The Federal Reserve, America's central bank, which sets interest rates, is expected to raise them on Wednesday for the first time in three years, and on Sunday the president who appointed its chair demanded the opposite in public. "We should be paying the lowest interest rate in the world, regardless of their formulas," Donald Trump told reporters in Ireland. Traders put the odds of a rise at 86%, with a second one expected before the end of the year (Bloomberg). The chair is Kevin Warsh, Trump's own pick, and a president leaning on his own appointee three days before a decision is not routine.

The set-up is unusual in three ways at once. The Fed has held rates steady for five straight meetings, and in July three of its officials (Hammack, Kashkari and Logan) voted against the majority in favor of a quarter-point rise, so the committee was split before Friday's inflation figures for August arrived (prices up 0.4% on the headline measure, 0.3% on the core measure that strips out the most volatile items). The Treasury Department, which borrows the money the government spends, spent last week buying back its own ten-to-twenty-year bonds at triple its usual size to hold down the rates on long-dated debt. And John Williams of the New York Fed says the central bank will stop shrinking its balance sheet, the list of what it owns and owes, in December and is "getting closer" to buying bonds again for what he calls technical reasons: the cash banks keep at the Fed has fallen to its lowest in 39 weeks ($2.895tn), while the Treasury's own cash pile there has risen to $967.9bn.

So: short-term rates are about to go up, the government's long-dated debt is being bought to push its rate down, and the president is asking for the opposite of both. The thing to watch is the extra interest lenders demand for tying money up for a long time when they cannot tell what policy will be (the term premium). A rate rise delivered by a chair under public attack from the man who appointed him can push long-term rates up rather than down, if investors decide the rise is a one-off made under duress while inflation is not, with Brent crude at $107 after the strike on the Saudi pipeline, tariffs, and a war in its seventh month. That is our inference; Warsh has not said it. The interest rate the government pays to borrow for thirty years closed Friday near a 19-year high (5.354%), before any of this.

Which of the three the market believes is the question Wednesday answers. The prices in front of us are Friday's; nothing in them has reacted to Sunday. This is an ignition story, not a fragility one: nothing has been added to the pile, but the moment that could set it off is three days away, and we cannot yet say which way it goes.

Marsh, the world's biggest insurance broker, says AI data centers are about to become the next big users of captive insurance, in which a company sets up its own in-house insurer instead of buying cover from someone else. That matters because these buildings are mostly built with other people's money, and a lender to a building normally relies on an outside insurer standing behind it. Companies have run captives for decades; what is new is that the buildings involved sit behind roughly $500bn of AI-related debt sold this year.

"What you're going to see is an explosive growth in the use of captives to take on the portfolio risks for data centers," Michael Serricchio, who runs Marsh's captive business in the US and Canada, told Bloomberg. "Inadvertently, some of the risk for build-outs, construction, surety, property and liability will end up in their captive." He would not name companies.

The reason is capacity. The facilities are larger than anything traditional property insurers have taken on, and many sit on cheap land that comes with tornadoes, floods and drought. When the market will not take a risk at a price the owner likes, the owner keeps it and calls the retained risk an insurer. Oil companies pioneered captives fifty years ago for exactly this reason, so the structure itself is ordinary.

Why it matters here is the debt. These campuses are increasingly financed through a separate company set up to hold the building off the parent's books, in which the tech giant owns about 20% and investment funds or bond investors put up the other 80%. Meta's Hyperion campus, a $27bn package arranged by Morgan Stanley with funds run by Blue Owl holding 80%, is the template (Reuters Breakingviews). Lenders to a building normally require insurance on it. If that insurance is written by a captive owned by the sponsor, the lender's protection against the roof coming off is, in the end, the sponsor's willingness and ability to pay itself. That is fine while the sponsor is Meta. It is a different proposition for a smaller AI cloud company or a bitcoin miner turned landlord.

What the reporting does not say: how much money these captives hold, whether the lenders to those separate companies are accepting captive cover, or how much of the physical risk behind the roughly $500bn of AI-related debt sold this year sits in structures like this. This is fragility, not a spark, and it is the pattern we already count: a risk that used to be priced by an outsider moves inside the structure, where nobody outside puts a price on it. It does not move the number today because it is the kind of risk the number already holds.

Two sets of figures published in the past week describe the same second quarter for BDCs, the funds that borrow money, lend it to mid-sized private companies, and pass the interest to shareholders, and they appear to disagree. The dollar total of loans that have stopped paying interest fell; the typical fund's share of such loans jumped. Both are right, and the reason is that the trouble is in smaller funds lending to smaller companies, which a dollar total barely notices.

The data firm Octus counted $9.5bn of debt on which interest has stopped arriving (non-accrual) across 173 BDCs, down 5% from $10bn in the first quarter, or 1.9% of what the loans cost. KBRA, a firm that grades borrowers, reported that for the median fund among those with a fixed life (the "non-perpetual" ones) the figure jumped to 2.75% of cost from 1.81%, a rise of 0.94 percentage points in a single quarter, which is a lot for three months. A Reuters cut of 44 BDCs put the total at 3.4% of cost, up from 2.5% at year-end, with portfolios valued at 97.57% of cost (matterfact).

The reconciliation is arithmetic. A dollar total is dominated by the largest funds, which lend to the largest borrowers; a median gives every fund one vote. The dollars fell and the median rose because the deterioration is in smaller funds lending to smaller companies. Houlihan Lokey's data make the same point from the borrower's side: among companies earning $10–20m a year, 12% of loans are now valued below 90 cents on the dollar, up from around 1%. Across the sector, 11.8% of loans by size paid at least some interest by adding to the debt rather than handing over cash; interest switched to that arrangement after the loan was made, because the borrower could not pay ("amended PIK"), was 1.6% of interest dollars.

The stock market has already sorted the funds the same way. The nine BDCs whose shares trade above the stated value of what they hold carry average non-accruals of 0.9%; the ten trading more than 30% below it carry 2.6%. Ares Capital and Blackstone Secured Lending are within 2.5% of their highs; Blue Owl's OBDC and FS KKR are 6–7% off theirs. Roughly $772m of interest income across the sector is owed by borrowers who are no longer paying it.

None of this changes our fragility number. It is June data reaching us in September: the world did not get more dangerous this week, our picture of it got sharper. What it changes is where to look. The system-wide headline can fall while the worst cases get worse, and the worst cases are where the smaller managers and the funds sold to ordinary savers with limited withdrawal rights live.

More American car loans are seriously behind on payments than at any point since the record began, including the depths of the last crisis: about one in every eighteen dollars of car-loan balances was at least ninety days late in the second quarter (5.5%, per the New York Fed, above the 5.3% peak of 2009) (The Money Overview). The stress sits where it always sits, among young and low-income borrowers. The lenders who assumed it would arrive are holding up better than the lenders who did not.

Fitch, one of the three firms that grade how likely a borrower is to pay you back, says the share of bundled loans to the riskiest car buyers (loans packaged and sold to investors) at least sixty days late touched 6.90% in January, a 32-year high. Borrowers aged 18 to 29 are seriously behind at 4.83%; those aged 60 to 69 at 1.77%. Fitch calls it a "K-shaped economy", with affordability pressure "weighing disproportionately on lower-income, highly leveraged borrowers". In other words, the squeeze is falling on the people with the least money and the most debt.

Two things in the same reporting cut the other way. The rate on those bundled subprime loans has come down from 6.5% at the end of 2025 to 5.8% at the end of June, and the ninety-day figure itself edged down to 5.49% from 5.60% in the first quarter. A ninety-day delinquency is a backlog: it counts loans that went bad months ago and have not yet been written off. Whether the pipeline behind it is still filling, these numbers do not say.

The share prices are not panicking about the borrowers you would expect. Credit Acceptance, which lends to the riskiest car buyers of all, closed 0.7% off its high and is up 3.6% in twenty days. Capital One is 8% off its high, Ally 6%, Synchrony 6%, Affirm 9%. Our reading, and it is inference: Credit Acceptance buys loans from dealers at a discount that assumes heavy losses from the start, so record delinquency is closer to its plan than a surprise, whereas the lenders to safer borrowers were priced for a customer who was going to be fine. The market is saying the stress is real but sitting where it was expected to sit.

That is reassuring only up to a point. The lenders who assumed the worst are holding up; the lenders who did not are being repriced. That is the pattern of the early stage of every credit cycle. This does not move the odds: it is second-quarter measurement of risk we already carry, and the household beat has been the quietest of our eight because the stress is visible, priced, and so far confined to the borrowers who were always marginal.

The Dutch central bank moved 86 metric tons of gold out of the United States and Canada this month, calling the move a hedge against "extreme systemic risks". Central banks moving their gold across borders is rare enough that each case makes news, and this one came with a reason attached. France has already taken all of its gold out of New York, Germany has debated it, and Norway's sovereign wealth fund, the world's largest, says it will cut its holdings of US government bonds amid the turbulence in that market (Axios). None of it has yet shown up in prices, which is the most useful fact in the story.

Read against that, a study from the New York Fed published the same week is almost a rebuttal. The dollar's share of the reserves central banks hold fell from 64% in 2015 to 56% in 2025, but the researchers find the decline is driven by a handful of countries: China and Russia in 2015–19, then China, Russia, Mexico and Morocco, which together account for nearly all of the 2.3-point drop between 2019 and 2023. Roughly as many countries increased their dollar holdings as reduced them (Bloomberg).

Both can be true, because they measure different things. Moving bars from a vault in Manhattan to one in Amsterdam does not sell a single dollar; it takes the bars out of US jurisdiction. That is a statement about trust in the custodian, not about the currency. Norway's plan is different: it reduces demand for US government bonds at the margin, in a week when the Treasury Department could not find sellers for the full $6bn of bonds it wanted to buy back. And the New York Fed's finding covers reserves through 2025; it says nothing about what allies decided in September 2026.

The prices side with the New York Fed for now. The dollar index, which tracks the dollar against a basket of other currencies, is flat on the week at 99.1 and down half a percent on the month. Gold, at $4,382, fell 2.1% in five days while long-term government borrowing rates went to 19-year highs, which is the opposite of investors running for safety. The yen is up 3.75% in twenty days on the prospect of Japan's central bank raising rates, which is an interest-rate story, not a flight. If the world were losing faith in the dollar as the currency it keeps its savings in, gold would have risen as bonds fell. It fell with them. This does not change the odds; it is one more reason to believe the slow part of the risk is still slow.

Traders in futures markets now put the odds of the Federal Reserve, America's central bank, raising interest rates next week above 85%, according to the FT, up from the 70% we quoted last run. Friday's data showed inflation stuck at 3.4%, with oil above $100 a barrel and federal debt past $40tn, and Governor Christopher Waller had already said a rise 'may be appropriate' if August's improvement proved fleeting. In the same week, the president of the Fed's New York branch, John Williams, said the Fed is 'getting closer' to buying bonds again. Raising rates is how a central bank tightens and buying bonds is normally how it loosens, so the two arriving in the same season is the thing worth understanding.

Williams describes the purchases as technical, with 'no implications for monetary policy', and says the Fed will stop letting its bond holdings shrink in December (sweep). The plumbing explains why both can be true at once. Banks keep cash on deposit at the Fed, called reserves, and that pile fell to $2.895tn in the week to September 2, the lowest in 39 weeks. It fell because the Treasury Department, which borrows the money the government spends, has built its own account at the Fed up to $967.9bn, a four-month high, and every dollar the Treasury holds there is a dollar not in the banking system.

There used to be a cushion. The Fed's overnight facility for spare cash, where lenders park money with the Fed against bonds put up as security, absorbed drains like this from 2022 to 2024. It is now down to $626mn, empty in all but name. With no cushion left, keeping overnight interest rates inside the Fed's target range means the Fed has to add to what it owns, whatever it decides about the policy rate.

So the plausible autumn picture is a Fed that raises rates, a Fed that buys short-term government IOUs, and a Treasury that buys back government IOUs that do not come due for twenty or thirty years at up to $14.5bn a week. None of that contradicts itself. The difficulty is what it looks like. Chair Kevin Warsh told the Jackson Hole conference two weeks ago that the Fed's balance sheet, the list of what it owns and owes, should be reserved for genuine crises and otherwise used sparingly, if at all. A restart labeled 'technical' three months later will be read by some as exactly what he said he would not do.

The market has priced half of this. Three-month government IOUs pay 3.91%, already above the 3.62% that banks pay to borrow cash overnight against bonds put up as security (SOFR), which is traders betting on the rise. What nobody has yet had to price is the Fed explaining purchases and a tightening in the same statement.

On the odds of a crash, our read is that this is the plumbing working: the Fed noticed the cushion was gone and is refilling it before anything seizes. It adds nothing to the pile of borrowed money and brings no spark closer, which is why both halves of the meter hold. Plenty of people we respect will read a central bank buying bonds while it raises rates as a central bank that has lost control of its own signal, and if the market takes that view next week, ignition is where it would show.

Oracle said on Saturday that Larry Ellison, its founder, will not sell any stock under a plan disclosed only the day before, which would have let him sell up to 50mn shares by the end of October, about $7.5bn at Friday's $150 close (FT). Plans like this are set up months ahead so a seller cannot be accused of trading on news; this one was set up in late June, and pulling it a day after disclosing it is not the usual sequence. The company declined to say why he changed his mind. A person close to him says he considers the shares undervalued.

Oracle's stock has fallen by more than half since it disclosed a $300bn deal with OpenAI, and Thursday's results showed revenue from its data centers up but profit margins still shrinking. It closed Friday down 1.8%, 7.5% off its recent high. Of the giant companies building the data centers behind AI, Oracle is the one paying for its buildout most heavily with borrowed money, which is why it is on our list of AI construction spenders.

The detail that belongs here is further down the story. Ellison owns 40% of the company and, as of September 2025, had put up 346mn shares as security for personal loans. A loan against shares has the same shape as any loan against a bet: the lender advances cash worth some fraction of the shares, and when the shares halve the borrower hands over more shares, hands over cash, or gets sold out. At Friday's price the block he pledged is worth roughly $52bn, by our arithmetic and on a count that is a year old. The shareholder filing does not say how much he borrowed against it or what fraction of its value the lenders advanced, and nothing published since says whether the count has changed. In December he also agreed to personally stand behind the $40bn of equity financing for his son's hostile bid for Warner Bros Discovery, now frozen until as late as June.

Two readings fit the facts. One is a founder signaling confidence by refusing to sell into weakness. The other is a holder who cannot easily sell 50mn shares into a market that would immediately ask about his loans. The reporting does not settle it. What it does show is concentration stacked on concentration: the AI builder most dependent on borrowed money has a controlling shareholder whose own borrowing rests on the same stock, in a form nobody has to put a public price on.

None of this is a spark. It is measurement: the pledge is a year old and the stock had halved before Saturday. It stays in the fragility pile because a lender demanding more cash from a 40% holder is the kind of forced sale that turns a falling stock into a collapsing one, and we cannot see from outside how far away that point is.

The shares of the firms that run private lending funds have fallen far harder than the funds they run, and that gap is the most precise statement about the business in front of us. Blue Owl is 16.6% below its recent high, Blackstone 13.9%, KKR 12.3%, Ares 11.9%, Apollo 9.8%. The funds those firms manage, which borrow money, lend it to mid-sized private companies and pass the interest to shareholders (the trade calls them BDCs), are down a fraction of that: Ares Capital 2.3% off its high, Blackstone Secured Lending 2.4%, Golub 4.6%, FS KKR 5.9%, Blue Owl's OBDC 6.7%. Shaky companies are paying 2.7 percentage points more than the government to borrow, which is not a level that says lenders are frightened.

The two prices measure different things. A fund owns loans, so its price reflects how many of those loans investors expect to go bad. A manager earns fees on the money it raises and runs, so its price reflects whether that pile keeps growing. The split says the market thinks the loans are roughly all right and the growth machine has stalled. This week's data supports the second half plainly: money raised by the private lending funds sold to retail savers is down 82% on a year ago, and Blackstone's BCRED has limited withdrawals to 5% a quarter for the second quarter running (sweep). Retail money was the engine. It has slowed sharply while money leaving runs at the limit.

The first half, that the loans are roughly all right, is more contested. KBRA, a firm that grades how likely borrowers are to pay you back, finds the typical fund now has 2.75% of its loans no longer paying interest, up from 1.81% a quarter earlier. Octus, counting in dollars across 173 funds, finds the total of such loans fell to $9.5bn from $10bn. Both can be true if trouble is spreading across more small funds while a few large ones improve, and the market has already sorted funds that way: nine that trade above the value they put on their own loans carry 0.9% of loans not paying on average, while ten trading more than 30% below that value average 2.6%. Houlihan Lokey finds 11.8% of loans took at least some of their interest as an addition to the debt rather than in cash in the second quarter. Among borrowers earning $10–20m a year before interest and tax, 12% of loans are carried by the funds at less than 90 cents on the dollar, from around 1% before. One first-quarter number we cannot ignore: 53 listed funds were, taken together, unprofitable, paying out more in dividends than they earned.

Every one of these figures reaches us through a single second-tier sweep, so hold the decimals loosely. The shape is the point. Either the fund prices fall to meet the managers, or the managers bounce, and one side is wrong. Which side settles the crash question. If the funds catch down, the market is saying the loans are worse than the funds admit, and that is fragility discovered rather than added. If the managers bounce, this was a story about fees and a stalled sales machine, which is a bad decade for the firms and not much for the pile.

Long-term government borrowing costs hit multi-decade highs in Britain, Japan, Germany and the United States in the same week, and the fact that they moved together is the story. Britain's thirty-year bond paid 5.948% on Friday and its twenty-year 5.895%, both the highest since 1998; its ten-year, at 5.378%, is the highest since 2007. Japan's ten-year touched 3% for the first time since 1996. German government bonds pay the most they have in fifteen years. Across the G7, two-year rates rose about 0.2 percentage points on the week and ten-year rates nearly 0.18, with France's ten-year up 0.24, and European government bonds had their worst week since March (sweep). In the US the ten-year closed at 4.975% and the thirty-year at 5.354%, and the Treasury Department's first expanded buyback found sellers for only $5.19bn of the $6bn it advertised. A buyback that cannot fill is unusual; the point of announcing one is that people queue.

The mechanism worth understanding is what happens when everything moves together. In a local selloff, a French budget scare or a wobble in British bonds, money moves into another government's bonds and the fall limits itself. When every country's long bonds sell at once there is nowhere for that money to go, so whoever is still buying demands more of the extra interest lenders want for tying money up for a long time, everywhere, and the usual ranking of safe places breaks down. That is why bonds issued by French banks and secured on their loans now pay less than the French government itself (our last headline), and why Bloomberg's credit desk is calling company debt 'a haven'. It also reframes the Treasury's buybacks: $6bn a week aimed at a domestic symptom of a global cause.

Japan is the leg to watch. The FT reports traders pricing a 75% chance the Bank of Japan raises rates to 1.25% next week and expecting another 0.5 percentage points by January; the yen has recovered from ¥164 to the dollar to about ¥154 on that and on the joint intervention of July and August. Our inference, not the FT's: a Japanese investor who can earn 3% at home has less reason to hold US government bonds at 5% and carry the risk of the currency moving against them, which is one reason the world's largest foreign holder of those bonds has been a seller (see the rumor mill). Norway's sovereign wealth fund has said it is reducing its holdings too (Axios).

What we do not have is any sourced read this week on whether the big banks that trade these bonds are running out of room to hold them, or on the technical gauges that would show it. Interest rates at multi-decade highs are a repricing, which is painful but orderly. Dysfunction is when a price cannot be found at all. The buyback that came up short is the only hint of the second so far, and it is why we watch this story for ignition rather than fragility: nothing new was borrowed this week, but the buyer that normally caps a move has stepped away everywhere at once.

The number of British first-time buyers borrowing 4.5 times their income or more jumped 66% to 45,800 in 2025, and loans above 5.5 times income rose tenfold to 4,628, according to data the FT obtained under freedom-of-information law from the FCA, which polices British financial firms. That happened because a brake was removed on purpose, and the borrowers who used it are the ones with the least room for error. Coventry Building Society this month started offering 6.5 times income to first-time buyers earning at least £30,000. Nationwide's six-times product saw a 57% rise in loans at five times income or more and a fivefold rise above 5.5 times.

The rule that changed is worth being precise about. In 2014 the Bank of England's committee for financial stability capped loans at high multiples of income at 15% of each lender's new mortgages. That was a limit on the flow, not a ban on any single loan: it stopped lenders competing on how many times your salary they would lend. In July 2025 it was relaxed for first-time buyers. Within a year the response is visible in the data, which is roughly what lifting a cap on the flow should produce. A broker quoted by the FT calls this 'positive'; another says buyers 'will do it if it means they can get on the property ladder'.

Two things follow. First, the highest multiples are being concentrated on the youngest borrowers with the smallest stake in their own homes, the group with the least room if income or rates move against them. Second, the timing. The British government's ten-year borrowing rate is at its highest since 2007, and British mortgage rates move with it. The data does not say what rate these loans were tested against, or how a borrower on 6.5 times income at today's fixed rate copes when the fixed period ends and they have to borrow again in a few years.

In absolute terms 4,628 loans is small. The direction is the point, and it cuts against our usual story. This is not risk migrating out of the regulated system into somewhere nobody prices it. It is a central bank deliberately putting household borrowing back onto the books of banks and building societies, where it can at least be measured. It is also a decision made before British borrowing costs went back to 1998 levels. On the meter this adds a little to fragility, slowly, and brings no spark closer; it is the kind of thing that decides how bad a downturn gets rather than whether one starts.

Investors are now lending to French banks, against pools of ordinary French home loans, at a lower interest rate than they charge the French government. These bonds, known as covered bonds, pay almost 0.3 percentage points less than ten-year French government bonds, the widest gap of its kind Bloomberg has on record. Four months ago the two paid roughly the same (Bloomberg). That is not supposed to happen: in every country's bond market the government is meant to be the cheapest borrower there is, and everyone else pays a little more.

Why that is the rule takes one sentence. The government taxes, and it prints money (or its central bank does), so every other borrower's cost is built as a step above its rate. A covered bond is a bank's promise to pay, with a ring-fenced pool of mortgages behind it if the bank fails. It is very safe. It is not supposed to be safer than the state that regulates the bank, guarantees its deposits and taxes the people paying the mortgages.

Bloomberg puts the move down to France's budget deficit, the rising cost of servicing its debt and the presidential election in April. Part of it is mechanical. Covered bonds come with regulatory privileges (they rank ahead of other creditors, they have assets pledged behind them, and payments are protected from interruption) that make them scarce and prized by banks, which must hold a stock of very safe assets, so demand can outrun supply for reasons that are not a pure verdict on the Republic. Nothing in the price separates the two.

But it lands in a week when the same question was put to three other governments. The US Treasury Department bought back its own long-dated bonds and its borrowing rate rose anyway while it was doing so; the rate Britain pays to borrow for thirty years touched 5.95%, its highest since 1998; Japan's ten-year rate hit 3% for the first time since 1996. The common thread is governments borrowing more just as energy-driven inflation pushes central banks toward raising rates. France is where the market has gone furthest in saying so, and Bloomberg calls it 'a harbinger of trends worldwide.'

For the odds of a crash, this is a measurement rather than new tinder. We already count France's finances as stressed; the market has now said the same thing in a form it never had before. What would make it benign: a normal gap reopening after the election. What would make it serious: the same inversion showing up in Italy or Britain.

Two life insurers owned by Mark Walter, Delaware Life and Clear Spring Life and Annuity, disclosed earlier this year that more than $20bn of loans on their balance sheets (the list of what they own and owe) should have been labeled as deals with companies connected to their own owner, and were not. The Justice Department and the SEC, the agency that polices American markets, are now investigating his businesses. Senator Elizabeth Warren has written to the National Association of Insurance Commissioners, the club of state officials who regulate insurers, asking how they police Wall Street firms' influence over the insurers they own and the private loans those insurers make (Bloomberg). The oddity is not the investigation. It is that the question came from a senator before it came from a supervisor.

The machinery here is the center of what this site watches. A life insurer sells annuities, policies that pay a saver a monthly income for the rest of their life, collects the money up front, and invests it because it will not have to pay out for years. If the insurer is owned by an investment manager, the tempting place to put that money is into loans the manager itself arranges: lending to companies by investment funds rather than banks, at rates the public market does not offer. Those loans are carried at what the insurer paid for them, not at what they could be sold for today; the regulator is a patchwork of state commissioners rather than the Federal Reserve, America's central bank; and the person holding the annuity cannot pull their money out. It is, structurally, where risk goes to avoid being measured.

Lending to a company tied to your own owner is not banned. The rules require you to say so, because that is the one loan where the lender's judgment is most compromised. $20bn undisclosed at two insurers is a failure to label, not a loss: nothing in the reporting says the loans have gone bad. Warren's question is the right one, and it is whether this is two firms or a pattern. The NAIC says it 'looks forward to sharing' how it oversees insurers' exposure to private lending.

The stock market has given its answer through one company. Apollo, the listed firm that most openly runs private lending through an insurance balance sheet, is down nearly a tenth from its high (9.8%) and 4.1% on the week, while the extra interest shaky companies pay over the government has barely moved. For a crash, nothing got more dangerous this week. The loans were signed long ago; what improved is our view of how little anyone was looking.

Kalshi, the platform where people bet on elections and other events, will file as soon as next week for approval to list the first perpetual futures on individual stocks in the United States, starting with contracts on Tesla, Apple and Nvidia. It wants two regulators, the CFTC, which polices derivatives, and the SEC, which polices markets, to oversee them jointly as a security futures product, and plans to add farm commodities and exchange-traded funds later (Bloomberg). Nobody in America has been allowed to offer this before, and the question of which regulator owns it has not been answered.

A perpetual is a bet made with borrowed money that never has to be closed. An ordinary futures contract settles on a fixed day; a perpetual never does, so a position can be held on borrowed money indefinitely with no expiry to force a reckoning. Bloomberg notes the contracts have become increasingly popular. That popularity was built on offshore crypto exchanges, where perpetuals became the standard way for small investors to bet with borrowed money, and Kalshi's chief executive Tarek Mansour says it is 'about time that these products come to the US, under a regulated umbrella.'

What matters for us is the route the money takes. Borrowing to buy stocks is governed by Federal Reserve rules and runs through brokers whose exposure regulators can see. A perpetual on Nvidia listed by a venue that began with election bets is the same borrowing routed through a derivatives exchange instead. The filing does not say how much borrowing it will allow, how it will demand extra cash when bets go wrong, or which firm would stand behind the contracts if one side cannot pay.

The context is a week in which JPMorgan cut off lending to the investment fund Situational Awareness and the fund rebuilt its borrowed position through options at a smaller broker (FT). That was one fund finding a new door after a bank shut one. Kalshi's proposal is the same thing for millions of ordinary investors.

This is a filing, not a product. It moves the odds of a crash by nothing today. What it does is show where borrowed money is trying to go, and that is the slow kind of risk, tinder rather than spark.

The Education Department is moving 7.5 million people with federal student loans off a repayment plan called SAVE, which no longer exists. Anyone who has not picked another plan within 90 days of being told will be dropped automatically into the standard plan: the whole balance split into fixed payments over ten years. Nearly half of the people on SAVE had incomes low enough to qualify for a $0 monthly payment. The first 90-day window closes on September 29, and so far about 18% have moved (Washington Post). That leaves several million people who have paid nothing for two years a few weeks from a bill set as if their loan were a ten-year mortgage.

Borrowers describe conflicting information, missing payment histories and a broken calculator for working out what each plan would cost; the department says the glitches are largely fixed. Three quarters have been notified; the rest will hear by year-end.

This arrives on top of a household sector that has just set its first record of this cycle. One in eighteen car loans is now three months behind, the worst in the New York Fed's records and above the peak of 2009 (5.5% in the second quarter), though we have that from a secondhand summary we have not been able to check against the original release. Among the riskiest borrowers, the share two months behind on a car loan peaked at 6.9% in January and has since eased to 5.8%.

The stock market has a view, and it is a specific one. Capital One is down 5.5% on the week, Synchrony 4.7%, Ally 3.5%, Affirm 3.9%. Credit Acceptance, which lends to the riskiest car buyers of all, sits 0.7% off its high. Read literally, that is investors saying the trouble is moving up the income scale, into people with decent credit, rather than down into the poorest borrowers, whose bad stretch is already priced in. A payment shock landing on people who were paying $0 fits that reading uncomfortably well. For a crash this is the slow kind, a little more strain on households, and the thing the announcement does not say is how many of the 7.5 million will simply not pay.

Companies building artificial intelligence have sold nearly $500bn of debt so far in 2026, through early August, by Goldman Sachs's count. That is about a fifth of all the bonds sold this year by better-rated American companies, up from 1% in 2024. The six largest cloud-computing companies alone have issued roughly $320bn of long-term debt and related financing. JPMorgan's Michael Cembalest converts that into the equivalent of about $303bn of ten-year bonds and notes it equals roughly 68% of the new long-term borrowing the US government itself has done this year (Reuters Breakingviews via web sweep). Half a dozen companies are borrowing on a scale that competes with Washington.

These are measurements, not new risk. The bonds were sold months ago and we have counted AI debt in fragility all summer. But the 68% figure names a mechanism we had not stated plainly. Long-dated company bonds and long-dated government IOUs compete for the same buyers: pension funds, insurers, foreign governments parking their reserves. When a handful of companies borrow two thirds as much long money as the US government, they are not just borrowing; they are bidding up the cost of long-term money for everyone, the Treasury included. The FT's account of this week's selloff lists 'a flood of supply from tech companies' alongside oil and record government borrowing as a driver of interest rates worldwide (FT). Amazon alone has raised around $100bn in the bond market this year.

The shape of the borrowing is shifting too. Goldman expects debt to fund 33% of AI building spend this year and 35% next. Reporting this week describes data centers financed through separate companies set up to hold them off the parent's books, in which the cloud company owns 20% and investment funds or bond investors supply 80%, with the lease ultimately resting on the parent. Whether those leases sit on the parent's balance sheet, the list of what it owns and owes, is the question the filings do not answer.

The circle closes at the top. The ten-year US government borrowing rate, at 4.975%, is the yardstick every AI profit forecast is measured against, and the borrowing that justifies those forecasts is one of the things pushing it up. Nothing here moves the odds today; it is tinder we already counted, now weighed more carefully.

Second-quarter figures for BDCs, funds that borrow money, lend it to mid-sized private companies and pass the interest to shareholders, now contradict each other. KBRA, a firm that grades borrowers, reports that at the typical fund the share of loans on which interest has stopped arriving jumped to 2.75% of what the loans cost from 1.81% in the first quarter. Octus, scanning 173 of these funds, finds the dollar amount of such loans fell, to $9.5bn from $10.0bn, or 1.9% of total cost versus 2.02% (web sweep). Both are secondhand summaries, so treat the exact figures with care. The lesson is in the disagreement itself, because in this market nobody is forced to say what a loan is worth, and the statistic you pick does the job a price would.

A bad-loan ratio can fall without a single borrower getting better, and there are three ways it happens. First, write-off: a loan marked down to 5 cents on the dollar, as Blue Owl's OBDC marked a company called Loparex, eventually drops out of the count entirely. Second, the denominator drifts: Ares Capital reports bad loans at 2.4% of what they cost but 1.4% of what it says they are now worth, and the gap exists precisely because bad loans shrink as a share of the portfolio once they are marked down. Third, amendment: a borrower who cannot pay cash interest is allowed to pay by adding to the debt instead, and the loan stays 'performing.' Houlihan Lokey's second-quarter data show 11.8% of these loans by size chose that route, with 'amended PIK', meaning the option was added after the loan was made, which is the closer sign of stress, at 1.6% of interest dollars.

The typical fund getting worse while the total gets better says the trouble is concentrated in smaller funds and smaller borrowers. Houlihan finds 12% of loans to companies earning $10–20m a year before interest, tax and depreciation are valued below 90 cents on the dollar, up from around 1%. Morningstar puts $772m of interest income at risk against $38bn of cash interest across the sector, which is small, so far.

We already carry these marks in fragility, so this moves the odds very little. What is new is the demonstration that the headline number a manager chooses to report can go either way from the same set of loans.

JPMorgan has stopped lending to Situational Awareness, the hedge fund run by Leopold Aschenbrenner, after the largest loss in that industry's history, the FT reports. A hedge fund is an investment firm that manages money for the wealthy and can bet in ways ordinary funds cannot, and the bank that lends it the money for those bets and holds them is called its prime broker. The fund has begun working instead with Clear Street, a specialist broker the FT describes as “far smaller” than JPMorgan. The bank walking away is the part that is supposed to happen. The fund finding a new lender within days is the part that matters. Morgan Stanley had talks with the fund since it launched and never opened an account. Goldman Sachs and Citigroup have had relationships with it.

Read that first as the system working. A prime broker is where a fund's borrowed money comes from and where it gets measured: the bank decides how much cash the fund must put up against each bet, sees the whole book and can shrink the loan whenever it likes. JPMorgan looked at a fund that went from more than $45bn to about $10bn in July and decided not to be the lender standing behind the next attempt.

Then read what happened next. Per CNBC, Situational Awareness spent late last week and early this week buying options, contracts that give the right to buy a share at a fixed price later, on AMD, Bloom Energy, CoreWeave, SK Hynix, SanDisk and the Roundhill Memory ETF, a fund that trades like a single share. CNBC could not establish whether that is new money or what was left after the fire sale to Citadel. Options are borrowed money with the borrowing built into the price of the contract, and the dealer who sells them protects itself by buying the underlying shares. So the exposure to those companies is back, routed through a smaller firm and the traders who sell such contracts rather than through a bank credit line.

Clear Street's terms, how much cash it demands the fund put up, are not public, and neither is how much of the remaining $10bn is at work. What can be said is that the best-capitalized lender in the chain stepped away and the position was rebuilt anyway. That is the migration this site exists to track, in miniature and with a name attached. It adds nothing new to the pile; it moves part of the pile to where it is harder to see, which is a fragility story rather than a spark. The names on the ticket are memory chips, and on those, see the next dispatch.

DeepSeek released V4.1 Flash on Thursday, a model it prices at a fraction of a cent per million tokens (the chunks of text a model reads and writes) and says beats Moonshot's Kimi K3. MiniMax and Z.AI fell more than 8% in Hong Kong, and Alibaba slid more than 2% (Business Times, citing Bloomberg). The share prices are not the reason to care. The reason is that DeepSeek says its model needs less of the one component the entire AI buildout is short of.

The damage in Hong Kong is the smaller story. The larger one is a design detail. DeepSeek says it has trimmed the model's KV cache, the working memory a model keeps while it is generating, which cuts the high-bandwidth memory and solid-state storage needed to run it. It also says the model, which has 552bn parameters (the internal settings that hold what it has learned), switches on only a tiny part of its capacity at a time.

High-bandwidth memory is the tightest bottleneck in the AI build. A chronic shortage of it has driven profits at Micron and SK Hynix, and it is one reason the money the biggest cloud companies spend building things is assumed to keep rising. Microsoft's spending hit $145bn in its last fiscal year, and Bloomberg reported plans to reach 38GW of data-center capacity by 2032, a figure Microsoft called inaccurate. A model that needs less of the scarcest input is a shock to demand aimed at the one place in the supply chain where the seller sets the price.

Two cautions, both ours rather than the reporting's. Cheaper AI has historically increased total usage, so less memory per token need not mean less memory in total. And DeepSeek's benchmark claims are DeepSeek's. But the fund in the previous dispatch bought options on SK Hynix, SanDisk and a memory fund in the days before this release. Nvidia is down 4.4% on the week. Anthropic and OpenAI are preparing to sell shares to the public into a price war in which “good enough and nearly free” is the competitor. The AI building spree has always required that computing power stay scarce. This is what a test of that assumption looks like, and the US memory companies have not yet had a trading day to price it. Nothing has broken, so this is not a spark. It is the first serious test of the assumption underneath a very large pile of debt.