Bad loans at a nine-year high, long bonds at 2007 yields

Archived reading, published Wed, 19 Aug 2026 06:44:32 UTC (7 days ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

+2 since the last reading

Fragility88

how much tinder is stacked up — moves slowly

Ignition46

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 37 sources and rewrites this page.

Why it moved: Up two, from 65 to 67. Median non-accruals at the 20 largest listed BDCs jumped from 2.0% to 2.8% of cost in a single quarter — the worst since 2017 — while the 30-year Treasury yield reached 5.33%, its highest since 2007, after two tailed auctions. Neither shows up in credit spreads, which sit at 270bp and haven't moved on the week; that is why it is two points and not ten.

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.

Private credit and BDCshigh

Two point eight per cent

Non-accruals are the earliest hard number in a market that otherwise reports its own valuations, and they moved a lot in one quarter.

The median share of loans on non-accrual at the 20 largest publicly traded business development companies hit 2.8% of cost in the second quarter, up from 2.0% at the end of March — the highest since 2017, when the industry was digesting the oil crash. That is the FT's analysis of data from Solve. (FT) Non-accrual is a specific thing, not a mood. It means the fund has stopped booking interest income on a loan because the borrower has stopped paying, or is expected to. It flows straight through to net investment income, which is what BDCs pay dividends out of. An 80 basis point move in the median in one quarter is fast. FS KKR reported 7.1% of its book troubled. David Golub of Golub Capital told investors there was "elevated credit stress" and added: "We're in a credit cycle. Others denied it for a while. I don't think there's a lot of denial any more." Fitch says private credit defaults hit a record in July. PitchBook LCD data show the biggest listed BDCs shrank again — repayments and sales outran new commitments at vehicles run by KKR, Blue Owl and Apollo's MidCap Financial. Now look at the prices. Ares Capital is up 3.7% over 20 days, Blackstone Secured Lending 3.8%, FS KKR 8.2%. Blue Owl Capital Corp trades 4.2% off its high. High-yield spreads are 270bp, unchanged on the week. So the reporting says a credit cycle has arrived and the market says it hasn't. The reason both can be true for a while: private loans are marked by the manager, not by a screen. A non-accrual is a disclosure event. A markdown is a decision.

The AI capex bubblehigh

Nvidia guarantees the rent

The financeability of the largest AI buildout now rests on one chipmaker's willingness to stand behind its customers' rent.

Nvidia has agreed to guarantee up to $105bn of OpenAI's lease payments on a data-centre campus in Pike County, Ohio being built by SoftBank-backed SB Energy. It is also putting $1.5bn of equity into SB Energy. OpenAI is contracting for roughly 8 gigawatts of capacity; the first 800MW is due in 2028. Nvidia's backing phases in from 2028 to 2030 and runs 20 years, and the $105bn covers only the first 4.25GW. It may take on the second phase too. (Bloomberg) Here is what the guarantee does. Whoever lends against that campus is lending against a stream of lease payments. Without the guarantee, that stream is OpenAI credit — a company with enormous revenue growth and no rating anyone would lend $100bn against for twenty years. With it, the stream is substantially Nvidia credit. Same building, same tenant, radically cheaper money. Nvidia's own numbers make the loop visible: OpenAI's chip purchases could represent about $600bn of Nvidia revenue through 2030, and Nvidia has separately signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR to mobilise more than $500bn of third-party compute financing, backstopping up to $125bn of it itself. Jensen Huang says this is not circular financing. Nvidia stresses it is not guaranteeing the whole site or all of OpenAI's obligations. What we do not know is the accounting. A guarantee of this type would normally sit as a contingent liability in the footnotes until a payment is missed. We have not seen the trigger, the coverage waterfall, or what Nvidia would recover if it paid. Bond desks are already trying to size roughly $70bn of what Bloomberg calls "shadow AI backstops." This adds a number an order of magnitude larger.

Hidden leverage and shadow bankinghigh

$1.65 trillion in the footnotes

The largest single addition of corporate leverage in a decade is being recorded in places that leverage ratios do not capture.

Nikkei went through the filings of Alphabet, Microsoft, Amazon, Meta and Oracle and totted up the obligations that do not appear on their balance sheets. The figure is $1.65tn — more than the $1.35tn of debt that does appear, and roughly eight times the level of four years ago. Meta's off-balance-sheet total is about $420bn, nearly triple its recorded debt. Oracle's is $273.3bn as of end-May, a more-than-thirtyfold increase in four years. (Nikkei) The mechanism is boring and that is the point. Under lease accounting, an operating lease becomes a balance-sheet liability when the lease commences. A signed twenty-year lease on a data centre that is still a field in Ohio is a commitment disclosed in a footnote. GPUs and servers under long-term purchase contracts that have not yet been delivered are the same. All of it is perfectly legitimate. So the period of maximum commitment — when the contracts are signed, the sites are being poured, and nothing has been delivered — is also the period of minimum balance-sheet disclosure. Leverage ratios during the buildout look like the leverage ratios of the old capital-light software businesses. They are not. The structures reinforce it. Meta's $27bn Hyperion financing sits in a special purpose vehicle backed by a twenty-year Meta lease promise, keeping the debt off Meta's books. Goldman estimates roughly $1.5tn of comparable leasing commitments across the sector, about $1tn of it invisible in financial statements. (Telegraph) Meta's shares are down 15.5% in twenty days. That may be the market doing the footnote arithmetic on its own.

Bond market dysfunctionhigh

Every long end at once

Term premium is rising everywhere at once, and the marginal holder of US duration is a repo-financed relative-value trade rather than a real-money investor.

The 30-year US Treasury yield touched 5.33% on Tuesday, the highest since 2007, having been under 5% at the start of last month. It did not happen alone. The 30-year Bund hit 3.78%, its highest since 2011. French 30-year yields reached 4.9%, the highest since 2008. UK 30-year gilts hit 5.86%. Japanese 30-year yields went to 4.16%, close to a record, and the 10-year JGB to 2.93% — its highest since 1996. (FT) Auctions are where you see whether this is repricing or indigestion. A $25bn 30-year sale cleared at 5.216%, the highest auction yield since 2001. A $42bn 10-year cleared at 4.683%, the highest since 2007. Both tailed. Three things are pushing. Brent above $90 with the Hormuz truce expiring, so an inflation premium. US debt near $40tn with no appetite to address it, so a fiscal premium. And, newly, AI. Investment-grade issuers have sold nearly $1.5tn of bonds this year, up 36%. Nomura reckons the roughly $200bn borrowed by the biggest tech companies alone equals about 25% of the Treasury's net issuance of notes and bonds to private investors — five times its share in 2025. (Bloomberg) What sits underneath is the part that worries us. Federal Reserve research puts the cash-futures Treasury basis trade at roughly $830bn as of September 2025 — about double the early-2020 peak, and around 35% of hedge funds' long Treasury exposure, financed almost entirely in repo. That is not a buyer of duration. It is a levered arbitrage that requires repo to stay cheap and margin to stay stable. The last time it broke, in March 2020, the Fed bought $1trn of Treasuries in three weeks.

The AI capex bubblemedium

The power trade cracks, the credit doesn't

The equity market has started to discount the AI buildout while the credit market is still funding it at investment-grade terms.

Tuesday was ugly for everything that sells shovels to the AI buildout. Talen Energy fell 10.96% in a day and is down 11.3% on the week and 20.6% from its high. Vistra fell 3.85% and is down 13.4% over twenty days. CoreWeave fell 12.07%. Applied Digital fell 8.48%. Core Scientific fell 6.4% and is 21.5% lower over twenty days. IREN fell 6.5%. Nebius fell 7.65% — after rising 28.7% in the preceding five sessions. The Nasdaq fell 1.33% and sits 3% below its June high. So this is not a market selloff. It is a specific repricing of the companies whose value is a levered claim on the buildout proceeding at the assumed pace: the merchant power generators signing long-dated supply contracts, the ex-bitcoin miners converting sheds into GPU halls, the neoclouds borrowing against chips. Now the other side of the ledger. High-yield spreads are 270bp, flat on the week. Investment grade is 81bp, three basis points wider over twenty days. VIX is 15.84. Blackstone-backed QTS priced a $3.9bn data-centre bond tied to Microsoft at about 7.63% — a junk-like yield on a Baa3/BBB- rating — and drew $10bn of indications, letting it upsize the deal by $1bn. (Bloomberg) Our reading, and it is inference: equity holders reprice every afternoon, and credit holders do not have to reprice until a coupon is missed or a lease is broken, which on twenty-year data-centre paper is years away. Equity is the fast instrument. That does not make it right. But when the two disagree this sharply about the same cash flows, one of them is going to move a long way.

Crypto and TradFi contagionhigh

The flywheel spins backwards

A levered vehicle whose only business is holding an asset has become a structural seller of it, which is how these things end rather than how they begin.

Strategy sold $333.7m of common stock last week — not to buy bitcoin, but to repurchase preferred shares and build a cash reserve. Over five weeks it has sold about $2.1bn of common, bought back about $347m of its STRC preferred, and sold $213.3m of bitcoin. It has not bought a coin since mid-June. The cash reserve is now $4.8bn against roughly $59bn of bitcoin. (Bloomberg) The original machine worked like this. Issue equity above the market value of the bitcoin per share; use the proceeds to buy bitcoin; bitcoin per share rises; the premium is justified; issue more equity. It is a perpetual motion device that runs on the premium and nothing else. Remove the premium and every arrow reverses. Issuing equity below NAV now destroys bitcoin per share, so the sensible thing to do with the proceeds is retire the liability that has a hard cash coupon — the preferred — rather than buy more of the asset that has none. When the equity market won't take the paper, the coins go instead. That is a deleveraging, executed in public, one Monday filing at a time. The prices tell the same story more precisely than either number alone. Bitcoin is at $64,124, down 0.9% over twenty days — essentially flat. Strategy is down 9.2% over the same period and fell 5.3% on Tuesday. Over a year the stock is off about 73% while bitcoin is off 45%. The holding company is being marked down relative to the thing it holds. The wider point: the largest marginal buyer of bitcoin for six years is now, in modest size, a seller. Bitcoin's 30-day realised volatility has fallen to a three-year low. Nothing here has broken. Something here has stopped.

Signs of the times

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

Anthropic's backers expect the five-year-old company to float in October at $2tn or more, which would be the largest IPO ever. One investor's arithmetic: "If Anthropic is growing 800 per cent a year, you'd think at the incredibly low end they would trade at 30 times revenue. That would make them a $3tn company."

Financial Times

Taiwan's economy grew 12.9% in the second quarter on semiconductor demand, filling the treasury so thoroughly that President Lai will hand every eligible resident NT$10,000 ($314) next year. It is the second such payout in two years; the stock market is up almost 60% in 2026.

Bloomberg

Mark Walter spent the summer negotiating a multibillion-dollar loan from Apollo secured against his stake in the Los Angeles Lakers, then sold the stake outright for $12.5bn instead. Insurers he controls have disclosed more than $20bn of loans to related parties whose connections they had not disclosed, after receiving federal subpoenas.

Financial Times

Leopold Aschenbrenner's $35bn July loss on leveraged AI stock bets now sits at the top of the all-time leaderboard of trading losses — ahead of every rogue trader, bank blowup and mishedged corporate on the list. Most of the top twenty-five were made by institutions, not funds.

Financial Times

The rumour mill

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

Partly true

“Free cash flow at the largest hyperscalers turns broadly negative by 2027, forcing them to fund data centres with debt.”

Reuters, using LSEG consensus, has Microsoft, Alphabet, Amazon, Meta and Oracle spending more on capex than they generate in free cash flow by 2027; Bank of America's wider AI-exposed aggregate swings from roughly +$180bn in 2025 to about -$64bn in 2026 and -$144bn in 2027. It is not true of all four today — Alphabet and Microsoft are the exceptions — but the direction is well documented.

Claimed by Meet Kevin

Confirmed

“Tokyo began intervening to support the yen on 30 July and the United States joined the next day — the first US participation since 1998.”

Japan's finance ministry confirmed coordinated yen-buying with the US on 31 July, and the Wall Street Journal and Reuters both date the previous joint operation to 1998. Estimates put the two days at roughly $85–88bn. Japan has since repurposed a COVID-era Fed repo facility to borrow dollars against its $1trn-plus of Treasuries, which is the part worth watching.

Claimed by Mark Moss

Confirmed

“A business financially tied to the Trump family works with a platform distributing Chinese AI models that the US government has restricted, and the family earns revenue from it.”

Reuters reports World Liberty Financial — 38% Trump-family owned — collaborating with Hong Kong-based WorldClaw, which offers 43 of 90 models from Chinese firms flagged by the Pentagon and accepts World Liberty's USD1 stablecoin as payment. Separately, the OCC has granted World Liberty preliminary approval for a trust bank licence, letting it issue USD1 — the fourth-largest stablecoin at about $4bn — directly.

Claimed by Gregory Mannarino

Earlier developments

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

Private credit and BDCshigh

The lender of last resort said no

Private credit's function in the system is to absorb debt nobody else will hold; when it starts declining, the marginal borrower has nowhere left to go.

Sophos, the cybersecurity firm Thoma Bravo bought in 2020, needs to refinance or extend more than $2bn of loans as soon as next month. It spent months trying to line up private credit backing. That effort faltered, and the company is now going back to its existing leveraged-loan holders, potentially offering a higher coupon, amortisation payments and a tighter covenant package to get them to stay (Bloomberg). Understand what that inverts. The entire pitch of private credit is that it is the buyer of last resort for corporate debt: it will lend where the syndicated market won't, on terms the syndicated market won't write, at a price. When a $2bn software LBO shops itself to private credit and comes back empty, the order of operations has reversed. The paper is going back to the public loan market — the one place it gets marked every day. Two further details matter. First, Thoma Bravo has told lenders it will not inject fresh capital, rebuffing a request from investors. The sponsor is declining to defend its equity, which means any loss stays with the creditors. Second, the stated worry is AI: lenders are now being asked to underwrite whether a security software product survives a technology shift, over a loan that outlives the forecast. This follows a pattern. In July, Thoma Bravo had to sweeten a $5bn Proofpoint refinancing after a lender revolt. In June, Blackstone-led creditors took control of its survey company Medallia. One firm's portfolio is not a cycle. But this is what a credit cycle looks like at the beginning: not defaults, refusals.

The AI capex bubblehigh

A $14bn building with $427m of cover

The credit quality of AI data-centre debt rests entirely on the tenant's lease, and the insurance market has now declined to stand behind the building itself.

Meta and BlackRock's Sopaipilla campus in El Paso is a one-gigawatt data centre costing about $14bn. On the advice of Marsh, the world's largest insurance broker, the venture has bought up to $427m of all-risk property cover during construction and $450m once it is running, rising 2% a year, for a premium of roughly $5m. Terrorism cover runs to $645m. Rent-abatement cover for construction delay is $218m. General liability is $50m per event and $50m in aggregate, for about $1m a year. The project is not insured against a total loss (FT). So roughly 3% of the asset value is covered. Lenders to smaller data centres normally demand full insurance; as deal sizes have grown, partial coverage has become acceptable, because the insurance market simply will not build that much aggregate exposure to a single site. That is a capacity constraint, not carelessness. But the constraint has to land somewhere, and where it lands is the debt. What we do not know is how the lease treats a casualty. If Meta keeps paying rent on a destroyed building, the bondholders are fine and the equity absorbs it. If the lease abates on total loss, the credit evaporates with the concrete. That single clause determines whether this is a footnote or a hole, and it is not public. This is the shape of the whole AI financing complex in miniature: the tenant's covenant is doing all the work, and the physical asset underneath it is thinly protected. Insurance is one of the few places where a genuinely independent party has to price the tail. It looked at this and said: not for $14bn, not in one spot.

Fed, Treasury and policymedium

Skin in the game, waived

The AI buildout is being financed through exactly the channels — SPVs, vendor guarantees, ABS without retention — that were designed after 2008 to be harder to use.

The SEC has exempted some data-centre owners from crisis-era rules on asset-backed securities, removing certain disclosure and risk-retention obligations. The change came after Latham & Watkins, a major adviser to the AI sector, asked the regulator to clarify whether data centres should be regulated as tightly as other ABS. The SEC's answer: data centres are not "financial assets" like mortgages and car loans (Telegraph). Risk retention is the Dodd-Frank rule requiring whoever assembles a securitisation to keep a slice of it, so that the person choosing the collateral eats some of the loss. It exists because in 2006 nobody who picked the loans kept any. Removing it from a market JPMorgan expects to run at $30–40bn a year, up from about $27bn in 2025, is not by itself dangerous — the market is small next to Morgan Stanley's $250–300bn estimate for total AI debt issuance in 2026. What it signals is the regulatory posture heading into the peak of the buildout. Cambridge's Raghavendra Rau put it plainly: "The SEC has been replacing regulatory constraints with a reliance on market discipline." Stack it with the rest of the week. Nvidia has signed memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build compute-financing platforms targeting more than $500bn of third-party capital, with Jensen Huang saying Nvidia could backstop up to $125bn, or a quarter of potential deals (CNBC). Meta's $27bn Hyperion financing sits in an SPV, off its balance sheet, behind a 20-year lease promise. Every one of these structures does the same thing: it moves the exposure to a place where no regulated entity has to hold capital against it, and no quarterly mark reveals it.

Hidden leverage and shadow bankingmedium

The market maker was a hedge fund

Market-making capacity and directional risk have merged inside firms that no regulator capitalises, which means liquidity disappears exactly when it is needed.

Jane Street lost $15bn in July — its first losing month in a decade — on directional bets, according to Bloomberg. The firm has still generated more than $40bn of net trading revenue this year, surpassing all of 2025. In its latest debt offering it described its strategy as evolving to include longer-term positions, in line with a hedge fund (Bloomberg). The FT, ranking it against history's biggest trading losses, put the July figure at $35bn and attributed it to Leopold Aschenbrenner's leveraged AI stock bets. The two numbers are not reconcilable from public information — one may be a strategy's gross loss and the other the firm's net monthly P&L. We do not know. Either way it belongs on the all-time list. Why this matters beyond one firm: Jane Street is a core liquidity provider in ETFs and options. Its balance sheet is what stands between a redemption order and a price. If that same balance sheet is also running leveraged directional exposure to AI equities, then in a selloff the entity you need to make prices is the entity taking losses. Banks are capitalised and supervised precisely to prevent that overlap. Jane Street is not a bank, so nobody sees the overlap until a month like July. It sits alongside a bigger number. Federal Reserve research puts the cash-futures Treasury basis trade at roughly $830bn as of September 2025 — about double the 2020 peak, some 35% of hedge funds' long Treasury exposure, financed almost entirely in repo. The liquidity of the two largest markets on earth now depends on leveraged non-banks whose positions are disclosed after the fact, if at all.

Bond market dysfunctionhigh

Japan's insurers are sitting on $194bn

If Japanese life insurers become sellers of duration rather than buyers, the global long end loses its most reliable price-insensitive bid.

Unrealised losses on domestic bonds at Japan's major life insurers reached ¥30.86tn ($194bn) at the end of June, up 60% year on year. For the first time they exceed the same insurers' unrealised gains on domestic stocks, at ¥30.03tn (Nikkei). The cause is straightforward: 30-year JGB yields reached the 3.9% range by end-June, about 2.7 percentage points above July 2023. The standard reassurance is that unrealised losses vanish if you hold to maturity, and for a life insurer matching 30-year liabilities that is usually right. Three things break it. First, accounting: if a bond's market value falls 50% below cost, impairment can be forced. Nippon Life booked ¥44bn of impairments in April–June; Meiji Yasuda booked ¥25.3bn. Bonds bought in the late-2010s at near-zero yields are now reaching that threshold. Second, duration mismatch. If assets are longer than liabilities, rising rates cut net assets, and the fix is to sell bonds — converting paper losses into real ones. Third, and most important, policyholders. Sony Life's lapse and surrender rate rose 0.2 points year on year to 1.4% in April–June, partly because yen weakness prompted customers to cancel foreign-currency policies. A surrender is a withdrawal. Enough of them and a hold-to-maturity book becomes a sell-to-meet-redemptions book. That is the same arithmetic that killed Silicon Valley Bank in 2023, running in slow motion on a much larger balance sheet. Japanese lifers are structural buyers of ultra-long JGBs and, historically, of foreign bonds including Treasuries. The 10-year JGB is at 2.93%, its highest since 1996. Watch the lapse rate, not the yield.

The AI capex bubblehigh

Investment grade, junk price

The bond market is now pricing top-rated data-centre debt like junk while the ratings say otherwise, and the equity of everyone standing behind the lease is falling.

Blackstone-backed QTS is selling roughly $3.9bn of five-year notes to fund a Microsoft-linked data centre, a deal nicknamed Project Odyssey. Initial price talk was a yield of about 7.63% — junk territory — on paper expected to be rated Baa3 by Moody's and BBB− by Fitch. The pricing worked: about $10bn of indications came in, and the deal was upsized by roughly $1bn (Bloomberg). Do the arithmetic. With two-year Treasuries at 4.19% and tens at 4.72%, a 7.63% five-year is something on the order of 300 basis points of spread. The investment-grade index option-adjusted spread is 81bp. This deal is being marketed at roughly four times the average — with an investment-grade rating attached. (That comparison is our arithmetic, not the bank's.) The ratings agencies are looking at a Microsoft-tied lease. The buyers are looking at concentration, obsolescence and a queue of similar deals, and are demanding to be paid for all three. When the rating and the clearing price disagree by this much, believe the price. The equity market went further on the same day. Talen Energy fell 11.0%, its second double-digit week; Vistra lost 3.9%; CoreWeave 12.1%; Applied Digital 8.5%; Core Scientific 6.4%, now down 21.5% in twenty days; IREN 6.5%; Nebius 7.7% after a 28.7% week. The Nasdaq 100 fell 1.3%. That split is the interesting part. Debt investors will fund the buildout at 7.6%, because they hold a lease from one of the most creditworthy companies alive. Equity investors in the power suppliers and the neoclouds own the residual — and they are marking it down hard.

Household credithigh

The same cars, pledged twice

Asset-backed lending is only as good as the verification of the assets, and the last two frauds both exploited the fact that nobody performs it.

The SEC has sued Daniel Chu, founder of subprime auto lender Tricolor Holdings, along with former CFO Jerome Kollar and former finance director Ameryn Seibold, over the firm's $1.9bn collapse. The regulator says they ran a multi-year scheme to prop up a failing business: false claims about financial health, faked documents, and pledging the same vehicles as collateral for multiple loans. Kollar and Seibold have pleaded guilty and are cooperating (Bloomberg). The SEC's enforcement director, David Woodcock, chose his words carefully: the defendants "defrauded investors based on bogus collateral and violated the integrity of our private credit markets." Double pledging is the oldest failure in secured lending, and it works for a simple reason: nobody physically checks. Warehouse lenders and ABS investors underwrite a servicer's data tape. The tape says there are 40,000 cars with 40,000 VINs. Verifying that is expensive, slow and — in a market competing on speed of execution — the thing that loses you the deal. Note the family resemblance to First Brands, where the collateral was factored invoices rather than cars. Two large blowups in a year, both in the part of the structure that was supposed to be safest: the asset behind the loan. The environment makes this surface. Subprime auto ABS 60-day-plus delinquency ran at 5.67% in June per Fitch, and KBRA put late-stage non-prime delinquency at 9.46% in July. When cash flow is fine, nobody counts the cars. When it isn't, someone does. Private credit's core defence is that it lends secured against real assets. The past year's two worst cases were both about whether the assets were real.

The AI capex bubblehigh

Nvidia now underwrites its customer's rent

The largest company in the world is now a contingent creditor to its own customers, and the exposure is concentrated in precisely the scenario where its own earnings fail.

Nvidia has agreed to guarantee up to $105bn of OpenAI's lease payments on a data-centre campus in Pike County, Ohio, being built by SoftBank-backed SB Energy. It is also putting $1.5bn of equity into SB Energy. OpenAI is contracting for up to roughly 8 gigawatts of capacity, with the first 800MW due online in 2028; Nvidia's backing phases in from 2028 to 2030 and runs for twenty years. The $105bn covers only the first 4.25GW. Nvidia may take on more to secure the second 3.75GW phase (Bloomberg). Here is the plumbing. A lease guarantee is credit exposure that does not look like debt. The people lending against that campus are not underwriting OpenAI, a company with no earnings history at this scale; they are underwriting Nvidia's promise to pay if OpenAI stops. That converts project risk into something that prices near investment grade, which is why the buildout can be financed at all. The problem is correlation. The guarantee is worth something only in states of the world where Nvidia is healthy — and Nvidia is healthy only if AI compute demand holds, which is the same condition that keeps OpenAI paying. The backstop is strongest exactly when it is least needed. What we do not know matters. Nvidia says it is not guaranteeing the whole site or all of OpenAI's lease obligations, and has not said what triggers payment, where it sits in the capital structure, or whether this is recognised as a liability or disclosed in a footnote. Jensen Huang denies this is circular financing; OpenAI purchases could represent about $600bn of Nvidia revenue through 2030 (The Economist). Separately, Nvidia has signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR to mobilise more than $500bn of third-party compute financing, with Huang saying Nvidia could backstop up to $125bn of it. Nvidia shares fell 2.4% on the day and sit 2.5% off their high. The market is not charging much for any of this.

Hidden leverage and shadow bankingmedium

The footnotes are bigger than the balance sheets

The most creditworthy borrowers in the world have arranged their AI buildout so that the debt appears on schedule rather than at signing, which makes leverage look lower today than it is contracted to be.

Nikkei went through the filings of Alphabet, Microsoft, Amazon, Meta and Oracle and put their off-balance-sheet obligations at $1.65tn — more than the roughly $1.35tn of debt actually on their balance sheets, and eight times the figure of about four years ago. Meta's is around $420bn, nearly triple its recorded debt. Oracle's reached $273.3bn at the end of May, a more-than-thirtyfold rise in four years (Nikkei). None of this is improper. Under lease accounting, a data-centre lease does not become a liability until the asset is available for use, and GPUs and servers ordered under long-term purchase commitments are not liabilities until delivered. So the obligation exists, is legally binding, is disclosed — and lives in the notes rather than the balance sheet. The mechanism worth understanding is the timing. These are not contingent risks that may never arrive. When each campus energises and each lease commences, the present value of twenty years of rent lands on the balance sheet as a liability, on a schedule that is already fixed. Leverage at these companies is not a forecast; it is a delivery queue. Moody's flagged ballooning not-yet-commenced leases in February and Morgan Stanley has written it up, per Nikkei. The structures compound the opacity. Meta funded its $27bn Hyperion project through a special purpose vehicle backed by a twenty-year lease promise, keeping the bond off its own balance sheet; Goldman estimates about $1.5tn of similar leasing commitments across the sector, roughly $1tn of it outside the statements. This is the migration Crash Lab keeps tracking — obligations moving from places where they are measured to places where they are disclosed in jargon. One market is noticing: Meta is down 15.5% in twenty days, worse than any other hyperscaler.

Private credit and BDCshigh

Non-accruals up 40 per cent, the managers up 15

The clearest measured deterioration anywhere in the system is in the asset class with the fewest observable prices, and the equity market is still paying up for the managers who hold it.

Loans placed on non-accrual at the twenty largest listed business development companies rose to a median 2.8% of cost in the second quarter, from 2.0% at the end of March — the highest level since 2017, when the industry was working through an oil crash (FT). Fitch says private-credit defaults hit a record in July. FS KKR reported 7.1% of its book troubled. David Golub of Golub Capital told investors: "We're in a credit cycle. Others denied it for a while. I don't think there's a lot of denial any more." A BDC is a listed fund that makes private loans, mostly floating-rate, to mid-sized companies owned by private equity. Non-accrual is the one number in this business that cannot be smoothed: it means the fund has stopped booking interest it is owed. Everything else — the fair value of a loan with no market price — is an estimate the manager makes. So an 80 basis point jump in a quarter is a hard signal inside a soft dataset. The second-order effect is the funds are shrinking. PitchBook LCD found repayments and loan sales outpaced new commitments at vehicles run by KKR, Blue Owl and Apollo's MidCap Financial. A shrinking book with rising non-accruals means less interest income against a fixed dividend, and a smaller base to earn fees on next year. Now the disagreement. Over twenty days the BDCs themselves are up — Ares Capital +3.7%, Blackstone Secured Lending +3.8%, Blue Owl's BDC +4.3%, FS KKR +8.2% — and the managers are up far more: Blue Owl +25.6%, Ares +16.1%, Blackstone +14.4%, Apollo +13.0%. The market is saying deteriorating credit is a fee-collector's inconvenience. The past five days are the first hint of doubt: Apollo −4.6%, Blackstone −4.5%, Blue Owl −4.3%, KKR −3.5%.

Bond market dysfunctionhigh

AI is now issuing a quarter as much as the Treasury

The AI buildout has stopped being an equity story and become a rates story: it is competing with the US government for duration, and the buyer of last resort is levered.

The 30-year Treasury yield reached 5.33% on Tuesday, its highest since 2007, having been below 5% at the start of last month. It is not just a US story: the 30-year Bund hit 3.78%, most since the euro crisis in 2011; the 30-year French yield 4.9%, most since 2008; the 30-year gilt 5.86%; Japan's 30-year 4.16%, close to a record (FT). The 30-year TIPS yield hit 3.09%, highest since 2008 — that is a real yield, so this is not purely an inflation story. The auctions are where the strain shows. Treasury sold $25bn of 30-year bonds at 5.216%, the highest auction yield since 2001 (FT), a day after $42bn of 10-years cleared at 4.683%, the highest since 2007. The new piece is who else is in the queue. Investment-grade companies have sold nearly $1.5tn of bonds this year, up 36%. Nomura reckons the roughly $200bn borrowed by the biggest tech companies alone equals about 25% of the Treasury's net note and bond issuance to private investors — five times the 2025 share (Bloomberg). Crowding out, with the hyperscalers doing the crowding. Who is absorbing it? Increasingly, leverage. Secondary summaries of Fed research put the cash-futures Treasury basis trade at about $830bn as of September 2025, roughly double the 2020 peak and around 35% of hedge funds' long Treasury exposure, financed almost entirely in repo. We cannot verify that directly. If it is close to right, the marginal buyer at these auctions is a repo borrower, and Bank of America's survey shows real-money investors at their lowest bond allocation versus benchmark and highest equity allocation since 2021. One calming note: the rate-vol proxy in our dataset is down 27% over twenty days. Multi-decade high yields, no panic. That is a repricing, not a run.

The dollar, gold and reserve statusmedium

Japan found a way to defend the yen without selling Treasuries

One of the most-cited crash channels — Japan dumping Treasuries to save the yen — has been quietly plumbed around, which lowers ignition risk in the Treasury market and raises it in Japanese funding.

The dollar has fallen to a three-month low, its weakest since 15 May, as traders cut the odds of a Fed hike next month to one-in-three from about 75% in late July; only 36 basis points of further tightening is priced (Bloomberg). Gold is up 7.9% in twenty days to $4,391. Long yields at multi-decade highs, a falling currency, and a rising gold price is the signature of investors demanding more compensation to hold the paper — not of a growth repricing. The more interesting development is plumbing. After the first joint US-Japan yen intervention in 28 years on 31 July, Tokyo is now armed with a repurposed COVID-era Fed repo facility that lets the Bank of Japan borrow dollars against its $1tn-plus of Treasury holdings as collateral, announced by Treasury Secretary Bessent on 3 August. Japan's top currency diplomat called the arrangement "a U.S.-Japan currency union"; local analysts are calling it a new Plaza Accord (Nikkei). Why this matters mechanically: the standard doom loop runs yen weakness → intervention → Japan sells Treasuries to raise dollars → US long yields rise → more pressure. The repo facility cuts that link. Japan can raise dollars against its Treasuries instead of liquidating them. The largest foreign holder of US government debt has been given a way to defend its currency without becoming a forced seller of the thing everyone worries about. It is not free. Repo borrowing has to be rolled, and it converts a solvency-style buffer into a funding position. Futures price a 100% chance of a 25bp BOJ hike by the 29-30 October meeting, which is the market betting Japan will have to raise rates to hold the line anyway.

Household credithigh

Ten quarters of elevated card delinquency, no break

Household credit stress is real, persistent and priced as spread income rather than as risk — which is defensible while employment holds and dangerous if it doesn't.

Pending US home sales fell 2.3% in July to an index of 71.2, the lowest since January and matching the second-worst reading in data going back to 2001. All four regions declined; the West fell 4.7%. The average 30-year mortgage rate is just below the one-year high it set in late July (Bloomberg). The consumer credit data tell a similar story of grinding rather than breaking. New York Fed figures for the second quarter put credit-card balances at $1.263tn, up $21bn on the quarter. The share of card balances flowing into 90-day-plus delinquency was 6.97% — the tenth consecutive quarter above the roughly 6.75% level that marked the onset of the Great Recession. About 12.9% of outstanding card balances are 90 days or more delinquent. Serious student-loan delinquency is 10.6%; auto is 5.5%. Overall household delinquency actually ticked down, to 4.7% from 4.8% (New York Fed). In securitisations, Fitch had subprime auto 60-day-plus delinquency at 5.67% in June and KBRA had non-prime at 9.46% in July. What the mechanism tells you is why this has not become an event. A card lender charging 25% on a performing book earns more from higher rates than it loses to a slow drift in charge-offs, as long as employment holds. That is what the equity market is pricing: Synchrony +12.1% over twenty days, Capital One +7.8%, Sallie Mae +6.7%. Deep subprime auto is the exception, with Credit Acceptance −7.7%. So: two and a half years of Great-Recession-level card delinquency without a recession. This is a chronic condition, not a trigger. It matters mainly as the thing that stops absorbing shocks when one arrives elsewhere.

Crypto and TradFi contagionhigh

Bitcoin is flat; everything built on it isn't

The market is repricing the wrappers and intermediaries rather than the asset, which is what the end of a premium-issuance cycle looks like from the outside.

Bitcoin closed at $64,542, up 0.06% on the day and up 1.0% over twenty days. Over the same twenty days Coinbase is down 16.9%, Galaxy Digital down 19.6%, and Strategy down 9.2% — Strategy fell 5.3% on the day alone. The coin is doing nothing. The businesses attached to it are being marked down hard. The mechanism is worth being precise about, because these are three different claims. Coinbase and Galaxy are volume-and-fee businesses: a flat coin means flat trading, and flat trading compresses earnings regardless of price level. Strategy is something else — a company whose stock was a levered claim on its own ability to issue equity above the value of the bitcoin it held. That premium is what funded the purchases. That machine is now running backwards. In the week to 16 August, Strategy sold $333.7m of common stock and used it to buy back preferred shares and build cash, not to buy bitcoin. It has bought no bitcoin since mid-June. Over five weeks it has sold about $2.1bn of common, repurchased about $347m of its STRC preferred, and sold $213.3m of bitcoin. Its cash reserve is $4.8bn against roughly $58bn of coin. The stock is down about 73% in a year (Bloomberg). Read that as plumbing: common equity is being issued to retire preferred, which means the residual holder is funding the fixed-claim holder's exit. That is a liquidation sequence, conducted in public and in an orderly way, at a pace set by the preferred dividend. Meanwhile the regulated perimeter widens: the OCC has given World Liberty Financial, co-founded by Donald Trump, conditional approval for a national trust charter to issue stablecoins and custody digital assets, with no authority to lend or take deposits (FT).

The AI capex bubblehigh

The $14bn campus that cannot be totalled

The catastrophic tail on the largest new industrial assets in America is being retained by lenders rather than sold to insurers, and it is not priced anywhere visible.

Meta and BlackRock's one-gigawatt data centre in El Paso — internally called Sopaipilla, cost about $14bn — has bought only hundreds of millions of dollars of insurance, on the advice of the broker Marsh, and is not insured against total loss, the FT reports. The specifics are worth reading slowly. All-risk property cover of up to $427mn during construction and $450mn once operating, for a premium of about $5mn a year rising 2% annually. $645mn against terrorism. $218mn of rent-abatement cover if construction runs late. General liability with a limit of $50mn per event and $50mn in aggregate, for roughly $1mn a year. Above those caps, the venture eats the loss. Why it happened is not a scandal, it is arithmetic. Insurers will write data-centre risk, but most are unwilling to build large aggregate exposure to a single site holding tens of billions of dollars of chips. Full cover for a project this size has become, in the FT's words, prohibitively expensive for lenders. Lenders to small data centres still demand full insurance; as tickets grew, partial coverage became acceptable. What we do not know is the part that matters most: how Sopaipilla's debt is structured, and whether lease payments from the tenant survive a catastrophic loss. Meta's $27bn Hyperion financing sat in a special purpose vehicle backed by a 20-year Meta lease promise. If El Paso is built the same way, the uninsured tail belongs to bondholders in an SPV, not to Meta's balance sheet — which is exactly the migration this site has been tracking: risk moved into a structure where nobody has to mark it until something happens.

Bond market dysfunctionhigh

Japan's insurers are sitting on $194bn of losses

The marginal global buyer of very long bonds is becoming a manager of its own losses, which removes demand precisely where supply is exploding.

Unrealised losses on domestic bonds at Japan's major life insurers reached ¥30.86tn ($194bn) at the end of June, up 60% year on year, according to Nikkei's survey of 13 of the 14 largest. That figure now exceeds their ¥30.03tn of unrealised gains on Japanese stocks. The mechanics matter more than the headline. Life insurers buy ultra-long JGBs to match obligations decades out; held to maturity, unrealised losses vanish. Three things break that. First, impairment: if a bond's market value falls 50% below acquisition cost, insurers may have to book the loss. Nippon Life took ¥44bn of impairments in April–June, Meiji Yasuda ¥25.3bn — bonds bought in the late-2010s zero-rate era are now hitting that threshold. Second, duration mismatch: where assets are longer than liabilities, rising rates cut net assets, and the fix is to sell long bonds, converting paper losses into real ones. Third, policyholders. Sony Life's lapse and surrender rate rose 0.2 points to 1.4%, partly because yen weakness prompted cancellations of foreign-currency policies. Surrenders force selling. And yields keep going. Japan's 30-year hit 4.16% on Tuesday, near a record; the 10-year touched 2.93%, highest since 1996, as the global long end sold off — US 30-year at 5.33%, German 30-year at 3.78%, the highest since 2011. The reason this matters outside Tokyo: Japanese insurers have been among the world's most reliable buyers of long-dated duration. An institution managing its own unrealised losses is not shopping for more of them.

Hidden leverage and shadow bankingmedium

The market maker was running a hedge fund

The firms that make prices in a stress are now taking the kind of risk that makes them sellers in one.

Jane Street lost $15bn in July — its first down month in a decade — on directional bets, Bloomberg reports. FT Alphaville puts a related figure at $35bn, the loss attributed to Leopold Aschenbrenner's leveraged AI stock positions, which would top the all-time leaderboard of trading losses. We do not know how the two numbers reconcile; the $15bn looks like the firm-level monthly hit and the $35bn like the strategy's gross damage, but neither outlet spells that out. The interesting part is not the size. It is what the firm told bond investors: in its latest debt offering, Jane Street described its strategy as evolving to include longer-term bets, in line with a hedge fund. It has generated more than $40bn of net trading revenue this year, so it absorbed the loss and is still on pace for a record. Here is the plumbing problem. A market maker of this scale is infrastructure — the price you get on an ETF or an option depends on it standing there. A hedge fund is a risk-taker whose capital shrinks when its bets go wrong. When the same balance sheet does both, liquidity provision becomes correlated with the positions being provided liquidity against. A bank running $15bn of directional risk would be a supervisory event. Jane Street is not a bank and is not measured that way. For scale on how much risk sits in this non-bank layer: Federal Reserve research cited this week estimates the Treasury cash-futures basis trade at about $830bn as of September 2025, roughly double its 2020 peak, 35% of hedge funds' long Treasury exposure and financed almost entirely in repo.

Private credit and BDCshigh

Apollo read the file and shorted it instead

In a market with no marks, the difference between a good loan and a fraud is one manager's reading of a footnote — and the loser is whoever bought the paper last.

A 112-page filing in the First Brands bankruptcy, written by the restructuring CEO Charles Moore of Alvarez & Marsal, says Katsumi Global — a Japanese trade-finance player — tried to sell part of its First Brands factored-invoice exposure to Apollo. Apollo declined, Moore writes, "after correctly identifying the numerous red flags discussed herein, even without the benefit of the specific disclosure known to Katsumi." Apollo then shorted First Brands debt. The filing quotes an internal Apollo presentation from April 2024 — eighteen months before the Chapter 11. Its key issues: "the cash flow of the business has never matched the Company's claims about its margins"; 20% operating margins are "difficult to believe" for a company acquiring sub-5% and 10% EBITDA businesses; and the company "is 100% controlled by CEO Patrick James," who "has controlled dozens of entities over the years, many of which have ended up in litigation or liquidation (or both). Simple Google searches reveal almost nothing about him." Two mechanisms here. Invoice factoring is the softest part of the credit stack: you are lending against receivables you did not originate, and if the invoices are not what they claim, your claim is on nothing. And a holder of that exposure trying to place it with a smarter buyer is what "risk transfer" looks like from the inside — the marks did not move, the paper did. The uncomfortable implication for private credit generally: the only control is diligence, because there is no price to warn you. Everything Apollo saw in April 2024 was in the file. It stayed in the file until default.

What would change our mind

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

A listed BDC cutting its dividend or taking an unscheduled NAV markdown, or high-yield spreads breaking 400bp — either would mean the credit cycle has stopped being a disclosure event and started being a price.

Would move the number

Another tailed 30-year auction with the yield above 5.6%, or evidence of basis-trade stress in repo rates and dealer balance sheets. That is the route from a bond repricing to a funding accident.

Would move the number

Nvidia disclosing how the $105bn OpenAI lease guarantee is accounted for — the trigger, the recovery assumption, and whether it lands on the balance sheet or stays in the footnotes.

Would move the number

Downward: Q3 BDC non-accruals stabilising, and the autumn AI issuance calendar clearing without concessions or upsized coupons. Both would take several points off ignition.

Would move the number