Wall Street wants OpenAI and Anthropic rated investment grade

Archived reading, published Tue, 08 Sep 2026 13:40:40 UTC (2 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility91

how much tinder is stacked up — moves slowly

Ignition37

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 23 sources and rewrites this page.

Status: Held at 64. The first day of trading after Labor Day looked like the quiet holiday days before it: the VIX, the market's gauge of how much turbulence traders expect over the next month, sits at 15.4, where anything under 20 is calm; the interest rate the government pays to borrow for ten years slipped to 4.76% from 4.78% and the thirty-year rate sits at 5.22%; the extra interest the shakiest companies pay compared with the government held at 2.65 percentage points on the last available reading; the S&P 500 closed Friday 1.0% below its high; bitcoin is down 1% overnight and 3.7% on the week. So ignition stays at 37. Britain selling £4.25bn of thirty-year debt at 5.82% was a test that passed, at a price. Fragility holds at 91: bankers lobbying for OpenAI and Anthropic to be graded safe enough for pension money is a plan for where the next slice of AI debt will sit rather than debt that exists yet, TWG's 42%-affiliated figure sharpens our measurement of a Walter problem that has been public since June, and nobody was forced to sell anything to pay down borrowed money.

Reporting from 3 Sep to 8 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.

The AI capex bubble8 Sepmedium

The label that opens the $11.7tn door

A safe-borrower label is the switch that would move AI debt out of the funds that lend to companies and into pension and insurance portfolios that neither trade what they hold nor reprice it.

Morgan Stanley and Goldman Sachs have spent recent weeks talking to the three firms that grade how likely a borrower is to pay you back, on behalf of Anthropic and OpenAI, arguing that both should get the top-tier stamp, investment grade, the moment they sell shares to the public, the FT reports. Both companies lose money and neither has said when that sale will happen. The prize is entry to the $11.7tn market for corporate bonds, where the biggest pools of retirement money are only allowed to lend to borrowers carrying that stamp. One senior credit analyst put the pitch bluntly: "Wall Street is trying to minimise their overall debt impact by arguing that these two companies will soon be flush with liquidity." In other words, the argument is that the two will shortly have so much cash on hand that their borrowing hardly counts. The letters matter more than the interest rate. A great many pension funds and insurers are limited, by their own rules or by regulators' capital rules, in how much debt they can hold from borrowers below that grade. So the stamp does not just make borrowing cheaper. It changes who is allowed to lend at all. That matters because of where the money already sits: Bloomberg notes that a record 46% of life insurers' debt holdings are already in loans made by investment funds rather than banks. A bond from an OpenAI carrying that grade is the instrument that would move the debt behind AI infrastructure out of the loans smaller cloud-computing providers are taking at 9% and into the pools of money that care least about price, the ones that hold a bond until it is repaid and almost never reprice what they own. There is one precedent, and it is recent. SpaceX went public in June and got the stamp immediately, the first large technology company ever to, helped by changes to index rules that pushed money from tracker funds into the shares on day one. That is not how it usually goes. Meta, Netflix and Tesla each waited a decade or more after going public for the same grade. The labs are not the only ones who would gain. Oracle and Nvidia stand behind the computing contracts and, in Nvidia's case, some of their customers' debt. If the firm on the other side of your deal can borrow more cheaply, your own credit looks better too. What we do not know: the grading firms have made no decision, neither company has laid out a path to a public listing, and both lose money. The filing that would let anyone test the "flush with liquidity" claim does not

Bond market dysfunction8 Sephigh

Britain borrows for thirty years at 5.82%

Every time old debt is replaced at today's rates, a move in interest rates becomes a permanent line in the budget, and the buyers who used to take these bonds without haggling over price are gone.

Britain sold £4.25bn of thirty-year government bonds on Tuesday at an interest rate of 5.82%, the highest since the office that sells its debt was created in 1998, the Telegraph reports. Governments sell debt every week; paying the most in a generation for it is what is new. Britain's interest bill already runs above £100bn a year, and every sale like this one adds to it for the next thirty years. The important word is "sold". An auction with no takers is a market breaking down. An expensive one is a price, and this one cleared. What it tells you is what a buyer of government IOUs that do not come due for twenty or thirty years now demands, and the number is not Britain's alone. The FT puts the combined interest bill of the OECD, the club of rich countries, above $2tn in 2025, 3% of everything those economies produce; Britain, France and the United States all now spend more on interest than on defense; the group expects to borrow a record $18tn this year; and the average ten-year rate across the seven largest rich economies is at 4% for the first time since 2008. Fidelity's Mike Riddell: "These massive sovereign debts are having to get refinanced at ever-increasing rates. Global investors are already starting to get scared." The mechanism is simple and slow. Debt borrowed at 1% or 2% comes due and is replaced at 5% or 6%. The interest bill rises every year even when the deficit does not, and the interest bill is itself paid with borrowed money. America is on the same path: the rate it pays to borrow for thirty years sits at 5.22% this morning and for ten years at 4.76%, with the national debt past $40tn. Two British dates worth having in the diary. On September 17 the Bank of England, Britain's central bank, sets next year's pace for selling off the government bonds it holds; with only around £30bn of those coming due on their own, FT Alphaville points out that keeping the shrinkage on schedule means selling long bonds into this market. On September 24 the debt office tests "switch auctions", a way of swapping bonds that come due far in the future for ones that come due soon. It is a test of the plumbing, no bonds change hands, but it signals that how far out Britain's debts fall due is now on the table. For the odds of a crash, this is the slow kind. The auction clearing means no spark. The price it cleared at is one more layer on the pile, and the pile is what turns a small fire into a large one.

Crypto and TradFi contagion8 Seplow

Bitcoin fell. The companies that hold it rallied.

When shares in the companies that hold bitcoin run far ahead of bitcoin itself, the gap is borrowed money the coin's price does not show, and banks are building the plumbing that connects it to their deposits.

Bitcoin fell this week and the companies whose whole business is holding it went up, and that gap is the story. The coin is at $78,301 this morning, down 1% overnight and 3.7% on the week, though up 13% over twenty days. Strategy, the largest of those companies, is up 12.1% on the week and 42.7% over twenty days, and sits 1.5% off its high. BitMine, which holds ether rather than bitcoin, is up 32.8% over twenty days; SharpLink 35.3%; Galaxy 30.6%; Coinbase 20.2%. Over the same twenty days the coin did a third of that. Over the last five, the coin went down and the companies went up. A company that exists to hold digital coins is a claim on those coins, made bigger with borrowed money, plus whatever extra the market is willing to pay for the company on top of what it holds. When the shares outrun the coins by three to one in a month, and rise while the coins fall, that extra is growing. This is our inference from the relative prices; we do not have the value of the companies' actual holdings in front of us and cannot say how large the gap is. But the extra is the engine. It is what lets these companies sell new shares, and bonds that turn into shares, to buy more coin, and it is the first thing that disappears when the mood turns. Prices are saying the market wants the company more than the thing inside it. Meanwhile the banks are laying rails. A sweep of the week's funding finds seven US regional banks, First Horizon, Huntington, KeyBank, M&T, Old National, SouthState and Glacier, putting $32.5m into Cari, a network for moving companies' bank deposits around as digital tokens; OpenReserve getting preliminary conditional approval for a banking license from the OCC, the regulator that supervises national banks; and Citi running live transactions in deposits turned into digital tokens on the ledger of Swift, the network banks use to send money to each other, on September 2 with First Abu Dhabi Bank and OCBC. Payments and banking infrastructure took 93% of the week's disclosed crypto funding. Twenty-one institutions including Citi, Bank of America, Goldman and Wells Fargo have committed this half to a company issuing a digital token meant to be worth exactly one dollar. That is the channel by which trouble in crypto would reach the banks, and it is being built on purpose: ordinary bank deposits as tokens on shared ledgers. For the odds of a crash, the first half of this story is more tinder, borrowed money hidden in a gap the coin does not show, and the second

Private credit and BDCs8 Sepmedium

Retirement savings in the owner's own loans: 42%, not 3%

This is the balance sheet the risk moved to: retirement savings invested in loans made by the insurer's own owner, valued by that owner, and supervised by nobody with a view of the whole country.

Two insurers controlled by Mark Walter's TWG Global failed to tell anyone that more than $20bn of what they held was backed by companies Walter himself controls, Bloomberg's editorial board writes. Once corrected, investments tied to the owner's own businesses were 42% of the insurers' assets, not the 3% reported. That is not a rounding error, and federal investigators are now looking. To begin getting out of those positions, Walter sold the Los Angeles Lakers for $12.5bn. Policyholders, Bloomberg notes, are "so far unharmed, if perhaps startled to learn their retirement income was in any way tied to an NBA franchise." We covered KPMG's warning about a Guggenheim unit yesterday. This is the other end of the same structure, and the structure is what our thesis is about. A saver buys an annuity, a policy that pays a monthly income for the rest of their life, and the insurer invests the money. When the insurer is owned by an investment firm, what it buys is increasingly loans that firm itself has made, lending to companies by investment funds rather than banks, and the price those loans are said to be worth is set by the people who made them. Bloomberg cites a record 46% of life insurers' debt holdings in such loans this year, and sales of annuities hit a quarterly record in the second quarter ($123.9bn). Oversight sits with fifty separate state regulators; there is no national one. Bloomberg's proposed fix is to declare parts of the annuity business important enough to the whole system that a federal regulator can set one set of standards. What is unusual here is the mislabeling, not the model. Apollo, which pairs lending by its own funds with an insurer's balance sheet, the list of what it owns and owes, is the archetype of that model, and the market is not treating the model as a problem: Apollo's stock is 6.6% off its high and down 1.1% on the week, while the listed funds that borrow money, lend it to mid-sized private companies and pass the interest to shareholders, its closest peers, sit within a point or two of theirs. The regulatory question is whether anyone can see a problem before it becomes one. A 3% that turns out to be 42% suggests not. For the odds of a crash, this is the third kind of story. Nothing got riskier this morning; these loans were sitting inside those insurers last year. What changed is our measurement of how much of this is out there and how little anyone checked. That sharpens the fragility reading rather than raising it.

Fed, Treasury and policy7 Sepmedium

The overflow tank is empty. Treasury buybacks start tomorrow.

With the cushion of spare cash at the Federal Reserve gone, every new government IOU now drains money straight out of the banks, at the moment the Treasury and the Fed are pushing interest rates in opposite directions.

The pool of spare cash that money-market funds park overnight at the Federal Reserve, America's central bank, has run dry: it held $725m on September 1 and $525m on September 2, which is effectively zero. That pool has been the cushion between the government's borrowing and the banks. With it gone, every new IOU the Treasury Department, which borrows the money the government spends, sells now has to be paid for with money that comes out of the banking system itself, and the price of the government's shortest debt is already showing it. Three more numbers from the Fed's own data show the strain. The overnight rate banks charge each other, which the Fed steers, sits at 3.63%, inside its 3.50–3.75% target range; the rate on cash borrowed overnight against bonds put up as security is 3.65%; and a three-month government IOU pays 3.76%, at the very top of the Fed's range. Here is how the cushion works. When the pool held a lot, new government borrowing could be absorbed by draining it, and the reserves banks keep at the Fed barely noticed. When it is at half a billion dollars, every dollar of new short-term debt the Treasury sells has to come out of those reserves or out of private hands, and a three-month IOU paying more than the Fed's own overnight rate is what that looks like. This is our reading of the mechanism; the Fed has said nothing about it. Into this, two arms of the government are pulling in opposite directions. The Treasury's expanded buybacks of long-dated debt take effect tomorrow, September 9: at least $4bn per operation of bonds due in ten to twenty and twenty to thirty years, doubled from $2bn, inside a $14.5bn weekly limit, running to November 4. Buying its own long bonds back props up their price and holds down the rate on them. Meanwhile Fed Governor Waller said on September 3 that he is inclined to leave rates alone at the September 15–16 meeting if inflation cooperates, which took the market's odds of a rate rise from about 63% to 50%. Holding down long-term rates while pushing short-term ones up has a name on trading desks, a twist, and here it is being done by two authorities that do not coordinate, while the president demands 1% rates, per our earlier reading. Scale check. The buyback program is $14.5bn a week. Japan's holdings of foreign securities fell $87.8bn in August, most likely short-dated US government debt sold to pay for propping up the yen. The buyer of last resort is a fraction of a single seller. For the odds of a crash, th

Signs of the times

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

Retirement income tied to the Lakers

Two insurers controlled by Mark Walter turned out to have 42%, not 3%, of their assets in companies tied to their owner, and to start getting out of them Walter sold the Los Angeles Lakers for $12.5bn. Holders of the insurers' retirement policies, Bloomberg observes, were unharmed but startled to learn their pensions were connected to an NBA franchise.

insuranceprivate creditannuities

8 SepBloomberg

The safe-borrower stamp before the share sale has a date

Bankers for OpenAI and Anthropic are lobbying the firms that grade borrowers for the top-tier stamp on two money-losing companies that have not published plans to sell shares to the public. Meta, Netflix and Tesla each waited a decade or more after going public for the same grade.

airatingsbonds

8 SepFinancial Times

$31.6tn of data centers

PwC estimates global spending on data centers will reach $31.6tn through 2050 to meet AI demand, about four-fifths of the entire US national debt, which passed $40tn last month.

aicapexscale

7 SepBloomberg

China bought gold into a 10% rally

The People's Bank of China, the country's central bank, added 650,000 ounces of gold in August, its biggest monthly purchase since 2023 and its 22nd straight month of buying, in a month when gold rose almost 10%. Central banks used to buy when the price dipped; this one is buying as it climbs.

goldreservesdollar

7 SepBloomberg

A$4m to keep a A$3.4bn builder alive

Bathla Group, the collapsed Australian developer that owes its lenders A$3.4bn, secured about A$4m of emergency funding from five lenders, roughly a tenth of one percent of what it owes, to keep limited operations going for two weeks.

private creditpropertyaustralia

7 SepBloomberg

The rumour mill

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

Confirmed

“US CPI inflation has reaccelerated this year, reaching a local high of 4.2% in May.”

The BLS, the agency that measures prices and jobs, reported inflation of 4.2% year on year in May 2026; it has since eased to 3.5% in June and 3.4% in July. The pickup is real, and it is why a rate rise in September is still a coin flip.

Claimed by Heresy Financial

Partly true

“Interest on the US government debt is now the largest federal expenditure at $1.1 trillion a year.”

Interest on the debt is running at around $1.0–1.1tn a year, but it is the second-largest item in the federal budget, behind Social Security, not the largest. It is still ahead of defense.

Claimed by Palisades Gold Radio

Partly true

“The 30-year Treasury yield has broken out to 5.26%, trading in ranges last seen in the late 1990s/early 2000s, with the 10-year near 5% at levels not seen since 2004-2007.”

The interest rate the government pays to borrow for thirty years has traded between 5.26% and 5.34% in recent weeks, levels last seen in 2007, and sits at 5.22% this morning. The ten-year rate is 4.76%, not near 5%.

Claimed by Heresy Financial

Earlier developments

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

Hidden leverage and shadow banking7 Sephigh

Deutsche Bank sells the first losses on its data-center loans

This is the whole pattern in a single deal: a regulated bank makes the loans to AI data centers, then sells the first losses to investors who never have to publish what they are worth.

Deutsche Bank is arranging a deal to pay outside investors to take the first losses on about €2bn ($2.3bn) of loans, including loans to data centers, Bloomberg reports. Deutsche has done deals of this kind before on other loans, but never on loans that finance the building of something. It is not alone in looking at data centers. Société Générale, ING and BBVA have considered or completed such deals that include data-center loans, and BNP Paribas, Royal Bank of Canada and Toronto-Dominion have explored ones aimed squarely at AI infrastructure. Deutsche's finance chief, Raja Akram, says the bank plans new platforms for these deals so that regulators will let it set aside less capital. The bank keeps the loans. It keeps the client, the fees and the interest. It pays other investors to take on the risk that its loans go bad, while keeping the loans on its own books; the trade has a name, significant risk transfer. In practice it sells a note to an outside investor, usually a hedge fund (an investment firm that manages money for the wealthy and can bet in ways ordinary funds cannot), an insurer or a fund that lends to companies in place of banks, and that note loses value as loans in the pool go unpaid. Because the first slice of losses now belongs to someone else, the regulator lets the bank hold less capital against the book. On the bank's balance sheet, the list of what it owns and owes, risk has fallen. In the system it has not moved an inch. It has changed address, from a lender that publishes stress tests to a buyer that often pays for the note with borrowed money and reports what it is worth once a quarter at best. Data-center loans are a natural fit. They run for years, the thing pledged to the lender is a building full of chips, and repayment depends on one tenant honoring one lease for the life of the debt. A bank that wants the fees from AI without the risk of the big loss at the end has every reason to sell that risk. The terms may change, Bloomberg says, and the report does not name the buyers, say at what level of losses the investors start paying, or say how much of the €2bn is data centers rather than other projects. Those are the numbers that would tell us whether this is careful hedging or the first losses on the AI build-out being quietly moved out of sight. Either way it is a small deal. What it signals is that European banks now see data-center lending as a risk worth paying to get rid of. This is fragility, not a spark. No new loan was ma

The AI capex bubble7 Sephigh

ByteDance asked for $20bn. Banks gave it $29.6bn.

The borrowing to pay for AI has reached Chinese tech, and banks offering half again more than was asked for tells you lenders are still competing to supply it.

ByteDance is finalizing a $29.6bn loan to pay for AI chips and data centers, Caixin reports. It asked banks for $20bn. They put in more than $30bn of orders, and it took nearly all of them. A borrower getting half again what it asked for is what a hot lending market looks like. The loan runs for three years, is coordinated by Citigroup and JPMorgan, and ranks as the second-largest dollar borrowing in Asia this year. It is not yet signed. The money has a destination. ByteDance budgeted nearly 160bn yuan ($23.8bn) for AI in 2026, more than half of it for buying chips, with another 50bn yuan for computing infrastructure. Its net profit fell by more than 70% in 2025 under international accounting rules (IFRS); the company says that is partly down to how it accounts for preferred shares and options rather than to operating costs. Revenue grew about 20% at home and nearly 50% abroad. At a meeting in August the chief executive, Liang Rubo, said the company's large language models were falling further behind foreign rivals. Put those together. A company whose models are losing ground is borrowing $30bn for three years to buy hardware in the hope of catching up, and the banks are so eager that the loan grew by half between launch and close. When lenders behave like that, the amount a borrower asks for is the floor, not the ceiling. One thing cuts against the argument we have been making. This is bank debt. It sits on the books of Citi, JPMorgan and the other banks in the group, where it will be measured, backed by capital and stress-tested. That is the opposite of the American side of the build-out, where much of the borrowing runs through funds that lend in place of banks or through chip suppliers financing their own customers. Whether it stays on bank books is another matter: Deutsche Bank's data-center deal, reported the same day, shows banks learning to make AI loans and then sell the first losses on them. This adds to the pile. It is $29.6bn more borrowed money riding on AI, and that is fragility, though we had it in last week's tally so the number does not move again today. The three-year term is what to watch. It is short for infrastructure and long for a race between AI models, and it means the question of borrowing again to pay this off arrives in 2029, whoever is holding the debt by then.

The dollar, gold and reserve status7 Sepmedium

Foreign money is leaving US government bonds for US stocks

The owners of the world's benchmark asset are shifting from central banks that never have to sell to funds running on borrowed money that can be forced to. That changes how a sell-off would spread, not just who holds what.

Foreign investors now put more money into American stocks than into American bonds. Deutsche Bank says that has never happened before outside the 2008 crisis. Per CNBC, Deutsche's George Saravelos writes that foreigners' share of US government bonds has fallen from more than 50% at its peak to about 30%, while the US drew a record $600bn of net foreign money into its stocks in the year to March 2026, outweighing purchases of government bonds and the debt of government-backed mortgage agencies 'by the largest margin in history'. It is happening while the government's own finances get worse: the national debt passed $40tn last month, and the deficit is heading for $2.1tn, more than 6% of everything the economy produces. Saravelos's explanation: 'the American private balance sheet is booming — AI, record profit margins — but the public sector balance sheet keeps worsening'. In other words, foreigners want more of a stake in American companies and less of a stake in Washington's borrowing. The official holders line up with that. China's holdings fell to $633.4bn in June from $731.4bn a year earlier. Norway's national fund has proposed cutting US government bonds from 34.1% to 21.9% of its bond portfolio. Japan's holdings of foreign securities fell by a record $87.8bn in August, most likely US government bonds sold to raise the money to prop up the yen. And China's central bank bought gold for a 22nd straight month (650,000 ounces in August, its biggest purchase since 2023). The pile of US government bonds does not shrink because foreigners want less of it. Somebody has to hold every one. The patient holders are central banks and national wealth funds, who never get a demand from a lender for more cash because a bet has moved against them. When they step back, the buyer who sets the price becomes whoever can raise the money to hold the position. Right now that is hedge funds, which hold $2.6tn of these bonds and pay for them with $3tn of cash borrowed overnight against bonds put up as security, and, as we reported over the weekend, Chinese commercial banks paying 3 to 4% for short-term dollar deposits. Holders who can be forced to sell are replacing holders who cannot. Two things soften this. Data from the IMF, which lends to countries in trouble, show the dollar's share of central-bank reserves ticked up to 57% in the first quarter, so this is a retreat from bonds that take decades to come due, not from the currency. And State Street's Masahiko Loo describes J

Household credit7 Sepmedium

Car loans are going bad at a record rate. The lenders' shares are near their highs.

Households at the bottom are missing payments at rates not seen since the last recession, but the lenders bundled and sold the loans, so the losses land on investors rather than on the lenders' shares.

About one in eighteen dollars of American car loans is now at least three months behind, the worst in more than two decades of records (5.5% of balances 90 or more days past due in the second quarter, per data from the Federal Reserve's New York branch that resurfaced this week). That is a tenth below the first quarter's 5.6%, so the record has stopped climbing for now. Americans owe $1.713tn on their cars, and 3.00% of that slid into serious delinquency during the quarter, up from 2.93% a year earlier. Credit card debts written off as lost edged up to 3.70% from 3.66%. The split by borrower is the story. Fitch, one of the three firms that grade how likely a borrower is to pay you back, tracks bundles of car loans to borrowers with weak credit that were sold on to investors. Its index shows payments 60 days late easing to 5.8% at the end of June from 6.5% at the end of 2025, then getting worse again in July. Another reading puts July at a record 6.13% for those weak-credit bundles, against 0.49% for bundles of loans to solid borrowers. A Federal Reserve survey has 34% of adults earning under $25,000 failing to pay all their bills last month, against 7% of those earning $100,000 or more. Borrowers with good credit are fine. The bottom of the income ladder is at levels not seen since the last recession. Now look at what investors think. Credit Acceptance, which lends to the weakest car buyers, closed Friday 0.4% off its high and up 4.3% on the month. Ally is 2.8% off its high, Capital One 3%, Synchrony 1.3%. Only Affirm, the buy-now-pay-later lender, has sagged, 7.9% off its high after falling 6.9% in a week. Both pictures can be right. A lender's shares are worth the gap between what it charges and what it loses. If it has raised the rate on new loans and bundled the old ones together and sold slices of the income to investors, the losses land on the buyers of those slices, not on the lender. That is the same migration we track everywhere else: the missed payments show up in a trust, and the stock does not flinch. What we cannot see from here is how much of the lending from 2022 and 2023 is still on the lenders' own books. If the answer is 'a lot', the market is wrong. None of this moves the odds today. The figures are two months old and this is not new stress; the tinder was already in the pile and already in our number. It is a picture worth redrawing while the lenders' shares sit at their highs, because the one thing that would change the picture, the lo

Fed, Treasury and policy3 Sephigh

Who pays for the power lines to the data center?

Households have been quietly paying, through their electricity bills, for the power lines that serve AI data centers. States are starting to make the developers pay instead, and the biggest tenant is fighting to keep the old arrangement.

Microsoft has asked Virginia's Supreme Court to strike down rules that make data-center developers pay upfront for the transmission lines built to serve them, the FT reports. It filed the appeal quietly, months after signing a pledge organized by President Trump to shield electricity customers from exactly those costs. The filing does not say which parts of the rules it objects to. This is a fight over who pays for the grid, and until this year the answer was households. Dominion Energy, the utility for Virginia's data-center belt, asked regulators this year to recover more than $1.5bn it had spent on transmission lines. Historically that money went into customers' bills and was paid back over decades. In July the State Corporation Commission, which regulates Virginia's utilities, ordered Dominion to bill data centers directly and upfront instead. Governor Abigail Spanberger's office had urged the commission to protect customers from 'severe stranded-asset risks' and from utility 'capital biases that prioritise shareholder profits over Virginia's ratepayers'. In other words, the state was worried that households could be left paying for wires to data centers that never got built or lost their tenant. Under the old arrangement the AI build-out had a hidden subsidy. The big cloud companies ordered the power, the utility borrowed to build the wires, and households repaid the debt through their bills. If the data center is canceled or its tenant walks away, the wires are stranded and the households still pay. Under the new arrangement the developer pays first, which raises the cost of every project and cuts its return. Microsoft, which canceled a 244-acre site in Caledonia, Wisconsin, after local opposition in October, evidently prefers the old one. This matters for how fast the bubble grows. Data Centre Watch counts more than $130bn of US projects blocked or delayed in the first quarter of 2026. Household electricity prices are up 15% since the president took office, according to the EIA, the government's energy statistics agency, two months before midterm elections being fought over the cost of living. The limit on AI spending may turn out to be neither chips nor money but permission, and whether the tenants will pay for their own power lines once someone else stops. On its own this moves our number very little, and we are including it for what it says about direction. A subsidy being withdrawn makes the marginal project less worth building, which slows how

The dollar, gold and reserve status7 Sepmedium

Tokyo's reserves fell $88bn. We can guess what it sold.

The biggest seller of US government bonds in the August rout was a friendly central bank doing something it had no choice about, and that is precisely the kind of selling no speech from the Federal Reserve can talk down.

Japan's Finance Ministry said on Monday that its holdings of foreign securities fell by $87.8bn in August, the largest monthly drop it has ever recorded. It had already admitted spending ¥15.4tn ($98.6bn) to prop up the yen in the month to August 26, also a record. The two figures are close enough that analysts quoted by Bloomberg conclude Tokyo paid for the rescue by selling US government bonds, most likely ones due to be repaid soon. A ministry official would say only that intervention was "a factor" behind the fall. The mechanics are not complicated. To buy yen you have to sell dollars, and Japan keeps most of its dollars in US government bonds. So the harder Tokyo defends its currency, the more of Washington's debt lands on the market, and this happened in the same weeks that the rate the US government pays to borrow for ten years (the 10-year) reached 4.82%, its highest since 2023, and Treasury Secretary Bessent doubled the department's purchases of its own long-dated bonds to at least $4bn per operation from September 9. None of this was hostile. The United States joined the yen operation on July 31, the first time the two countries acted together on the currency since 1998. It is still one of the largest holders of US government bonds selling into a falling market. What makes it worth watching is that it feeds itself. High American interest rates weaken the yen; Japan sells US bonds to buy yen; the supply of US bonds goes up. Sales of bonds due soon do less damage to long-term borrowing costs than the headline suggests, and we do not know what mix Japan sold. But Japan's own government now pays more than 3% to borrow for ten years, the first time since 1996, which gives Japanese money one less reason to stay in dollars at all. Put that beside the rest of the ledger. Norway's sovereign fund has proposed cutting roughly $80bn from the $215bn of US government bonds it holds, though not before 2027. Chinese commercial banks are buying, paying 3–4% on the dollar deposits that fund the purchases. The Treasury Department starts buying back its own bonds today. The sellers are governments. The buyers are either borrowing to do it or are the issuer itself. That is why this sits in fragility rather than ignition: nothing caught fire, but the natural buyers are leaving.

Bond market dysfunction7 Sepmedium

Two debt offices, one problem: too many long bonds nobody wants

When both Washington and London are pulling their longest-dated bonds out of the market by hand, those bonds have lost their natural buyer, and the government is trading the risk that prices fall for the risk that it cannot borrow again when the debt comes due.

The Treasury Department, which borrows the money the US government spends, begins the active phase of buying back its own debt today, September 7, with a weekly limit of $14.5bn. From September 9 through November 4 the operations aimed at the longest bonds, the 10-to-20-year and 20-to-30-year ones, double from $2bn to at least $4bn each. Bessent calls this support for orderly trading, not a sign of strain in the public finances. A government buying back its own bonds is not odd in itself. Doubling the size mid-program is a decision, not a routine. Across the Atlantic the same problem is being attacked from two directions at once. On September 17 the Bank of England, Britain's central bank, sets the pace at which it will actively sell the government bonds it holds over the coming year. Only about £30bn of those holdings come due in that period, so keeping its balance sheet, the list of what it owns and owes, shrinking at the same pace means selling bonds into a market where Britain paid 5.89% to borrow for thirty years last week (the 30-year gilt), the most since 1998. Then on September 24 the Debt Management Office, which does the British government's borrowing, tests a "switch auction", a device last used at the turn of the century that pulls bonds of one maturity out of the market and hands back bonds of another. The office insists no bonds will actually change hands this quarter. FT Alphaville reads the test as a signal that it is getting ready to reshape what it owes, and points out who the largest unhappy holder of long British government bonds is: the Bank of England itself. What a buyback or a switch actually does is worth holding onto. The state takes debt that does not come due for twenty or thirty years out of private hands and sells shorter debt to pay for it. Investors are relieved of the risk that those bond prices keep falling. The government takes on a different one: more debt coming due, more often, to be borrowed again at whatever the short-term rate happens to be then. That works right up until the short-term rate is where the trouble is. In the US the short end is where the Federal Reserve, America's central bank, may raise rates next week, and where, according to Bloomberg, Japan was selling. For the odds of a crash this is pile, not spark. Nothing is on fire; two debt offices are showing us they know the buyers have thinned out. One caveat: the US buyback detail comes from a single market note, and we have not seen the Treasury Departm

Private credit and BDCs7 Sepmedium

An auditor's warning inside Mark Walter's empire

The structure that lets lending by investment funds grow, a manager owning an insurer that buys the manager's assets, only works if someone outside the group can see what the insurer holds. Here nobody could.

KPMG, the accounting firm that audits Guggenheim's books, told the firm last year that a subsidiary of its $367bn asset-management arm had a problem with how it counted revenue, according to the FT. A whistleblower had reported in April 2025 that Guggenheim Private Investments was booking $275mn of revenue in a questionable way. KPMG first flagged a "control deficiency" while auditing the unit's 2024 accounts, then raised it to a "material weakness". Auditors do not use that second phrase lightly. It means the flaw is serious enough that the numbers themselves may not be reliable. On its own that would be an accounting story. It does not stand on its own. Insurers controlled by Mark Walter, through his holding company TWG Global, revealed in June that they had lent billions of dollars to other parts of his empire and described those loans as something else in their filings to regulators. US prosecutors are now investigating the insurers. The reason we keep returning to this is what a life insurer is for. It takes in premiums and must hold assets against what it will one day owe policyholders. Regulators see those assets only through the filings. If the assets are loans to the owner's other businesses, and the filings call them something else, then policyholders are financing the owner's empire and the people whose job it is to notice cannot. The FT's point is that the strain now shows up on the asset-management side too, months before any of it was public. This is our running thesis in miniature: risk that would be measured on a bank's balance sheet, the list of what it owns and owes, sitting on an insurer's instead, valued by the people who hold it, and surfaced by a whistleblower rather than a regulator. Nothing here proves wrongdoing. The reporting rests on people familiar with the matter, and the filing gives no figure for the loans between group companies. Apollo, whose entire model is an asset manager feeding an insurer, closed 0.7% lower on Friday and sits 6.6% below its high. The market is not treating Walter's problems as an industry problem. Not yet. For the meter this is the third kind of news: the risk did not grow this week, our measurement of it improved.

The AI capex bubble7 Sepmedium

The borrowers rally while the customers sag

If the market is right that the borrowers are the safe end of the chain building AI computing capacity, then the risk that hundreds of billions of dollars of chips are worth less than expected in a few years has been handed to the customers, and their share prices are starting to say so.

Look at who went up on Friday. IREN, a bitcoin miner that now rents out AI chips, closed 7.3% higher, is up 26% in five trading days, and sits less than 1% off its high. Nebius rose 7.5% and CoreWeave 5.6%. Oracle, the most debt-funded of the giant cloud companies, sits at its high after an 8% month. Now look at who went down. Alphabet is 10% off its high, Amazon 9%, and Microsoft fell 2.1% on the day. The companies borrowing to build are being rewarded; the companies that will have to pay for what gets built are drifting. The week's reporting explains what the first group is doing. Blue Owl led $2.4bn of lending secured on IREN's chip purchases on August 28. Nscale closed $3.05bn of secured loans for sites in Texas and North Carolina. ByteDance's data-center loan, sold to a group of banks, started at $20bn, drew more than $30bn of orders and closed at $29.6bn. Broadcom is reported to be in talks for more than $60bn of debt to finance chips for Anthropic, on top of a $35bn package with Apollo and Blackstone. Reuters puts new AI borrowing in 2026 at up to $400bn and AI-related debt at about 7% of the indices that track America's safest corporate bonds. Morgan Stanley estimates $800bn will sit in lending by investment funds and in separate companies set up to hold things off a parent's books by 2028. One reading: the builders have signed customers. IREN says customer prepayments cover about half of what it spends on chips and lenders most of the rest. So they trade like a utility with a tenant on a lease, while the cloud giants carry the real question, which is whether the computing earns its keep. The market is pricing the lease as safer than the tenant. The other reading is momentum. The structural point is that a lease to Microsoft is only as good as Microsoft's appetite to renew it when the chips are three years old, and these loans are sized to the first term. That is fragility, not a spark: more borrowed money resting on an assumption nobody has tested. The caveats are real. The tally is second-hand, the Broadcom deal is talks rather than a signing, and the sources on IREN's borrowing cost disagree: one report says 9% fixed on $2.4bn, another about 6% on $3.6bn of financing tied to Microsoft.

Hidden leverage and shadow banking7 Sepmedium

The Federal Reserve's overflow tank is empty

The cushion that used to absorb a cash shortage before it reached the cost of borrowing overnight is gone, so the next shock lands directly on $3tn of US government bond bets that have to be funded again every night.

Spare cash parked at the Federal Reserve, America's central bank, through the facility that takes overnight deposits from money-market funds, fell to $525mn on September 2, from $6.73bn the day before. For most of the past few years this was where those funds left money they had no better use for, and its balance was the system's spare change. At half a billion dollars it is effectively empty. Why this belongs on a page about borrowed money: the New York Fed published a framework this week describing one corner of the market for cash borrowed overnight against bonds put up as security, the part where dealers lend to clients, as dealers lending cash to "levered accounts, most of which are hedge funds executing Treasury relative-value strategies". In other words, funds that borrow heavily to profit from small gaps between nearly identical prices. It is the most expensive of the market's three segments, priced around the 75th percentile of SOFR, the benchmark rate for that overnight borrowing. That corner is where hedge funds' $3.0tn of overnight borrowing lives, the money behind $4.0tn of US government bond positions held by roughly fifty funds, per the June Fed paper that Reuters revisited on September 3. With the tank drained, any shortage of cash shows up directly in the overnight rate. SOFR was 3.66% on September 3 against 3.63% for the rate the Fed actually targets, a gap of 0.03 percentage points. That is small and not a stress reading. But the basis trade, hedge funds borrowing heavily to profit from a small gap between two nearly identical prices, earns a few hundredths of a percentage point a year on enormous positions. A squeeze of a few hundredths in the overnight rate is the whole margin, and the position has to be borrowed again every night. When the margin goes negative the trade closes by selling US government bonds, into a market where Japan just sold $88bn and the Treasury Department is buying back $14.5bn a week to keep it orderly. None of this is new this week, and we are not raising the number for it. It is the tinder, described more precisely, going into a Federal Reserve meeting that markets give 60% odds of producing a rate rise.

The dollar, gold and reserve status6 Sepmedium

The new buyer of America's debt is funding it month to month

Who holds US government bonds, and with what money, decides how those bonds behave when the market is under strain. The holders are shifting toward hands that have borrowed short-term and are betting with borrowed money.

Chinese commercial banks have spent the past few months buying US government bonds, paying for them by raising the interest they offer on dollar deposits, according to people who spoke to Reuters. Nobody had reported it before. It matters because the buyers America relied on for years are leaving, and their replacement is holding a ten-year IOU with money that can walk out the door in months. The Big Five state lenders have capped what they pay on dollar deposits at 2.8% since 2023, which is the ordinary state of affairs. What is not ordinary is the rest: customers with more than $50,000 have been able to negotiate above 3% since June, and some smaller and foreign banks have offered close to 4% since August, a few of them advertising on social media to find the money (Business Times). The arithmetic is simple. The interest rate the US government pays to borrow for ten years has risen more than 0.3 percentage points since June (to 4.76%); the rate China's government pays to borrow for ten years has fallen to its lowest in over a year; and the gap between them briefly reached its widest since 2006 on Wednesday (about 3.13 percentage points, per Nikkei). A bank that gathers dollars at 3.5% and holds US bonds paying 4.75% pockets the difference without ever touching the yuan. That matters, because the sources say the banks are reluctant to change currency themselves while regulators are watching. What the story does not say is how much. Reuters could not establish the amount, or whether it changes China's overall holdings of American debt. The mechanism is what interests us. The buyers the US bond market has traditionally depended on are leaving or shrinking. Japanese investors have sold a net ¥3tn ($18.7bn) of foreign bonds this year through August 22, the biggest outflow at this point in a year since 2022, because Japanese government bonds now paying above 3% are pulling money home (Reuters analysis via Astro Awani). Norway's sovereign fund is proposing to cut roughly $80bn. Their replacement, on this evidence, is banks borrowing short from depositors who can renegotiate or leave, to own a bond that pays a fixed rate for ten years. Anyone who watched the brokerage Schwab in 2023 knows what happens to that trade when the cost of the deposits moves. For the meter, none of this is a spark. A buyer showing up is better than no buyer, and the money is real. But the holder stepping in is smaller, less patient and more reliant on borrowed money than the one steppin

Fed, Treasury and policy6 Sephigh

Beijing writes checks to the insurers it used to prop up its stock market

When a state has to put fresh capital into the companies it used to hold up its own stock market, it is telling you what those companies were absorbing.

China's finance ministry is putting a combined $54bn into state-owned insurers and banks, the companies said on Sunday. China Life gets 35bn yuan ($5.2bn), China Taiping 7bn, China Export & Credit Insurance 10bn, China Re 3bn, and PICC plans to sell the ministry up to 15bn yuan of new shares. Separately, three state lenders announced 290bn yuan of fresh capital between them, including 30bn for the Export-Import Bank (Business Times). A government does not put sums like that into companies that are fine. The plan itself is not new; it was unveiled at the parliamentary session in March. What is new is the money actually landing, and the reasons given for it. The report says the state insurers were 'directed to support the stock market with medium- and long-term funds', that years of low interest rates have eaten into the industry's profits, and that many small and mid-sized insurers are reporting a shrinking cushion of capital against their promises to policyholders. The injection also positions the big groups to 'help regulators manage' the weaker ones. Read that as plumbing. Beijing used the insurers' balance sheets, the list of what each company owns and owes, as the buyer that shows up for shares when nobody else will. Those balance sheets, sitting on low domestic interest rates and on the shares they were told to buy, have worn thin. So the state is now writing checks to the institutions it used as shock absorbers, and asking them to absorb the smaller insurers as well. This is the Crash Lab thesis in state-directed form. In the West, risk has drifted onto insurers through lending to companies by investment funds rather than banks, and through reinsurance; in China it was put there by instruction. Either way the insurer is the layer where losses are slowest to be admitted in the prices firms put on their holdings, and where the rescue, when it comes, is a policy decision rather than a market event. For the meter, capital going in is capital going in, and it does not raise fragility. But look at the order of events: banks reported more bad loans to households last week, and insurers get their capital topped up this week. Our read is that this is measurement of strain that was already in the system, not new strain. The world did not get more dangerous on Sunday; we got a clearer look at it.

Crypto and TradFi contagion6 Seplow

Twenty-one banks decide to issue the dollar token that has been draining them

The banking system is copying the product that has been pulling deposits out of it, while bitcoin quietly becomes something you can pledge to buy a house. Both widen the pipe between crypto and the regulated system.

Twenty-one financial institutions, among them Bank of America, Capital One, Citi, Goldman Sachs and Wells Fargo, say they intend to issue a digital token meant to be worth exactly one dollar, a stablecoin, in the first half of 2027. Thirty-eight state bankers' associations have announced a separate network, the BankChain Alliance, for deposits turned into digital tokens and for stablecoins, with the same launch year (TheStreet). The coverage is explicit about the motive: to stop competitors that are not banks pulling deposits out of the banking system. Banks have spent years watching that happen; deciding to build the thing themselves is the change. That motive is the whole mechanism. A stablecoin is a bank deposit that has left the bank and become a short-term US government bond held by somebody else. Tether, one of these issuers, has more than $183bn of tokens outstanding, backed by reserves parked in US government debt; Circle minted 5bn of its USDC token in a single week; and products that turn US government bonds and money-market funds into tokens passed $16bn in August, according to congressional testimony cited in the same sweeps (Bitcoin.com). Every dollar of that is a deposit a bank no longer has to lend against. The banks' answer is to build the drain inside their own house. Two other items this week matter more for contagion than for competition. Coinbase and Better have launched mortgages backed by bitcoin, letting borrowers pledge their coins as security for the down payment rather than sell them; a waitlist reportedly points to more than $260m of demand. And on August 27 Virtu, M1X and Tradeweb completed a loan of cash borrowed overnight against bonds put up as security, with every step recorded on a blockchain ledger (the Canton network) and a government's digital bond as the security. A fall in bitcoin that becomes a demand from a lender for more cash, right now, against somebody's house, and the market's short-term plumbing rebuilt on a ledger that has never lived through a March 2020, are new channels from crypto into places it did not previously reach. Share prices say the market is not worried: Coinbase fell 4.2% on Friday but is up 20% in twenty days, and Strategy, a listed company whose shares are largely a wrapper around bitcoin, is up 43% over the same stretch against bitcoin's own 24%. All of this reaches us through secondary sources rather than filings, and none of it is under stress yet. For the meter this is wiring being laid, n

The AI capex bubble6 Sepmedium

Investors are rewarding the AI builders who borrow to build

Once the AI buildout is paid for with debt rather than profits, a shortfall in demand stops being a fall in a share price and becomes loans that do not get repaid, with named lenders holding them.

Oracle spent about $56bn building data centers and the like in its 2026 financial year and, by one account circulating this week, covered a $23.7bn hole in its cash by raising $43bn from bond investors and $5bn from selling shares (Inforcapital, via web sweep). We have not checked those figures against Oracle's own filings and the source is secondhand, so treat the shape, not the decimals, as the fact. The shape is that one of the biggest builders is paying for the AI boom with borrowed money, and the stock market is rewarding it for doing so. The same sweeps put the big technology companies' AI-related debt above $350bn and this year's new AI-related corporate borrowing at roughly $220bn, alongside a week of fresh secured loans further down the chain: Nebius enlarging a deal to $5bn, Nscale closing up to $3.05bn of loans backed by its projects in Texas and North Carolina with lenders who stand first in line, and Lambda's $1bn facility to buy GB300 chips that Microsoft will lease (Yahoo Finance). Morgan Stanley is quoted estimating some $800bn of lending from investment funds and separate off-the-books companies for this buildout by 2028. Now look at Friday's close. Oracle, the most debt-funded of the giants, sits at its high for the period, up 8% in twenty days. Amazon is 9% off its high and Alphabet 10%; Microsoft fell 2.1% on Friday. Investors are currently rewarding the builder that borrows to build and marking down the ones paying out of their own earnings. That is not irrational on its own terms. Borrowing does not hand new shares to outsiders, and if the computing power earns more than the debt costs, existing shareholders keep all the upside. But it changes what kind of story this is. Spending your own cash is a risk to shareholders: if demand disappoints, they take the loss. Spending borrowed money is a risk to lenders: if demand disappoints, bondholders, the firms on the other side of the leases and whoever bought the paper issued by those separate companies take the loss first, and the question becomes who admits it and when. For the meter this is fragility, and it makes a crash more likely, not less. A share price at its high is the market saying that question will not be asked. The $43bn is the market saying there are already a great many people to ask.

Household credit6 Seplow

Household debt: write-offs up a touch, late payments down a touch. Fine.

The consumer is the part of the system that would turn a problem in the market's plumbing into a recession, and this quarter's data says that is not happening yet.

The second-quarter numbers on American household debt, as relayed in this week's sweeps of data from the Federal Reserve, America's central bank, and its New York branch, are not alarming, and it is worth saying so. The share of credit card debt banks gave up on rose a fraction (from 3.66% to 3.70%). The share of card payments running late eased (from 2.80% to 2.74%). Household debt at least a month behind fell (from 4.8% to 4.7%), seriously overdue debt fell (from 3.4% to 3.3%), and seriously overdue car loans fell too (from 5.6% to 5.5%, per Motoring Chronicle, citing NY Fed data). The page carrying those numbers is headlined as a 32-year high in late car payments, which the figures on it do not support; we are using the data, not the headline. The pattern, write-offs up a touch and late payments down a touch, is what steadying looks like. A write-off is a loan the lender has given up on. A late payment is a loan that has just started to go wrong. When the first rises while the second falls, lenders are cleaning up last year's late payers while fewer new ones fall behind. That is the back end of a deterioration, not the front end of one. The share prices mostly agree. Capital One, Synchrony and Ally, all big lenders to consumers, are up on the week and within 3% of their highs; Credit Acceptance, which lends for cars to people with the weakest credit, is 0.4% off its high. The exception is Affirm, down 6.9% on the week and 7.9% off its high. Its buy-now-pay-later customers skew younger and poorer, and the sweep found no fresh loss data on that kind of lending at all, so the market is guessing without us. What is still standing from earlier: a January reading of roughly one in every fourteen car loans to weak-credit borrowers two months behind (6.90%), and roughly 3.6m federal student-loan borrowers in default since the payment pause ended. Those are piles of stress we continue to carry. But nothing in this quarter's flow adds to them. For the meter this adds nothing to the pile, and we say so plainly; the household beat is one place where the machine is, for now, running as designed.

Private credit and BDCs6 Sepmedium

Four in five software loans are worth less than the books said

The individual loans we have watched lose value one at a time are now a number for the whole sector, and four in five software loans have started moving the way Loparex's did.

Reuters went through the filings of ten American funds that borrow money, lend it to mid-sized private companies, and pass the interest to shareholders, and found that four in five of their software loans are now carried at less than they cost (81%, against 40% of loans to every other kind of business). Software was the industry these lenders liked best. Loans where the borrower has stopped paying, or is unlikely to pay in full, have risen to about one in thirty from one in forty at the start of the year (3.4% of cost at June 30, from 2.5% at the end of 2025), and the share carried at less than 80 cents for every dollar lent has quadrupled from where it sat for three years (about 4% now, around 1% from 2023 to 2025). Every one of those figures is the lenders' own admission that the loans are worth less than the books said, and until this week we had only seen it one company at a time. A wider cut of 44 such funds puts the value they place on everything they hold at $92.88bn, against $95.19bn it cost them: a discount of $2.31bn, or about 2.4%. That number is small, and it is built to be small. These funds decide for themselves what their own loans are worth, and until a borrower actually misses a payment there is little that forces them to admit a loss. That is why the 81% figure is the one to hold on to. A software loan that has been written down is still, in most cases, carried at 80 or 90 cents on the dollar. Loparex was carried at 88 cents in December and at 5 cents last week. The sector data says the queue behind it is long and concentrated in one industry. Software was the collateral private lenders liked best: subscription income that arrives every month, little need to spend on buildings or equipment, and easy to borrow against. The reporting does not say why it has turned, and we will not guess. Reuters also found 88 companies where two funds valued the same loan more than 10 cents on the dollar apart, which tells you these valuations are opinions. And KBRA, a firm that grades how likely borrowers are to pay you back, found the typical rate of stopped-paying loans among the funds set up with an end date jumped in a single quarter (its median rose to 2.75% from 1.81%). The market is not pricing a queue. Shares in two of the big listed lenders of this kind closed within a whisker of their highs (Ares Capital 0.8% off, Blackstone Secured Lending 1.1%). The firms that manage the funds are softer (Blackstone down 4.4% on the week and 8.8% off its high,

Hidden leverage and shadow banking4 Sephigh

The world's largest bond manager sees 2007 in two products

The largest bond manager in the world is saying, on the record, that two of the structures moving loan risk from banks to insurers are being priced the way the middle slices of bundled subprime mortgages were in 2007.

Christian Stracke, president of Pimco, the largest bond manager in the world, told Bloomberg this week that some buyers of two structures now carrying loan risk out of the banks are not getting paid enough for the risk, and that the structures share "at least some similarities" with the subprime mortgage paper of 2007–08. Fund managers compare things to 2007 for a living, but Pimco is the firm whose whole business is deciding whether a bond pays enough, so it does not say so lightly. He also said the problem is not systemic and that Pimco is not bracing for a crisis. Both halves matter. Here is how the two products work. The first takes stakes in private investment funds, the ownership slice that gets paid last when anything goes wrong, pools a lot of them together, and cuts the pool into slices ranked by who gets paid first. The top slices get a safe-looking grade from the rating firms and are sold, mostly, to insurers, whose rules let them hold rated debt cheaply but not fund stakes. That trade has a name, a collateralized fund obligation, and Bloomberg puts the wider market these sit in at over $1tn. The second runs the other way. A bank keeps a portfolio of loans on its own books but pays an outside investor to take the losses on the middle slice, and the amount of its own money the rules make it hold against those loans falls. That one is called a significant risk transfer, and the investor has, in effect, sold insurance on the middle of a loan pool. "Very similar to what we saw before the GFC," Stracke said. "Examples of people under-pricing the risk to get leverage." In other words, like the run-up to the 2008 crisis: people accepting too little pay for a risk because it lets them borrow more against it. This is our whole thesis in two products. The second moves loan risk out of the bank, where regulators charge for it and stress-test it, and into a fund or insurer whose exposure appears in no bank filing. The first turns fund stakes that cannot easily be sold into something an insurance regulator counts as a bond. Neither is priced against a market, because there is not one. What the interview does not tell us, and what nobody publishes, is how large the pile of those middle slices held by insurers is, or what the holders say it is worth. Stracke's "not systemic" is a judgment about size. Ours is that the size cannot be established from outside, which a regulator would call the point. This adds nothing new to the pile. It is one more credible perso

The AI capex bubble6 Sepmedium

Whose money is actually in the chips?

When the owners put in almost nothing and the thing pledged to the lender wears out, the lenders funding the AI buildout are the ones betting that a three-year-old chip will still cover their loan.

The chief executive of IREN, one of the companies filling data centers with Nvidia's Blackwell Ultra chips, said this week that its customers' advance payments can cover about half the cost of those chips and that lenders can fund most of the rest. Read that twice. A company is buying some of the most sought-after hardware in the world mostly with its customer's money and its lender's money, and what its own shareholders own is whatever is left over. Funds run by Blue Owl arranged $2.4bn for IREN's British Columbia campus, split between a $1.2bn loan and $1.2bn of bonds, both of which get paid first and have the assets pledged against them. Chinese-language coverage picked up by our sweep put the interest rate at a fixed 9% and said the money covers about 90% of the related chip spending. We have not seen the term sheet. The same stack is everywhere. Nebius raised its debt sale to $5bn. Nscale closed up to $3.05bn of loans, secured on its sites in Texas and North Carolina, and Bloomberg reports it wants $3.5bn more before listing its shares: $1.5bn of bonds that can turn into shares and about $2bn from Nvidia, the vendor whose chips the money buys. Lambda's $926m loan sits in a separate company set up to hold the servers off the parent's books, pays an interest rate that floats 3 percentage points above the market's benchmark overnight rate, comes due in December 2030, and is secured on the servers and the cash they earn. Big Tech's AI-related debt has passed $350bn by one estimate. The mechanism to hold on to is this: a chip is collateral that loses value on a schedule nobody agrees on. If the customer pays half up front and the lender funds most of the rest, the party carrying the risk of what the chips are worth at the end is whoever holds the bond when the contract ends and the hardware is three years old. The debt is secured. The security wears out. The market's answer this week was to buy the shareholders' slivers. IREN rose 26% in five days and sits 0.6% off its high; Nebius rose 7.5% on Friday, CoreWeave 5.6%. Oracle, which has raised $43bn of debt against $55.7bn of building, closed at its high. Prices say the question of what used chips are worth has been answered. The filings say it has not yet been asked. This is fragility, and the slow kind: nobody is going to be forced to sell a chip this month. But every one of these loans rests on a price for used hardware that nobody has had to name, and the more of them there are, the more it matters when

What would change our mind

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

An auction of long-dated US, British or Japanese government debt that goes badly or fails outright, meaning a collapse in how many dollars of bids arrive for every dollar on offer rather than just a high rate, would move ignition sharply.

Would move the number

A grading firm awarding OpenAI or Anthropic the safe-borrower stamp on the strength of money raised from shareholders rather than cash the business earns would raise fragility; so would a second large open-ended fund that lends to companies limiting withdrawals alongside BCRED.

Would move the number

The extra interest the shakiest companies pay compared with the government moving past 3.5 percentage points, or the very lowest-graded companies past 12 percentage points, with the VIX, the market's turbulence gauge, above 25, would tell us ignition has arrived in prices rather than in the reporting.

Would move the number

BCRED paying every withdrawal request in full next quarter, or the share of loans that lending funds have stopped counting interest on falling in their third-quarter filings, would mean a valuation nobody had tested was holding, and would lower fragility.

Would move the number