CoreWeave sold a set of bonds due in 2032 four months ago at full price, 100 cents on the dollar. On Wednesday they traded below 78. At that price the bonds pay anyone who buys them an interest rate above 15%, and the cost of insuring the company's debt against it failing to pay now implies a better-than-even chance that happens within five years. The shares fell as much as 17%, the most since February, and at least seven bonds backing data centers tied to CoreWeave have traded lower this week (Bloomberg). A bond losing more than a fifth of its value in four months is not a bad quarter. It is the market saying it doubts the money comes back.

The money financing CoreWeave sits in layers. At the top are loans backed by its chips and by contracts with named customers; the $8.5bn loan tied to Meta won a top safety grade from the rating firms in March, the grade that lets insurers and pension funds hold it. Below that are bonds backed by nothing in particular, like the 2032s, and below those the shareholders. The plan, as co-founder Brannin McBee described it to the FT this week, was to keep winning high safety grades for the chip-backed debt and eventually for the company itself, which would unlock "all the money in the world that sits in the insurance tranche." In other words, move the financing to life insurers, who hold a loan until it is repaid and never have to show a daily price for it. That is this site's whole thesis, stated by someone on the inside.

The public market got there first. The bonds backed by nothing in particular are the one layer where a price is set every day, and that price now says the shareholder cushion protecting the chip-backed lenders is in doubt. Gil Luria of D.A. Davidson, who has called CoreWeave "the subprime tranche" of AI, meaning the riskiest slice of the boom's borrowing, put the arithmetic plainly: "They've run out of the ability to borrow even at 9%," and a company borrowing at 9% to earn 5% does not fix that by getting bigger.

What we cannot see: how the insurers and funds holding the chip-backed loans value them. Those loans do not trade. No borrower has been reported to break a promise in its loan contract and no lender has been reported to withdraw a credit line. But 15% is not a rate at which a company renting out AI computing power can borrow again to pay off what is coming due, and the rest of the sector closed Wednesday as if it knew: Nebius down 10%, IREN 12%, Core Scientific 13%, Applied Digital 9%, Oracle 9%. This is ignition, not new tinder. The debt was already there; what moved is how close the match is.

South Korea's main stock index, the Kospi, fell nearly 8% on Thursday, taking it 21% below its September high, past the line traders call a bear market. The exchange stopped trading for 20 minutes as the selling deepened. SK Hynix and Samsung Electronics, the two memory-chip giants, each lost more than 9%; Japan's Nikkei 225 dropped almost 5%, with Ibiden and Kioxia among the worst hit; and the main US index of chipmakers has now shed $1.6 trillion of value (Bloomberg). This is the third day this week Korea has been hit, and the hardest. Trading halts in a major market are rare; they exist for days like this one.

Why Korea, when the borrower in trouble is in New Jersey? Because Hynix and Samsung are the two most easily sold stocks on earth whose profits depend directly on how much the giant cloud companies spend building things. When a global fund decides at 9am to cut its AI exposure, it cannot sell a loan backed by CoreWeave's chips or a stake in a private data center; nobody is standing there to buy. It can sell memory makers in size, in an hour, with buyers waiting on the other side. "Korea is becoming the shock absorber for the global AI selloff," said Vantage Global Prime's Hebe Chen. The IMF, which lends to countries in trouble, made the same point in advance in its stability report this week, naming the stock markets of rich Asian countries as the most sensitive to whether AI pays off (Enterprise).

The chain runs like this: CoreWeave's bonds break, US inflation comes in hot, bets on the Federal Reserve raising rates rise, and the bet that gets closed is the one that can be closed. South Korea's won and Taiwan's dollar are leading the currency losses for the same reason. Nothing we can see suggests Hynix or Samsung have a money problem of their own; they are being sold because they are sellable.

The US figures in our tables are too narrow to show the scale. The Nasdaq 100 closed Wednesday down 1.9% and 3.3% off its record; the S&P 500 is 1.7% off. Korea has just lost a fifth of its value in a month. That lopsidedness is the point: the losses are landing on the suppliers, in the market where selling is easiest, before they land on the lenders. For the odds of a crash this is ignition moving closer, not fragility growing. Forced sellers have appeared; the question is whether they run out of easy things to sell.

On the same Wednesday that CoreWeave's bonds fell into distress, Anthropic told prospective investors at its San Francisco headquarters that its revenue in the third quarter reached $16.8bn, that it made a profit in the quarter by its own adjusted measure, that more cash is coming in than going out, and that after the direct cost of serving customers it keeps 85% of its revenue, with more than 11% left after all running costs. It plans to start formally selling shares to the public as soon as the week of November 9 and to be trading before Thanksgiving (Bloomberg).

Take the numbers at face value for a moment and they answer the question the whole building boom rests on: is anyone actually paying for this? One model company, at least, is being paid handsomely. Last week OpenAI told investors its yearly run rate of revenue was $20bn below what it had signaled; this week its rival says it is generating cash.

But look at where the money is kept. Keeping 85% of revenue after direct costs means the computing power that serves the model is a small fraction of what customers pay. That computing power is what CoreWeave and its peers borrowed to build. The model company keeps the margin; the infrastructure company keeps the debt, and at Wednesday's prices earns, by Gil Luria's estimate, 5% on money it borrows at 9% or more. If Anthropic's figures are right, the AI boom is not without economics. They are concentrated at the top of the chain, and the bottom of the chain is financed as if it owned them.

The caveats are real. These are adjusted figures, told to prospective buyers of the stock by people familiar with the presentation, not audited accounts. The company does not say what "adjusted" leaves out, and its commitments to buy computing power are the obvious candidate. Nor does it say how fast the $16.8bn is growing. The formal document it must file before selling shares will. Until then the day's most interesting disagreement stands: the customer reports a profit, and the supplier's bonds trade at 78. On its own this nudges the odds of a crash neither way. It tells us where the money is, which is something we did not know this morning.

The interest rate the US government pays to borrow for ten years fell by 0.1 percentage points to 5.09% in Asian trading on Thursday, its lowest in three weeks, and the rate on every other length of government debt fell at least 0.07 points. The Bloomberg dollar index, which tracks the dollar against a basket of other currencies, dropped with it, and government bonds in Japan, Australia and New Zealand rose. A gauge of what US government bonds have returned is now up almost 0.7% in October after a 2.2% loss in September, which if it holds would be the best October for those bonds since 2014 (Bloomberg). Our own tables have the ten-year rate at 5.05% this morning, from 5.32% on Tuesday.

This happened on the day US consumer prices were reported to have risen 0.6% in September, the fastest since April. Austan Goolsbee of the Federal Reserve's Chicago branch, the central bank that sets American interest rates, called the figure "discouraging" on air, pointing to inflation in services above 5%, the core rate above 4% for the month, and "very little that we're hearing that sounds like downturn." He named the AI data-center buildout as a possible source of overheating (Yahoo Finance). Normally a hot inflation number pushes government borrowing costs up, not down. That is why the direction matters more than the size.

For three weeks the defining feature of this selloff was that stocks and bonds fell together. That is the condition in which a second bet placed to limit the loss on the first stops working and forced sellers appear; it is why Chile's pension funds are closing out an interest-rate bet that has lost $7.7bn. On Thursday the old relationship came back: shares fell, government debt was bought. "The market has finally moved back to the old flight to quality," said ITC Markets' Kit Lowe. In other words, when investors got scared they ran to government bonds again, which is what they used to do.

What the rally is betting on is less comfortable than it looks. A 12% rout in chip stocks and a Korean bear market are being read as something that cools inflation, or as something the Federal Reserve cannot raise rates into. Goolsbee's logic runs the other way: the building boom is itself the overheating, so a central bank that raises rates to cool it raises the cost of the debt that funds it, CoreWeave's 15%. The IMF counts 14 of 33 major central banks tightening since the second quarter. The rush into safe assets is real and it is the best thing in today's data. It makes a crash less likely, and we say so plainly. It is also a bet that the Federal Reserve will blink before the borrowers do.

Goldman Sachs and UBS are among the global banks becoming more selective about the specialized financing they give to Australian fund managers who lend to companies, according to people familiar with the matter. New lending is going to the largest, top-tier managers; smaller firms and those heavily exposed to Australian real estate are being steered away from. Existing commitments are largely being kept (Bloomberg).

The plumbing matters here. A fund that lends to companies in place of a bank does not only lend; it borrows too. Banks provide the loans at the fund level that let it put money into deals before investors' cash arrives and, the important part, pay back investors who want out without selling loans. That is how a fund holding property debt nobody could sell quickly can tell its investors they may withdraw every quarter. Take the bank's money away and the fund has two options: freeze withdrawals, or sell loans into a market with no buyers, which forces it to admit a price.

Australia is a small test case with the right ingredients. About half of its A$200bn ($140bn) market for this kind of lending is tied to real estate. Bathla Group, a Sydney developer, went bust in August owing A$3.4bn, largely to these lending funds, and the regulator has been halting fund offerings. "The froth is coming out of the market," says Corval Avenue's Belinda Bible.

The lenders are global, and the behavior is the tell. The New York branch of the Federal Reserve last week counted $1.5tn of US bank lending to firms that are not banks; this is that pipe being squeezed at the small end, by the same institutions whose examiners have been asking about it. The publicly traded firms that run these funds keep falling with no American event to explain it: Blackstone down 3.5% on Wednesday and 21% off its high, Blue Owl down 5.4%, Prospect Capital 26% down in a month. Meanwhile the last reading we have of the extra interest shaky companies pay compared with the government, 3.16 percentage points on October 13, has barely moved. The shares of the lenders are being repriced. Their loans are not. This stacks tinder rather than lighting it; the loans still have not been tested at a real price, and the day they are is the one to watch.

Cash Today Inc., a consumer lender in Edmonton founded in 2005, reported more than 44,000 outstanding loans worth about C$229m at the end of last year. When the court-appointed receiver, MNP, inspected its offices last month it found records of 256 active loans worth less than C$900,000. Its creditors and the holders of its IOUs, about 270 of them, have potential claims of more than C$613m, a figure MNP says may be overstated by at least C$190m, which still leaves a gap no honest balance sheet, the list of what a company owns and owes, can close (Bloomberg).

The structure is the smallest possible version of the one this site writes about every day. Investors bought IOUs, called debentures, from a high-cost consumer lender outside the banking system. Some of those agreements required the money to sit in separate accounts until it was used to buy portfolios of loans. Several investors have shown the receiver records of their money going instead to an account belonging to a related company. Cash Today told the receiver it had no bank accounts in its own name, and the bank records agree. MNP cannot yet establish whether investors' money was used as the agreements required, or what the company's assets and debts actually are. One of the two men who controlled it died in the week of September 21; the other resigned as a director on October 1.

C$613m is not a number that threatens the system. What it shows is a mechanism: the thing pledged to the investors was a reported count of loans, and nobody checked the loans existed until the money stopped arriving. Canada's lightly regulated lending industry is where this one surfaced. The same question, who has actually seen the assets behind the income, is the one being asked this week of a far larger set of private lenders: in Australia, in New Jersey, and in the Federal Reserve's visits to the Wall Street banks that lend to funds. This changes the odds of a crash hardly at all. We include it because it is the pattern in miniature.

Chile's seven private pension funds, which hold about $229bn and are the only retirement savings most Chileans have, are sitting on a $7.7bn loss from a bet that US interest rates would fall, according to figures the pension regulator released on Tuesday, Bloomberg reports. A year ago the same bet showed a $1.4bn gain. The loss is roughly 3.4% of everything the system holds, and the funds are not allowed to wait for it to come back: new rules taking effect early next month force them to close the bet, in stages, whatever the price.

The mechanism matters more than the number. Over the past two years the funds (known in Chile as the AFPs) built up a large position in contracts tied to US dollar interest rates. The usual version is a swap, in which the fund agrees to receive a fixed interest rate for years to come; if rates fall, that fixed rate becomes worth more than the going rate and the contract gains. Rates rose instead, hard, through the late-August selloff, and the contract went the other way. Ordinarily a long-term investor can sit on a loss like that and wait. The AFPs cannot: the new regulations force them out, in stages, against interim thresholds. To exit they have to pay Bank of America, Citigroup and the other banks on the opposite side the full amount the contracts are now worth against them. "They're going to have to realise the loss at some point," Euroamerica's Felipe Alarcón told Bloomberg.

This is what a forced seller looks like when the force is a calendar rather than a demand from a lender for more cash because a bet has moved against you. The banks know the deadline, know which way the flow has to go, and have their own offsetting bets to unwind as the contracts are torn up. That, by our inference rather than anything in the report, means more of the bet that rates rise, and more selling of US government bonds that do not come due for a decade or more, into a market where the government's ten-year borrowing rate is already back at 5.32%. How much is already closed and how much remains is, Bloomberg says, unclear.

It also cuts across the usual story of hidden borrowing. These are not hedge funds, the investment firms that manage money for the wealthy and can bet in ways ordinary funds cannot. They are the regulated, conservative, compulsory savings plan of a middle-income country, and the regulation meant to make them safer is the thing turning a paper loss into a real one. On the meter this is a spark, not tinder: nothing new was added to the pile, but the thing that could set it off just moved closer.

New bonds sold by companies the rating firms consider safe did worse, relative to the rest of that market, in the ten days to October 8 than at any time since the depths of the pandemic, six and a half years ago, Bloomberg reports, citing a JPMorgan note dated Tuesday (new bonds lagged JPMorgan's index of such debt by 0.27 percentage points). One company caused it. Skydance's bonds, sold on September 30 to pay for its takeover of Paramount, have been falling since the moment they started trading. That matters because the companies building AI data centers are about to sell tens of billions of dollars of the same kind of bond.

The honest caveat is in the note itself. Leave Skydance out and new bonds actually did slightly better than the market (by about 0.04 percentage points). So this is not yet a market-wide reset. But the market for new bonds does not run on averages; it runs on the last deal. Investors buy a new bond because it comes a little cheaper than existing ones and usually rises afterward. When a $41bn deal does the opposite, the next buyer demands a bigger discount, and the company, seeing the discount, waits. That is exactly what JPMorgan says happened: companies sold far less debt last week than dealers expected, even though money was flowing into bond funds and the swings in interest rates were calming down.

For scale, in April 2020 the extra interest these companies paid compared with the government was 2.7 times what it is now. Last week's problem opened with that gap narrow (0.83 percentage points as of October 12, unchanged on the day), which is why the strategists call it "far worse by comparison". The gap says nothing is wrong; the behavior of new deals says buyers have stopped trusting them.

Why this matters beyond a media takeover: the AI build-out is paid for through this window. Oracle, Broadcom and SpaceX are lining up tens of billions in the same kind of debt this quarter. A market that is choosy about the discount it demands is a market where some of that borrowing does not get done, or gets done at prices that change the math. On the meter this is tinder rather than a spark, and a small piece of it so far: one deal, but the deal everyone is pricing the next one off.

Character Technologies, the company behind the Character.ai chatbot app, filed for bankruptcy protection in Delaware on Monday to sell most or all of what it owns, the Wall Street Journal reports. Disney has set the opening bid: $20m, for a license to some of the technology that never expires and is not exclusive, plus help with the handover from former Character.ai staff who already work at Disney. Two years ago this company was one of the stars of AI. If nobody beats Disney's offer by late November or early December, the service may simply be shut down.

In 2024 Google paid $2.7bn to license Character's technology and hire its original co-founders. In September Disney hired its then chief executive, Karandeep Anand, as its first chief technology officer, and several technical staff went with him. What is left is the shell: the users, the lawsuits (families in at least four states sued over teenage self-harm, since settled) and whatever stake the remaining investors still carried.

This is the mechanism behind the valuations of the venture-funded AI companies, shown from start to finish. The big company does not buy the start-up. It buys the people and a license, which keeps the deal away from the regulators who review mergers and leaves the liabilities behind. The headline number, $2.7bn, gets reported as what the company was worth. The filing does not say how much of that reached shareholders, and nothing in the record tells you what the business itself was ever worth. Now we know: a license Disney will pay $20m for, and will not even get exclusively.

Character.ai is small, and on its own it changes the odds of a crash very little. The reason to pay attention is that the same structure, talent and license out, shell and liabilities left behind, has been used at far larger scale across the sector, and this is the first time one of the shells has gone through a court-supervised sale where somebody has to name a price. The world did not get more dangerous this week. We just found out what one of these shells was worth all along.

SoftBank is paying the highest interest rate in its history, 9.75% a year, on last month's record $11bn sale of bonds in the category reserved for companies too shaky to be considered safe, and its finance arm has now explained, on the record, why that is fine. Alex Clavel, chief executive of SoftBank's investment unit, told Reuters that the rate is "reasonable" given how much its stake in OpenAI has risen in value. "We've borrowed money in large part because we don't want to sell assets that we believe will keep rising in value," he said. The company is borrowing real money, at a steep rate, against a gain nobody has cashed in.

Take that sentence apart. SoftBank's backing for OpenAI totals $64.6bn. It used to pay for bets like that by selling things it owned, Nvidia shares, T-Mobile US shares, which turns a paper value into cash. Now it borrows against the paper value instead. That value is a private one, set in funding rounds, which nobody has tested by selling. The interest is cash: on our arithmetic, roughly $1.07bn a year on this one bond sale, before the rest of what SoftBank owes.

This is the loop Crash Lab keeps coming back to, in its purest form. OpenAI's valuation rises; SoftBank borrows against the rise to put more money into OpenAI; the new money holds up the valuation. The debt sits with investors in shaky-company bonds who have been told, explicitly, that the thing making the interest rate reasonable is a paper gain on an unlisted company whose annual revenue figure was revised down by $20bn last week.

Reuters notes the sale "added to a sharp surge in corporate borrowing this year" alongside Amazon, Alphabet and Tencent. The difference is that those three borrow at the low rates offered to companies considered safe, against the cash their businesses actually bring in. SoftBank borrows at 9.75% against a paper valuation. If the valuation holds, the trade is brilliant. If it does not, the interest still falls due, and what Clavel describes as not wanting to sell becomes not being able to. On the meter this is tinder: more borrowed money stacked on an assumption nobody has tested.

Hong Kong, which fixes its currency to the US dollar, is running low on the cash cushion it uses to defend that fix. The pool of cash its banks hold with the city's monetary authority, called the aggregate balance, has fallen to HK$54bn, about $6.9bn, a level crossed only once before, in 2023, during an earlier stretch of currency defense, according to a Bloomberg analysis of data from the Hong Kong Monetary Authority, the HKMA, which plays the role of the city's central bank. On Wednesday the Hong Kong dollar traded a hair from the weakest point the fix allows (within 17 pips, tiny fractions of a cent, of 7.85 to the US dollar). This is what the US bond selloff looks like when it lands somewhere else.

The plumbing is simple and unforgiving. Higher interest rates on US government bonds make traders prefer US dollars; to hold the fix the HKMA buys Hong Kong dollars; every purchase drains cash from the banking system. With HK$54bn left, even modest interventions can send the rate Hong Kong's banks charge each other, Hibor, jumping, and that rate feeds straight into mortgages. DBS's Carie Li sees the trend in local rates as "set to be upward" and "volatile" if the balance hits its previous lows. The HKMA has already told the public to manage interest-rate risk on property purchases and borrowing. The alternative, the HKMA putting cash back in by not rolling over the short-term bills it has sold, which OCBC expects if the drain gets too fast, would, in Bloomberg's words, be "counterproductive" to the intervention it is meant to support.

India is fighting the same shock with different weapons and more of them. Even after raising $133bn through a deposit scheme aimed at Indians abroad and lifting interest rates for the first time in nearly four years, the rupee slid toward 97 per dollar, with analysts eyeing 100. The Reserve Bank of India, the country's central bank, has moved its response out of the markets and into the economy: state-owned oil refiners, who need about $10bn a month for crude, are being taken out of the currency market altogether, and it has been made more expensive for everyone else to lock in a dollar price in advance. Importers' advance purchases of dollars are running at a record $64bn a month, up 44% on last year, Bloomberg reports. Governor Malhotra calls the market "unduly pessimistic". That bought a 0.5% rally on Tuesday.

Two mechanisms, one source. A 5.32% borrowing rate for the US government is a Hong Kong mortgage rate and an Indian import bill. Neither is a spark on its own; both are the pile getting bigger in places far from where the fire started.

Payward, the company that owns the Kraken crypto exchange, is buying Saldo Bank, a Lithuanian lender that has held a European Union banking license since 2021, the Financial Times reports. No crypto company has bought a bank in Europe before. Saldo takes deposits locked up for a fixed term and lends to consumers and businesses; the price was not disclosed. What makes it worth your attention is what Kraken says it wants the bank for: lending against crypto with a bank's own deposits.

The stated purpose is not holding customers' coins or moving payments. Co-chief executive Arjun Sethi told the FT that customers want to "borrow and lend against their assets without selling", with cash, crypto, tokenized stocks and credit "all under one system". Read plainly: a deposit-taking bank, inside the EU's regulatory fence, lending against crypto pledged as security.

This is the step the US has so far refused. Coinbase and Ripple have conditional approval from the OCC, the regulator that supervises national banks, for a narrower kind of charter that lets them hold assets and process transactions but, as the FT notes, not take deposits. Community banks are suing the OCC this month over even that, calling it "a side door into the banking system". Payward has gone through the front door in a country that allows it.

Why Crash Lab cares: until now the link between crypto prices and the banking system has run through exchanges, the reserves behind stablecoins (digital tokens meant to be worth exactly one dollar) and the occasional overexposed bank. A licensed bank lending depositors' money against bitcoin makes the link direct. The loan is fine while the bitcoin is worth more than the loan. On Tuesday night bitcoin dropped from $83,000 to $81,000 and back inside hours, and $800m of bets made with borrowed money were forcibly closed on the way. A bank running the same loans would not close them. It would hold the loan and the bitcoin and carry the gap. Which supervisor puts a value on that, and how, is now a live question in Vilnius.

One bank, size undisclosed, in Lithuania. Small, and it moves the odds very little today. But it is the template, and Sethi says so: "we're just going to be new versions of banks".

JPMorgan and Goldman Sachs are trying to sell a $5bn loan package for Volta Infrastructure Holdings, a new company that rents out computing power for AI. They have promised to deliver $1.3bn of it, an ordinary company loan, whether or not other investors buy. The other $3.7bn, secured on the chips Volta is buying, carries no such promise: it is being offered at an interest rate of about 11%, among the highest paid by any heavily indebted company right now, and it has no grade from the firms that rate how likely a borrower is to pay you back (Bloomberg). The two banks that arrange most of this kind of debt have just said which half of it they will not put their own money behind.

The difference between the two slices is the whole story. When a bank underwrites a loan, it promises the borrower gets the money whether or not anyone else buys the debt; if investors refuse, the bank keeps the loan and swallows the loss when it eventually sells at a discount. That is what happened in 2022, when rising interest rates left banks holding billions of loans they could not pass on. Selling the chip slice on a best-efforts basis, meaning they will look for buyers and walk away if there are none, is the banks saying, as plainly as a loan document can, which part of an AI company's balance sheet, the list of what it owns and owes, they are not prepared to own at today's prices.

The missing grade shuts the usual door. The biggest buyers of loans like this are pooled funds that package hundreds of company loans together (collateralized loan obligations), and they generally cannot hold debt that has no rating. So the natural buyer of an 11%, unrated, chip-backed loan is not one of those funds and not a bank. It is a fund that lends to companies in place of a bank, or the investment pool an insurer uses to back the policies it has sold, which is where we have argued all year the risk has been heading. That is our inference. Bloomberg's sources do not say who has been offered the loan.

For the loan to work, two things have to stay true: the chips have to hold their value as security for as long as the debt runs, and Volta's customers have to pay enough for it to cover 11% a year. Whether the $3.7bn finds buyers, and at what price, is the single most useful piece of evidence the AI debt market will produce this month. On the meter this is ignition, not fragility. The debt was always going to exist; what changed is that a named lender declined to stand behind it, and a buyer stepping away is how a spark gets closer.

Samsung Electronics and SK Hynix, the world's two largest makers of memory chips, fell as much as 8.8% and 9.4% in Seoul on Tuesday morning. Companies that size do not usually move that far in a morning. Together they are about half of the Kospi, Korea's main stock index, which dropped as much as 7%; the pair have lost almost 15% in five days. SoftBank, which owns more than 10% of OpenAI, fell nearly 8% in Tokyo, and the Nikkei, Japan's main index, fell 2.7%. Micron lost almost 4% in New York on Monday (Financial Times). The firms that sell the AI buildout its parts were repriced in a few hours, and the people who were supposed to fund the buildout had already started backing away.

Two things set it off. CXMT, a Chinese chipmaker, claimed advances in memory chips, which threatens the pricing power the two Korean companies have enjoyed as nearly the only suppliers. And Hock Tan, chief executive of the chipmaker Broadcom, told an event in Singapore that "a lot of investors have no clue what they are investing in" and set expectations so extreme that "it's almost like they are in a casino". The Kospi more than tripled between early 2025 and its July peak and was the world's worst-performing market in the third quarter. Markets do not triple without a crowd piling in, and this is the crowd leaving, not a surprise.

The share-selling side of the machine was already retreating. The Australian Financial Review reports that Coatue Management, already a major backer of Firmus, had been expected to buy up to $1bn of the AI data center group's share sale on the Australian exchange and instead took under $100m, the cold feet that sank last week's listing (AFR). An existing backer cutting its order by nine tenths is a vote, not a negotiation. And AirTrunk, the data center operator owned by Blackstone, has pushed into next year its plan to list a property fund in Singapore (a REIT, a listed company that owns buildings and pays the rent out to shareholders), which would have raised more than $1bn, citing turbulent markets and uncertainty over interest rates; VITRO REIT in the Philippines put its own listing off to 2027 earlier this month (WSJ).

The mechanism is simple. Data center builds were to be paid off by selling shares and property-fund units to a public willing to accept a low return. If the public will not buy, the private owners and their lenders hold the asset for longer, at interest rates that have risen a full percentage point since the plans were drawn. On the meter this is the fast-moving kind. The debts and the buildings were already on the pile; what moved on Tuesday is how close the market is to deciding they are worth less than the plans assumed.

Australia's markets regulator, the Australian Securities and Investments Commission, has ordered a 21-day halt on one class of units in the Ares Global Credit Income Fund, run by Fidante Partners with Ares Management, a big lender to companies that borrow from funds instead of banks. For three weeks Fidante may not deal in the product, hand out its disclosure document or recommend it to ordinary savers (Bloomberg). Nobody has said the fund lost money. The regulator has said the fund is being sold to people it may not suit, and in this corner of the market that is the more serious charge.

The order is about a document, not a loss. Every product sold to ordinary Australians must carry a statement saying who it is designed for (the target-market determination). ASIC questioned three things in this one: whether a fund that lends money is suitable for investors who want to preserve their capital, whether the time it takes investors to get their money back is appropriate, and a clause allowing a person to put as much as 50% of their investable assets into the fund.

That last point is the one to sit with. A fund of this kind holds loans that do not trade on any market, so its unit price is the price the fund says its holdings are worth rather than a price anyone paid, and the money it can hand back is whatever the manager can find from new money coming in and loans being repaid. Telling an ordinary saver such a thing preserves capital and is fit for half their wealth is the selling end of the migration we keep describing: risk leaves banks, where regulators make them hold a cushion of their own money against it, and arrives in a wrapper sold to a retiree.

ASIC warned last month of tougher enforcement after the collapse of the Sydney property developer Bathla Group. We do not know whether the fund's loans have gone bad; the regulator has not said so and the order does not require it. Ares' shares have fallen from their recent high (down 15.6%), which says the market has already priced some of this in. Whether Fidante rewrites the document or the order is extended will show how hard the regulator intends to push. On the odds of a crash this changes little. Nothing was added to the pile and nothing got safer; we learned that a regulator is now looking at the end of the pipe where the risk lands, which, if it keeps looking, is how that risk gets smaller.

Investors have pulled money out of US funds that hold mortgage bonds (bonds made by bundling home loans together and selling the income) over the past three weeks at a pace that ranks among the seven worst stretches since the start of 2021. The main index of those bonds had its second-worst September on record, and its return so far this year is the second-most negative in the index's 50-year history (Bloomberg). This is the market that was supposed to be dull. The bonds are backed by the government, every bank and insurer owns them, and only one year in the last fifty has been worse.

Mortgage bonds get hurt twice when interest rates jump. The price falls, as any bond's does. Then the homeowners whose loans are in the bundle stop refinancing, because a new mortgage costs more than the old one, so a bond that was expected to be paid off in a few years now runs for decades. How badly its price falls when rates rise gets bigger at exactly the moment that hurts most. That is why a fund whose investors want out has to sell more of it into a market with fewer buyers. And the biggest natural buyers, the banks, are thinning out: Erica Adelberg of Bloomberg Intelligence notes that the Federal Reserve, America's central bank, keeping money tight, and a smaller gap between short-term and long-term rates, reduce how much banks want to buy.

The other end of the same pipe is the house. The average interest rate on a 30-year mortgage reached 7.4% on October 8, up from 6.66% in late August. The number of homes going under contract, by Redfin's count, fell 4% in September, the biggest drop in three years, to the lowest level since 2014 outside the first months of Covid; it fell in all 50 big metro areas at once for the first time since September 2023. Nearly one in five sellers cut their asking price, and 14.5% of signed purchase agreements fell through, the most ever recorded (Bloomberg).

The consolation Adelberg offers is real: fewer purchases and no refinancing means fewer new mortgage bonds created into the end of the year, so there is less for the market to absorb. But supply falling because nobody can afford a house is not the kind of support a market wants. On the meter this adds nothing to the pile; the bonds were already there and already counted. What it does is open a route by which selling forces more selling, and that is the ignition side, the fast one.

Amazon and Alphabet, Google's parent, have between them borrowed about $70bn from bond investors since July 1. Over the same period, the governments of developing countries sold $45bn of bonds to international investors, half of last year's $91bn; the safest of those governments have borrowed 56% less, $29bn, and none at all in October. The average interest rate on their dollar debt has risen almost a full percentage point this half-year to its highest since April last year. Two companies borrowing more in a quarter than every developing-country government put together is what crowding out looks like when you can see it. Merveille Paja of Bank of America puts it directly: "The risk that hyperscalers actually are crowding out emerging markets really is a risk for our asset class." In other words, the giant cloud companies may be taking the money that would otherwise have been lent to poorer governments. Those governments have $27.7bn of dollar and euro debt falling due by the end of the year, Saudi Arabia ($5.2bn) and Turkey ($3bn) first (Bloomberg).

The mechanism is the oldest there is: when the safest borrower in the world pays 5.2% and the biggest companies in the world pay a little more, the last spare dollar does not travel to Kampala. Uganda's currency, the shilling, fell to 4,143 per dollar on Monday, its weakest since Bloomberg's records begin in 1993 and the worst performer of 144 currencies since the selloff in US government bonds began on August 25, down more than 10%. Citi estimates more than $200m has left Ugandan markets in two weeks; last week's government bond auction was carried by local banks and pension funds, which is what it looks like when outside buyers stay away. President Museveni has ordered the central bank not to sell dollars to prop the currency up, a government choosing to let its money fall rather than spend its reserves (Bloomberg).

The IMF, which lends to countries in trouble, published its stability report on Tuesday. It lists a spike in government borrowing costs and an AI selloff among the "plausible near-term catalysts", notes that markets are struggling to absorb the volume of government and company borrowing, and urges tighter limits on borrowed money in AI; "resilience is not necessarily safety" (Reuters). The poorest borrowers are where that sentence is being tested first. On the meter this is fragility, and most of it we already count. What is new is that a whole class of borrowers has lost its buyers, and $27.7bn coming due with nobody to lend is the slow pile edging toward the fast kind of problem.

Getty Images, the photo agency, will file for bankruptcy in the coming weeks with the deal already agreed, and come out of it as a private company owned entirely by Koch Equity Development and members of the Getty family, who are putting in $533m of fresh money. Holders of 83.5% of its unsecured bonds have agreed to cancel $270m of what they are owed. Holders of about 89% of its secured bonds and loans, and every bank behind its credit line, have agreed to cut the interest and push back the dates the money is due. Koch will lend up to $143m to keep the business running while this happens, which is meant to be finished by the end of the year (WSJ). Nothing in Getty's business broke. Its debt was sized for cheap money, and cheap money is gone.

A prepackaged bankruptcy is bankruptcy with the fighting done in advance: the lenders' votes are collected before the filing, so the court part is short and the business keeps trading. The ordinary kind can take months of argument. Read who gets what and it is a tidy map of who bears an interest-rate shock. Public shareholders get nothing. Lenders with no security behind them accept less than they are owed. Lenders who hold security, something Getty pledged that they keep if it cannot pay, give up interest and time rather than the money itself. And a family investment office with patient money buys the whole company for the price of putting it back on its feet.

This is the undramatic version of the thing we track. Getty is not an AI story and its lenders are, as far as the filing says, ordinary ones. It is simply a company whose debt was sized for a world where money cost less, meeting a world where the US government pays more than 5% to borrow for ten years. JPMorgan counted $65bn of loans to heavily indebted companies in deep distress last week, priced as if the lenders expect losses, the most since March 2020, the month the pandemic hit markets; Getty is one of the orderly ones. The disorderly ones are the ones without a Koch. On the meter this is the rare case that takes something off the pile: one small, overborrowed company is being sorted out without a fire sale, and days like that deserve the same care as the other kind.