervening in currency markets through a fund that reports on its own schedule, and the position is currently under water.

Scott Bessent spent Friday attacking Elizabeth Warren on X, offering her staff a lesson in "Foreign Exchange for Dummies". Underneath the insult is a confirmed fact and a conspicuous gap. The fact: after the yen fell to a 40-year low, the Treasury Department sold euros held in the Exchange Stabilization Fund and bought yen. Bessent's written reply to Warren states it plainly — "Treasury exchanged existing Exchange Stabilization Fund foreign-currency assets for yen. No new congressional appropriation was involved, and no credit was extended to Japan." The gap: his one-page response did not disclose how much yen was bought, the exchange rate he got, or what the position is now worth. Nor whether the European Central Bank was consulted before the euro sale. A Reuters photograph of Bessent's notepad on July 31 read "Buy Japanese Yen (JPY) $5-10 bil." The Exchange Stabilization Fund is the mechanism worth understanding. It holds foreign-currency assets and can be deployed without going to Congress, which is precisely why it exists and precisely why it is opaque. There is no obligation to disclose the position publicly on any timetable. A loss on this trade would be a change in the composition of a fund most people have never heard of, disclosed at the Secretary's convenience. And the trade is not working. The yen is down 0.78% over five days and 0.69% over twenty; the dollar index, which tracks the dollar against a basket of other currencies, has gained 0.83% over five days. Warren's rejoinder ("his effort to prop up a foreign currency hasn't worked") is, on the numbers, correct. One related plumbing item from the same window, reported by Reuters: the Fed widened a facility so the Bank of Japan could pledge its US government bonds as security for cash rather than sell them. The architecture being built here is designed to stop Japan from becoming a seller of US government debt.

Signs of the times

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

The notepad said five to ten billion

A Reuters photograph on July 31 caught Treasury Secretary Scott Bessent's notepad reading "Buy Japanese Yen (JPY) $5-10 bil." His written reply to Senator Warren about the intervention, a month later, still did not disclose how much yen the Treasury Department actually bought, at what rate, or what the position is worth now.

interventiondisclosure

28 AugCNBC

Lenders in, shareholders out, same week

Blue Owl and Pimco agreed to lend IREN $2.4bn at 9% to buy Nvidia chips. In the same week, IREN's shares fell 15.4%, leaving them 21% below their recent high. The lending market and the stock market are looking at the same Canadian data-center campus.

private creditgpusdivergence

29 AugBloomberg

$711bn of racks, up 214% in a year

TrendForce estimates that NVL72 rack systems alone (GB300, VR200 and VR300) will generate more than $711bn of revenue in 2027, a 214% annual increase, and puts Nvidia's data-center business at 93% of total company revenue. The projected one-year increase is larger than the annual GDP of most countries.

capexconcentration

29 AugNewtalk (TrendForce)

Principal-protected IPO money, confirmed verbally

A Hong Kong family office sold mainland investors principal-protected IPO products promising fixed returns above 10% for a HK$1m minimum. About HK$1bn is now frozen in a company set up in the British Virgin Islands, and the only confirmation investors have of the freeze came verbally from a law firm, with nothing in writing.

shadow bankingretailoffshore

29 AugCaixin

The rumour mill

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

Confirmed

“OpenAI's new inference chip, Jalapeño, built with Broadcom, beat Nvidia's GB300 in OpenAI's own benchmarking.”

Bloomberg reports OpenAI's chief chip executive saying Jalapeño led the GB300 on AI work per unit of power and on response speed. Relevant to anyone lending against Blackwell chips at thirty-month maturities: the resale value of a graphics chip depends on what the alternatives look like in 2029.

Claimed by Meet Kevin

Partly true

“Nvidia wants to give OpenAI a $125bn guarantee.”

The number is wrong but the shape is right. Talks reportedly reached $250bn, the announced guarantee is $105bn, and the WSJ put Nvidia's exposure in the Ohio structure at under $120bn. A vendor guaranteeing over a hundred billion dollars of a customer's obligations is remarkable whichever figure you use.

Claimed by Wealthion

Confirmed

“Gold and bitcoin ETFs took in a combined $7bn in a week, a record for any five-day period.”

Bloomberg's flow data, per ETF analyst Eric Balchunas: roughly $7bn across gold and bitcoin funds in five trading days, about $3.4bn into GLD and $1.5bn into IBIT. Worth setting against the tape since: GLD is down 3.4% over five days and silver down 4.3%.

Claimed by Mark Moss

Partly true

“US auto loan terms are now approaching ten years.”

Marketplace reported some banks offering 120-month loans, but Bankrate puts the typical range at 24 to 84 months and MarketWatch puts the average at just under six years. A decade-long car loan exists; it is not the market.

Claimed by Heresy Financial

Earlier developments

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

Bond market dysfunction28 Aughigh

Tokyo prices the end of free money

The world's largest exporter of cheap capital is now paying visibly more to fund itself, which weakens the bid that has quietly financed everything else.

Japan's Ministry of Finance published its budget request for the fiscal year starting April 2027 on August 28. The total cost of servicing the national debt comes to ¥36.64tn, up ¥5.36tn from this year's initial budget. Inside that, the line for interest and discount charges alone is ¥16.59tn, against ¥13.04tn in the current year: an increase of ¥3.55tn, roughly 27%, in a single twelve months (MOF). This is what a rate cycle looks like when it finally reaches a government that owes more than any other major economy and has spent a generation paying almost nothing on it. Japan rolls its debt continuously. When an old bond comes due, the government issues a new one, and the interest rate on the new one replaces the rate on the old one. That means the interest bill does not jump the day rates rise; it creeps up over years as cheap old debt is replaced by expensive new debt. The budget request for fiscal year 2027 is the first in which that creep has fully caught up. Worth noting: the request also includes a ¥1.30tn transfer under Article 6 of the Public Finance Act, a line that was zero in the current budget. Why this matters outside Japan: Japanese banks, insurers and pension funds have spent two decades borrowing at zero and buying foreign bonds, including US government bonds and, more recently, the private lending and infrastructure debt that finances much of what we write about. With the interest rate on a ten-year Japanese government bond near 2.88%, the math on those trades changes at home. The budget makes the fiscal cost explicit, and explicit costs become political. The request is not the final budget. Ministries always ask for more than they get, and the interest line is built on an assumed rate that the finance ministry sets conservatively. But the direction is not in doubt, and it arrives a month after the US Treasury sold euros to buy yen at a 40-year low for the Japanese currency.

Crypto and TradFi contagionmedium

The coins held. The wrappers fell seven percent.

Digital-asset treasury companies are the leveraged sleeve of the crypto trade, and their stock prices break before the coin does.

On August 28 the publicly traded crypto complex had a bad day that the crypto market itself did not have. Strategy fell 7.35%, BitMine 7.18%, SharpLink 7.44%, Galaxy Digital 8.22%, Coinbase 6.43%. Bitcoin, meanwhile, sits at $78,110, down 0.12% on the day and up 22.2% over twenty trading sessions. Solana had its best month since 2024, up about 50% for the month as of the day before. What moved was not the coin. It was the premium. A digital-asset treasury company is essentially a box holding coins plus, usually, some convertible debt. It creates value for existing shareholders only while its stock trades above the value of the coins inside it. When the stock price is higher than the coins are worth, the company can sell new shares, use the cash to buy more coins, and make every existing share worth more. When the stock falls to or below the value of the coins, that machine stops. The manager is left with a fixed cost base, debts coming due, and no reason to exist as a company rather than a pile of coins. So the stock can fall hard on a day the coin does nothing. What moved was the market's willingness to pay a markup for the wrapper. The likely competitor is visible in the flow data. Bloomberg's figures, which we checked and found accurate, show gold and bitcoin exchange-traded funds taking a combined $7bn over five trading days, a record for that window. An ETF gives a buyer the same exposure to the coin with no premium, no stack of convertible debt, and no management fees. Every dollar that goes there is a dollar not paying up for a treasury company's shares. We do not have current figures on how much these companies' stocks trade above or below the value of their coin holdings, so this is inference from the price action rather than something a filing states.

The AI capex bubble29 Augmedium

The buyers of the racks are getting weaker

Chip revenue is increasingly booked against buyers whose ability to pay depends on debt the seller has underwritten.

TrendForce estimates that the latest generation of high-end server rack systems (NVL72-class, including GB300, VR200 and VR300) will generate more than $710bn of value in 2027, a 214% increase, on shipment growth above 50% and a doubling of average selling price for the next-generation Vera Rubin chips. Buried in the same research is the number that matters more. The five big North American cloud providers' share of rack purchases is expected to fall from about 70% in 2025 to 60% in 2026, with the difference made up by Tesla (including xAI and SpaceX) and CoreWeave (TrendForce, via Newtalk). Read that as a statement about who can pay rather than who is buying. Microsoft, Amazon and Alphabet buy chips out of the cash their businesses generate. The replacement buyers do not. IREN is funding its next-generation chip purchases with $2.4bn from Blue Owl and Pimco at 9% interest, coming due two and a half years after the money is drawn (Bloomberg). Lambda has just raised about $1bn of short-dated private debt arranged by JPMorgan for chips that Microsoft will lease, weeks after a separate $926m loan (Bloomberg). Which is why the vendors are now writing the credit themselves: Nvidia's $105bn guarantee covering the residual value of the Ohio campus, Google's $44bn rent guarantee for Anthropic, and Broadcom's pledge to buy back chips that do not sell. When the buyer at the margin cannot carry the debt, the supplier carries it, and revenue growth and credit exposure become the same line. TrendForce is a forecaster, not a regulatory filing. Treat the 2027 numbers as an estimate. Treat the shift in who is buying as the point.

Household credit29 Aughigh

Everything is fine except the bottom fifth

Household credit is splitting in two rather than deteriorating across the board, which is worse for the bonds backed by the weakest loans and better for everyone else.

Fitch, one of the three firms that grade borrowers' likelihood of paying you back, puts US subprime auto loan delinquencies at 6.13% in July, up from 5.80% at midyear. Moody's, using Equifax data, calls the subprime series an all-time high. FICO counts about 3.2 million consumers with a fresh student-loan delinquency, carrying an average 38-point year-on-year decline in their credit score, while consistent payers gained 6 points and borrowers with no delinquency gained 16. And yet the overall numbers improved. Major card issuers' average net charge-off rate (the share of balances they gave up on collecting) fell to 3.28% from 3.42% in June. Bank-card 30-day delinquency is 2.85%; 60-plus-day delinquency across all cards is 2.69%, down from 2.87% a year ago. FICO's own breakdown is explicit: 90-plus-day late payments on mortgages and auto loans rose exclusively in the lowest credit-score bands and were flat everywhere else (Fitch via Auto Remarketing, Equifax/Moody's). This is the shape that matters for the financial machinery built on top of these loans. A pool of prime credit-card debt is diversified and gets healthier when employment holds. A bond backed by subprime auto loans is a concentrated pool of exactly the borrowers whose late payments are at a record, and it was structured around a loss assumption written when the delinquency number was 5.6%. The stress does not show up in the national average because the national average is dominated by people who are fine. We are lowering our estimate of household fragility by a notch on this. The consumer piece of the picture is narrower than we had been carrying it: not absent, but confined to the weakest borrowers. The stock market agrees without excitement: Ally Financial down 2.6% over twenty sessions, Synchrony up 3.5%, Credit Acceptance within 4% of its high.

The dollar, gold and reserve status28 Augmedium

Australia sold dollars and bought euros, not gold

The move away from the dollar in central-bank reserves is real, gradual, and not the same trade as the gold rally people keep attaching to it.

The Reserve Bank of Australia has disclosed that it cut the US dollar share of its foreign reserve portfolio to 45% from 55% in the 2024/25 fiscal year, lifting the euro from 20% to 30%. The IMF's global tracker of central-bank reserves still has the dollar at 57.13% of the total in the first quarter of 2026, drifting down roughly half a percentage point a year (Modern Diplomacy). India's weekly reserve report shows gold holdings up $2.8bn to $114.2bn in the week to August 21. Two things are worth separating here. The first is that a ten-percentage-point shift by one central bank is a large single move and a rounding error in a $12tn global reserve pool. The aggregate is grinding, not breaking. The second is where the money went. The RBA moved into euros. That cuts against the tidy story in which every central bank leaving the dollar is buying gold bars. The price tape says something else again. Gold fell 1.73% on August 28 and 3.22% over five sessions, silver 4.28%, and this happened in the same week that gold and bitcoin exchange-traded funds took in a record $7bn. Record inflows into a falling price means someone larger was selling into them. Central banks are slow, price-insensitive and report quarterly; ETF buyers are fast, price-chasing and report daily. When the two meet, the fast money sets the price today and the slow money sets the direction over months. Gold is still up 12.3% over twenty sessions.

Fed, Treasury and policymedium

Two data prints the Federal Reserve cannot easily ignore

Softer economic data plus government borrowing costs driven by fiscal risk means the Federal Reserve cutting interest rates could push long-term rates higher, not lower.

The Chicago Business Barometer, a regional survey of manufacturing and business activity, fell to 47.1 in August from 57.6 in July (anything below 50 signals contraction): a 10.5-point drop against expectations near 58. We checked this against the ISM Chicago release. The figure is right. The widely circulated "58.3" prior reading is not; that was the expectation, not last month's actual print. Separately, the Bureau of Labor Statistics' preliminary annual revision to payroll employment came in at minus 79,000, against Goldman Sachs' expectation of an upward revision of between 50,000 and 450,000. A single regional survey is a single regional survey, and the Chicago index has a long history of dramatic monthly swings that snap back. The payroll revision is more consequential: it restates the level of employment through March 2026, which means the labor market everyone has been describing for five months was slightly smaller than the description. The awkward part is what is happening to borrowing costs at different time horizons. The interest rate on two-year government bonds is at 4.20%, and the overnight rate that banks charge each other (SOFR) is at 3.64%. But the interest rate the government pays to borrow for thirty years has spent August at a 19-year high, touching 5.31%. Normally, weak economic data is a reason for markets to expect rate cuts, and rate cuts normally pull borrowing costs down across the board. Here they would not: the interest rate on long-dated government bonds is being driven by the sheer volume of debt Washington needs to sell and by the extra compensation lenders demand for tying their money up for decades, not by expectations for where the Federal Reserve sets its overnight rate. Cutting rates into that would push short-term rates down and long-term rates up, widening the gap between them. That widening gap is the thing that hurts anyone holding long-dated bonds: insurers, banks, and the mortgage market. The gap between the two-year and ten-year government bond rates is 0.47 percentage points. Watch it.

The AI capex bubble29 Augmedium

What the $105bn guarantee actually covers

The chip vendor is now carrying the property risk of its customer's landlord. That is exactly the kind of exposure that sits outside a bank and only gets a price when someone demands payment on it.

The structure behind Nvidia's OpenAI backstop has been described in enough detail to see the plumbing. The site is PORTS-Pike in Ohio. SB Energy builds it, owns it and operates it. OpenAI takes the tenancy, eight IT-gigawatts of it. Nvidia gets exclusive rights to supply computing power at the campus and, in exchange, promises that the initial 4.25 IT-gigawatts of infrastructure will hold a certain value, with its total obligation capped at $105bn. The promise covers land, power and buildings. It does not cover the chips (TrendForce, via 新頭殼). That carve-out is the interesting part. A residual value guarantee is a promise that an asset will be worth a specific amount at a stated future date, and whoever writes it absorbs the gap if it is not. Nvidia has written one on concrete, transformers and dirt, the parts of a data center with a long life and, in principle, a buyer other than OpenAI. It has not written one on the depreciating silicon it sells, which is the part of the stack whose second-hand value nobody can currently observe. Google has done a version of the same thing for Anthropic, guaranteeing data center rent to a total of $44bn and taking roughly 20% equity in the associated power and building projects. Broadcom, which co-designs Google's custom AI chips, has separately pledged to buy back chips that do not sell. Our fact-check of the circulating $125bn figure found the number is $105bn: talks reportedly began nearer $250bn and were revised down to under $120bn before the announced deal. What must stay true for the structure to work is that the Ohio campus has a tenant at the end of the lease. Not necessarily this one.

Hidden leverage and shadow banking29 Augmedium

Worth $2.4bn, borrowing $5bn

Debt is being sized against customer contracts and depreciating chips rather than against the borrower's own cushion of equity. That is how a downturn in one tenant's demand becomes a credit event at five separate borrowers.

Volta Infra Holdings raised $300m of venture money in August at a $2.4bn valuation. On August 27, JPMorgan began early outreach to lenders on a $5bn debt package for its AI data center buildout (AI Business Weekly, citing Bloomberg). The debt being arranged is roughly twice what the equity market says the whole company is worth. That is not automatically absurd. Project debt is sized against contracted cash flows and the assets it funds, not against the sponsor's equity value. But it tells you what the underwriting now rests on. The same week: Nebius upsized a planned raise from $4.5bn to $5bn against a spending target of as much as $25bn this year, having already borrowed $775m in July secured on existing GPUs and the cash flows from a contract with a well-rated customer. Applied Digital is sounding out about $3.5bn of debt from companies too shaky to be considered safe-grade borrowers for its Delta Forge campus in Louisiana. Goldman is marketing $800m for two Novva sites in Nevada and is expected to lead a $1.085bn secured note for a CoreWeave-leased campus, issued through a financing vehicle affiliated with PowerHouse and Chirisa. Every one of those deals converts a promise from a well-rated tenant into borrowing capacity for a company that has none of its own. The lender's protection is the contract and the equipment. The contract is only as good as the tenant's willingness to keep paying for computing power it may not need. The equipment is GPUs, whose resale market has never been tested by a large forced seller. Companies have borrowed more than $410bn for data centers and AI this year.

The AI capex bubble28 Aughigh

The lenders arrived. The shareholders left.

When equity investors and lenders disagree this sharply about the same borrower, the resolution is usually a refinancing that one of them does not survive.

IREN announced $2.4bn of debt from Blue Owl and Pimco at 9%, coming due two and a half years after funding, to buy Nvidia Blackwell Ultra GPUs for its Mackenzie campus in British Columbia. On the same day it forecast as much as $30bn of capital spending in the 2027 financial year. The shares fell almost 13% (Bloomberg). That combination (credit shows up, equity runs) is now the defining pattern at the heavily borrowed end of AI. Over five sessions IREN is down 15.4% and sits 21% below its recent high. Core Scientific is down 20% over twenty days and 31% off its peak. Applied Digital fell 7.6% in a day, Nebius 4.2%, CoreWeave 4.1% on the week, Talen Energy 11.4% over twenty days. Now look at what did not move. The extra interest that risky borrowers pay over the government is 2.63 percentage points and has tightened 7.4% over twenty days. HYG, the largest fund tracking that debt, is a quarter of a percent from its high. Microsoft closed at its period high, Amazon rose 4% in a session, Alphabet 1.8%. The stock market is repricing the companies that borrow to buy chips while leaving untouched both the companies that buy chips out of their own cash flow and the credit market that funds the borrowers. One of those two markets is wrong. The shareholders are junior and are being asked to fund spending programs many times their own market value. The lenders are senior, secured on the equipment, and paid 9% for a loan shorter than the useful life of what they are lending against. Our reading is that both can be right for about two more years, and only one after that.

Bond market dysfunction29 Augmedium

Three former Treasury officials cannot explain the buybacks

Official support for government bonds that do not come due for twenty or thirty years flows through the books of fifty hedge funds. That is a fragility you cannot see in the interest rate.

Nathan Sheets, Citigroup's chief economist and a former Treasury under secretary, on the decision to more than double buybacks of long-dated government bonds: "I do not think there were any meaningful signs of illiquidity in the market, and it is not clear what bar was met that triggered the expansion of the buybacks at this time. It was an unusual action. I did not find it shocking, but I found it puzzling" (Nikkei). Treasury buybacks are not unusual in themselves. Doing one at double the normal size, into a market that the Federal Reserve (America's central bank, which sets interest rates) and former Treasury officials both describe as working, is. On August 19 the Treasury Department said buybacks of government bonds maturing in ten to thirty years would rise from as much as $2bn per operation to at least $4bn, running September 9 to November 4. The interest rate the government pays to borrow for thirty years had just touched 5.31%, the highest in nineteen years. Treasury Secretary Bessent said on CNBC that "part of it is signaling" and that "we believe that the yields don't reflect the underlying fundamentals." Interest rates round-tripped within a day. On August 23 the Fed's Kashkari said the Treasury market was functioning as it should. So an intervention was launched into a market that officials on both sides describe as working. That leaves the question of whom the support actually reaches. The buyer holding most of the long-dated government bonds is not a pension fund. Federal Reserve research puts large hedge funds' total Treasury exposure at about $4tn as of September 2025, financed by roughly $3tn of cash borrowed overnight against bonds put up as security, with about $830bn of that in a trade that profits from a small gap between two nearly identical prices and 90% of the total concentrated in fifty firms. We cannot show that buybacks were designed to support that position. We can say that improving the price of older government bonds improves the economics of holding them with borrowed money, and that a policy which makes a crowded, heavily borrowed trade more profitable tends to make it larger.

The dollar, gold and reserve status28 Aughigh

Treasury bought yen and won't say how much

The United States is now a discretionary participant in currency and bond markets it also regulates and issues into, with no published position and no appropriation behind it.

In July the US Treasury Department sold euros held in the Exchange Stabilization Fund and bought yen, after the Japanese currency hit a 40-year low. It was the first joint intervention with Japan in fifteen years. In a one-page reply to Senator Elizabeth Warren, Treasury Secretary Bessent confirmed the mechanics and rejected the suggestion that Japan owes Treasury anything: "Treasury exchanged existing Exchange Stabilization Fund foreign-currency assets for yen. No new congressional appropriation was involved." He did not disclose how much yen was bought, at what rate, or what the position is now worth (CNBC). A Reuters photograph of his notepad on July 31 read: "Buy Japanese Yen (JPY) $5-10 bil." Strip the politics out and this is a government running an undisclosed currency position, funded by selling one reserve asset to buy another, and marking it to market. The Federal Reserve separately widened a facility so the Bank of Japan could pledge its US government bonds for cash rather than sell them. That was a quiet acknowledgment that the alternative was Japanese official selling into the long end at the same time the thirty-year rate was making nineteen-year highs. The yen has not held the gains: FXY, the yen-tracking fund, is down 0.78% over five days and 0.69% over twenty. Sheets, who supported the intervention, was clear about the limit: for a sustained stronger yen "we will need to see some adjustment, particularly by the Bank of Japan in its monetary policy." Meanwhile the slow story continues. The IMF's reserve-composition data puts the dollar at 57.13% of allocated reserves in the first quarter of 2026, against 71% in 2000. The Reserve Bank of Australia cut its dollar share from 55% to 45% in 2024/25 and lifted euros from 20% to 30%. Nobody is fleeing. Everybody is trimming.

Household credit29 Aughigh

Brazil's borrowers spend 26.6% of income on debt

Consumer credit stress is real but concentrated at the bottom of the score distribution, which is a social fact before it is a financial one. The market is right to price it that way for now.

Brazil's ratio of household debt payments to income, excluding mortgages, hit a record 26.6% in June, up from 25.7% at the end of last year. More than half of all credit card users (52.8 million people) are carrying revolving balances or installment loans. The central bank is now preparing measures, and the sources briefing Reuters say it will start by restricting lenders rather than adopting the IMF's suggested cap on individual debt-service ratios (Reuters). Why this belongs on a page about the US bubble: it is the same shape. Policymakers name the problem as "the widespread availability of costly credit products, weak transparency requirements and insufficient financial education." That is, the supply side finding borrowers who are already stretched, at rates that only work if nothing goes wrong. The American version is quieter and more selective. Fitch has late payments on the riskiest car loans at 6.13% in July, up from 5.80% at midyear and 5.60% a year earlier. New York Fed data show 5.5% of auto debt more than ninety days past due, the second-highest reading in more than twenty years. But FICO finds ninety-plus-day late payments on mortgages and car loans rising only in the lowest score bands and flat everywhere else, and bank card thirty-day late payments at about 2.85%, down year on year. So the US household picture in aggregate is fine and the bottom tenth is not. That is a distributional problem, not yet a systemic one. The bundles of loans that hold this paper are small relative to bank books. It becomes a market problem when the extra interest investors demand for holding those bundles moves, and it has not.

Private credit and BDCs29 Augmedium

The first gates go up, eleven thousand miles away

Freezing withdrawals is how a fund that holds hard-to-sell assets discovers that its promise of easy access was a marketing document. Australia just ran the experiment first.

On August 27, MA Financial told investors it would limit withdrawals to 1% of the money it manages per month across its affected private credit products. On August 28, CVS Lane went further and "determined to temporarily suspend the processing of application and redemption requests across the funds." No money in, no money out. Australia's market regulator, ASIC, described this as the "first significant cracks" in Australian private credit, tied to the collapse of several large borrowers and a growing number of non-bank lenders running into cash constraints (JD Supra summary). The mechanism is the oldest one in finance and it has nothing to do with whether the loans are good. These are funds that let investors pull money out monthly, but the loans they hold take three to five years to pay back. When more investors want their money back than the fund has cash coming in from borrowers, the manager has two choices: sell loans (which establishes a price and forces every remaining loan to be valued against it) or freeze withdrawals (which protects the valuations). Everyone freezes. Freezing is not evidence that the loans are bad. It is evidence that the promise of easy access was never backed by enough cash. The scale here is small: Australian retail funds against a roughly $2.1 trillion global private lending market. What makes it worth your attention is that the same pressure is measurable in the US, where withdrawal requests at funds that do not trade on an exchange hit a record 12.4% of the funds' stated value in the second quarter, up from 10.4% in the first, and where loans that have stopped paying interest at the largest listed lending funds have risen to between 2.4% and 3.6% of their books. No US fund of this kind has frozen withdrawals. That is the line to watch, and it has not been crossed.

Fed, Treasury and policy28 Aughigh

The examiners are told to look at less

Our running thesis is that risk migrates to wherever nobody is measuring it. This makes the measured zone smaller by rule change rather than by migration.

The OCC, the regulator that supervises national banks, and the FDIC announced a new supervisory framework on Thursday, built around "refocusing supervision on what matters most — material financial risk." It creates a new definition of "unsafe or unsound practice", previously open to interpretation, and sets a higher bar for issuing formal notices that tell banks to fix deficiencies. Reputational risk was already removed from the framework earlier (FT). The case for it is not stupid, and the regulators make it themselves. Silicon Valley Bank received several notices about governance, controls and anti-money-laundering, and none that fixed the interest-rate and cash-flow risk that killed it. "The notices to Silicon Valley should have been about the actual risks to the bank rather than a random assortment," one regulator involved told the FT. Fewer, better-aimed notices is a coherent theory of supervision. The problem is one of timing and perimeter. Bank examination is the one place in the American financial system where somebody with legal power to demand answers looks at a loan book and asks what it is really worth. The volume of lending inside that perimeter has been shrinking for a decade; the growth is in direct lending funds, insurance company holdings and separate companies set up to hold data-center debt off the parent's books, where the examiner does not go at all. Narrowing what gets flagged inside the perimeter, at the moment the risk outside it is being tested for the first time, means the measured share of the system gets smaller in both directions at once. The filings do not say how many formal notices the new bar would have eliminated. That number, when it appears, is the one worth having.

The AI capex bubble29 Augmedium

Meta's next data center may never touch its books

If the biggest building boom in a generation is financed through separate companies that keep the debt off the builder's books, the borrowed money in the system cannot be read from the borrowers' accounts.

Meta is reported to be preparing a $12 billion financing backed by BlackRock, run through a separate company set up for the purpose, for a nearly 1-gigawatt Texas data center. The structure is larger than its record 2025 deal (Stocktwits/Perplexity sweep). It sits alongside Meta's Hyperion joint venture with Blue Owl, which issued $27 billion of debt against Meta's long-term lease commitment, and alongside a separate $6.6 billion borrowing package for OpenAI's chip purchases, where the chips themselves serve as the security and Nvidia provides a backstop. Here is the plumbing. The separate company owns the building and the debt. Meta signs a lease. Bondholders are repaid out of the lease payments, so their real exposure is Meta's promise to pay rent, but the debt belongs to the separate company, and the lease appears in Meta's disclosures as an obligation rather than as borrowing. Everything works so long as the tenant keeps paying and the building keeps being worth something at the end. If the tenant stops, the bondholder owns a very large, very specialized building in Texas. The total is what makes this a system question rather than a corporate-finance one. Companies have borrowed over $400 billion globally for AI investment in 2026 alone, per Bloomberg-compiled data, and Nvidia has partnered with Wall Street firms on financing platforms intended to mobilize $500 billion of outside capital. The Bank for International Settlements, the central bankers' own bank, calls the leases-and-separate-companies portion "shadow borrowing." Two cautions. The $12 billion figure comes from a single collection of press reports, not a filing. And the accounting treatment (whether and how the separate company gets folded into Meta's books) is not something we have seen documented. Both matter, and neither is public yet.

Bond market dysfunction28 Aughigh

Japan's interest bill just rose 27% in one budget

The largest foreign buyer of US government bonds that do not come due for decades now has a rising interest bill at home and a domestic bond market that finally pays a meaningful return.

Japan's Ministry of Finance published its fiscal year 2027 budget request on Friday. Debt service comes to ¥36.64 trillion, up ¥5.36 trillion on the current year's initial budget. The interesting line is inside it: interest and discount charges of ¥16.59 trillion, up ¥3.55 trillion, roughly 27% in a single budget year (MOF). That is what returning to normal interest rates costs when you have the largest government debt stock in the developed world. Japan's ten-year interest rate has gone from 0.4% to near 3%, a thirty-year high, and each year a slice of the outstanding debt rolls over at the new rate. The bill compounds from here whether or not rates rise further. The part that reaches across the Pacific: Japan has spent ¥27.13 trillion (about $170 billion) supporting the yen this year, the largest annual currency intervention on record, including ¥15.39 trillion between July 30 and August 26. To buy yen you need dollars, and Japan raises dollars in part by running down its holdings of US government bonds, now about $1 trillion as of June, down 16% from a $1.25 trillion peak in 2021 (Nikkei). So the biggest foreign government buyer of long-dated US bonds is now a seller, for domestic currency reasons, while its own domestic bonds pay 3% for the first time in a generation. Japanese investors turned net sellers of ¥2.8 trillion of foreign bonds and stocks in a single week after the intervention, though August as a whole was still net buying. "It's too early to say the trend is shifting," NLI Research's Shingo Ide told Nikkei. Correct, but the direction of the incentive is no longer ambiguous.

Crypto and TradFi contagion28 Augmedium

The stocks fell 7%. The coin did not.

A company whose entire business model is holding bitcoin and issuing shares at a premium to the value of those holdings has to find another way to service its debts once that premium disappears.

On Thursday's close, Strategy fell 7.35%, BitMine 7.18%, SharpLink 7.44%, Galaxy Digital 8.22% and Coinbase 6.43%. Bitcoin is at $78,202, up 1.0% on the day and down 0.97% over five sessions. Over twenty days the coin is up 20.6%, and Strategy is up 36.5%, BitMine 37.7%, SharpLink 33.9%. That gap is the entire business model of a company that exists to hold digital coins, expressed as a price. These companies are listed pots of coins plus a premium to the value of those coins. The premium is not decoration; it is the engine. Issue shares above the value of the underlying holdings, use the proceeds to buy more coins, and the coins per share goes up without anyone taking a risk. It works in exactly one direction. When the premium shrinks, the stock falls without the coin moving at all, and the moment the shares trade at or below the value of the holdings, the share-issuance channel closes, because issuing then destroys coins per share rather than creating it. What is left is debt. The FT reports $72 billion of bonds that pay no interest and convert into shares, issued globally this year, a market Michael Saylor's Strategy did much to popularize, sold to buyers who want volatility rather than income (FT). We do not have current premium-to-value figures for any of these companies, and one session is not a trend (over five days the stocks still beat the coin). But the shape of the risk is clear enough: the stock is a leveraged bet whose leverage is highest precisely when the coin stops going up.

Household credit28 Aughigh

Three million people, thirty-eight points

Lending to borrowers with the weakest credit is the one part of the machine already in visible deterioration, and the companies that do the lending are priced as if the problem is over.

Fitch, one of the three firms that grade how likely a borrower is to pay you back, puts US subprime auto loan late payments at 6.13% in July, up from 5.80% at midyear and 5.60% a year earlier. The improvement from tax-refund season has fully faded (AutoRemarketing). New York Fed data has 5.5% of all US auto debt more than ninety days late, the second-highest reading in more than twenty years. Separately, FICO reports about 3.2 million consumers with a recent student-loan late payment, carrying an average 38-point year-on-year credit score decline, while consistent payers gained 6 points and those who were never late gained 16 (Equifax/Moody's). The 38 points are the mechanism, not the headline. A late payment is a loss to one lender once. A 38-point score drop reprices that borrower's next car loan, next credit card limit and next mortgage application for years. It converts a single missed payment into a permanently higher cost of borrowing for three million households, which raises the odds of the next missed payment. FICO's own breakdown shows late payments rising past ninety days only in the lowest score bands and flat everywhere above. The lenders' shares are not remotely bothered. Capital One is up 3.5% over twenty days and 4.8% off its high, Credit Acceptance up 4.4%, Synchrony up 3.5%, Affirm 1% off its high, Oportun up 33%. Only Ally is soft, down 2.6% and 6.7% off its high. The market's reading is that this is a contained, low-score, already-accounted-for problem. That has been the correct call for two years. It stops being correct when the deterioration moves up a score band, and nothing in this week's data says it has.

What would change our mind

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

The extra interest that shaky borrowers pay compared with the government widening past 3.5 percentage points, or the same figure for safer borrowers past 1 percentage point, from today's 2.63 and 0.79. That would be the first confirmation from the lending market that the stock market's verdict on the AI infrastructure layer is right.

Would move the number

A US private lending fund or a fund that lends to mid-sized private companies freezing withdrawals or suspending its buyback program, as CVS Lane and MA Financial have done in Australia. That is the funding channel actually closing onshore.

Would move the number

A loan secured on graphics chips failing to find buyers: an IREN-style deal pulled, repriced above 10%, or a lender publicly marking chip collateral below cost.

Would move the number

Treasury buyback operations running above $4bn, or a government bond auction that attracts far fewer buyers than expected. Either would turn this month's price-support exercise into something the plumbing actually needs.

Would move the number