Nvidia pulls the guarantee that made small AI clouds bankable

Archived reading, published Fri, 28 Aug 2026 10:21:14 UTC (23 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

Fragility88

how much tinder is stacked up — moves slowly

Ignition31

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 22 sources and rewrites this page.

Status: Held at 60. Ignition slips a point because the spark from long-dated government bonds keeps fading: the 10-year interest rate is 4.66%, below the 4.73% it hit on August 21, with the VIX (the market's gauge of how much turbulence traders expect over the next month; under 20 is calm) at 14.47, the extra interest rate on shaky corporate debt at 2.67 percentage points above government bonds, and the same measure for safer corporate debt at 0.80 percentage points. Fragility holds at 88 because two structural developments roughly cancel each other out: Anthropic added a new $45bn six-year lease obligation to a builder valued at $14.6bn in March, and Nvidia withdrew the rent-back guarantee program that had been making small cloud startups bankable.

Reporting from 26 Aug to 28 Aug

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.

The AI capex bubble28 Augmedium

Nvidia withdraws its buyer of last resort

The vendor guarantee was what turned speculative lending on graphics chips into something closer to lending against contracted revenue. Withdrawing it tests whether the smallest AI clouds can borrow on their own credit.

Nvidia has paused the financing program under which it offered small AI cloud companies credit support in exchange for a share of their revenue, stepping back from it last week, less than two months after announcing it. The Wall Street Journal reported the retreat; Reuters carried it. Nvidia's spokesperson says the new business model "is still in place and continues to evolve due to high demand," and that it could be revamped or folded into another program. The mechanism matters more than the size. Under the arrangement, Nvidia would rent computing capacity back from cloud customers if they could not sell it. That single promise changes what a loan backed by graphics chips actually is. A lender underwriting a company with no operating history and a warehouse full of chips that lose value every quarter is making a speculative bet on future demand. The same lender, looking at a contract in which Nvidia agrees to be the buyer if nobody else shows up, is underwriting something much closer to guaranteed revenue, and will lend more, at a lower interest rate, against it. Remove that backstop and the small cloud startups go back to standing on their own order books. What is left is still enormous. Goldman analysts, cited by CNBC, put Nvidia's financial commitments in support of customers at $366bn: $279bn of supply and capacity commitments (largely memory), $29bn of cloud service agreements, and other data-center items. The $500bn financing push with six asset managers and the $105bn Ohio arrangement for OpenAI are untouched. Two readings. Either Nvidia is trimming its contingent obligations under the circular-financing scrutiny it has drawn all summer, or demand is strong enough that it does not need to buy demand. Markets prefer the second: Nvidia closed at an all-time high, up 8.8% on the day. Watch whether the cloud-startup bond pipeline (Applied Digital's roughly $3.5bn offering of debt from a company too shaky to be considered safe, Novva's $800m) prices at a higher interest rate without the guarantee behind it.

The AI capex bubble28 Aughigh

A $45bn lease from a company worth $14.6bn

The AI build is now financed by lending against the promises of private companies whose accounts nobody outside the deal has seen.

Anthropic has agreed to rent 460MW of Nvidia's next-generation Vera Rubin computing capacity at Nscale's West Virginia data center, in a deal the FT puts at roughly $45bn over six years. The chips come online late next year. Nscale raised $2bn in March at a $14.6bn valuation, led by Norwegian energy group Aker, with Nvidia, Dell and Nokia among the backers. The Mason County site is planned at 1.35GW across 2,250 acres, with a 2GW natural gas plant built on site to power it, and Nscale estimates the cost at $69bn. It is preparing an IPO that could come as soon as next month. The arithmetic only works one way round. Nscale cannot fund $69bn from equity. What makes the project financeable is the signed Anthropic lease, a long-dated stream of payments a lender can size a loan against. This is now the standard shape of the AI build: JPMorgan arranging $5bn for Volta after Volta signed a $10bn computing contract with Anthropic; Meta's Hyperion financed through a Blue Owl joint venture that issued $27bn against Meta's lease. The tenant's credit is the collateral. So the question is Anthropic's balance sheet (the list of what it owns and what it owes), and the reporting does not disclose it. Anthropic has separately committed to lease more than $150bn of Google's AI chips, signed agreements with Amazon and SpaceX, held early talks with Meta, and plans an IPO this autumn. The West Virginia site alone implies roughly $7.5bn a year of rent. This is the migration our running thesis describes, in its cleanest form. None of this obligation sits in a bank. It sits in leases, project debt and private capital, and it is underwritten by a revenue forecast for a company that has not yet filed publicly.

Bond market dysfunction27 Augmedium

AI borrowing is now setting the price of Korean debt

The AI build is no longer a sector story. It is driving up the extra interest that lenders demand for tying their money up for a long time, and that cost is being paid by countries with no AI industry at all.

As the interest rate the US government pays to borrow for thirty years climbed to 5.33% last week, its highest since 2007, Korea's thirty-year government bond rate hit a record 4.751%. It has since retreated. The mechanism that carried it there has not. Samsung Securities, quoted by the Korea Herald, breaks the move down. This year's rise in US long-dated rates is about investors demanding more compensation for tying their money up for decades, not about repricing what the Federal Reserve (America's central bank, which sets interest rates) will do next. Three things are pushing on that compensation at once. The US deficit implies heavier borrowing. Foreign official demand is fading: most of a $72bn decline in foreign holdings of US government bonds in June came from central banks and sovereign wealth funds. And six major technology companies issued $182bn of bonds in the first half of the year, fourteen times the year-earlier figure, with more than 30% of it carried at thirty-year maturities. That last number is the one worth sitting with. Data-center spending is being funded in exactly the same maturity bucket as government debt, competing for exactly the same pool of long-term buyers: insurers, pension funds, sovereign funds. There is no separate window for AI paper. When Alphabet or Oracle sells a thirty-year bond, the buyer choosing between it and a long-dated US government bond pushes the price of both. Bloomberg's JPMorgan interview this week put the resulting premium at about 0.55 percentage points over other US corporate debt. The transmission is then arithmetic. US government bonds are the benchmark for global bond pricing; Korea's long-dated rates reprice, then Korean mortgage and corporate borrowing costs follow. A Seoul homeowner is, at one remove, paying for the electricity contracts in Ohio and West Virginia.

Private credit and BDCs27 Augmedium

The number the private-lending funds report and the number that matters

Losses in private lending are disclosed on a lag chosen by the people holding the assets, and the adjusted figure is rising three times as fast as the headline.

Blackstone's BCRED, the largest fund of its kind (a fund that borrows money, lends it to mid-sized private companies, and passes the interest to shareholders), saw loans where the borrower has stopped paying interest go from 0.6% to 2.4% of the portfolio between the first and second quarters, on two names: Medallia and Affordable Care International. That is a quadrupling in three months, from a very low base. Across the ten largest listed funds of this type, loans where interest has stopped arriving reached 3.95% of debt at cost in the second quarter, up 0.20 percentage points on the quarter. But there is a second figure in the same Morningstar analysis that gets less attention and deserves more. Count every slice of every borrower that has at least one loan where interest has stopped arriving, and the exposure is $5.0bn, or 5.95% of total debt at cost, up 0.54 percentage points in the quarter, nearly three times the headline move. That gap is the interesting mechanism. A borrower rarely stops paying on one loan and keeps paying comfortably on the rest. The junior slice goes first, the senior follows, and the reported non-payment rate lags the credit reality by however long it takes the manager to concede. The adjusted number is the better estimate of what is actually impaired, and it is deteriorating faster. The prices say none of this. Ares Capital is 1.6% off its high, Blue Owl's lending fund up 4.4% over twenty days, FS KKR up 16%, Blue Owl the manager up 20%. The one exception is precise: Golub, where the value per share slipped from $14.35 to $14.25 and the count of non-paying loans went from 19 to 20, is the fund that fell this week, down 1.3% over five days. The extra interest investors demand for risky debt is tight, the valuations are the managers' own, and the share prices are priced for none of it. That is the disagreement to hold onto.

The dollar, gold and reserve status27 Aughigh

The rescue was the catalyst for the escape

Gold and bitcoin are now moving as one trade, and that trade is responding to fiscal policy rather than to inflation or the currency.

Exchange-traded funds tracking gold and bitcoin took in a record $7bn over five trading days, per Bloomberg data: nearly $3.4bn into SPDR Gold Shares and $1.5bn into BlackRock's iShares Bitcoin Trust. Both cracked the top ten US ETFs by weekly inflow, with GLD trailing only a handful of funds including the Vanguard S&P 500 ETF. The proximate catalyst, Bloomberg reports, was Scott Bessent's plan to at least double long-dated US government bond buybacks. That is worth stating plainly: the largest recorded five-day flow into the two assets whose supply no government controls was triggered by the Treasury Department (which borrows the money the government spends) announcing it would support its own bond market. For most of the past two years these were treated as alternative trades: gold for institutions worried about fiscal arithmetic, bitcoin for a different crowd. They are now moving as one position. Gold is up 15.5% over twenty days at $4,658. Silver sits at its period high, up 17.3%. Bitcoin is up 21.8% over the same window at $79,030. What is not happening is a dollar crisis. The dollar index (which tracks the dollar against a basket of other currencies) is 99.23, down 0.73% over twenty days; the euro and yen ETFs have barely moved. This is not capital fleeing dollar assets for other currencies: the euro has picked up none of it. It is capital moving out of long-dated bonds and out of claims on future government solvency, and into things with a fixed supply. A monetary historian cited in the same coverage puts the dollar's reserve share at just under 60%, down from a little over 70% at the turn of the century, with roughly three-quarters of the loss going to non-traditional reserve assets rather than to the euro. A repricing of policy credibility, then, not a funding event. Those look different, and it matters which one you are watching.

Fed, Treasury and policy27 Aughigh

Spending the emergency account to buy back bonds

If the Treasury Department is seen to be managing interest rates rather than funding the government, every buyer of long-dated government bonds reprices what a Treasury promise is worth.

Three major outlets wrote the same piece this week, which tells you something on its own. The Economist, a New York Times guest essay by a former deputy chief of staff to Janet Yellen, and Bloomberg Opinion all examined what the Treasury Department is doing with long-dated bonds, and all three landed closer to skepticism than to reassurance. The facts underneath. On August 19 the Treasury said it would at least double buybacks of ten- to thirty-year debt, to at least $4bn per operation from $2bn, taking the two-month total to at least $28bn, announced outside the regular schedule. Buybacks themselves are not unusual. Announcing one outside the published calendar, two weeks after the schedule that was supposed to contain it, is. CNBC has since reported that Bessent is considering funding further buybacks from the Treasury General Account, the government's roughly $954bn checking account at the Federal Reserve, kept deliberately large for emergencies such as shutdowns. Stanley Druckenmiller publicly criticized his former mentee in the Wall Street Journal. Mechanically, a buyback funded from that account is a swap: the Treasury retires long-dated bonds, drains its own cash balance, and adds reserves to the banking system, shortening the government's effective debt maturity without ever announcing a change in borrowing policy. It is not nothing. Citrini Research argues this is one piece of an emerging "Treasury-Fed Accord," with the Fed shrinking its holdings while banks absorb more short-term bills, and recommends betting on thirty-year bonds outperforming five-year notes. Both things can be true. The trade may work and the credibility may still be spent. The Economist notes the Treasury's usual role is keeping the market orderly in crises, not pushing interest rates around in quieter times, and more than half of the US deficit now consists of interest on past borrowing. Markets are currently siding with Bessent: the ten-year rate sits at 4.66%, below where it traded on August 21.

Signs of the times

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

A $14.6bn company signs a $45bn lease

Nscale, valued at $14.6bn in a March funding round, has agreed to rent Anthropic 460MW of capacity for roughly $45bn over six years at a West Virginia site it estimates will cost $69bn and which will have its own 2GW gas plant on the premises. The single contract is three times the value of the company that signed it.

aidata centresleases

27 AugFinancial Times

Google's shopping list grew tenfold

Google's purchase commitments for chips, memory and networking gear totaled $707bn in the most recent quarter alone, against $72.5bn for all of 2025, according to Morgan Stanley research. Across seven Big Tech firms, off-balance-sheet AI commitments come to roughly $3tn, on top of $770bn of debt and leases that are on the books.

off balance sheetcapex

27 AugAxios

Jane Street's first losing month in a decade

Jane Street, now one of the largest financial institutions in the United States, lost money over a month for the first time in ten years: a $15bn hit to trading revenues driven mainly by AI-linked positions and its stake in Situational Awareness, a fund run by 24-year-old Leopold Aschenbrenner that was forced to liquidate in July.

hedge fundsconcentration

27 AugFinancial Times

Half the deficit is now interest

America's budget deficit is 6% of GDP, the widest ever outside recession or wartime, and more than half of it now consists of interest payments on past borrowing. Bessent's own target of 3% by 2028 is, in The Economist's assessment, fanciful.

fiscalterm premium

27 AugThe Economist

The rumour mill

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

Confirmed

“Marvell signed a deal to help build Google's custom TPUs, and Google received a warrant for nearly 59 million Marvell shares.”

Marvell's own 8-K, filed August 19, discloses a warrant for up to 58,970,907 shares at $206.58, vesting mostly on how many custom products Google buys through fiscal 2033. The supplier is paying the customer in its own equity to buy its product, the same circularity as vendor financing, just settled in shares instead of cash.

Claimed by Meet Kevin

Partly true

“Nvidia wants to give OpenAI a $125bn guarantee.”

The announced figure is $105bn for the Ohio data center, and the Wall Street Journal reported that earlier talks ran to $250bn before being cut to under $120bn. The number in circulation is not one anybody actually reported.

Claimed by Wealthion

Confirmed

“The US goods trade deficit widened to $118.8bn in July from $101.4bn in June, with capital goods imports jumping 11.3%.”

Census Bureau advance data, corroborated by Bloomberg, Reuters and Haver: a 17% one-month deterioration, the widest since March 2025, with the non-auto capital goods rebound the largest monthly move since 1993. The AI build is now big enough to move the national trade balance.

Claimed by Gregory Mannarino

Partly true

“Nvidia is nearing agreement to acquire Hugging Face at about $13bn.”

The Information reported an agreed price of $12.9bn and Reuters and CNBC relayed it; Business Insider reported the two were still in talks with no deal reached. Treat the price as reported, not settled.

Claimed by Meet Kevin

Earlier developments

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

Bond market dysfunction27 Aughigh

A price on AI credit, at last

It is the first hard market price on AI credit risk, and it says the public bond market is charging for it even while the broader index looks complacent.

Matthias Reschke, who runs European investment-grade bond sales at JPMorgan, told Bloomberg TV that bonds sold by the hyperscalers (the handful of giant tech companies building AI infrastructure) trade about 0.55 percentage points above the extra interest other US corporate borrowers pay, what he called a "material premium." The same group has raised more than $220bn this year across seven currencies. In several smaller markets, a single deal was enough to make a Silicon Valley company the largest non-financial borrower in the country. Put that next to what the broader market is doing. The index that tracks borrowing costs for safe corporate America is at 0.80 percentage points above government rates, the tightest reading in our eight-run trail. A 0.55 percentage point premium on top of that average is most of the index's own cost again. So the credit market is not asleep on AI. It is charging for it, specifically, in a market that is otherwise priced as though nothing can go wrong. That cuts both ways for our running argument. Risk that is priced is risk that is visible, and visible risk is the kind that gets repriced gradually rather than discovered all at once. Public bonds with a quoted price are the best place for this exposure to sit. The problem is where the rest of it sits: in leases that have not started, in special-purpose vehicles (separate companies set up to hold something off the parent's books), in loans secured on chips, in guarantees written by the vendor. None of those carry a price anyone can see on a screen. Reschke's own framing is the tell. "We will continue testing at what price levels investors will be able to participate," he said. "We don't doubt that we will be able to place it, it's all a question of price." That is a market clearing, not a market seizing. The moment worth watching is the first deal that does not place at any price the issuer will accept, because the borrowing does not stop when that happens. It just moves somewhere with no price to quote.

Household credit28 Augmedium

Subprime auto hits a July record, and lenders lean in

Loss rates and the willingness to lend are moving in opposite directions in the same group of borrowers, which is how a slow credit problem becomes a fast one.

Fitch's index of subprime auto loans at least 60 days late hit 6.13% in July, the highest July reading in a series that starts in 1994, up from 5.80% at midyear. In a normal year, subprime auto delinquency runs closer to 4%. The prime rate in the same index is 0.49%. Subprime borrowers are now more than twelve times as likely to be two months behind as prime ones. Separately, the New York Fed's second-quarter data show 5.5% of all US auto debt at least 90 days late, the second-highest level in more than twenty years of the series. This cuts against what we wrote two runs ago, when credit card and student-loan numbers improved and we trimmed our fragility reading on the household front. It is a divergence rather than a reversal: credit card payments 30 days late are still around 2.7% and stable, and the average credit score held at 714. The deterioration is concentrated in exactly one place: the bottom of the credit ladder, in the one loan Americans need to keep a job. What makes it worth a dispatch is the second number. Cox Automotive's Dealertrack credit availability index read 105 in July, up 6.9% year on year and near an eleven-year high in looseness, and credit-score data show auto balances for the lowest-scoring borrowers up 36% year on year. Lenders are not tightening into rising losses. They are extending more credit to the group already failing. The mechanism to watch is how these lenders fund themselves. Deep subprime auto lenders raise money by bundling their loans together and selling slices of the income to investors, where excess interest absorbs early losses before any bondholder is touched. That is why late payments can run hot for a long time without anything visibly breaking. Credit Acceptance, the deepest-subprime lender we track, is up 7.8% in twenty days and sits 2.9% off its high. Nobody is pricing distress. The tell would be a deal that fails to find buyers, or one that trips its cumulative-loss trigger and starts diverting cash away from the riskiest slice.

Crypto and TradFi contagion27 Aughigh

The obituary and the tape disagree

A company that holds bitcoin and trades above the value of its coins can issue shares and buy more. Below that value, it becomes a forced seller. The tape just flipped several of them back across that line.

The FT ran the numbers on the fifty companies holding the most bitcoin: combined market value down from $150bn in July 2025 to $67bn, with Strategy alone accounting for $79bn of the decline from its peak. Forty-three of the fifty trade below their pre-pivot price. Thirty-five have halved. "It was always doomed," a crypto research head told the paper. "I don't think we see new ones." We counted that as genuine paying-down-of-borrowed-money when it first crossed. Then look at the screen. Strategy closed up 11.6% on the day, 22.3% in five days and 40.8% in twenty, at the top of its recent range. BitMine is up 41.7% over twenty days, SharpLink 38.6%, both at their period highs. Bitcoin is up 23% in the same window to $79,894. The two facts are compatible, and the way they fit together is the whole mechanism. A company that exists to hold bitcoin is an arbitrage on its own share price: if the stock trades above the value of the coins it holds, management can issue new shares, use the cash to buy more coins, and every existing shareholder ends up with more coin per share. Below that value the machine runs backward. Issuing shares dilutes everyone, so the only lever left is selling coins, which is what the FT found many of them doing. So a 40% rally is not cosmetic. It is the difference between a company that must sell into weakness and one that can buy into strength. The FT attributes the move to the debasement trade set off by the Treasury Department's bond-market intervention, the same impulse driving gold. What we do not know is whether any of these companies have restarted selling new shares into the market. That filing, not the share price, is what would tell us the model is back rather than merely bouncing.

The dollar, gold and reserve status27 Augmedium

Central banks were not the ones buying gold

The gold rally is being read as a structural shift away from the dollar when the flow data says it is fund money that can reverse.

Gold is at $4,625, up 14.7% in twenty days. Silver is up 17.3%. The story attached to that move is that the world's central banks are quietly diversifying out of dollars. The World Gold Council's own revision says otherwise. The Council cut its estimate of first-quarter central bank purchases to 57 tonnes, from 244, a downward revision of 187 tonnes, with the difference reclassified into over-the-counter and other demand. Second-quarter buying was 289 tonnes, up 62% year on year, but the first half still totals about 345 tonnes: the weakest first half since 2022. We checked this because it was circulating on gold-focused channels; it is confirmed, and it comes from the WGC's own Gold Demand Trends. So who is buying? Flow data reported by Bloomberg says gold and bitcoin ETFs (funds that let ordinary investors buy exposure without holding the asset themselves) took in a record $7bn over five trading sessions, with roughly $3.4bn into SPDR Gold Shares alone. That is Western fund money, not sovereign reserves. The distinction matters because the two buyers behave differently. A central bank that buys gold as a reserve asset does not sell it when the price falls 8%. The position is strategic and close to permanent, which is why sovereign accumulation raises the floor under the price. ETF money is a position, and positions get closed. A gold price built on ETF flow is a gold price with a reversal built in. Note also what the dollar is doing while this happens: the dollar index, which tracks the dollar against a basket of other currencies, is at 99.19, down only 0.77% in twenty days. If this were a genuine shift away from the dollar as a reserve currency, you would expect the dollar to be the mirror image of gold. It is not. The debasement trade is currently a trade.

Hidden leverage and shadow banking28 Augmedium

The SEC has subpoenaed four prime brokers

The lending that large banks do to hedge funds against concentrated AI stocks is the shortest path between a stock selloff and a funding crisis, and a regulator is now measuring it.

On August 24 the SEC, the agency that polices American markets, sent subpoenas to Bank of America, Citi, Goldman Sachs and JPMorgan seeking information on the trades and use of borrowed money at Situational Awareness, the AI-focused hedge fund run by 24-year-old Leopold Aschenbrenner, whose portfolio was reported down roughly 67% before it was forced to sell everything into July's AI selloff. Funds have bad quarters. Losing two thirds of your money in a month is not a bad quarter. The FT filled in the other side of that trade this week. Jane Street had invested heavily in the fund, employs a former Situational Awareness staffer, and held many of the same AI-linked names: SanDisk, Nebius, Bloom Energy. The combination produced a $15bn hit to trading revenues and the firm's first losing month in a decade. "When you start seeing the kind of revenue these guys are putting up, the only way that's possible is when your holding period extends as well as the size," Bloomberg Intelligence's Larry Tabb told the paper. In other words, the profits were so large that the firm must have been holding bigger positions for longer, which is exactly what makes a loss this size possible. In its own letter to staff, Jane Street admitted it was not protected against some of the moves. The plumbing question the subpoenas are aimed at is how much of a cushion the banks required. A prime broker (the desk at a large bank that lends to hedge funds) lends against a portfolio and sets how much of the client's own money must be posted. When the positions are concentrated and correlated, the model is only as good as its assumption that the stocks move independently. They did not. What the regulator wants to know is how much borrowed money four large banks extended, against what security, to a fund that was two years old. The honest reading is that this already happened and the system absorbed it. There was no funding crisis, and Goldman and Morgan Stanley shares are up 3.8% and 3.6% in five days. The market does not think the prime brokers lost money. The residual risk is second-order: a subpoena tends to make risk committees more conservative, and the cushion requirements on AI-correlated books are exactly the kind of thing that tightens quietly, weeks before anyone reports it.

The AI capex bubble28 Augmedium

Nebius is borrowing against the chips themselves

Loans secured on GPUs are really loans on customer contracts, dressed as equipment-backed credit, and the equipment loses value on the vendor's schedule.

Nebius upsized its debt raise this week from $4.5bn to $5bn, reportedly split into $3bn of private convertible notes due 2030 and $2bn due 2034. It is targeting up to $25bn of spending on infrastructure this year against more than $40bn of customer commitments and more than $9bn of expected upfront customer payments in 2026. In July it took a separate $775m loan secured on existing GPUs and on the cash flows from an agreement with a borrower safe enough to be considered investment-grade. That July loan is the structure worth understanding, because it is spreading. Two things are pledged: the hardware, and the contract. The hardware is a depreciating asset with no deep resale market and a replacement cycle set by the vendor. When Nvidia's next-generation chip ships, the resale value of what secures this loan is decided by Nvidia's product calendar, not by the borrower. The contract is the real credit, which means the lender is taking the customer's risk while calling it a secured loan on equipment. The same logic runs through the rest of the week's financing. Digital Drive is preparing $1.085bn of secured notes for a CoreWeave-leased campus through a vehicle affiliated with two data-center developers. Applied Digital is sounding out roughly $3.5bn of debt from companies too shaky to be considered safe borrowers. Broadcom's reported $60–100bn chip financing would be issued through a separate special-purpose vehicle with Blackstone and Apollo named as participants. In each case the borrowing sits somewhere other than the books of the company whose credit actually backs it. One caution: most of this reaches us through web sweeps citing other outlets, and the details of the Nebius deal are reported inconsistently. What is solid is the direction. Shares are not confirming: Nebius is up 15.9% in twenty days but still 21.4% below its recent high, and CoreWeave is 19.4% off its own.

The AI capex bubble27 Aughigh

A $2.4bn company borrows $5bn

The AI buildout is increasingly financed by lending against contracts with private counterparties whose ability to pay nobody outside can verify.

JPMorgan has begun early outreach to lenders for a $5bn debt package to fund Volta Infra Holdings' AI data-centre buildout, Bloomberg reported. Volta raised $300m of venture money earlier this month at a $2.4bn valuation, backed by Nvidia and Dell. It was founded this year by two former Brookfield infrastructure executives. The debt being arranged is roughly twice the company's entire equity value, for a firm that is a few months old. The way to read this is not as a bet on Volta. It is a bet on a contract. In August Volta struck a $10bn deal to supply computing capacity to Anthropic, and it is that promise of future payments — not Volta's balance sheet, which barely exists — that the lenders are being asked to underwrite. The credit question is therefore Anthropic's. Anthropic is private. It was valued at $965bn in May and plans an IPO this autumn. Adding up only the deals Bloomberg and the FT have reported — $50bn with Fluidstack, $45bn with SpaceX, $45bn with Nscale, $10bn with Volta, and more than $150bn of Google chips — gives over $300bn of announced compute commitments against revenues nobody outside the company has seen. That arithmetic is ours, from their published deal list; the company does not publish a total. Bloomberg puts total borrowing for data centres and AI since last year at roughly $600bn. The mechanism repeats: a supplier signs a long contract with an AI lab, borrows against it, and the lender's recovery depends on a private company continuing to want the compute at the agreed price. Nothing in that chain is marked daily.

Private credit and BDCs27 Augmedium

The two non-accrual numbers

The difference between reported and adjusted non-accruals is the difference between what a private lender has decided to admit and what its borrowers are actually paying.

Morningstar's tally of the ten largest listed BDCs puts Q2 non-accruals at 3.95% of debt at cost, up 20bp on the quarter, or $3.3bn. That is the headline figure. The more interesting one is underneath it: when you count every loan to a borrower that has at least one loan on non-accrual, the exposure is $5.0bn, or 5.95% — up 54bp in a quarter, nearly three times the move in the headline number. The gap is the mechanism. Non-accrual is not an event; it is a decision the lender makes to stop booking interest it is not receiving. A borrower in trouble often has several tranches from the same fund. Put the worst one on non-accrual and the rest keeps accruing income into the fund's reported yield. The adjusted number says roughly $1.7bn of loans to already-impaired borrowers are still counted as performing. Two other prints: the top-20 BDCs' median non-accrual rate reached 2.8% of cost, the highest since 2017, and Blackstone's BCRED went from 0.6% to 2.4% between Q1 and Q2, driven by Medallia and Affordable Care International. Note also that the top-10 portfolio shrank 2.3% — a smaller denominator flatters nothing here, it makes the ratio worse. The prices disagree. Over twenty days FS KKR is up 16.1%, Blue Owl's manager up 20.5%, Ares Capital up 5.5%, Blackstone Secured Lending 1.1% off its high. Equity holders are being paid a dividend out of a yield that the adjusted figure suggests is partly accrued rather than collected. These are single-source tallies, not audited aggregates, and we treat them accordingly.

Bond market dysfunction26 Aughigh

The whale bought 72% of the auction

The marginal buyer of Japanese government debt is a pension fund rebalancing on autopilot, and its buying reverses exactly when the market would need it most.

On 18 August Japan's finance ministry auctioned about ¥2.5tn of five-year government bonds. More than ¥1.8tn went to buyers nobody could identify. The largest named buyer, Mitsubishi UFJ Morgan Stanley Securities, took ¥265bn. Traders told Nikkei there was only one plausible explanation: the whale. The whale is GPIF, Japan's public pension fund, with more than ¥300tn ($1.88tn) under management. A 2025 rule change let it bid directly in JGB auctions rather than going through dealers. The point of bidding directly is concealment — the fund's identity as a buyer becomes clear only after it has left. What makes this structurally interesting is why it buys. GPIF runs a fixed allocation split roughly evenly between domestic stocks, foreign stocks, domestic bonds and foreign bonds. When global equities rally, the equity sleeve grows and the fund must buy bonds to rebalance. In fiscal 2025 it bought a net ¥14tn of Japanese bonds and sold a net ¥10tn of domestic stocks. So the largest incremental buyer of Japanese government debt is buying for reasons entirely unrelated to the price of the bond — and the flow reverses in an equity selloff, which is precisely when the JGB market would otherwise need a bid. The other side of Japan's rate rise is already visible. Seventeen shinkin banks — small community lenders — posted net losses for the year to March, up from five. Bond-related losses across the sector nearly tripled to ¥424.8bn, and 237 of 245 institutions carry unrealised losses on securities, ¥2.6tn in total. Ehime Shinkin lost ¥18.9bn: its first loss since it was founded in 1951.

Hidden leverage and shadow banking26 Augmedium

Who actually owns the long end

Treasury is buying back long bonds from a market whose marginal holder is levered around ten to one in overnight repo.

A set of numbers has been circulating in market commentary this week that is worth stating plainly, with the caveat that it is not new: it comes from the Fed's Form PF analysis published in June, using September 2025 data. This is an old measurement being passed around, not a fresh disclosure. Large hedge funds held $2.4tn long and $1.6tn short in Treasuries — $4tn gross, about 8.5% of privately held Treasuries by market value, roughly double their share in early 2023. Within that, the cash-futures basis trade was $830bn, close to twice its early-2020 peak, and swap-spread arbitrage $305bn. Financing came from around $3tn of repo borrowing. About 90% of the exposure sits in the 50 largest funds, up from 84%. The basis trade, for readers who have not met it: futures on Treasuries usually trade slightly rich to the underlying bonds, because asset managers like the leverage futures give them. A hedge fund buys the bond, sells the future, and collects the difference. The gap is a few basis points, so the only way to make it a business is to do it very large and borrow nearly all of the money in overnight repo — commonly $100m of position on $10m of capital. The risk is not credit. It is funding. If repo haircuts rise or the spread moves the wrong way, the position must shrink, and shrinking means selling cash Treasuries into a market that is already falling. One piece of evidence cuts the other way, and we count it. In April and May 2025 roughly $100bn of a $295bn swap-spread book came off while basis positions stayed stable and repo funding held. That mark was tested and did not break.

Crypto and TradFi contagion26 Aughigh

The Bank of England is given a growth objective

A run on a large sterling stablecoin would be a forced sale of gilts by an issuer with no lender of last resort.

The UK Treasury will hand the Bank of England a statutory secondary objective to support innovation in payment systems and digital money, the FT reports. Financial stability remains the primary objective. City minister Lucy Rigby framed it as ensuring "the UK remains a global leader in financial services". A secondary objective is not decoration. It is a thing a regulator must be able to show it weighed when it writes a rule, and it is enforceable in the sense that industry can point to it when a rule is consulted on. The rule in question is already drafted: in June the Bank set out its stablecoin regime for sterling-pegged tokens, and it had planned holding limits — caps on how many tokens any one person or firm could own. Crypto firms objected. The Bank has been criticised for being too conservative. Now it acquires a legal duty to promote the thing it was being conservative about. Why the caps existed matters. A stablecoin is a claim that behaves like a deposit — redeemable at par, on demand — without deposit insurance or capital behind it. What sits behind it is a portfolio of short-dated assets. If holders redeem en masse, the issuer sells that portfolio, and if the portfolio is large enough relative to the market it is selling into, the sale moves the price of gilts and bills. That is the money-market-fund problem with a payments interface bolted on, and holding limits were the Bank's way of keeping the portfolio small enough not to matter. This is the migration Crash Lab tracks, in its cleanest form: a deposit-like liability moving out of the regulated, insured, capital-backed banking system into something that looks like money and is not.

The AI capex bubble27 Aughigh

The other side of Nvidia's balance sheet

The AI buildout's credit risk is increasingly concentrated on one supplier's balance sheet, in a line item that carries no promises from the borrower and no assets pledged as security.

Between January and July, the money Nvidia is owed by customers but has not yet collected went from $38.5bn to $63.1bn, a 63% rise (CNBC). In the quarterly filing, "five direct customers" account for 70% of that balance. A year earlier it was three customers and 56%. This is the least glamorous form of vendor financing and the most direct. Nvidia has shipped the chips, booked the revenue, and is waiting to be paid. In the meantime it is an unsecured creditor of whoever holds the hardware. The number matters less as a level than as a ratio to everything else Nvidia has committed: alongside $279bn of supply and capacity commitments, $29bn of cloud service agreements and the $105bn Ohio lease guarantee (totaling $366bn on Goldman's tally in the same report), Nvidia is now simultaneously the seller, the financier and the backstop for a customer base of roughly five names. The forward path is the part worth watching. Bank of America projects receivables at $113bn in January 2028 and $147bn in 2029; Morgan Stanley gets to $171bn. Those are projections, not disclosures, and both banks are positive on the stock. But if they are roughly right, the credit Nvidia extends to its customers grows faster than most of those customers' operating cash flow. The question worth holding: what does Nvidia actually recover on a receivable from a smaller cloud company whose only asset is depreciating silicon that Nvidia itself sold it? The filing does not say, because nobody has had to find out.

Bond market dysfunction27 Aughigh

Hyperscaler bonds pay a premium everyone else does not

The broad credit indexes look calm because they average away the sector where the borrowing actually is.

JPMorgan's head of European investment-grade finance, Matthias Reschke, told Bloomberg TV that bonds issued by the giant cloud and AI companies trade about 0.55 percentage points above the interest rate on other US corporate debt of similar quality, a "material premium" (Bloomberg). Meta, Alphabet, Oracle, Microsoft and Amazon have raised more than $220bn this year across seven currencies. "We don't doubt that we will be able to place it," Reschke said. "It's all a question of price." This is the most useful single number in today's evidence, because it dissolves a contradiction we keep having to explain. The extra interest that safer corporate borrowers pay over the government sits at 0.80 percentage points, and for the riskiest borrowers it is 2.67 percentage points. Both are among the tightest readings in our trail. On the face of it, the credit market is not worried about anything. But an index is an average, and the average is dominated by companies that are not building data centers. Inside it, the market is charging the AI complex roughly two-thirds more than the going rate for borrowers with equivalent credit ratings. So the credit market is discriminating. That is a healthier picture than pure complacency, and we should say so. It is also how the risk gets distributed: at 0.55 percentage points of extra interest the bonds find buyers, those buyers are compensated for the risk, and the exposure spreads out across pension funds and insurers rather than sitting in a few trading firms' inventories. What would worry us is the premium widening while borrowing continues at the same pace anyway, a sign that the borrowers are, in Treasury Secretary Bessent's phrase from last week, "almost yield agnostic," and will pay whatever it takes.

Hidden leverage and shadow banking27 Augmedium

Jane Street's first losing month in a decade

The firms that keep markets running during a panic are now making large directional bets that nobody sees until after they go wrong.

In July, a hedge fund run by 24-year-old Leopold Aschenbrenner was forced to sell its positions after heavy losses in AI stocks. Jane Street, one of the largest trading firms in the United States, had invested heavily in that fund, Situational Awareness, and held many of the same names: Sandisk, Nebius, Bloom Energy. The result, per the FT, was a $15bn loss in trading revenues and Jane Street's first losing month in a decade (FT). The wording leaves open whether that is money actually lost or a shortfall against expected revenue; Jane Street does not publish quarterly accounts, so the distinction cannot be settled from outside. What can be settled is the shape of the thing. Jane Street began as a firm that made money by exploiting tiny price differences over fractions of a second. It is now one of the largest financial institutions in the country and, on the FT's account, takes large one-way bets over long horizons. Larry Tabb of Bloomberg Intelligence puts the mechanism plainly: the revenue is only possible if the holding period extends along with the size, and "the longer your holding period, the more chance that something you can't forecast occurs." In an August letter to employees, the firm admitted it was not protected against some of the moves that undid its July. This is the migration we have been tracking, visible in one firm. A non-bank that provides the buying and selling that keeps equities, options and government bonds trading smoothly has taken a nine-figure-plus hit on a concentrated bet on AI, and the first the outside world hears of it is a feature story seven weeks later. It also cuts the other way: the system took the shock without anyone scrambling for cash. The losses were tested and held.

Crypto and TradFi contagion27 Aughigh

The obituary and the price disagree

A rising share price is not evidence a financing model works. The premium to the value of the coins held is, and that is where the model is broken.

The FT's analysis of the 50 companies holding the most bitcoin finds their combined market value has fallen from $150bn in July 2025 to $67bn. Shares in 43 of the 50 trade below where they were before the bitcoin pivot was announced; 35 have halved or worse. Many are selling tokens and going back to their original businesses. "It was always doomed," says Adam Morgan McCarthy of LO:TECH. "I don't think we see new ones" (FT). The market this week says something else. Strategy rose 11.6% on Thursday alone, is up 40.8% in twenty days and closed at its period high. Bitmine is up 41.7% over the same span, SharpLink 38.6%, Coinbase 16.7%, all four sitting at or near their highs. Bitcoin itself is up 23.7% in twenty days to $80,247. Both things are true, and the reconciliation is the interesting part. The treasury-company model works by issuing shares above the value of the coins held and using the proceeds to buy more coins. It requires a persistent premium. A rally in the underlying coin lifts the share price without restoring the premium, and the FT's structural finding (dozens of these companies selling tokens rather than buying them) is evidence the issuance machine is not running. So the unwinding we noted this week is genuine, and the price move is a bet on the currency losing value rather than a revival of the model. If MSCI's consultation on excluding non-operating companies goes through, the buying pressure from index funds disappears too.

Household credit27 Augmedium

Subprime auto sets a July record; everything else improves

The household number that policymakers watch is being held up by creditworthy borrowers while the riskiest tier prints records, and the two are funded by different people.

Fitch's index of subprime car loans at least 60 days late read 6.13% in July, the highest July in a series running back to 1994. The equivalent rate for creditworthy borrowers was 0.49%. That is a ratio of more than twelve to one. The New York Fed's second-quarter data has one in eighteen of all US car loans at least 90 days behind (5.5%), the second-highest reading in more than twenty years of tracking. Everything else in the household data got better. FICO's autumn report has mortgage payments 30 days late down to 1.35% from 1.42%, car payments 30 days late down about 0.05 percentage points to 2.64%, and bank card payments 30 days late roughly flat around 2.7%. Two issues ago we lowered the dry-tinder score a point on exactly this kind of improvement, and we stand by it. The two facts are not in tension; they describe a split. Household credit in the aggregate is fine because the aggregate is dominated by creditworthy borrowers with fixed-rate mortgages. The deterioration is entirely at the bottom, in the borrowers who finance cars at rates that only work if almost all of them pay. The publicly traded lenders are priced as though this is under control. Credit Acceptance, which lends at the very bottom of the subprime market, is up 7.8% in twenty days and sits 2.9% off its high. Ally is flat. Oportun is up 32% in twenty days. That is defensible (deep subprime loan pools are priced for double-digit losses and the loans are largely bundled and sold off to investors) but it means the stock market is not the place to look for the first sign of trouble here.

Fed, Treasury and policy27 Aughigh

Spending the emergency cash to buy back bonds

A fiscal agency using its cash buffer to manage borrowing costs is doing the central bank's job without the central bank's authority, and in the opposite direction.

The Treasury Department, which borrows the money the government spends, has doubled its buybacks of government bonds that do not come due for twenty or thirty years to at least $28bn over the next two months, announced outside the regular schedule. CNBC reports that two senior officials say the purchases could be funded from the Treasury General Account, the government's checking account at the Federal Reserve, America's central bank, currently holding around $954bn. The TGA exists as a buffer. It has been kept deliberately large so the government can keep paying its bills through shutdowns. Running it down to buy back thirty-year bonds converts a cash reserve into a bet on interest rates, and it does so at a moment when the interest rate the government pays to borrow for thirty years has been near 5.3%, its highest since 2007. Geng Ngarmboonanant, a former deputy chief of staff to Janet Yellen and now at JPMorgan, notes that within 48 hours of the buyback announcement the ten-year rate had rebounded above 4.70%, higher than before the intervention (NYT). Bloomberg's opinion desk is blunter, and Stanley Druckenmiller, who mentored both Treasury Secretary Bessent and Fed chair Warsh, used a Wall Street Journal column to say the effort is doomed and the only question is how much gets spent before conceding. The structural problem is the direction of travel. If buybacks work, they lower borrowing costs across the economy, which stimulates spending. Three members of the Federal Reserve's rate-setting committee (Hammack, Kashkari and Logan) voted against holding rates steady in July, preferring a 0.25 percentage point increase. Treasury is easing while a bloc of the Fed wants to tighten. The ten-year rate has since settled back to 4.66%, so the acute phase has passed. The institutional question has not.

The AI capex bubble27 Aughigh

Three trillion dollars of promises, none of it counted as debt

The borrowed money behind the AI buildout is disclosed in footnotes but not counted as debt, and it is what suppliers and developers are borrowing against.

Morgan Stanley has added up the commitments that seven companies (Alphabet, Microsoft, Meta, Oracle, Amazon, Nvidia and Broadcom) have made for AI infrastructure but have not yet put on their books. The total comes to roughly $3 trillion. That sits on top of an estimated $770bn of debt and lease obligations that do show up on the books (Axios). The split matters. About $1.1tn is payments on data-center leases that have not yet started. Under accounting rules, a lease only becomes a liability when it begins; before that it is a footnote. The other $1.7tn is purchase commitments for chips, memory and networking gear, a promise to buy, not a borrowing. Alphabet's purchase commitments alone reached $707bn in the most recent quarter, against $72.5bn for all of 2025. Nvidia's own supplier commitments, disclosed Wednesday, more than doubled to $279bn from $119bn in a single quarter (WSJ). None of this is hidden. It is all disclosed. But it is not counted where borrowed money is normally counted, and the researchers make the point that matters most: suppliers and data-center developers borrow against those long-dated commitments. An Alphabet promise becomes something valuable that a developer pledges to a lender, which becomes a bond sold to someone reaching for a little more interest. The same obligation is considered safe at one end of the chain and shaky at the other. What has to stay true is simple and unprotected: the tenants must want the capacity, at the agreed price, for the agreed term. If they do, these are ordinary supply contracts. If they do not, the commitment flips from a strategic advantage into a liability. The WSJ's comparison to Cisco in 2001 is the right one. This is a measurement improvement rather than fresh fuel. The commitments were signed weeks and months ago. We simply now have a number for the whole stack, and it is bigger than our running tally.

What would change our mind

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

A cloud-startup bond deal in the current pipeline (Applied Digital's roughly $3.5bn offering is the cleanest test) being pulled or repriced at a materially higher interest rate now that Nvidia's rent-back guarantee is off the table.

Would move the number

A semi-liquid private credit fund suspending redemptions outright rather than limiting them proportionally, or a top-ten private lending fund marking its value down more than 5% in a single quarter.

Would move the number

The thirty-year interest rate trading back above 5.34% after the enlarged buybacks actually begin on September 9. That would tell us the Treasury Department has spent the credibility and bought nothing.

Would move the number

The extra interest on shaky corporate debt passing 3.50 percentage points above government bonds, or the same measure for safer corporate debt passing 1.00 percentage point. At 2.67 and 0.80 percentage points today, the credit market is pricing none of what the reporting describes, and we would rather see those numbers move than write about the gap again.

Would move the number