France's government paid more than 5% to borrow for ten years on Thursday morning, the most since 2002 (the ten-year rate reached 5.0601%), and the extra France pays compared with Germany reached its widest since 2011 (1.64 percentage points, per LSEG data cited by the Wall Street Journal). Germany's borrowing cost went the other way, down 0.017 points to 3.407%. So did America's: the interest rate the US government pays to borrow for ten years fell 0.059 points to 5.088% during European trading, and the rate for thirty years hit a four-week low of 5.376%. By the time we built this edition the US ten-year rate was 5.04%.

Read those three moves together. France, Germany and the United States borrow in the same world, and this morning investors sold one of them to buy the other two. "Today's tentative risk-off, with U.S. Treasurys supported and struggling equities, argues for continued pressure on eurozone government bond spreads," Commerzbank's Hauke Siemssen wrote. In other words, when investors sell anything risky and buy American government bonds instead, the weaker European governments are the ones left paying more. The Journal attributes the rush into US bonds to falling stock prices and "concerns about the sustainability of debt-funded investments in artificial intelligence infrastructure", and notes that worries about growth now outweigh worries about inflation.

That is the mechanism. For most of October the story was a rate shock: US government bonds sold off, and everything priced from them repriced. Today those bonds are the refuge, and the pressure has moved to the one large borrower that cannot print the currency it owes. The gap over Germany is what investors demand for lending to France rather than Germany in the same euro; at 1.64 percentage points they are demanding more than at any time since the euro crisis, and the Journal names budget worries and social unrest as the reasons.

What it does not yet show is anything breaking. The Journal reports no failed auction, and no sign that French banks or insurers, the natural owners of French government bonds, are being forced to sell. But a major government paying 5% raises the cost of every French mortgage and company loan priced from it, and the rush into US bonds happened because investors fear the AI boom, not because they stopped fearing American debt. Yesterday we recorded the French gap above 1.5 percentage points for the first time since 2012. Today it is 1.64. That is a step, not a repeat, and it is ignition: nothing new was stacked on the pile, but a spark in a core European bond market got closer.

Meta's quarterly filing on Thursday puts a number on what it has promised to pay: about $825bn of future spending commitments, much of it for computing power and AI infrastructure, according to Bloomberg. In July the figure was almost $700bn. That is roughly $125bn of new obligation signed in a single quarter, and adding that much every three months is not a routine thing even for a company of Meta's size.

What it is made of matters more than the total. $355.7bn is leases not yet on Meta's balance sheet, the list of what it owns and owes: data centers and energy infrastructure whose leases start between now and 2036. Another $454.5bn is contracts Meta cannot cancel, many for computing rented from other companies, energy for data centers and Reality Labs hardware. Meta is also obliged to buy up to $14.7bn of cloud computing over five years, an amount that may fall if the provider sells to other customers.

A lease that is not yet on the balance sheet is a building someone else is paying to put up on the strength of Meta's signature. The developer borrows, the investors who lend to the developer look through to the tenant, and the debt sits with them rather than with Meta. Last week the bonds funding data centers rented to Meta, Nvidia and Oracle were trading well below the price they were sold at; this filing says the pipeline of such promises is $355.7bn deep and runs for a decade. That is the migration this site keeps describing: the obligation is Meta's, the borrowed money is somebody else's, and nobody has to admit a loss until a lease payment is late.

On Wednesday Meta raised its spending plan for the year and announced a new class of shares that carry no vote, which could be used to pay for acquisitions and other expenses. Shares without votes let a company raise money by selling stock without the people in charge giving up control, and a company whose shares are 10% off their high does not invent a share class for fun. The filing does not say how much of the $825bn falls due in 2027, nor how much of the leased capacity has been financed by lenders who would lose if Meta's plans change. Those are the numbers that would tell us where the risk actually sits. For the meter this is fragility: $125bn more borrowed on Meta's word in a quarter, stacked where nobody has to price it.

Nebius, which builds data centers and rents out Nvidia chips, said on Thursday it has secured a $3.5bn loan from Citi to pay for construction and chip purchases, Reuters reports. The loan has to be repaid in 364 days and starts at 2.75 percentage points over the benchmark rate banks use for short-term lending (roughly 6.6%, with that benchmark, term SOFR, at 3.88% today). The company already owes about $10bn on bonds that can be swapped for shares, after selling $5.75bn of them in August, and its shares are down 24% in a month and 31% from their high.

Look at the repayment date. A loan due in a year is a bridge, not a source of funding: it buys chips that lose value over several years, and it will be repaid, on the company's own account, from "further fundraising to support growth in 2027 and beyond", customer prepayments and the cash the business throws off. Chief operating officer Ophir Nave said Nebius could also borrow against more than $40bn of long-term contracts with customers graded safe to lend to, including Microsoft and Meta. That is the same thing Meta's filing describes from the other side today: the contract that is a cost for the tenant is what the landlord pledges to its own lender.

The reason a one-year bank loan is doing this job is that the stock market has stopped funding these companies. Bloomberg reports that Nscale, the Nvidia-backed British company in the same business, is grappling with delays, equipment failures and power constraints at some sites, and with investors balking at the $30bn value it wants to list at; Jensen Huang said on Thursday he told the company to "work around the clock". Firmus Grid abandoned a $30bn stock listing in Australia and Elice had a difficult first day of trading in Korea. Huang said demand was high and these companies had excellent customers.

He may be right about demand. The question the Nebius loan raises is who carries the risk between now and the next round of borrowing. Until this week the answer was investors in those swap-for-shares bonds and buyers of new listings. For the next year it is Citi, a regulated bank, the one place this site's thesis says risk had moved out of. A bank lending where the stock market will not is either a vote of confidence or a bridge to nowhere, and the repayment date is when we find out which. For the meter this changes little in total: the borrowing is not new, it has changed address, and the new address is one regulators can see.

SoftBank shares fell 6.9% on Thursday, taking the fall from their June 2 record to 50%, Bloomberg reports. Four months ago the group briefly overtook Toyota as Japan's most valuable company; at ¥24.7tn ($157bn) it is now smaller than Hitachi and Recruit. On Monday the cost of insuring its debt against the company failing to pay over five years reached a four-year high ($4.32 a year for every $100 insured, or 4.32 percentage points).

SoftBank matters here because it is the biggest listed bet on the AI boom made with borrowed money: a holding company whose loans are secured on what it owns, and what it owns is stakes in AI companies. Kazuyuki Muramatsu of Nagomi Capital put the loop plainly: if the mood on AI sours, the Vision Fund has to admit its holdings are worth less, which raises what lenders charge SoftBank the next time it borrows. Those higher borrowing costs are what the shares are now pricing in. This is what makes holding companies built on borrowed money dangerous in a downturn: the thing pledged to the lender and the borrower are the same thing, so a fall in the asset tightens the credit at exactly the moment the asset needs patient money.

The other side of the bet is now crowded too. Goldman's desk that lends to hedge funds says those funds have bet on falls in about five times as much stock as they have bought this month. The borrowed money behind those bets against stocks is larger than in 99 of every 100 months over the past five years (the 99th percentile), while the net exposure of funds that bet both ways is near a three-year low. Unwinding has been expensive: funds focused on technology, media and telecoms are down 7.4% in October. Goldman's note floats "short squeeze risk alert" with a question mark, meaning a scramble by those bettors to buy back stock if it bounces, and Bloomberg observes that the bets against were added during a sell-off with identifiable real-world causes.

So two things are true at once. The biggest borrowed-money bet on the trade has lost half its value, and the bets against the trade are as heavy as they have been in five years. Positioning like that does not tell you direction. It tells you the next move, whichever way, will be forced. That is ignition rather than fragility: no new debt was added, but the chance of a sudden, involuntary move got closer.

The property empire of Mahdi Al Tajir, the billionaire owner of Highland Spring, has been hit by loan defaults and by creditors moving to seize assets, the Financial Times reports. The family is looking to sell chunks of a portfolio that runs from luxury developments in London to estates in Perthshire; last month the 94-year-old put a 620-acre golf resort near Gleneagles up for sale, a course finished in 2009 that never opened to the public, with a hotel and housing that were never completed. National Bank of Kuwait lent £12.75mn against it, according to the bank's 2024 accounts.

The names of the lenders are the point. The FT says the family's property companies have borrowed heavily in recent years from Gulf banks and from US institutional investors including Apollo Global Management. Apollo is not a bank. It is a fund manager lending alongside an insurance company's own money, and the price its loans are carried at is set by the lender, not by a market. A default on a prime London development or a Scottish estate is therefore the kind of event this site's thesis cares about: a loss in a place that does not report it the way a bank would, surfacing only because a creditor has moved to collect.

Trophy property is the kind of asset most dependent on cheap money. It produces little income, it is hard to sell quickly without cutting the price, and its value rests on the next buyer accepting an even smaller return. With the US government paying more than 5% to borrow for ten years, and France now with it, that buyer is scarce, which is why a resort that was never opened is the thing being sold. The FT's account does not give the size of the total debt or how much sits with Apollo rather than the Gulf banks. What it does show is an owner of prime assets, bought with borrowed money, selling under pressure from creditors with a private lender among them: the first visible test of such a valuation in that corner of the market this month. For the meter this does not make a crash more likely. It is the other kind of story, where we find out how bad something already was.

Argentina's dollar bonds have been the worst performers among developing countries' debt over the past two months, Bloomberg reports, with the interest rate the country would pay pushed to nearly 12% earlier this week and the extra it pays over the US government at its highest in more than a year. The country owes roughly $25bn in payments on debt in foreign currency next year, and the presidential election is a year away.

The move surprised the people who own the bonds. Vontobel, Carmignac, Schroders and ProMeritum had all said the debt was too cheap to ignore. "Most large investment funds were overweight and surprised by the magnitude of the selloff," Ramiro Blazquez of StoneX told Bloomberg. "Argentina is one of the first places investors cut exposure when they reduce risk because it's a higher-risk credit, and that magnifies the move." Overweight means they held more Argentina than their benchmarks called for; they were betting on it.

Two mechanisms are at work. The first is arithmetic: a government paying 12% cannot sensibly borrow again to pay off $25bn in the bond market at that price, so the money has to come from somewhere else, reserves, official lenders or a deal to pay creditors less, and the market is pricing the uncertainty about which. The second is who owns it. Bloomberg reports that foreign money managers hold substantial positions while Argentine investors have been cutting theirs, leaving fewer buyers when the appetite for risk drops. A large position concentrated in foreign hands is the shape that falls fastest, because the holders all want to reduce at once and the local buyers who would normally absorb them have gone.

This is how an AI sell-off in New York and Seoul becomes a funding problem in Buenos Aires. Nobody has reassessed Milei's handling of the budget this week; the bonds fell because they are where global funds cut first. For the meter this is ignition, and a small step: a borrower with $25bn due next year has lost its market price for borrowing again, and that is worth remembering when we look at the next borrower that depends on a crowd of foreign holders staying put.

Insuring $100 of Oracle's debt against the company failing to pay cost as much as $3.43 a year on Tuesday, more than it ever has, and the same measure for Meta rose as much as 0.063 percentage points to an all-time high of 1.148 percentage points, according to ICE data cited by Bloomberg. Companies worldwide have already borrowed more than $450bn this year to fund AI. Meta and SpaceX report earnings this week, and once the big tech names come out of the period in which they cannot sell new debt, bondholders expect them to come straight back for more. Lenders are already charging for borrowing that has not arrived.

The insurance moves first for a reason. A bondholder who owns so much of a bond that selling it would push the price down buys protection instead, so the insurance is where fear shows up before the bonds themselves. Pimco's Lotfi Karoui put it plainly: "We have yet to find some kind of a landing zone for the level of debt issuance that the market needs to absorb." In other words, nobody yet knows how much new borrowing the market can take before buyers stop turning up. Bonds that already exist are being repriced for bonds that do not. Oracle is the barometer because, of the giant cloud companies, it funds its building most heavily with debt, and it borrows against contracts it has signed rather than cash it has earned.

Notice that the stock market disagrees. Oracle shares rose 1% on Monday and the S&P is within a percent of its record, on the same day insuring Oracle's debt cost more than ever. Shareholders are pricing the demand for AI; lenders are pricing who pays for it.

One more layer. The Wall Street Journal reports, from lien filings in California, that Larry Ellison called off a $7.5bn sale of Oracle stock this summer and instead borrowed more from Citigroup, HSBC, JPMorgan, UBS and BNP Paribas, pledging Oracle, Tesla, SpaceX and Paramount shares as security, to fund his son's bid for Warner Bros. Oracle disclosed in September that 413 million of his shares, worth around $58bn, are pledged against his personal debts, up from 346 million a year earlier. That is not a new risk; it was standing already, and what changed is that we can see it. It is worth knowing that the founder's security and the company's credit are the same stock, now 13% below its high.

On the meter, the sum of this is a spark creeping closer, not more tinder. The debt was all there last week. What is new is that the market has started charging for the next round before it is sold, and if that price keeps climbing through this week's deals, it is the first place a wave would break.

Anthropic's draft papers for selling shares to the public, seen by the Wall Street Journal, say the company could spend up to $750bn in the coming years on computing capacity, including deals it is still negotiating. That is more than every company in the world has borrowed to fund AI this year put together ($450bn). Revenue in the first half of 2026 was about $16.3bn, up from $1.1bn a year earlier, and spending on computing and infrastructure was nearly $9bn, up from $2.3bn. The loss for the half was more than $80bn, against $13bn a year earlier, a figure the filing says largely reflects an accounting adjustment tied to financing agreements.

Take the three numbers in order. The revenue growth is extraordinary, and we mean that literally: almost no company of any size has ever multiplied its sales like that in a year. Around a fifth of it came from just two customers, and the filing says a "significant portion" will come from a limited number of customers for the foreseeable future. That matters well beyond Anthropic, because this is the demand behind every data-center bond we have written about this month. The lease payments that back those structures are, several steps removed, these two customers' invoices.

The $750bn is the other side, the supply. It is not a debt yet: it is a ceiling that includes deals not yet signed, and the filing lists uncertain access to computing among its risks. What has to stay true for a number that size to look prudent rather than reckless is that revenue keeps compounding at this pace.

The $80bn loss is the one we cannot explain. The filing attributes it to financing agreements, which suggests a paper charge on instruments issued to investors rather than cash leaving the building, but that is our inference and we may be wrong. Until the final filing is public, the honest statement is that Anthropic's accounts can swing by tens of billions on how its own financing is valued, which matters for a company aiming at a $2tn valuation.

On the meter this is tinder rather than a spark, and it is mostly tinder we are measuring rather than tinder being added. The spending is prospective. What the filing gives us is a fact we did not have: the demand at the top of the whole AI financing chain is, for now, two customers wide. The pile did not grow today, but we now know how narrow its base is.

NextDC, the data-center builder listed on the Australian stock exchange, lost as much as 45% of its value in Sydney on Wednesday after cutting its earnings forecast by more than a third, halting all new development and land purchases, and saying it is considering whether its contracts allow it to claim force majeure, the clause that excuses a company from delivering because of events outside its control, according to Bloomberg. Companies cut forecasts all the time. A builder stopping every new project and asking whether it can be excused from its own contracts is not that. The board is also weighing selling development sites. "We are not going to continue any work on any sites that don't have an existing customer commitment," said chief executive Craig Scroggie.

The cause is not a shortage of customers. It is power: community opposition delayed electricity to its M3 site in Melbourne, and the company could not deliver capacity it had already sold. Commonwealth Bank estimates Australia's buildout could reach A$150bn by 2030, on a grid retiring coal plants faster than it adds storage. A builder claiming force majeure against its own customers is telling the market that the physical limit has beaten the financial one.

The same day brought two more signs that the stock market's door is narrowing for these companies. DayOne, the Singapore data-center operator backed by the investment firms Coatue and Hillhouse, is weighing delaying its US stock market listing or cutting it to $2.5-3bn, roughly half the $5bn it had planned at a $20bn valuation (Bloomberg). And Nvidia is in talks to invest hundreds of millions of dollars in Firmus, the Australian company that rents out AI computing and abandoned what would have been one of the country's biggest stock listings this month, The Information reported via Bloomberg.

Put them together. Public investors declined to fund Firmus at the price offered; the company whose chips Firmus buys is now proposing to fund it instead. That is the supplier replacing the market as the source of its customer's money, and it has a name, vendor financing. On the meter it is tinder moved rather than removed. A listed share is priced every day and a loss on it shows up that afternoon; a stake sitting on Nvidia's balance sheet, the list of what it owns and owes, is priced when Nvidia says it is. The risk did not shrink. It went somewhere harder to see, and that makes the pile bigger, not smaller.

France closed Tuesday paying more extra interest than Germany to borrow for ten years than on any day since January 2012 (the gap closed above 1.5 percentage points), Bloomberg reports. We have noted the gap passing that line during the trading day in the last two editions; a close above it is the threshold crossed, because the closing price is where the market settles rather than where it flinches. January 2012 was the depth of the euro crisis, and France is a far bigger borrower than the country at the center of that one.

Sonal Desai, who runs bond investing at Franklin Templeton, said: "I don't think anyone is really ready for a eurozone debt crisis version 2.0, and this one could be significantly worse if it were to come to pass." Her arithmetic: France is a €3tn economy carrying about €3.6tn of debt. "France has more than 10 times the debt that Greece had. A simple bailout is not on the cards." Franklin's funds lean toward Italian and French bonds that come due soon rather than in decades, which is what you hold when you want to be out before anything breaks.

The mechanism that matters is the one that nearly broke the euro in 2011. Banks hold their own government's bonds, so when those bonds fall in price the banks' cushion of capital shrinks, and weaker banks need rescuing by the same government that is getting weaker. Desai herself lists the differences: France is still graded a safe borrower by the firms that rate them, its banks hold more capital than in 2010, and the European Central Bank has tools it did not have then. Those are real. But the premium investors demand to hold France over Germany is where it was when Greece was restructuring, and it got there without a single rating firm downgrading France in the window.

This is a spark getting closer, not new tinder; France's debt was all there last week. What we are watching is not the size of the gap but whether French banks' own borrowing costs start to follow it. That is the loop closing, and it is the thing that would move ignition rather than fragility.

America's Car-Mart, a used-car dealer that lends to buyers with poor credit, is working with its lenders on a sale that would run either through bankruptcy court (Chapter 11, the process that lets a company keep trading while its debts are sorted out) or a shareholder vote, Bloomberg reports. The company failed in June to pay what it owed on a $300m loan from Silver Point Capital, a loan it had taken to pay off a credit line secured on its car loans, and has been living on week-long reprieves from its lenders since; the latest runs to October 22. Living a week at a time on a lender's patience is not how a company survives. It is how one is sold.

The interesting part is what survives. The company's bundles of car loans, sold in slices to bond investors, are expected to be left untouched, and are quoted at around 98 cents on the dollar. That is close to whole, and the price implies there is more car loan inside the bundles than there is bond, a cushion built in at the start. So the structure does what it was designed to do. The loans were carved out of the company years ago, the bondholders keep them, and the corporate shell that made them fails around them. The loss lands on shareholders and on Silver Point, which in a court sale typically ends up as the owner. A fund that stepped in as lender of last resort in June is the likely proprietor in November.

Whether this is one weak lender or the start of something is not yet visible in prices: the shares of Credit Acceptance and Ally, two other car lenders, are down 4% and 3% over 20 days, nothing like Car-Mart's path.

The backdrop is the other end of household borrowing. Pending US home sales fell 7.3% in September to the lowest level since records began in 2001, lower than in 2008, as the interest rate on a thirty-year mortgage climbed above 7% to its highest since November 2023, the Financial Times reports. Housing starts fell to 1.22mn. A government paying five percent to borrow is no longer a story about bond traders; it is the number on a family's mortgage quote.

On the meter, Car-Mart is a small piece of tinder being resolved in an orderly way, which takes a little off the pile. The housing numbers are the opposite kind of fact and the larger one: evidence that the cost of long-term borrowing has reached households, which is where a slow problem turns into a fast one.

Three companies built on borrowed money moved toward being owned by their lenders in the same 24 hours. None of them is a crash. Each is how debt actually leaves the system when it goes well: in a court or across a negotiating table, one company at a time.

Optimum, Patrick Drahi's US cable business, is closing in on a restructuring that would hand control to its lenders and end the litigation between them, with the gap between what Drahi demanded for his stake and what creditors would give him having narrowed; a deal could close before the end of November, Bloomberg reports. This is the fight in which Optimum accused its lenders of breaking competition law for banding together, and in June warned them they risked triggering more than $4bn of tax liabilities. The lenders held together anyway, and that is the point: a borrower's best weapon is splitting its creditors, and it did not work.

Getty Images filed for bankruptcy protection in Delaware with the terms already agreed with its creditors, listing $1-10bn of assets and the same range of liabilities, after skipping an interest payment due September 1 on its unsecured notes; up to $533m of new money from the Getty family and Koch Equity Development takes it private, with suppliers paid in full (Bloomberg). The business was undermined by customers using AI-generated images instead of stock photos. Getty is an AI loser rather than an AI borrower.

Telesat has about US$1.71bn coming due on December 6 and has said its cash flow cannot cover what it owes. Lenders including Silver Point, GoldenTree and Sound Point tabled a second proposal on October 14 that would repay them in full through a $1.7bn sale of new shares to existing holders at $31.50 and a $1.5bn loan; Telesat has acknowledged the proposals but not negotiated, a source told the Globe and Mail. The shares fell 9.8%.

Silver Point appears in two of today's workouts and in Car-Mart's, which tells you how few hands these outcomes now sit in. The price of capital for the strong is rising too: Bank of America sold $2bn of notes that never come due at 7.2%, against 6.25% for the same instrument in July 2025, and did so days after raising $8bn (Bloomberg). On the meter, the three handovers take tinder off the pile, a little at a time, and that is worth saying plainly. The bank number is the part that does not. When one of the best-rated borrowers in the country pays a point more than it did a year ago, the cost of everything beneath it is rising with it.

Elice, a Seoul company that rents out computing power for AI, fell as much as 38% in its first session on Korea's second-tier stock exchange on Tuesday and closed down 32%, the worst opening day for a sizable Korean stock-market debut in Bloomberg's records. The first day is supposed to be the easy part, and this sale was priced as if it would be. The company raised 201bn won ($151m) at 90,500 won a share, the top of its range, with the slice set aside for ordinary savers subscribed 400 times over. It closed at 61,500 won (Bloomberg).

On Monday we noted the trillions of won of savers' money chasing this sale in the week the AI trade broke. This is what that money bought. When a sale is 400 times oversubscribed, almost everyone who applied got a sliver and most of the cash went back, so the loss falls on the few shares actually handed out, and the real signal is about the buyer who never turned up. Demand at that scale was demand for a first-day jump, not for a share of a data-center operator founded in 2015 by three KAIST doctoral students as a coding-education platform. When the jump did not come, there was no second buyer.

That matters beyond Korea because selling shares to the public was the exit the whole trade was built on. The plan was that money raised from public shareholders would carry the debt that pays for the building. Bloomberg lists the rest of the queue: Firmus scrapped its listing, AirTrunk is delaying, Nscale faces pushback on the value it wants, and CoreWeave has more than halved since May with its bonds trading as if it might not pay. Each door that closes pushes the next company further into the hands of the private lenders and the equipment suppliers extending credit, who were already there. On the meter this is tinder, not a spark: a buyer the boom was counting on has left.

The one consolation is small. Goldman's desk that lends to hedge funds says they cut technology shares at the fastest pace in its records in the five sessions to October 15, and the borrowed money behind their bets is at a one-year low (Bloomberg). The sellers using borrowed money may have mostly sold. Korea's savers, it turns out, were on the other side of them.

Insurance against AI companies failing to pay their debts has become a market of its own. More than $9bn of these contracts tied to Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia changed hands in an average week in September, up from under $2bn a week in July, according to Morgan Stanley Research cited by the FT. The price of that insurance hit record highs this month, not only for CoreWeave, a renter of computing power built on borrowed money, but for Microsoft. Record prices to insure the debt of a company like Microsoft are not a verdict on Microsoft. They are a verdict on everything it is tied to.

The mechanism is simple and new. Companies have borrowed an estimated $466bn so far this year for AI, against $101bn in the same period of 2025. Fund managers who cannot sell their Nvidia or Microsoft shares without missing the gains if the boom runs on are instead paying for insurance on those companies' bonds. The contract pays out if the company fails to pay what it owes, and its price is a running vote on whether it will. Morgan Stanley's Jim Caron calls it a "cheap hedge" in a market where "the risk is very concentrated". In other words, a cheap second bet to limit the loss on the first, in a market where everyone's first bet is on the same few names.

Two things sit underneath. First, SoftBank, which has borrowed more than anyone else to back the boom. Insuring $100 of its debt against the company failing to pay over five years cost $4.32 a year on Tuesday, the most since November 2022, and more than 30 cents above where it stood when it crossed $4 only a week earlier (the contract moved more than 0.3 percentage points, to 4.32). Bloomberg Intelligence sees at least $6bn of cash leaving the group in the year to March 2027, with its debt as a share of what it owns heading toward a 25% ceiling it set for itself, as it weighs buying Allied Universal; the shares are about 50% off their June high (Bloomberg). The market is now pricing SoftBank as what it is.

Second, why the insurance is needed at all. S&P, one of the three firms that grade how likely a borrower is to pay you back, has a head of digital infrastructure, Mark Habib, who said in London that it is "safe to assume that there is more (debt) out there that we are not seeing": guarantees of data-center construction and of smaller AI firms that sit outside what the giant technology companies disclose and outside the deals S&P grades. The six largest will spend more than $7tn to 2030 and will burn more cash than they bring in until at least next year (Reuters). S&P has no estimate of how much extra debt there is. That is our thesis, stated by a grading firm: the debts are real and nobody can see them. Insurance is what investors buy when they cannot see. On the meter that is tinder being measured, not added; the risk did not grow this week, the count got better.

Nscale, which builds data centers for AI companies, is close to a second deal with Anthropic worth roughly $45bn, covering a second 460MW building at its Monarch campus in West Virginia, on terms resembling the $45bn agreement the two signed in August for the first. The stated purpose, per the FT, is to shore up confidence ahead of a planned $35bn share sale in New York, after Microsoft walked away from its British data center at the weekend. OpenAI had already dropped out in April. A company does not usually need a customer contract to rescue its listing; this one does.

Set the numbers side by side. Nscale lost $1bn on revenue of $141m in the first half of 2026. Last year ByteDance supplied 73% of its $33m of revenue. Against that, it is about to carry some $90bn of commitments from one customer, and the point of the second commitment is to support the valuation that will raise the shareholders' money that builds the capacity the customer has contracted for. The customer's promise props up the share sale; the share sale funds the promise. If the listing slips, and Bloomberg reported pushback on the valuation this week, the capacity still has to be paid for by someone, and at what lenders now charge these companies that someone is not the public bond market.

The physical side is no further along. State water permits and the air permit for the gas-fired plant that will power the campus are still pending, though a letter of intent to approve the air permit has reportedly been issued. The campus will run off its own gas plant rather than the grid, which is both the reason it can promise a date and the reason the permit matters.

The deal is not finalized, checks on the second building are under way, and the FT says it could fall apart. If it closes, the right reading is not that demand for computing power is strong. It is that one well-funded buyer is willing to sign for capacity that does not exist, from a company that cannot yet pay for it, because the alternative is a smaller queue of suppliers. That adds to the pile rather than lighting it: more promises resting on a share sale that may not happen.

A $175m bond sale backed by the mortgage on 475 Fifth Avenue, a 24-story tower near Bryant Park that houses Penske Media, publisher of Rolling Stone, has been postponed. Even the safest slice, a $74.12m piece graded AAA, could not find enough buyers at the price offered (about 1.8 percentage points more interest than US government bonds), and it was supposed to price on October 6. Mesirow Financial is the manager; RFR and Pacific Coast Capital Partners sponsor the deal (Bloomberg). The buyers who exist to take the safest piece of a Manhattan office loan have decided they do not want it at that price.

The mechanism matters more than the building. The point of a commercial mortgage bond is to move a loan off the lender's balance sheet, the list of what it owns and owes, into the hands of insurers and funds who buy the slices. If the AAA buyer, the least demanding buyer in the whole structure, will not take 1.8 percentage points over government bonds for a single Manhattan office loan, the loan stays with whoever was holding it while the bond was built. That lender's books now carry an asset valued at the price it was made, which the market has just declined to confirm. Nobody has to admit it is worth less. Nobody can sell it either.

Bloomberg says this is the latest in a growing number of deals backed by home and commercial loans put on hold; last week a unit of Figure Technology Solutions shelved a $346m bond backed by home mortgages for lack of demand. The deals "may eventually be completed". That is true, and it is also what was said about the loans that got stuck in the last cycle.

What is driving it is the same thing that slammed the door on AI companies' junk-rated borrowing last week: the interest rate the US government pays to borrow for ten years sat at 5.11% on Tuesday, the thirty-year rate above 5.4%, and the extra interest riskier borrowers pay has widened behind them (the latest reading we have for the shakiest companies, October 16, is 3.31 percentage points, up 23% in 20 days). Real estate is not an AI story. It is the same window, seen from a different building. When the price of money rises, everything financed with it gets harder to sell, and that is fragility, not a spark.

Three governments faced the people who buy their debt on Tuesday and got three different answers. France is being charged more to borrow than Germany by the widest margin in weeks, Britain has found it must borrow more than it planned, and Japan drew the strongest demand for its 20-year debt on record. At today's rates lenders are no longer taking whatever is offered. They are choosing.

France. The extra interest France pays to borrow for ten years compared with Germany went back above 1.5 percentage points, to 1.53, on Tuesday, the widest since October 2 and close to this month's peak of 1.59, after a weekend of attacks on public buildings and shops over prices, followed by curfews (Bloomberg). France's own borrowing rate barely moved, at about 4.94%; the gap widened because investors moved into German debt and pushed Germany's rate down. France owes more than it produces in a year, its deficit is among the euro area's largest, an election is six months away, and the government cannot spend its way out of the protests. JPMorgan's Karen Ward put it plainly: bond markets "have much more choice over who they allocate to". In other words, lenders can pick, and they are not picking France.

Britain. The banks that buy British government debt directly now expect it to sell £249.6bn of bonds this fiscal year, £3.4bn more than the April plan, because the rate Britain pays to borrow for ten years has risen to 5.4%, from 4.23% before the Iran war. That has cut the Burnham government's room to maneuver under its own spending rules by a third, to £15.2bn, even with higher taxes on banks (Bloomberg). The budget on October 28 will set the borrowing plan. Borrowing is still below last year's £304bn, which is the thin good news.

Japan. The sale of 20-year government debt drew 5.84 yen of bids for every yen on offer, the strongest on record, against 4.01 at the previous sale and a twelve-month average of 3.73. The interest rate on 20-year debt fell 0.06 percentage points to 3.71% and the thirty-year rate 0.05 points, after both had reached levels last seen in the mid-1990s or never seen at all (Bloomberg). State Street's Masahiko Loo: "JGBs are investable again." JGBs are Japanese government bonds.

The common mechanism is the one Ward described. At these rates buyers are no longer taking what they are given; they pick. Japan, with a sales-tax cut still unfunded, got a record bid because domestic savers decided they are paid enough to wait. France, with riots and an election, did not. That is not a market breaking. It is a market discriminating, which is the opposite of what a market flooded with sellers does. For the meter it cuts both ways: a buyer came back in Japan, and the pressure moved to France rather than going away.

The People's Bank of China, the country's central bank, will set up an emergency lending facility for investment funds and other non-bank institutions facing large-scale withdrawals, Governor Pan Gongsheng said at the Financial Street Forum on Monday. Central banks have always stood behind banks in a panic; standing behind funds is new, and announcing it before any panic has started is newer still. The stated purpose is to break the loop in which withdrawals force sales, sales deepen the stress, and stress triggers more withdrawals.

It will be built with the banking and securities regulators, switched on only under specified conditions, require top-quality assets pledged as security, and be open to the wealth-management arms of banks, mutual fund managers, their parent banks and brokerages. The central bank is also opening its Standing Lending Facility, which has lent for one to three months against good security since 2013, to every deposit-taking institution in China, including locally incorporated foreign banks (Caixin).

This is a central bank building a lender of last resort for funds, not banks, before the run arrives. Wealth-management products and mutual funds hold bonds, promise savers their money back any day they ask, and have no deposit insurance and no emergency window at the central bank. In a bank that mismatch is backstopped; in a fund it is not. Pan is extending the backstop to where the risk has moved. Demanding good security is the price charged so nobody treats the facility as free money.

For what the alternative looks like, see Turkey. Its market regulator chief, Mahmut Sutcu, resigned on Monday citing health, five months into the job, as the justice minister describes a "Ponzi-like scheme" among high-profile investment funds. Tera and Pusula, which had produced unusually high returns, failed to meet withdrawal requests in September; the regulator has since ordered 131 funds with nearly half a million investors wound up, dozens of executives are jailed pending trial, the deputy finance minister has put in his retirement papers, and the Borsa Istanbul 100, Turkey's main stock index, fell as much as 2.8% on the day (Bloomberg).

Same mechanism, no facility: withdrawals met by selling thinly traded shares into their own decline. The lesson China is drawing is the correct one, and on its own it makes a crash there a little less likely, since a known problem now has something standing behind it. The question for everywhere else is who backstops the funds that now hold the risk.