Jobs beat 162,000. Fed leans to hike. Trump demands 1%.

Archived reading, published Sat, 05 Sep 2026 00:40:28 UTC (6 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility91

how much tinder is stacked up — moves slowly

Ignition37

how close a spark is — moves fast

Crash Lab watches the machinery under the current boom: the debt paying for AI data centres, private credit, the leverage that has moved out of banks into places nobody has to mark, the bond market and the dollar.

Every four hours it reads the day’s reporting from 28 sources and rewrites this page.

Status: Held at 64. The close on September 4 was a shade softer than the reading we took at 13Z: the VIX, the market's gauge of how much turbulence traders expect over the next month, rose to 14.53 from 14.14 (under 20 is calm; it passed 80 in March 2020), the S&P 500 finished 1.0% below its record from 0.6% earlier, and the Nasdaq sat 2.2% below its own. But the extra interest the shakiest companies pay compared with the government is unchanged at 2.65 percentage points, and the interest rate the government pays to borrow for ten years is flat at 4.78%. So ignition edges up by only one, to 37. The reason it moves at all is that a report showing 162,000 jobs added has pushed the Federal Reserve, America's central bank, toward raising rates on September 15, on the same day the president demanded 1% rates and threatened trade embargoes. That is a policy collision with a date on the calendar, not a seizure in the markets where trading desks borrow their cash. Fragility holds at 91: Nvidia's $2bn into Nscale is more of the vendor financing we already count; the warning from S&P, one of the three firms that grade how likely a borrower is to pay you back, that the big cloud companies' financings are 'less transparent' is a measurement rather than new stock; Norway's proposed $80bn cut to its US government bonds leaves its dollar exposure unchanged; and nothing was sold off in a hurry.

Reporting from 2 Sep to 4 Sep

Crash points

Eight places it could go, scored 0–10. Tap one for the explainer.

What’s moving the needle

What changed in the last few hours, and what each one says about the plumbing.

Fed, Treasury and policy4 Sephigh

Ten days to a Federal Reserve meeting nobody agrees on

An interest-rate decision argued over in public ten days before it happens is the nearest thing on the calendar to a policy mistake in progress, and the rates the government pays to borrow for twenty or thirty years are already the highest in years.

The August jobs report showed 162,000 jobs added, nearly three times the roughly 55,000 that forecasters had expected, and July was revised from a loss of 23,000 jobs to a gain of 21,000. A miss of that size matters because it pushes the Federal Reserve, America's central bank, toward raising interest rates at a meeting ten days away, at the very moment the president is demanding it cut them. Within hours he posted: "LOWER THE RATE OR I'LL STOP TRADING WITH COUNTRIES WITH WHICH WE HAVE A DEFICIT", named Switzerland, Mexico and the EU as the targets, and said the Fed should be "at 1% or half-a-percent" rather than 4% (Bloomberg). The collision now has a date. The Fed's chair, Kevin Warsh, told the central bankers' annual gathering at Jackson Hole on August 28 that "two percent is a firm and fixed target" and that unless inflation is clearly heading there "we still have work to do". Traders read that as a signal that rates could go up rather than down, and put the odds of a September rise above 50%. The meeting starts September 15. On the same afternoon as the president's post, his own economic adviser Kevin Hassett argued the other way, saying consumer prices over the past three months are rising at an annual pace of 1.6%, and that a favorable inflation reading would be a "strong argument" to hold rates where they are. The part that matters most is what the government pays to borrow for a long time. It pays 4.78% to borrow for ten years and 5.25% for thirty, and the Treasury Department's enlarged bond buybacks, the program Scott Bessent announced to fix "especially poor" trading conditions in thirty-year bonds, had not carried out a single operation as of August 31. Ordinarily a rate rise into a market worried about inflation and deficits lowers the long-term rate, because bondholders need less compensation for the risk that inflation eats their money. The president rejected that explicitly: "It doesn't reassure the bond market." So three outcomes. A rise defies the president and tests the Fed's independence in the open. A hold, with Hassett's inflation number as cover, leaves the market asking whether it was the data or the post. A cut is implausible on this data, and the market for long-term government bonds would say so. So far investors have priced the Fed doing what the data says: rates flat, the dollar up 0.16%, and the VIX, the market's gauge of expected turbulence, at 14.5, which is calm. Where it fits our argument: policy is the one area where risk

The AI capex bubble4 Sepmedium

Nvidia lends $2bn to a customer that says it has $103bn of contracts

Each new computing company Nvidia funds adds another loop to the circle in which the chipmaker finances the demand for its own chips, and the contract backlogs that justify it rest on customers who are themselves still raising money.

Nscale, a London company that rents out computing power for artificial intelligence, is in talks to raise as much as $3.5bn before it lists its shares on the stock market: up to $1.5bn of convertible notes, debt that can later be swapped for shares, led by Daniel Loeb's Third Point, and about $2bn of financing from Nvidia, according to Bloomberg. Nvidia makes the chips Nscale will buy with the money, so the supplier is funding its own customer's order. The listing could raise another $3bn on top. Nscale is telling investors its contracts total about $103bn, following a $45bn agreement to supply computing power to Anthropic, and that it could generate about $18.1bn of annual revenue. Take the pieces in order. Nvidia is the company whose chips Nscale will buy with the money; Nvidia is also putting up more than half of the money. That is vendor financing, a supplier lending its customer the money to buy its product, and it is the pattern we track from CoreWeave to Lambda to IREN. Nvidia's share price closed at a record on Thursday. The convertible notes are offered at the listing price minus a double-digit percentage discount, with the discount adjusted up to a $30bn valuation. That means the lenders who get in before the listing pay less than the public will, and their money counts as debt until they choose to turn it into shares. Then the $103bn. Contract value is not revenue and it is not cash. It is a promise by customers, chiefly Anthropic, to pay for capacity over years, contingent on that capacity being built, powered and delivered on time, and on the customer still being able to pay. Anthropic's ability to pay $45bn depends on its own fundraising, which depends in part on the same giant cloud companies and chipmakers. Bloomberg's sources do not say what the cancellation rights are or how much of the $103bn the customers must pay whether or not they use the capacity; we cannot tell. None of this is unusual by 2026 standards, and that is the point. It is a company valued at $30bn whose customer list depends on the same handful of balance sheets, the lists of what each company owns and owes, that are financing it. The smaller computing-rental companies rallied on the day (Nebius +7.5%, IREN +7.3%, CoreWeave +5.6%) while Microsoft fell 2.1% and Alphabet 1.1%. The market is paying more for the borrowers and less for the customers. For the odds of a crash this is fragility rather than a spark: one more loan in a circle that holds only as long as everyone in

The AI capex bubble4 Sephigh

The credit-rating firm that has to guess

When the firm whose grade decides who may own the debt behind $7tn of spending says the financings are getting harder to see into and the returns are undisclosed, the safe-to-hold label is doing more work than the analysis behind it.

S&P Global Ratings, one of the three firms that grade how likely a borrower is to pay you back, has told investors in its latest "temperature check" on the biggest cloud companies that it cannot see the returns on the $7tn of spending it is grading, as quoted by the FT. Rating firms do not usually write sentences like this one: "Every time we take a deep dive into this sector, we find that capex is rising faster than we anticipated, financings are becoming more complicated and less transparent, and that returns on investment will take years to realize." In other words, the companies are spending more than expected on building things, borrowing in ways that are harder to see into, and the payoff is years away. The numbers behind it: S&P assumes the six largest US cloud giants (hyperscalers, in the industry's word) will spend more than $7tn in the five years through 2030, and reckons half the growth generated by the US private sector over the past year was linked to AI-centric activity. And the piece of information the ratings most depend on is missing. "The big six hyperscalers don't quantify their returns on investment on AI," the report says, either because it is hard, or because "entities just choose not to share that data." Why this matters for the machine: a credit rating is the grade that decides who may hold a bond. Insurers, pension funds and cautious lenders can own these companies' debt (roughly $219bn of top-grade new borrowing this year by one asset manager's count, an increasing share of it in euros) because the agency has signed off. S&P says it is filling the gap with an assumption: that the largest companies in the world probably know what they are doing. It is drawing guardrails around commitments that behave like debt without being called debt (leases, agreements to buy electricity for years ahead, ventures held off the companies' own books such as Meta's $27bn Hyperion project with Blue Owl), measured as a multiple of yearly operating earnings (EBITDA), above which ratings would come under pressure. This is a fairly exact description of the thesis this site runs on. The obligations are real, they are increasingly structured to sit outside the corporate balance sheet, the list of what a company owns and owes, and the institution whose job is to measure them says it cannot see the return that justifies them. Nothing broke on Thursday. Microsoft and Alphabet fell about 2% and 1%; Meta and Oracle sit at their highs. The market's assumption an

Bond market dysfunction4 Sephigh

Norway's oil fund wants $80bn less in US government bonds

The most patient buyer in the world is saying it holds US government bonds because they are easy to sell, not because they pay, which is the same judgment the market for long-term government debt has been reaching since the summer selloff began.

Norges Bank Investment Management, which runs Norway's $2.3tn oil fund, has proposed cutting government bonds from 70% to 50% of the bond portfolio it measures itself against, in a letter to the finance ministry signed by Governor Ida Wolden Bache and chief executive Nicolai Tangen (FT). The FT estimates the change would remove about $106bn of government bonds, most of it US debt (roughly $80bn), with the money going mainly into bundles of American home loans guaranteed by government-backed agencies (agency mortgage-backed securities). It matters because this is the most patient large investor in the world deciding that US government bonds no longer pay enough for what they are. First, what this is not. It is not an exit from the dollar: the fund says its dollar exposure would be "essentially unchanged", and the mortgage bundles are guaranteed by Fannie Mae, Freddie Mac and Ginnie Mae, so the exposure to a US government default barely moves. Nor is it a retreat from bonds whose prices fall hard when interest rates rise; the mortgage bundles do that too. What it is: the world's largest sovereign fund concluding that US government bonds are worth holding for one reason only. The letter says 50% "will be sufficient to cover the liquidity needs, including in periods of turbulence in financial markets", and that the rest of the bond index should "provide exposure to more sources of risk premiums". In other words: we hold US government bonds because they can be sold quickly in a panic, not for the return, and we need fewer of them for that. The mortgage bundles pay a little more because the holder bears the risk that homeowners repay early. The plumbing point is who is on the other side. The buyer Bessent hoped the stablecoin issuers would become is shrinking (see below). The holder who now sets the price at the margin is hedge funds, investment firms that manage money for the wealthy and can bet in ways ordinary funds cannot, holding $4tn of US government bonds funded by $3tn of cash borrowed overnight against those bonds put up as security, according to the Fed paper Reuters cited this week. That is money that leaves the moment its lenders stop lending. Norway is a patient buyer using no borrowed money stepping back in the same season the borrowed-money buyers step forward. $80bn is a proposal, would be phased in over years, and is small against the whole market for US government debt. But the direction is the same as every other data point in this area this s

Crypto and TradFi contagion4 Sepmedium

Bessent's trillion-dollar buyer shrank by $3bn

The policy bet that crypto would deliver a new permanent buyer of US government debt depends on a source of demand that shrinks exactly when markets get nervous.

The two largest digital tokens meant to be worth exactly one dollar, stablecoins, shrank in the first six months of the year, Bloomberg reports: Tether's USDT contracted by nearly $3bn to about $184bn, on track for its first annual decline since the 2022 crash, and Circle's USDC dropped by a similar amount to roughly $72bn. This matters because Scott Bessent has said the stablecoin market could grow tenfold to $3tn by the end of the decade, making the coins' issuers, who hold short-term US government debt as reserves, trillion-dollar buyers of that debt. The base he is counting on is going the other way. The mechanism is simple and the direction is the problem. Stablecoins are where traders park cash between crypto trades. Demand for them rises and falls with crypto trading volume, which dried up after prices crashed late last year. Their holdings of government debt are therefore a buyer that appears when investors feel bold and disappears when they do not, the opposite of what a government bond market under pressure wants from the buyer who sets the price. A combined total around $256bn was never going to fund the deficit; what Bessent was selling was the growth rate, and the growth rate has gone negative. Two things cut against the gloom. A crypto industry recap says Circle created 5bn USDC in a single recent week, which does not square neatly with Bloomberg's half-year figure; we cannot reconcile the two. And 21 banks including Citi, Bank of America, Goldman, UBS, Deutsche and MUFG say they will launch a dollar stablecoin that complies with the GENIUS Act, the law governing such tokens, in the first half of 2027. If the banks do it, the demand becomes a substitute for a bank deposit rather than a trading balance: steadier, but it is money that would otherwise have been a deposit funding a loan. Meanwhile the listed companies that exist to hold crypto fell on a day bitcoin did not: BitMine -5.5%, SharpLink -3.3%, Strategy -1.5%, Coinbase -4.2%, against bitcoin -0.2%. The coins are flat; the wrappers around them, companies that borrow to hold coins, were sold. Small moves, but the premium those wrappers trade at is where the borrowed money in this area lives. For the crash odds this is a slow subtraction from the pile of buyers rather than a spark: fragility, up a little.

Private credit and BDCs4 Sepmedium

The managers fall, the funds hold

When the market sells the firms that earn fees but not the funds that hold the loans, it is betting on a fundraising drought rather than on loans going bad, a bet that only holds if the prices the funds put on their loans are right.

On Thursday the listed firms that manage private lending funds fell and the listed funds they manage did not. Blackstone fell 1.1%, is down 4.4% over five days and sits 8.8% below its high; KKR fell 1.8%, Ares 1.0%, Apollo 0.7% and Blue Owl 0.8%. Their lending funds, vehicles that borrow money, lend it to mid-sized private companies and pass the interest to shareholders, barely moved: Ares Capital rose 0.1% and is within 1% of its high, Blackstone Secured Lending rose 0.3% and is 1% off its high, and Golub, FS KKR and Blue Owl Capital Corp were all flat to up. The extra interest shaky companies pay compared with the government was unchanged at 2.65 percentage points. The split matters because it tells us what the market thinks is wrong: not the loans, but the business of raising money to make more of them. Two markets, one set of loans. The lending fund's price is a bet on the loans themselves, and the loans look worse: loans on which the lender has stopped counting the interest, because it stopped arriving, are 3.4% of cost across ten listed funds, up from 2.5% at the end of 2025; 81% of software loans have been marked down this year; there is a roughly $15bn industry-wide backlog of investors waiting to get their money out; and BCRED is limiting withdrawals to 5% for a second quarter against 10% requested (Bloomberg). The manager's price is a bet on the fee stream: raising the next fund, keeping the open-ended vehicles growing, earning the performance cut. A withdrawal limit does not lose the manager money on the loans; it stops new money coming in. So the market is saying: the loans will mostly be fine, the fundraising machine will not. That is a coherent view, and there is one data point for it. A secondary source reports that KKR Income Trust I, which capped withdrawals in the second quarter, saw requests fall to about 2.5% of the fund's stated value in the third quarter from more than 5%; we cannot independently confirm it, but if right, the panic is not spreading fund to fund and the withdrawal limits did the job they were built for. It is genuinely one of the calmer facts of the week. It is also the view that requires the funds' own valuations of their loans to be correct, and those valuations are the thing under dispute. If the loans going bad keep compounding at the pace of the last two quarters, the fund prices are wrong and the manager prices are early. Blackstone's slide also carries its property exposure; the share price does not let us separ

Signs of the times

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

Stock bets on crypto exchanges: from 0.5% to 23.5% in nine months

Contracts on stocks, indexes and commodities made up 23.48% of the trading in never-expiring futures on the major crypto exchanges in August, from 0.5% in November; monthly volume went from $23.5bn to $778bn, a 33-fold rise in nine months. The reason given by one fund manager: 'the volatility of some stocks is multiple times higher than Bitcoin.'

perpsleverage24-hour markets

4 SepBloomberg

$7tn of spending, no return disclosed

S&P Global Ratings, one of the three firms that grade how likely a borrower is to pay you back, says the six largest US cloud giants will spend more than $7tn through 2030 and that half of US private-sector growth last year was AI-linked, while noting that none of the six 'quantify their returns on investment on AI', so the agency grades them on the assumption that the largest companies in the world know what they are doing.

ratingshyperscalersdisclosure

4 SepFinancial Times

The Dutch moved 59 tonnes without moving them

Of the 86 tonnes of gold the Dutch central bank shifted out of North America, 59 were sold in New York and bought back in London, and 27 were physically moved to a vault in Zeist; New York's share of Dutch reserves fell from 30% to 18%. A central bank treating the vault of the Federal Reserve, America's central bank, as a question of how quickly its gold could be sold, to be traded around.

goldreservescrisis preparedness

4 SepAxios

A mortgage secured on bitcoin

Coinbase and the online lender Better have launched mortgages with bitcoin pledged as security after a waitlist that projected more than $260m of demand, according to a crypto industry recap. A house loan resting on an asset that moved 27% in twenty days.

crypto collateralhousingleverage

4 SepBitcoin.com (web sweep)

The rumour mill

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

Partly true

“The August jobs report showed 162,000 jobs added against a median forecast of about 55,000 — a roughly four-standard-deviation beat across 76 forecasters, with the highest estimate at 125,000.”

The 162,000 jobs figure and the roughly 55,000 forecast are right and matter for the Federal Reserve's meeting on September 15. The '76 forecasters', '125,000 top estimate' and 'four standard deviations' flourish are unsourced; the documented top of the Reuters range was 121,000.

Claimed by Meet Kevin

Partly true

“Bessent has doubled the rate of Treasury bond buybacks, yet long yields have kept rising anyway.”

The cap on each buyback was indeed at least doubled to $4bn per operation on August 19, and long-term rates did rise. But Bessent himself said on August 31 that not a single enlarged buyback had been executed yet. The rates rose before the program started, which is a different and arguably worse fact.

Claimed by Arcadia Economics (Chris Marcus)

Partly true

“Fed chair Kevin Warsh delivered his most hawkish comments yet at Jackson Hole, and some fellow policymakers are pushing to raise rates.”

Warsh's 'we still have work to do' was read by Reuters and CNBC as a signal that rates may rise, and pushed the odds of a rise above 50%. No reputable coverage documents other members of the Fed's rate-setting committee publicly pushing for a rise; that half is inference presented as fact.

Claimed by GoldSilver (Mike Maloney)

Earlier developments

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

Private credit and BDCs4 Sepmedium

Twelve times the loans that stopped paying, and no discount on the shares

When a fund's loans go bad faster than its share price falls, either the market knows something about how much of the money comes back, or it has not read the filings. We lean to the second.

Blackstone Secured Lending Fund, Blackstone's listed fund that lends to mid-sized private companies, has stopped counting the interest on 3.6% of its loans, because the interest has stopped arriving. In the first quarter of 2025 that figure was 0.3%, according to a Jefferies analysis summarized in this window. A lender's book does not move twelvefold in five quarters in an ordinary cycle. The fund's shares are trading near their record high anyway, and that gap is the story. It is not alone. The same table has Ares Capital going from 1.5% to 2.4%, Blue Owl's fund from 1.4% to 2.8% and Golub from 1.2% to 2.9%. Across ten listed funds of this kind, loans that have gone quiet reached about 3.4% of what was paid for them at the end of June, against 2.5% at the end of 2025. A stricter monitor puts it at 3.95%, or 5.95% if every loan to a borrower with one bad loan is counted as troubled. A loan the lender has stopped counting the interest on is a specific thing. It means the fund has decided it no longer expects to be paid. These funds exist to borrow money, lend it to mid-sized private companies, and pass nearly all the interest to shareholders as dividends, so every loan that goes quiet is a dividend cut waiting to happen, unless new lending, or interest paid by adding to the debt rather than handing over cash, fills the hole. The prices the funds say their holdings are worth tell the same story from the other side: 81% of software loans have been marked down this year, and roughly 4% of loans are valued below 80 cents on the dollar. Now the share prices. Blackstone's fund closed September 3 1.4% off its high, and up 2.2% over 20 days. Ares is 0.9% off its high. The stock market is saying these books are fine. The books are saying they are not. Two honest cautions. The fund-level figures reach us through secondhand summaries of a broker's report, not the funds' own filings, and the three aggregate numbers use different yardsticks (cost versus current valuation, loans versus borrowers), so they are not contradictory, just differently strict. There are ways to square prices with books: a loan that stops paying is not the same as a loan that loses money, lenders first in line often get most of their money back, and shareholders may be paying for the manager's skill at swapping bad loans for good ones. But set this beside BCRED, the unlisted version sold to ordinary savers, where investors who want out are being told to wait. The listed funds have an exit door th

Hidden leverage and shadow banking3 Sephigh

JPMorgan lent to a rival, then cut it off

The business of standing ready to buy and sell US government bonds has moved to firms that borrow their working money from the very banks they compete with, and those banks can pull it for any reason.

JPMorgan sharply cut the money it lends Jane Street for trading bonds last year, after the trading firm started standing ready to buy and sell US government bonds, the IOUs Washington sells to cover its deficits, in direct competition with the bank, the Financial Times reports. The cut was about 5% of everything Jane Street borrows across banks for bond trading, and it made no material difference to the firm's 2025 revenue. What matters is not the size but the reason. The firms that now keep the government bond market moving do it with money rented from their rivals, and one of those rivals has just taken some of it back for no reason other than that it did not like helping a competitor. The scale of the business is the first thing to hold. Jane Street traded more than $900bn of bonds last year. Trading firms of its kind took 10% of the industry's revenue from trading bonds, currencies and commodities in 2025, per Crisil Coalition Greenwich. That is not a fringe. It is a tenth of the market. The mechanism is the second. Firms like Jane Street trade their own capital, made bigger with borrowed money, and the borrowed money comes from the banks whose trading business they are taking. That is how non-bank firms became a large part of what keeps government bonds easy to buy and sell: they rent room on the incumbents' balance sheets, the list of what a bank owns and owes. The FT says JPMorgan's own traders were unhappy about helping to build a competitor, and that the bank has also trimmed services to Citadel Securities for the same reason. So the funding beneath the non-bank government bond market is given at the lender's discretion, and this time it was withdrawn for a commercial reason. In March 2020 similar lines were pulled for reasons of risk, and hedge funds that had borrowed heavily to profit from a small gap between two nearly identical prices, the basis trade, were forced to sell out of it. The current version of that trade is larger. A paper from the Federal Reserve, America's central bank, cited this week puts hedge funds' gross bets on government bonds at $4trn and their cash borrowed overnight against bonds put up as security at $3trn, 90% of it in the 50 biggest funds. The rule that would make those funds report how much borrowed money they use has been pushed to 2027 by the CFTC, which polices futures markets. None of this says stress is here. SOFR, the rate at which cash is borrowed overnight against government bonds, is 3.66%, and the gap betw

The AI capex bubble4 Sepmedium

Chips in a box, Microsoft's lease, and money at 6.7%

The risk that today's AI chips are worth little in five years is being packed into separate companies and sold to loan funds and to investment funds that lend to companies rather than banks, which is where the losses would land if the leases are not renewed.

On August 27 Lambda, which rents out computing power for AI, borrowed $926m against a pile of Nvidia chips, at an interest rate of roughly 6.7%, through a separate company set up to hold the chips off Lambda's own books. Separately it has raised about $1bn of debt, arranged by JPMorgan, to buy Nvidia hardware that Microsoft will lease. The lenders are not really betting on Lambda. They are betting that Microsoft keeps paying to use those chips until the loan comes due at the end of 2030, and on what four-year-old chips are worth if it does not. The terms are these. It is a loan of the kind that is sold on to investors rather than kept by banks (a senior secured term loan B), priced at SOFR plus 3 percentage points, sold at 99.5 cents on the dollar, and due December 31, 2030. SOFR, the rate at which cash is borrowed overnight against government bonds, is 3.66%, which is how you arrive at about 6.7%. Lambda's bank credit line, meanwhile, has gone from $275m in August 2025 to $1bn this May. This is the structure to understand. The separate company owns the chips. The lenders lend to the chips and to the stream of lease payments Microsoft has agreed to make on them, more than to Lambda itself. If Microsoft pays, the loan performs. If Microsoft's appetite for that capacity fades before 2030, the lenders own four-year-old Blackwell-generation hardware and have to find out what it is worth then, a bet on the leftover value that none of the reporting prices for us. A loan of this type typically ends up in loan funds and in bundles of company loans sliced up and sold to investors (CLOs), so the leftover-value risk on Nvidia silicon is being spread to the same buyers who hold loans to heavily indebted companies generally. It is not alone. IREN closed $2.4bn from Blue Owl and PIMCO, split as a $1.2bn loan and $1.2bn of secured notes, for Blackwell Ultra chips in British Columbia, with the equipment bought in stages. CoreWeave added a $2.6bn loan it can draw in stages on August 7, on top of $8.5bn and $3.1bn earlier. One tally puts $110bn of corporate debt into AI infrastructure in 30 days. That is a great deal of money into one thing in one month. The pattern fits the thesis we have been tracking exactly: the customer (Microsoft) signs a lease, the supplier (Nvidia) guarantees the chips' leftover value elsewhere in the chain, and the debt sits in a separate company financed by loan funds and by investment funds that lend to companies rather than banks. No bank carrie

Bond market dysfunction4 Sepmedium

A third of new long-term bonds are paying for data centers

The AI building spree is borrowing in the one corner of the bond market where lenders are demanding the most extra pay, and the cost arrives through the general level of interest rates rather than through any penalty on the companies, where it is easy to miss.

The five companies building America's AI data centers, Amazon, Microsoft, Alphabet, Meta and Oracle, borrowed roughly $200bn from bond investors in the first half of 2026, according to a note from the fund manager Neuberger Berman. Technology now accounts for a fifth of all new borrowing by American companies considered safe (20%) and more than 30% of the new debt that does not come due for decades. That is happening in the same weeks that lenders around the world are demanding more to tie money up for that long than they have in years, and in Britain and Japan more than they have since the 1990s. The two facts belong in one sentence, and almost nobody puts them there. The rest of the tally. Financing for the big cloud companies, data centers and chipmakers reached about $165bn before the midpoint of the year, $27bn more than in all of 2025. Oracle alone plugged its AI spending gap with $43bn of debt and $5bn of new shares. Now the rates. The interest rate the US government pays to borrow for ten years is 4.79%, its highest since January 2025; for thirty years it is 5.25%. Britain's thirty-year rate hit 5.89% on September 1, the highest since 1998, and Japan's ten-year went through 3% for the first time since 1996. The New York Times reports that demand at the government's debt auctions looked normal and that trading was orderly. That distinction matters: this is investors deciding they want more pay for lending for a long time, not the machinery of the market failing, and we think it is real and worth holding on to. The link between the two is that the AI buildout has become the biggest new corporate borrower of long-term money at exactly the moment the world wants more compensation to lend it. The extra interest safe companies pay compared with the government has not moved (0.81 percentage points on the last reading), so the cost is arriving through the general level of rates, not through any penalty on these particular borrowers. LQD, the big fund that holds safe companies' bonds, is down 0.8% in five days. Every extra quarter of a percentage point on the thirty-year rate is a quarter point more that data-center leases running into the 2040s have to earn to be worth what was paid for them, and it lands harder on Oracle than on Microsoft. Meanwhile the Treasury Department, which borrows the money the government spends, has doubled its purchases of old bonds that come due in ten to thirty years, to at least $4bn per operation from September 9, described a

Fed, Treasury and policy3 Sepmedium

Britain built its debt to last fourteen years, and lenders no longer want to wait

Britain is the clearest case of a government being charged for the shape of its debt rather than the size of it, and the US Treasury is now reaching for the tools Britain already uses.

Britain pays more than 5% to borrow for ten years and around 5.8 to 5.9% for thirty, while France and Italy borrow at roughly 4%. On the raw numbers that ought to be the other way round: Britain expects a deficit of 3.6% of national output this year with debt near 95% of it, against France running a 5.1% deficit and Italy carrying debt at 137%. So the market is not charging Britain for how much it owes. It is charging for how the debt is built, and that is the question Washington has just started to face in its own market. Reuters Breakingviews' explanation is structural rather than political. About 24% of British government debt pays interest tied to inflation, against 9% in France. The average British government bond does not come due for 14 years, nearly twice as long as in France, Italy or Japan. Both features make the debt hurt more when inflation surprises and when investors decide they no longer want to lend for long. Those choices were made for a reason. The Debt Management Office, the arm of the British government that sells its bonds, built the market around company pension schemes that had promised retirees a fixed, inflation-protected income for life and so needed long, inflation-protected assets to back it. That buyer is shrinking as the schemes close and move into safer holdings. Britain is left with the most long-dated, inflation-exposed debt of any big rich country (the G7) at the moment global investors are refusing to lock money up and inflation shocks keep arriving, with £252bn of bonds to sell this year including the replacement of old ones, per the Telegraph. This is why Britain matters to a site about American financial plumbing. It is the live test of what happens when a government's debt outlives its natural buyer. America's debt is built the other way, shorter and mostly not linked to inflation, but Washington has just started doing what the Bank of England does: stepping into the market for its own long-dated bonds, with purchases doubled to $4bn per operation from next week. Britain shows the price a market charges when it decides the officials got the maturity wrong. The sources here are commentary, Breakingviews and a Telegraph opinion column, and the £6bn of lost room under the government's own Budget rules is that columnist's estimate, not an official one. For the meter this changes little on its own. It is a reading of pressure we already count under fragility, and we include it because it shows the mechanism at full size in

Crypto and TradFi contagion3 Seplow

$433bn of bets on American stocks, settled in digital dollars

Borrowed-money bets on American shares are now being sold offshore, around the clock, on crypto exchanges: a route by which a wave of forced selling in crypto could set stock prices before regulated markets open.

Binance, the crypto exchange, traded about $433bn in August of a kind of bet that tracks ordinary shares and other conventional assets without ever expiring, roughly 15 times what it did in January, according to a weekly roundup in this window. It has also launched options on more than 1,000 American stocks and funds. A fifteen-fold rise in eight months is not a market growing; it is a market appearing. It sits outside American trading hours, American borrowing limits and American regulators, which means a bad night in crypto can now put a price on Nvidia before Wall Street wakes up. It is not the only door opening. Coinbase has put tokens tracking Apple, Nvidia, Meta and Alphabet on its Base network for users outside the US, tradeable around the clock. Kraken has opened more than 7,000 US-listed stocks to eligible Europeans. Coinbase and the lender Better have launched mortgages backed by bitcoin after a waitlist projected more than $260m of demand. Schwab added three more coins. And twenty-one banks including Citi, Goldman, BofA and Deutsche have agreed to form an issuer of a digital token meant to be worth exactly one dollar, a stablecoin, for 2027. The path we usually watch is crypto trouble leaking into conventional finance. This is the reverse: conventional assets being traded on crypto plumbing. A never-expiring bet on Nvidia offered by an offshore exchange sits outside Reg T, the American rule that caps how much you may borrow to buy shares, outside US market hours, and outside anything the SEC, the agency that polices American markets, can see. When such a bet moves against its holder, the exchange sells them out automatically and at once. A sharp drop in those Nvidia bets at 3am on a Sunday would set a price before the New York Stock Exchange opened. The bitcoin mortgage is the household version of the same thing: a home loan with a demand for more cash, right now, built into it, triggered if the coin falls. Bitcoin is down 2.4% overnight to $79,300 after a 26% month. The waitlist figure is projected demand, not loans written, and we do not know how much the lenders will advance against the coin or at what point they take it. The companies that hold bitcoin as their main asset remain their own story (Strategy up 17.6%, BitMine up 14.7%, Galaxy up 10% in a single session on September 3), and the shape of that rally against the coin is one we described yesterday. What is new is how many doors between the two systems opened in one week, all of them

Private credit and BDCs3 Sephigh

Half out, at a lower price

Limiting withdrawals two quarters in a row at the largest fund of its kind means the line of people trying to leave is a fixture rather than a one-quarter scare, and everyone in that line is being paid at a valuation that keeps slipping.

Blackstone Private Credit Fund, known as BCRED, told shareholders on Thursday that it would give money back to only half of those who asked for it. Holders of 10% of the fund's shares wanted out; the fund is paying 5%, the second quarter in a row it has capped withdrawals at that level. At $77bn it is the largest fund of its type that does not trade on an exchange: it borrows money, lends it to mid-sized private companies, and passes the interest to shareholders. Funds have quarters in which some investors leave. A tenth of the shareholders trying to leave at once, twice running, is not that. Bloomberg reports that anyone who lined up over the last two quarters will have got about 75% of their money out within roughly 90 days. The price the fund says its holdings are worth has fallen since January (from $24.68 a share at the end of January to $23.64 at the end of July, a 4.2% decline), and this quarter's payout will be set at whatever that figure is at the end of September (Bloomberg). The mechanics matter more than the headline. The fund holds loans that cannot easily be sold. The manager decides what those loans are worth, and investors can leave only through a once-a-quarter window capped at about 5% of the fund's value. Shutting that window is not a malfunction. It is the design. What the design does is turn a request to leave into a bet on a future valuation. If you asked to leave in July and are paid in October at the end-of-September figure, you carry three more months of the manager's valuations, and those valuations have been falling. BCRED says it is "well capitalized" and that loan repayments and new money coming in have outpaced the money it has handed back. It also points to returns of 9% a year since it started, nearly 3 points more than loans to heavily indebted companies have paid over the same stretch. Both may be true, and neither answers the question the queue is asking, which is what the loans would fetch if they had to be sold. The only real sale prices we have seen for this kind of paper came from a $90m offer to buy shares in five similar funds, at an average of 26% below what the funds themselves said they were worth. We covered it earlier this week. That answer is well below the fund's. Investors in funds that do trade on exchanges see it differently. Ares Capital, a similar lender with a listed share price, is 0.9% below its high, and Blackstone Secured Lending is 1.4% below its high. Blackstone itself, the manager, is down 4.2% o

Private credit and BDCs3 Sephigh

Never lent to it, closed the exit anyway

A fund that lent nothing to the failing builder has limited withdrawals anyway, which is the first case we can point to this cycle of investors fleeing lending funds because they look alike, not because they have lost money.

Bathla Group, a Sydney property developer that borrowed about A$3.3bn from more than 40 investment funds that lend where banks used to, is trying to raise A$20m to keep operating for five weeks, and its restructuring adviser says he needs "a deal and a commitment in place" from lenders who are, in his word, divided (Bloomberg). One sentence in that report matters more than the rest this morning: a lending fund with no money out to Bathla at all has put new limits on how much its investors can withdraw. A fund restricts withdrawals when it fears a run. This one is being run on for a loan it never made. That is how contagion works, and it is worth spelling out. Investors in Australian property-lending funds cannot see inside the loan books fast enough to know which funds lent to Bathla. So they pull money from anything that looks similar. A fund with clean books that pays everyone out has to sell good loans at a bad time to raise the cash. A fund that limits withdrawals protects those who stay but confirms, in public, that the fear was reasonable. Either choice spreads the stress. This is the same problem Blackstone's BCRED is managing in the US, except that here the trigger is a single borrower rather than a whole sector. The list of lenders is also reaching past the familiar names. The Australian Financial Review reports that RMBL, a $2.8bn lender based in suburban Melbourne that has grown over six decades with almost no public profile, sits in the pile alongside La Trobe Financial and Centuria Capital (AFR). ASIC, the agency that polices Australian markets, has said it sees risk accumulating. Colonial First State, which manages A$179bn of Australians' savings, says it is "watching this very carefully." We covered the collapse itself on Thursday morning. What is new is the shape of it: one over-borrowed builder failing has become a cash squeeze for lenders who never touched it. The Bloomberg headline calls it the "first real test" for this kind of lending in Australia, and that is fair. The test is not whether Bathla's loans are worth what was lent (they are not). It is whether the funds wrapped around such loans hold together when the people who bought into them lose faith at the same time. This moves the odds in the ignition direction, though from a small base and a long way from New York. Nothing was added to the pile of debt this week. What happened is that the thing that sets a pile off, a run that spreads by resemblance, showed up in the wild, and we

Crypto and TradFi contagion2 Sepmedium

Bitcoin fell. Strategy rose 18%.

Companies that exist only to hold bitcoin are the one place you can see directly whether investors will pay more for a coin wrapped in a share than for the coin itself, and on Thursday the share prices said yes while the reporting said no.

Strategy, the company that invented the business of being a listed pile of bitcoin, closed up 17.6% on September 3. BitMine rose 14.7%, SharpLink 10.7%, Galaxy Digital 10.1% and Coinbase 10.2%. Bitcoin itself was down 0.5% overnight at $81,160. A company whose only asset is a coin should move with the coin. Rising 10 to 18% on a day the coin fell is not that, and it is not a one-day fluke: over twenty days bitcoin is up 28.8%, Strategy 49.7%, BitMine 44.3% and SharpLink 42.7%, and all five names are at their highs for the period. Set that against what the reporting says. Bloomberg on Tuesday described "a broader rout in digital-asset treasury companies that has wiped out the premium investors once paid to own Bitcoin through publicly traded vehicles." Its case study is Nakamoto, David Bailey's company, which raised about $760m to buy bitcoin and is down around 99% from its peak. It now trades for less than the coins it holds are worth, which means, in Bailey's words, "the financing engine has broken down" (Bloomberg). The engine is the whole point. One of these companies works by selling its shares for more than the coins behind them are worth, buying more coins with the proceeds, and reporting that each share now stands for more bitcoin than before. It is a machine for turning a stock-market premium into a growing pile of something that produces nothing on its own. When the premium goes, the machine stops, and the company is a fund with a headquarters and a payroll. Nakamoto's move into "businesses that generate cash" is what that looks like. So a day on which the big names add 10 to 18% while the coin does nothing is one of two things. Either investors are paying the premium again, the engine restarting at the top of the sector while it stays broken at the bottom, or traders who had bet against a crowded set of stocks bought with borrowed money were forced to buy them back in a hurry, which pushes prices up regardless of what anyone thinks the shares are worth. The evidence in front of us does not tell us which, and we cannot see from here how far above its coins Strategy is trading today. The tell will be whether Strategy goes back to selling new shares. If it does, the froth we count in this area is back and the pile gets a little taller. If it does not, this was a trading scramble, not a financing engine, and it changes little.

Fed, Treasury and policy3 Sephigh

Bessent wants Japan to raise rates. Japan raising rates pushes up what Washington pays to borrow.

The Treasury Department is asking Japan to do the one thing that pushes up America's long-term borrowing costs while spending money to push those same costs down, and Thursday's calm came from the Federal Reserve, not from either effort.

The US Treasury Department, which borrows the money the government spends, is now publicly telling Japan to raise interest rates. Its under-secretary for international affairs, Erin Browne, told NHK that Scott Bessent, who runs the department, said to Finance Minister Katayama and separately to Governor Ueda of the Bank of Japan, Japan's central bank, that Japan's "next step should be to hike." Katayama says they did not discuss monetary policy. The interest rate the Japanese government pays to borrow for ten years touched 3% for the first time since 1996 (Bloomberg). Governments do not usually tell one another, on the record, what to do with their interest rates. This one is doing so about a rate that is already at a level Japan has not seen in three decades. Now read what the same official told Nikkei a day earlier: "Japan is such a large buyer of sovereign debt ... It tends to have an impact, particularly to U.S. markets." Japanese institutions hold more than $1tn of US government bonds, the IOUs Washington sells to cover its deficits. When Japanese bonds pay more, the pull to bring that money home grows, and the interest rate the US government pays to borrow for ten years, already the highest in 20 months when she spoke (4.8%), goes up (Nikkei). The loop runs like this. The US helped prop up the yen in July, which made a sliding yen Washington's problem. The cleanest way to support the yen is for the Bank of Japan to raise rates. Higher Japanese rates push Japanese bond yields up, which the Treasury itself says pushes up what Washington pays to borrow for the long term. The Treasury is fighting those long-term borrowing costs by buying back its own bonds, a program it has doubled to at least $4bn per operation, starting September 9. One arm of Treasury policy is creating the pressure the other arm is spending money to absorb. The Bank of Japan meets on September 17 and 18. Thursday's calm came from a different building. Governor Waller of the Federal Reserve, America's central bank, said that if August inflation comes in hot he would consider raising rates, but that there is "little cost" to waiting one meeting. Traders' odds of a rate rise slipped back toward a coin flip, the ten-year rate fell to 4.76%, the VIX, the market's gauge of how much turbulence traders expect over the next month, fell to 14.3, well inside the range that counts as calm, and the S&P 500 closed within 0.7% of its high. That is a central banker buying two weeks, not the Treasury

Household credit3 Sephigh

Six for six: every big Chinese bank says more households are not paying

For the first time this cycle every one of China's big banks reports household loans going bad in the same direction, and the reason they give, people owing several lenders at once, is the same one behind America's credit card numbers.

All six of China's big state-owned banks reported at the end of June that a larger share of their loans to individuals had gone bad than at the end of 2025. Bank of Communications was worst, with the bad share of its household lending up 0.44 percentage points to 2.02%; ICBC rose 0.19 points to 1.77%; the other four sit between 1.31% and 1.58%, each up between 0.05 and 0.27 points. Among ten listed banks below the big six, nine got worse, and China Bohai Bank's share of bad household loans jumped 1.13 points to 4.93%. Listed banks raised the money they set aside for loans expected to fail by 17.8% from a year earlier (Caixin). Banks in any country report bad loans rising and falling. Every large bank in a country reporting the same direction in the same half-year is the unusual part. Two things are going on, and Caixin is careful to separate them. One is a stricter standard for classifying credit card debt, introduced in 2026, which relabels existing loans as bad without anything new having gone wrong. That is a change in the measuring stick. The other is what Xu Mingjie, chief risk officer at China Merchants Bank, called a wider erosion in households' ability to repay: "Retail asset quality is currently the main factor affecting our overall asset quality." In other words, loans to ordinary people, not to companies, are now the main thing going wrong on his bank's books. He named a specific mechanism, which the Chinese call co-debt: borrowers owing several lenders at once, so that a missed payment to one spreads to all of them. That mechanism is not Chinese. It is what the Federal Reserve figure we quoted on Thursday, 12.8% of US credit card balances 90 days or more behind, is made of. The difference is where the risk sits. Chinese consumer debt sits on the banks' balance sheets, the lists of what they own and owe, where it is valued and reserved against, which is why the number can be seen at all. A growing share of American consumer debt has moved to lenders that are not banks, to buy-now-pay-later services, and to bundles of loans sold off in slices to investors, where no single equivalent line exists in one place. The shares of American consumer lenders are not worried. Synchrony is 1.5% off its high, Capital One 2.6%, Credit Acceptance 1.2%. The exception is Oportun, which lends to borrowers with weaker credit, at 8.5% off its high and down 8% in twenty days. We do not read a Chinese bank filing as a signal about American default rates. We do read two

Hidden leverage and shadow banking3 Sephigh

The auditor flagged it a year before the prosecutors

It is the clearest single case of the pattern we watch for, risk parked inside an insurer, funded by a government-backed mortgage bank, valued by the man who owns it all, and the paper trail shows the warnings came long before the regulators did.

KPMG, the accounting firm, found a "control deficiency" in how Guggenheim Private Investments recognized revenue when it audited the firm's 2024 accounts. Guggenheim is part of the $367bn investment empire of Mark Walter. The finding followed a whistleblower who in April 2025 raised concerns about $275m of revenue. In the 2025 audit KPMG upgraded that to a "material weakness." It signed both sets of accounts as clean (Financial Times). An auditor finding a weakness and still signing off is not rare. What makes this one worth your time is who else has been looking at the same empire since. On its own that is a story about corporate housekeeping. It matters here because of what surrounds it. In June, insurers controlled by Walter revealed they had lent billions to other parts of his empire and had described those loans wrongly in filings to regulators; US prosecutors are now investigating. Bloomberg reported this week that the same insurers, in a unit called Group 1001, lent more than $20bn to affiliated companies while borrowing more than $6bn from the Federal Home Loan Bank of Indianapolis, a government-sponsored bank created to keep mortgage lenders supplied with cash. That borrowing has grown fast (up 36% in six months and roughly double the level at the start of 2025) (Bloomberg). Walter has been raising cash, agreeing last month to sell his majority stake in the Lakers at a $12.5bn valuation. Put the pieces together and you have the structure this site exists to describe. Money that people paid for insurance sits in an insurer. The insurer lends it to businesses owned by the insurer's owner. Part of the cash comes from a government-backed bank built to support mortgages. The owner's investment firm has a material weakness in how it books its revenue. And the first party to see the problem was the auditor of a private company, whose only duty was to tell management, which it did, while signing clean opinions. Nothing here is new borrowing. Each disclosure this week describes the same structure from a different angle: the loans in June, the funding this week, now the accounting. Our fragility number already carries it, so this does not by itself make a crash more likely. What changes is how much we trust the valuations. An owner who described loans within his own group wrongly in regulatory filings, and whose investment firm could not properly count $275m of revenue, is not an owner whose insurers' valuations of what they hold we would take at face value

The AI capex bubble3 Sepmedium

Nvidia's three guarantees

A supplier helping customers pay for its products has gone from lending them money to promising their lenders that the customers will pay, that the equipment will hold its value and that the business will earn enough, three promises that all come due in the same bad year.

Nvidia, the $5.4tn chipmaker, has in the past two months made three separate kinds of promise to stand behind the customers who buy its chips. The Economist's briefing this week, which calls the company "the central bank of AI", lists them. In mid-August it agreed to backstop up to $105bn on an Ohio data center that will use its chips. A week earlier it announced a plan to mobilize more than $500bn with six Wall Street firms by guaranteeing the value of the equipment it sells to those projects, underwriting, the paper says, as much as a quarter of the cost of some investments. And in July it helped some customers finance data centers "by promising to top up their income if it misses set targets" (The Economist). Companies routinely help customers pay for what they sell. A supplier guaranteeing its customers' lenders against three different kinds of loss at once is not routine. These are three different risks, and it is worth naming them. The Ohio backstop is a promise to pay a debt: if the tenant cannot pay the lease, Nvidia does. The equipment guarantee is a promise about resale value: if the chips are worth less than assumed at the moment a lender needs to seize and sell them, Nvidia covers the gap. The income top-up is a promise about earnings: if the data center does not make enough money, Nvidia pays the difference. A lender normally carries one of these. Nvidia is carrying all three, for the same set of customers, who are buying the same product. The problem is that all three pay out in the same scenario. If demand for computing power grows more slowly than expected and chip prices fall, the tenant struggles to pay, the equipment is worth less, and the data center misses its income target, all at once, and just as Nvidia's own profit margins shrink. Jay Goldberg of Seaport, a research firm, says the company is walking a line between "enabling demand" and "creating it" and is "getting pretty close." The comparison with Cisco and Lucent, the cautionary tale of the last technology boom, writes itself, and the Economist duly writes it. What the briefing does not say, and what no filing we have seen says, is how any of these promises are accounted for. We have only the Economist's word for the July income top-ups and have not seen the terms. Nvidia's shares closed at their high for the period on Thursday, 0% off. The market's view is that a guarantor this large cannot be called on. That is, historically, the view people hold about guarantors right up unti

The dollar, gold and reserve status3 Sepmedium

The protection that isn't there

With only 41% protected, an exit from US assets, if it comes, shows up in the currency market before it shows up in the bond market, and the sellers would be the same institutions that buy the government IOUs that do not come due for twenty or thirty years.

Pension funds and insurers in six markets, Japan, Canada and Taiwan among them, had protected only 41% of what they hold in foreign currencies against those currencies falling, as of June 30. That is the smallest share since at least 2015, according to Bloomberg's review of their filings (Bloomberg). These are the largest owners of US government bonds and US shares outside America, and the dollar is down about 2.3% this quarter. A falling dollar with the protection off means the losses land on the very institutions that hold the most. A Japanese pension fund that owns US government bonds is holding two bets at once: the bond, and the dollar the bond is priced in. Protecting the second bet means agreeing today to sell those dollars back into yen at a future date. That costs money, roughly the gap between short-term interest rates in America and in Japan, and for most of the last decade it was money wasted. The dollar tended to rise when markets fell, which cushioned the bond losses on its own. So the protection came off. The arithmetic that made that sensible needs two things to stay true: the protection stays expensive, and the dollar stays the thing that rises in a panic. Bloomberg's point is that both are being tested at once. Money-market brokers put a 94% probability on the Bank of Japan, Japan's central bank, raising its interest rate to 1.25% on September 17–18 (Nikkei), which narrows the gap between the two countries' rates and so cuts the cost of protecting. And the European Central Bank's own data say gold overtook US government bonds as the world's largest reserve asset last year (FT), which is not what you would see if everyone still trusted the dollar to be the thing that goes up. The mechanism that matters for us: when an unprotected holder decides to protect itself, it does not sell the bond. It sells the currency. Laura Cooper of Nuveen, a large asset manager, puts it plainly: given the scale of foreign holdings, "it doesn't take a dramatic change in positioning to matter." In other words, these institutions own so much that even a small shift toward protecting themselves moves the market. The dollar index, which tracks the dollar against a basket of other currencies, is 99.2 this morning, down 0.35%. The market data does not show this yet, and we are not pretending it does. This belongs in fragility, the slow-moving pile of tinder: a large, concentrated, unprotected position that nobody had to report until the filings came due.

The AI capex bubble2 Sephigh

Dell booked $130bn. Its buyers are borrowing to pay.

The AI demand showing up in suppliers' results is real. The question we track is whether the debt paying for it can be borrowed again when it comes due, and the share prices of the companies renting out AI computing say investors are less sure than Dell is.

Dell raised its forecast for AI server sales in its 2027 fiscal year to $74bn from $60bn, reported a record quarter that beat what analysts expected (sales of $47bn against a $44.9bn estimate), and said it has booked more than $130bn of AI server orders in the past twelve months from 6,500 customers (Channel NewsAsia). The shares rose about 8% in after-hours trading and have more than tripled this year. The orders are real and the cash is real. What is not settled is where the cash comes from, because the companies placing the orders are paying with borrowed money. Start with what is fine. This is revenue, booked when the servers ship and paid in cash. None of it is a projection, and none of it is a backlog in the sense SB Energy uses the word. The AI buildout is buying real boxes from a real company. Now follow the cash upstream. Dell's chief operating officer named the buyers: companies that borrow to build data centers and rent the computing power out, a group the industry calls neoclouds, with CoreWeave and Nscale as examples; governments; and large corporations. The neocloud share of that is funded almost entirely with borrowed money. In the last fortnight alone Nebius enlarged a debt deal to $5bn and sold $5.75bn of bonds that can convert into shares, JPMorgan began marketing $5bn of debt for Volta, a company seven months old, Blue Owl arranged $2.4bn for Iren earmarked specifically for Nvidia's Blackwell chips, and Lambda closed a $926m loan (Bloomberg via sweep). Dell's sales are what comes out the far end of two markets: lending to companies by investment funds rather than banks, and the debt of companies too shaky to be considered safe. The prices say investors know this. Nvidia is up 7% on the week and 1.6% below its high. Dell has tripled. Their customers are moving the other way: CoreWeave is down 8% in five days and 25% below its high, Applied Digital 20% below, Nebius 26% below. The sellers trade like winners and the buyers trade like the ones carrying the risk. That is the correct ordering, since the seller gets paid on delivery and the buyer has to earn the money back over the life of the debt. It is also exactly the shape of a boom in which the buyers borrow to pay the sellers. Dell's order book is a fact about demand. It is not a fact about whether that demand can keep borrowing to pay for itself. For the odds, this adds nothing new to the pile today; the borrowed money was already there. What it does is confirm how much of the pile the

Bond market dysfunction3 Sepmedium

Five companies, a third of all new long-term borrowing

The AI building boom is not only a stock-market story. It competes with the government for the same buyers of debt that does not come due for twenty or thirty years, which is a direct line from companies' borrowing into what Washington pays to borrow.

Amazon, Microsoft, Alphabet, Meta and Oracle borrowed roughly $200bn from bond investors in the first half of 2026, and technology companies accounted for a fifth of all new borrowing by companies rated safe this year and more than 30% of the new debt that does not come due for twenty or thirty years, according to a note from Neuberger Berman, an asset manager, that turned up in our sweep (nb.com). Financing for the big cloud companies, data centers and chipmakers reached about $165bn before the midpoint of the year, $27bn more than in all of 2025. Five companies are now competing with the US government for the same lenders, in the same maturities, and the government is paying more because of it. This is the mechanism behind a number we have been quoting all week without fully explaining: the interest rate the government pays to borrow for ten years is 4.80% and for thirty years 5.27%, the highest since 2007, just before the last financial crisis. Long-dated company bonds and long-dated government bonds compete for the same buyers, pensions and insurers, the very institutions in the dispatch above. When a third of the new thirty-year paper on offer carries Microsoft's or Oracle's name, the Treasury Department, which borrows the money the government spends, has to pay more to place its own. Mark Cabana of Bank of America told Bloomberg the market for government debt "has not been able to hold any type of significant rate decline" (Bloomberg). In other words, every time rates start to fall, something pushes them back up. One reason is that every dip meets another data-center bond. Against that supply, the Treasury's answer, buying back at least $4bn of its own old bonds per operation starting September 9, is a rounding error. The government sold $797bn of debt in a single week in late August. We want to be precise about what this is and is not. Every source we have, including Kashkari at the Federal Reserve, America's central bank, and the dealers Reuters spoke to, says the market for government debt is working: no auction has finished with buyers demanding a sharply higher rate than expected, and none of the gauges of stress between closely related interest rates has widened. This is the market finding a price, not the machinery failing. But the price being found is the cost of financing the AI buildout and the deficit at the same time, in the same maturities, to the same buyers. That is fragility, not ignition: nothing is about to break, but the government

Crypto and TradFi contagion2 Sepmedium

Fifty-eight banks, two new tokens, $12.5m of takers

Digital dollars issued by banks would connect the market for short-term government debt directly to crypto trading, inside banks large enough that their failure would matter to everyone. It is a new channel, not yet built, that nobody has tested under stress.

Twenty-one financial institutions, Goldman Sachs, Bank of America, Citi, Deutsche Bank, UBS, Wells Fargo and MUFG among them, said they will set up a company this year to issue a digital token meant to be worth exactly one dollar, a stablecoin, in the first half of 2027, with a euro version to follow (Reuters via Livemint). They will compete with Qivalis, a separate group of 37 institutions planning a euro token this year. Some banks, including BBVA, are in both. Until now this business belonged to companies outside the banking system; this is the largest banks in the world deciding to run it themselves, which moves the risk inside the buildings regulators watch. What a stablecoin is, mechanically: a deposit that has left the bank and become a token, backed by a reserve of short-term US government debt, Treasury bills, whose interest the issuer keeps. Tether, based in El Salvador, has issued more than $180bn of them and made billions doing exactly that. That is $180bn of demand for government bills and $180bn of deposits not sitting in any bank. The banks' move is best read as defensive. If ordinary and institutional money is going to become tokens under the new American and European laws written for them (the GENIUS Act and MiCA), the banks would rather issue the tokens than watch the deposits leave. This cuts against our thesis in one respect, since it is risk moving toward the regulated banks' balance sheets, the lists of what they own and owe, not away from them. It confirms the thesis in another: the same week saw Circle create about 5bn of its USDC tokens, Coinbase and Better launch mortgages secured on bitcoin off a $260m waiting list, and Vanguard and Wellington test tokenized shares in money-market funds as something a lender will accept as security (sweep). The pipes between crypto and the market for government bills are being laid faster than anyone is measuring them. The contagion question is simple. A run on one of these tokens, everyone asking for their dollars back at once, is a forced sale of government bills. Done by Tether, that is a market event. Done by a vehicle owned by 21 banks, it becomes a question about the banks. For today this adds nothing to the pile, because nothing has been built; the day it carries real money is the day we move fragility. And there is one data point on demand: Société Générale, the only major bank that has actually issued a dollar token, has $12.5m of it in circulation.

Fed, Treasury and policy2 Sephigh

Two dates: September 9 and September 17

A Japanese rate rise on September 17 would test whether $4bn buybacks can hold up the price of thirty-year US government bonds while the single largest foreign holder of them has a new reason to bring its money home.

The market for US government bonds that do not come due for twenty or thirty years now has a calendar, and the two dates on it pull against each other. On September 9 the Treasury Department, which borrows the money the government spends, runs the first of its enlarged buybacks, at least $4bn per operation in bonds with ten to twenty and twenty to thirty years left to run, continuing to November 4 (sweep). On September 17–18 the Bank of Japan, Japan's central bank, meets, and money-market brokers price a 94% chance it raises its rate to 1.25%. Its governor, Ueda, said on Tuesday a rise will be "discussed thoroughly in every meeting, including the next one" (Nikkei). The first date is Washington trying to hold its long-term borrowing cost down; the second is the country that owns more of that debt than any other being handed a reason to take its money home. The two are connected by scale. Bessent, the Treasury secretary, announced the expanded buybacks on August 19. By September 1 the interest rate the government pays to borrow for thirty years was back at 5.27%, exactly where it stood before he spoke, and the ten-year rate was a tenth of a percentage point higher, at 4.80% (Bloomberg). The program is presented as help for older bonds that rarely trade, not as an attempt to hold rates at a level, and $4bn is what the Treasury sells in about an hour of a normal week. It is a backstop for stale bonds, not a defense of a price. Japan is the larger force. Japanese institutions hold more than $1tn of US government bonds. Japan's own government now pays 3% to borrow for ten years, the first time since 1996, and that is roughly the point at which a Japanese life insurer can earn a return above inflation at home without taking on any currency risk. Add the arithmetic on currency protection from this morning's dispatch and the route home for that money is open. Bessent's own undersecretary said as much on Tuesday. Bessent himself called Japan's era of stimulus over and christened the new one "Takaichinomics". This is policy pulling in two directions. The Treasury is spending its own balance sheet, the list of what it owns and owes, to hold long-term rates down, while its officials in effect endorse a Japanese rate rise that pushes them up. Warsh, who chairs the Federal Reserve, America's central bank, has meanwhile called 2% a "firm and fixed target" and offered no help. None of this is a mistake yet. It is a policy with no lever attached. For the odds, this is the

What would change our mind

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

The outcome of the Fed's September 15 meeting and the reaction: a rate rise met with a rally in long-term government bonds would lower ignition; a hold or cut followed by the thirty-year rate breaking above 5.4% would raise it sharply.

Would move the number

The extra interest shaky companies pay compared with the government going through 3.5 percentage points, or a listed lending fund (ARCC, BXSL, OBDC) trading 10% below the value it puts on its own loans while the loans that have stopped paying rise for another quarter. That would say the market has joined the reporting on private credit.

Would move the number

A financing for one of the smaller computing companies failing to sell, whether Nscale's convertible notes repriced or an Nvidia-backed deal pulled, or a ratings action on one of the cloud giants from S&P or Moody's citing the debt-like commitments it now says it cannot see through.

Would move the number

A weak auction of long-term government bonds, or a Treasury buyback operation that fails to narrow the price gap between older bonds and the newest ones once operations start on September 10; we would also lower fragility if the stablecoin total and withdrawal requests at the lending funds both fell for a second consecutive quarter.

Would move the number