The euro fell as much as 0.8% on Monday to its weakest level since May 2025 ($1.1161, in Asian trading). The trigger was reports that Spanish officials are preparing for an early election, landing on a French bond market that had closed Friday paying its biggest premium over Germany to borrow since 2011 (Bloomberg). A government's bonds selling off is that country's problem. Its currency selling off belongs to everyone who shares it.

On Saturday we wrote about risk managers at banks and funds ordering traders to cut their borrowed bets on French bonds. That was one market, one country, one group of forced sellers. What happened overnight is the next stage. Fast-moving funds in Asia sold euros for dollars outright, and the fall pushed the currency through price levels that set off automatic selling by traders who had written insurance against exactly that move, according to traders Bloomberg spoke to. A borrowing-cost problem in Paris has become a currency problem for the whole bloc. French bond futures, the contracts traders use to bet on where prices go, slipped on Monday while German ones edged higher. The market is not fleeing Europe. It is sorting it.

The thing worth understanding is why Spain matters. France on its own can be told it is a one-off. Two of the euro area's large, heavily indebted members in political trouble at the same time cannot be, and Bloomberg points to the debt loads of France, Italy and Belgium as the particular weak point 'in a world of rising borrowing costs'. The 2011 comparison is the one everybody in the room remembers, which is why Lombard Odier's Homin Lee talks about an 'erosion in the country's fiscal anchor'. In other words, investors are losing faith that France will keep its borrowing within limits.

The other side of this trade is the dollar. The dollar index, which tracks the dollar against a basket of other currencies, is up 0.3% on the day and 3.1% over twenty days, and gold is 0.4% firmer ($4,179). So the dollar is being bought as a safe place to hide at the same moment the US government is paying 5.30% to borrow for ten years. That is not a contradiction. When the alternative is a euro coming apart, the question of whether American debt is safe gets postponed rather than answered. For the meter this is ignition, not fragility: nothing new was borrowed this weekend, but a fire that has reached a currency has far more to catch than one confined to a single bond market.

Shareholders holding almost three quarters of Partners Group Private Equity Ltd, a trust listed on the London stock market, have voted to have its investments sold and the cash handed back, Bloomberg reports (Bloomberg). The board has dropped the plan it floated last week to split the trust into two classes of share and will instead ask investors on October 7 to approve an orderly wind-down. That is the holders overruling the manager, which is not how these things usually go.

On Saturday we covered the split as a way of letting some holders stay. The holders have answered: most would rather have the money. This is a different thing from the gates we have been counting all month. A gate is a manager refusing to let investors out. A wind-down is investors insisting on being let out, and getting their way.

Why it matters is the test it forces. Trusts like this one almost always trade for less than the value the manager says the holdings are worth, and Bloomberg notes the gap is a chronic 'source of frustration' for the individual savers who typically own them. The discount is the market's verdict on the manager's valuations. A wind-down settles the argument. The portfolio has to be sold to real buyers for real money, and the proceeds either match the published value or they do not. This trust has already cut its valuations of Emeria and Ammega and written off at least one holding this year, so the first answers may not be flattering.

The context is a manager that in June stopped withdrawals from one of its funds designed to stay open indefinitely, as wealthy clients asked for their money back. A spokesperson says several institutional investors 'will continue to invest with Partners Group in other solutions'. That is probably true. It is also the point: the institutions are staying and the individuals are leaving, in the one vehicle where the individuals had a vote.

On its own the wind-down is too small to change the odds of anything. What we will watch is the price the assets actually fetch against the last published value, because that number travels to every other fund holding similar assets at similar valuations. If the sales come in close, that would be the first real reason to think the hidden pile is smaller than we fear. If they come in well short, nothing new was added to the pile; we would simply learn how big it already was.

Three of the world's biggest pension funds now hold less in American shares than the global benchmarks they are measured against, according to Financial Times research: Australian Retirement Trust, which manages about US$260bn, Canada's La Caisse (US$388bn) and Britain's People's Pension (£45bn), all citing stretched valuations on AI companies and record concentration in a handful of stocks (Financial Times). A Marsh survey of 430 institutions with more than $5tn between them found a third plan to cut their American share holdings over the next year, double last year's share. The money that is built never to flinch has started moving.

This is not commentary. It is allocation by the slowest, largest, least excitable money in the system, and that money changing its mind is rarer and heavier than any forecast. ART's Jimmy Louca does not expect a 'dotcom-style bust'; he simply judges American companies to be 'more than fully priced' and has moved money accordingly. La Caisse's Vincent Delisle calls AI risk 'the single most important factor determining how we want to position ourselves'.

The mechanism is about who the next buyer is. More than a third of the S&P 500 is in giant companies tied to the AI building boom, the FT notes, so what looks like 'geographic exposure' has become, in Marsh's words, 'sector and factor concentration'. When pension funds trim, they do it over quarters, and the index barely notices; the S&P is 0.9% off its record this morning. But the buyers replacing them are by construction shorter-term: the fast money that the stock-listing bankers elsewhere in this edition describe as chasing first-day jumps.

Strictly, this is the kind of thing that should lower fragility over time, a crowded position being spread across more hands. It is not doing so yet, because the position is passing from patient hands to impatient ones and the price has not moved. We record it as a measurement of who holds the risk, not a change in how much risk there is. If the next bout of selling comes, it will meet a market with fewer natural buyers than the index level suggests.

Firmus, a data center operator backed by Blackstone and Nvidia, set the price of its shares at A$11 last Thursday, valuing the company's equity at US$30.6bn, nearly three times the US$10.5bn it was valued at when it raised money in early August (Business Times). The US$5bn share sale would be the second-largest Australia has ever seen. Orders are taken from Tuesday and trading starts October 23. A company nearly trebling in value in two months is not how valuations usually move, and this one is trying to lock the number in just as the market for new listings shuts.

The arithmetic is the story. Firmus carries around US$30bn of debt, roughly six times its forecast earnings according to Morningstar's Lochlan Halloway, who sees 'hallmarks of a boom phase'. That is a heavy load for any company, and here it rests on five planned data centers plus two existing ones producing some US$5bn a year within five years. Merlon Capital's Kirit Hara put it plainly: his process 'prevents us from effectively buying into, kind of, the hopes and dreams'. Even so, early indications of interest exceed the size of the deal.

It arrives as the American door shuts. EG Group, Holtec, Oura and SoftBank's SB Energy have all paused their listings, the FT reports; SB Energy had been aiming for about $50bn 'despite not bringing a single facility online' (Financial Times). Technology companies that did list this year have fallen about 23% on average since their first day of trading, per BCA. One banker's verdict: 'New investors know they're getting a raw deal, all the gains are for the private holders.'

Why this matters for the machine: a listing is the moment a valuation set privately meets a price set in public. When the door is open, a valuation that trebled in two months becomes somebody else's problem. When it is closed, the valuation stays on the sponsor's books untested, and the US$30bn of debt stays with whoever lent it against the hopes and dreams. Firmus is trying to get through before that happens. Whether the orders hold at A$11 by October 23 is the test, and this is fragility rather than a spark: it is one more untested number in the pile.

BT has bought TalkTalk and PlatformX Communications out of administration, Britain's version of bankruptcy protection, after attempts to sell the consumer and wholesale businesses to anyone else fell through (Bloomberg). It puts the total cash cost for its 2027 fiscal year at £400m ($529m). BT says it acted 'in the public interest', and chief executive Allison Kirkby called it 'a genuinely unprecedented situation, where millions of citizens and businesses were at risk'. Britain's fourth-biggest broadband provider ran out of road and found no buyer except the company it was built to undercut.

This is a buyout funded with borrowed money failing in the plainest way. TalkTalk was bought and taken off the stock market in 2021 by Toscafund and Penta Capital, and Bloomberg says it has struggled under its debt ever since. Its business was buying wholesale access to BT's Openreach network and reselling it cheaply, and competition plus inflation ate the margin. Octopus Investments walked away from the wholesale division last month. Epiris did not do a deal. Nobody else came.

The part Bloomberg's report does not say is what the lenders got. A pre-arranged administration normally wipes out the owners and leaves creditors with whatever the buyer pays. The £400m is BT's cash impact, not necessarily the price, and what TalkTalk's lenders recovered is not disclosed. Whoever held that debt, and a 2021 deal of this kind would have placed much of it with investment funds rather than banks, has just had the value they put on it tested against reality. We do not know the answer. We know a test happened.

Two things to note. First, BT was TalkTalk's wholesale supplier as well as its rescuer, which is the pattern we keep seeing in AI: the firm with most to lose from a customer's failure ends up as its lender of last resort. Second, BT shares rose 1.5% and it kept its targets. The market read this as a cheap acquisition, not a bailout. It may be both. For the meter, one over-borrowed company reaching the end adds no tinder and lights no spark; it is a loss that already existed finding its way onto somebody's books.

Investors in Blue Owl's two unlisted lending funds asked to take out $4.2bn in the third quarter, and the funds will hand back a fraction of that. At Blue Owl Credit Income Corp the requests covered 16.8% of shares ($3.1bn); at the technology fund they covered 39% of shares ($1.1bn). Both will pay out 5%, which the Wall Street Journal notes is the industry standard. Nobody is being stiffed, and everybody who wants out is being made to wait.

These are funds that borrow money, lend it to mid-sized private companies, and pass the interest to shareholders, sold to wealthy individuals through their financial advisers. They have no stock-exchange listing, so the only way out is a quarterly offer to buy shares back, and that offer is capped. The cap is deliberate. It stops a stampede forcing the fund to sell loans cheaply, because a cheap sale would set the price every other fund has to admit its similar loans are worth. It also means the queue never clears. A 39% request met at 5% a quarter is, on our arithmetic and assuming nobody else joins, roughly two years of exits.

The signal is split. Requests at the flagship fell for a second quarter, from 21.9% in the first quarter to 18.8% to 16.8%. The tech fund's did not, and the fund blames the high requests on its "specialized investment mandate", which means lending to software companies, the exact place where doubts about AI meet lending by investment funds rather than banks. This is not one firm's problem. Blackstone's flagship had requests for 10% ($4.3bn) last month and Cliffwater's largest fund around 16%; both paid 5%.

Partners Group chose a different tool. On Friday it said it would restructure Global Value, a €6.6bn open-ended fund that buys companies with borrowed money and that capped withdrawals earlier this year, into two sub-portfolios under one umbrella (Wall Street Journal). The account stops there. A split of this kind usually separates what can be sold and returned from what cannot be sold at the value on the books, a quarantine ward under a new name. That is our reading, not the firm's; the announcement does not say which assets go where or whether the two halves will be valued differently.

Nobody is pricing these loans, because nobody can. The market is pricing the managers instead. Blue Owl's shares are down 23% in 20 days, Blackstone's 19%, KKR's 18%, while the listed funds holding similar loans are off only a few percent. The money that pays the fees is leaving; the loans stay where nobody has to say what they are worth. On the odds of a crash, this is tinder rather than spark, and most of it is tinder we had already counted. The pile did not grow this week so much as we got a clearer look at how it is being managed.

America added 29,000 jobs in September against 90,000 expected, and August's figure was cut to 133,000. Traders cut the odds of an interest-rate rise in October to about 21%. Shares rose, the Nasdaq by 1.2% and the S&P 500 by 0.7%. And the interest rate the government pays to borrow for ten years, after dipping when the report landed, closed higher at 5.28%, with the thirty-year rate at 5.63% (Nikkei/Reuters). That last sentence is the one that is not supposed to happen.

In an ordinary year a jobs miss is good for bonds: fewer rate rises coming, so lower interest rates on debt of every length. On Friday it helped only the shortest-dated debt. Across the week the Federal Reserve's own deputies had been pulling in the same direction. Before John Williams spoke on Tuesday the market put the odds of a rate rise at the October 27–28 meeting at 70%; by the time Philip Jefferson finished on Thursday it was 25%, with soft inflation figures in between. Bloomberg notes there is no evidence the two coordinated, with each other or with a chairman who has said he will not signal his intentions. Evercore called the joint message "authoritative" anyway.

So the expected path for the central bank's rate came down all week, and the cost of long-term borrowing went up regardless: the ten-year rate from 5.18% last Friday to 5.28%, the thirty-year from 5.50% to 5.63%. Short rates falling while long rates rise is unusual, and it says the extra interest lenders demand for tying money up for a long time is rising for reasons the central bank's rate cannot reach: energy prices from the Iran war, the sheer quantity of long-dated bonds the government must sell, and, as Reuters puts it, "already stretched public finances".

Company debt followed. On Wednesday the extra interest paid by companies too shaky to be considered safe stood at 3.24 percentage points over the government, from 2.80 a week earlier and 2.65 a month ago. For companies considered safe it is 0.86 percentage points, 9% wider on the week. The stock market read Friday as relief that rates will not rise. The bond market read it as the economy weakening while the government's need to borrow does not. Both readings cannot survive the inflation report due October 14. For the odds of a crash this is ignition, not tinder: nothing new was borrowed, but the thing every leveraged investor relies on to cushion a fall has stopped cushioning, and that moves the spark closer.

Interest rates on short-dated French government debt swung by as much as 0.4 of a percentage point over Thursday and Friday, among the largest moves in years, as hedge funds that had borrowed to own French bonds were forced to sell. In a market that size, a move like that in two days is not a bad week; it is the sound of positions being closed under orders. "It's been a slow bleed until this week," said Hank Calenti of SMBC, "which suggests people were starting to get tapped on the shoulder by risk managers and told it's time to vacate" (Wall Street Journal). In other words, the funds did not change their minds; the people who police their losses made them sell.

The trade is simple. Borrow cheaply and short-term, either as cash borrowed overnight against bonds put up as security or in yen, buy French bonds that pay a little more interest than German ones, and use borrowed money to make that small difference several times bigger. France was the preferred venue because its €3.5tn market, Europe's largest, is deep enough to get in and out of in size. The weakness is that the bet is sized to how calm prices are. A fund's risk limit shrinks when prices start to move, so the position must be cut whether or not the manager still believes in it. The cut moves prices, which shrinks the next fund's limit. The Journal describes "a self-fueling bout of losses and selling", and Pimco's Konstantin Veit says plainly that hedge funds are closing out these borrow-cheap-lend-dear trades "of which France has been part".

Who is left holding the bonds matters. More than half of French government debt, $3.9tn, is owned by foreigners, according to MUFG. The Japanese buyers who filled that order book for a decade can now earn comparable interest at home or on German bonds, and a week ago one large Japanese manager said it had sold every French bond it owned. On Thursday the gap between what Italy and Germany pay to borrow for two years doubled, as the stress moved south.

French shares and the euro rose on Friday, and the European Central Bank stands behind the eurozone's bond market as a last resort. But a forced seller in the government bond market of one of the seven largest economies is the thing this site watches for, and this week one appeared with its machinery visible: borrowed money, not anyone's judgment, set the price of French debt. That is ignition. Nothing was added to the pile; the spark got closer.

Paramount Skydance sold $52bn of loans and bonds in a single week to fund its $110bn takeover of Warner Bros. Discovery, clearing the way to complete the deal by October 6. The buyers' first experience of the debt was losing money on it. Bloomberg reports "instant losses and angry phone calls" at Bank of America and Citigroup as the new debt fell below the price it was sold at, and Paramount's shares fell almost 10% on Thursday as the deal was finished. A deal this size getting done is normal. Getting done only by paying buyers to be angry is not.

The mechanism is the bridge. When a takeover is agreed, banks promise to lend the money and plan to sell that promise on to investors before the deal closes. That works if the market holds steady between the signing and the sale. Here it did not: borrowing costs rose through the summer and the deal was delayed, so the same debt now costs Paramount an estimated $250m to $500m a year more in interest. The company plans $6bn of annual cost cuts to keep its borrowing under control. The banks could price the debt wherever investors would take it, or keep it themselves. They priced it to sell, and the buyers absorbed the loss the banks declined to.

Selling at any price is expensive in a week like this. Companies considered safe borrowers now pay 0.86 percentage points more than the government, 9% wider than a week ago, and LQD, the largest fund holding that kind of debt, is down 3% in 20 days. New bonds that fall below their sale price on day one tell the next borrower what to expect, and the next borrower is already in the room: Broadcom's banks began assembling $60bn this week to finance chips for Anthropic.

Paramount's finance chief, Dennis Cinelli, calls the merger "a strategic, long-term investment in the reshaping of the media industry". That may be right. The bond market's message was narrower: the biggest buyers of long-dated company debt have had their fill, and they will show it in the price. For the odds of a crash this adds a little tinder, since one more very large borrower has been locked into dearer debt, and it tells us something about ignition we did not know on Monday, which is how much appetite is left for the next giant deal.

The rate on a thirty-year mortgage posted its biggest weekly rise in four years as the sell-off in government bonds passed through to home loans, and Morgan Stanley's housing strategist Jim Egan says buying a home is now harder to afford than at any point this year. The monthly payment on a median-priced home is up $300, or 15%, since rates bottomed in February (CNBC). Mortgage rates move every week; they do not move like this.

The number that explains the housing market is the one Egan gives next. Someone who bought in 2016 and refinanced in 2020 or 2021 pays 7–8% of income on the mortgage. Someone buying the same house today pays 23–25%. More than 90% of American mortgages carry a fixed rate, so the rout does not touch the existing owner at all; it lands entirely on the newcomer. That is why the number of homes for sale is 20% below 2019 and why, on Egan's data, national price growth has started to pick up again over the past four months despite the rates: nobody who holds a 3% mortgage is selling.

There is a loop here that reaches back into the bond market. On Wednesday we described forced protective selling by holders of mortgage bonds as a driver of the government bond sell-off. The reason is that when rates rise, homeowners stop refinancing, so a mortgage bond expected to be paid off in five years now runs for ten. Its holders must sell government bonds to protect themselves against the extra years of exposure to rising rates, which pushes the government's borrowing costs up, which pushes mortgage rates up, which locks in more homeowners and stretches the bonds out further.

The strain shows at the edges, not in the middle. Credit Acceptance is down 12% in 20 days, Ally 13%, Capital One 12%: these are the lenders to households that cannot refinance anything. The housing market is fine for most of the people in it, and a shut door for the rest. On the odds of a crash this is tinder. Nothing about the loop is new this week, but each turn of it adds a little more borrowed money at the bottom of the household ladder and a little more forced selling at the top of the bond market.

Tokyo Electric Power's grid subsidiary changed its usage rules on Thursday to fight what it calls "capacity squatting": data-center developers reserving electricity they then do not use. Contracts will be cancelled if the operator fails to pay its share of the infrastructure costs within three months of signing, or if it changes the agreed amount or timing of the power. From April 2027, operators who delay switching on can be charged for the grid additions sitting idle (Nikkei). A utility rewriting its rulebook to penalize customers for not buying enough is not something that happens often.

The mechanism behind the rule is the mechanism behind the case for all the money being spent building AI capacity. A data center cannot be built without a power agreement, so developers sign for power first and design the building later, often, Nikkei reports, before the details of the facility are settled. In the busiest areas a connection takes years to a decade. The utility builds substations and lines against the contract; when the project is scaled back or shelved, the capacity sits unused and the cost falls on the grid and on the customers queuing behind. The Japan Data Center Council supports the change. Some operators object that penalizing under-use "is likely to stymie investment", and one says it is "only natural businesswise" to make investment decisions once you know when power can be secured.

That objection concedes the point. The demand forecasts that justify the building boom are, in part, reservations rather than orders. A utility writing cancellation clauses is saying out loud that the chance of a reservation turning into real demand is less than certain. Hydro-Québec told data centers last week that the era of cheap surplus power was over; TEPCO is now asking them to pay for capacity whether or not they turn it on. Ninety percent of Japan's data centers sit around Tokyo and Kansai, which is where any wasted cost would land. For the odds of a crash this changes little on its own, and we include it because it is the kind of small, unglamorous document that tells you what insiders believe: the people who build the power lines have decided the AI demand numbers need a discount.

The gap between what Italy and Germany pay to borrow for two years almost doubled on Thursday to 0.55 percentage points, the biggest one-day jump at the close since 2020. France's equivalent gap widened by as much as 0.22 percentage points, the most since 2012. German bonds rallied hard. In an ordinary selloff every country's bonds fall together. The core being bought while everyone else sells is what the euro area called fragmentation in 2011, and Barclays' Rohan Khanna said as much: the price action is "reminiscent of periods when bond market fragmentation was a major concern, such as during the European sovereign debt crisis" (Bloomberg).

The trigger was a French budget that tries to narrow the deficit by holding down spending and raising taxes, and may topple the prime minister in the attempt. The odd part is what happened to expectations for interest rates at the same time. Traders stopped fully betting on three more quarter-point rate rises from the European Central Bank; on Tuesday they had bet on at least four. Lower rate expectations normally lift bonds everywhere. On Thursday they lifted German and Dutch bonds only. The market is treating higher borrowing costs as a growth problem for the weaker countries rather than a policy choice.

Across the Channel the interest rate Britain pays to borrow for thirty years hit its highest since 1998 (6.03%), and British bank shares fell 4.1% on the FTSE 350 banks index, with NatWest down 5.2%, after Sky News reported that bank chiefs had been summoned by the chancellor ahead of the October 28 budget (Reuters).

The relief this bought the United States is real but borrowed. The Journal describes money flowing into US government bonds as a place to hide while hedge funds, the investment firms that manage money for the wealthy and can bet in ways ordinary funds cannot, closed out crowded European bets (WSJ). The ten-year rate, which touched 5.34% during the day, is 5.19% this morning; the S&P 500 is 0.8% off its high and the VIX is 15.6, which is calm. "Markets have a faint smell of a crisis in the making," said Columbia Threadneedle's Ed Al-Hussainy. Fidelity's Mike Riddell: "France has been slowly but steadily breaking. But today feels like the first day that broader financial markets have noticed."

One session cannot tell a reshuffle from a new regime. The 2012 episode ended when the European Central Bank promised to stand behind the weaker countries. Nothing in Thursday's reporting suggests it has yet been asked. Our read is that this is an ignition story, not a fragility one, and it moved sideways: the spark got closer in Europe by exactly as much as it moved away from the United States.

The interest rate to borrow the current ten-year US government note, by lending cash overnight against it, traded as low as 2.70% on Thursday and closed at 3.75%, according to Curvature Securities. The rate for borrowing any old Treasury traded at 3.86% and closed at 3.92% (Bloomberg). Normally the two sit close together. That gap is the sound of a crowd all wanting the same bond.

To bet that a bond will fall you have to sell one you do not own, which means borrowing it first. You borrow it by lending cash overnight to whoever owns it, with the bond put up as security. When many traders want to borrow one particular note, its owners can demand a cheaper cash loan in return: the rate falls, and the note is said to trade "special". A rate 1.2 percentage points below the general one means the people betting against the ten-year were paying, in effect, more than a percentage point a year for the privilege.

What makes this odd is that the note is not scarce. Roughly $92bn of it exists, the Federal Reserve, America's central bank, holds about $10.6bn of it, and the Treasury Department will sell $39bn more of it at an auction on October 7. "There's a deep short base in the market and no surprise there are a lot of shorts in the 10-year," said Curvature's Scott Skyrm, who expects the note to stay volatile for two weeks. In other words, a great many traders have bet against this bond, and he is not surprised.

Yesterday we described the forced selling of government bonds by holders of mortgage bonds as a loop that runs one way. A deep crowd of bets against the bond is the same machine wired in reverse. Thursday's rush into safety took the ten-year rate from 5.34% back to 5.19% this morning, exactly the move that hurts a crowded bet against it and forces those traders to buy, which is one reason a US bond rally built on French distress can run further than its cause deserves. A note trading special is not a cash crisis; the general rate is still in line with SOFR, the benchmark overnight rate, at 3.87%. But a market that is simultaneously being sold by forced hedgers and squeezed by forced buyers is not one where interest rates are telling you much about the economy. For the odds of a crash this is tinder rather than a spark: nothing has broken, but the world's benchmark rate is being set by who is trapped where.

The extra interest companies worldwide pay to borrow compared with governments widened by about 0.05 percentage points this week, the biggest move since March, leaving it at its widest in six months on a Bloomberg index. The equivalent gap for Asia's safest corporate borrowers opened Friday another 0.02 to 0.04 of a percentage point wider (Bloomberg). A twentieth of a point is not a rout. The reason it matters is the trade it undoes.

For the past year the fashionable view was that a company like Microsoft was a safer borrower than most governments, and company borrowing costs stayed close to government ones even as government bonds slumped. That worked while the supply of new company debt was manageable. It is not now: Paramount Skydance's $41bn of bonds sold to pay for a takeover fell as soon as they began trading, SoftBank ran a huge sale of debt too shaky to be considered safe to fund its AI push, and record new borrowing is colliding with investors who have started worrying about what companies owe. "It's that choppiness that keeps people away from the market," said T. Rowe Price's Sheldon Chan. Last week the Bank of England put AI-related borrowing at $450bn for the year to September; that supply has to be absorbed by someone, at some price.

The same selloff reached the market that is supposed to be asleep. American state and city bonds, measured by the Bloomberg Municipal Bond Index, lost about 4.4% in September, their worst month since Lehman collapsed in September 2008, with the interest rates on them at their highest since at least 2011 (Bloomberg). Thursday brought a second day of relief, the ten-year rate on those bonds falling 0.07 percentage points to about 4% and the thirty-year to 5.2%, as cash came back into the funds that hold them.

Our own numbers say the same thing more slowly: the extra interest paid by the shakiest companies stood at 3.12 percentage points on Wednesday against 2.7 in mid-September, wider every day this week. That is still less than half the 8 percentage points a crash trigger needs, and the move is orderly. But the one corner of the market that had been pricing no stress at all has begun to price some. That nudges the odds up on the fragility side: the buyers the AI borrowers need have just become choosier.

Broadcom's Wall Street bankers are about to invite lenders into a $42bn senior slice, the part that gets paid first and is backed by security, and Blackstone is leading an $18bn junior slice, the part that takes losses first, putting in $9bn from its own funds and selling the rest to others. The $60bn package, which has been taking shape for weeks, exists to let Anthropic and other companies buy Broadcom chips and data-center equipment (Bloomberg).

This is vendor financing: the seller's bankers arrange the debt that pays the seller. It is an old trick, telecoms equipment makers did it in 1999, and the question is always the same. Who eats the loss if the equipment is worth less than the loan? The reporting does not say who the borrower is, whether an Anthropic company or a vehicle that owns the chips, what the senior slice is secured on, or whether Broadcom stands behind any of it. Those three facts decide whether this is a loan to an AI lab, a loan against silicon, or a debt Broadcom may one day have to pay itself.

The structure does tell you one thing. The senior piece goes through banks to be sold on; the junior piece, the part that absorbs the first losses, sits with Blackstone's funds and whoever they sell it to. That is the migration this site exists to track: the riskiest slice of a $60bn AI loan lands outside the banking system, in vehicles whose investors will not see a price put on it until something forces one.

The same day, Bloomberg reported Anthropic aims to start marketing its stock-market listing the week of November 9 and trade before Thanksgiving, with prospective investors valuing it at $1.8tn to $2tn (Bloomberg). The prospectus that leaked earlier this week showed a $42bn net loss against $518bn of commitments to buy computing power. The lenders behind the $60bn are, in effect, betting on the listing: the cushion of shareholders' money beneath their chips is a valuation that does not yet exist. This adds to the pile. It is fragility, not a spark, and it is exactly the kind we count.

Amazon has been sounding out investors on moving about $8bn of Nvidia Grace Blackwell chips into a separate company set up to hold them off Amazon's books, which would borrow from outside investors and lease the chips back to Amazon. The chips are already running in more than a dozen US data centers across five states. Amazon would offer up to 10% of the vehicle's shares to others and keep none itself (FT).

This is selling something and renting it back, and the something loses value every year. The FT is explicit about the purpose: a balance sheet, the list of what Amazon owns and owes, that looks lighter and so protects its credit grade, alongside the promises about leftover value that tech companies have been giving lenders instead of borrowing directly. Investors expect the vehicle to be graded safe enough for the most cautious buyers, borrowing off Amazon's AA grade, which opens it to insurers and pension funds.

That grade is the tell. A box full of chips Amazon does not own cannot borrow cheaply on its own merits. It can only do so if Amazon's lease payments are, in substance, an Amazon debt. So the risk has not left Amazon; it has been renamed. The piece does not say how the lease is accounted for, how long it runs, or who bears the leftover value when Grace Blackwell is superseded by Nvidia's Vera Rubin, which the FT notes is imminent. If the vehicle's investors bear it, they are taking the risk that the technology goes out of date in exchange for the low interest a safe borrower pays. If Amazon bears it through the lease, the balance sheet is not lighter at all.

Eight billion dollars is small for a company that has committed up to $83bn to Anthropic alone. What makes it notable is who is doing it. The small AI cloud companies borrow against chips because they must. Amazon borrows against chips because it would rather not show the borrowing, and the people being invited to hold the debt are the insurers and pension funds this site has watched absorb loans from investment funds, insurer loans and data-center debt all year. The shares were down 0.4% on Wednesday; the market has not yet decided this is a statement about anything. For the odds, this is risk moving rather than risk growing, so it does not raise the number. It does make the next measurement harder.

The average thirty-year fixed mortgage rate reached 7.28% on October 1, up a quarter of a point in a week, the largest weekly rise since October 2022 and the highest level since late 2023, Freddie Mac said (FT). In late February it had dipped below 6% for the first time since 2022, days before the strikes on Iran. The round trip is the government bond selloff arriving at the kitchen table: US government bonds had their worst month in four years in September, and mortgage rates are set off them. KB Home, the builder, reported third-quarter revenue down 20% and homes delivered down 19%, and said conditions had weakened since June.

The same afternoon the housing regulator moved in the opposite direction. The FHFA, which oversees the two government-backed mortgage giants, plans to tell Fannie Mae and Freddie Mac to accept credit reports from two bureaus rather than the current three, an announcement that could come as soon as October 12 and take effect within one to three months, having already decided to treat VantageScore 4.0 as equivalent to Classic FICO, the standard credit score (Bloomberg). TransUnion fell 4.7% and Equifax 5.9% after hours. The stated aim is lower closing costs; the Biden administration shelved the same idea over concerns about judging risk accurately, and eight Senate Banking Republicans have asked Pulte, the regulator's director, to abandon it.

The mechanism is simple. Fannie and Freddie guarantee the loans and sell the bonds; the credit report is how they decide what to charge. Less data at the moment a loan is approved does not change how many borrowers fail to pay, only how well the guarantor sees them coming. Doing that at 7.28%, with Credit Acceptance, a lender to the riskiest car buyers, down 3.7% on Wednesday and 16% in a month, is a bet that the affordability problem is a fee problem. On the evidence of the market, it is a rates problem. This is fragility, slowly added: more loans made with less known about them, and the taxpayer standing behind the lot.

Some of Australia's largest retirement savings funds (superannuation funds) have been unable to take their money out of Partners Group's $12bn Global Value Fund for more than four months, the Australian Financial Review reported overnight. The Swiss manager, which runs more than US$185bn, paid out 51% of the withdrawals requested in May and none at all in June, July or August, according to investors who spoke to the paper. The reason given is worry about how the fund's unlisted holdings are performing and what they are worth. Funds like this promise investors their money back on a schedule. Four months of silence is not a delay, it is a door being held shut, and the industry word for it is a gate.

On Wednesday Metrics Credit Partners, one of Australia's biggest firms lending fund money to companies rather than bank money, froze withdrawals from two more funds and paused new lending for 90 days, a day after admitting three of its funds listed on the Australian stock exchange were worth a combined A$160m less than it had said. Metrics had already shut its $6bn fund for professional investors on Monday, after its auditor refused to sign off on the prices it put on its loans. The lending pause is the new fact, and it is the one that travels. A lender that stops making loans for a quarter has pulled away from borrowers who were counting on it to let them borrow again when their current debt comes due.

The mechanism in both cases is the same. A fund with no end date promises to let investors out at set intervals, against holdings that cannot be sold quickly. That works for as long as nobody doubts the price the fund says its holdings are worth. Once they do, everybody with a right to leave wants out at today's stated value before it is cut, and the manager's only way to avoid proving the price wrong by selling into a falling market is to stop paying out. The retirement funds, meanwhile, carry the holding in their members' balances at whatever price Partners Group gives them.

This is our running argument about risk migrating to places nobody checks, played out in one afternoon: two sets of untested prices, one at a firm that buys companies with borrowed money and one at a fund that lends to them, being tested by the people trying to leave. It does not add to the pile. It tells us how much was already there, which is a fragility story and not an ignition one. The publicly traded firms in the same business had already made up their minds. Apollo, KKR, Ares and Blue Owl fell a further 2% or so on Tuesday; Blue Owl is 26% off its high.