$14bn data centre, $450m of insurance, lenders hold the tail

Archived reading, published Tue, 18 Aug 2026 18:19:28 UTC (8 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

Fragility84

how much tinder is stacked up — moves slowly

Ignition42

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 27 sources and rewrites this page.

Why it moved: Up two: the market began discriminating inside the AI complex — CoreWeave -11.6%, Applied Digital -9.0%, Nebius -8.7%, Talen -7.8% on the day while the S&P fell only 0.65% — and a Blackstone-backed, investment-grade-rated data-centre bond had to pay 7.63% to clear. Fragility ticks up on the SEC's risk-retention carve-out for data-centre securitisations and the disclosure that Meta and BlackRock's $14bn El Paso campus carries no cover against total loss.

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 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.

Fed, Treasury and policyhigh

The rules loosened; the market charged more anyway

Less mandatory equity in data-centre securitisations plus junk pricing on the highest-quality version of the same risk tells you where the loss will land if the leases wobble.

Two facts from the same week, pointing opposite ways. First: SEC staff agreed with a request from Latham & Watkins that some data-centre debt is not an asset-backed security, and therefore not subject to Dodd-Frank's risk-retention rules — the requirement that a sponsor keep skin in the game. "It gives them the opportunity over time to push down the required equity in the deal," a securitisation lawyer at Orrick told CNBC. Lawyers at Katten put it plainly: the guidance "gets around" rules written after 2008 in response to securitisations of badly underwritten mortgages. It is a staff opinion, not a rule, and can be withdrawn. It arrives as Nvidia signs memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR aimed at mobilising more than $500bn of third-party capital for compute, with Nvidia willing to backstop up to $125bn of it. Second: Blackstone-backed QTS launched a $3.9bn five-year bond for a Microsoft-tied data centre — "Project Odyssey" — at initial price talk of about 7.63%. The notes are expected to be rated Baa3 by Moody's and BBB- by Fitch. That is an investment-grade stamp at a junk-bond yield. The deal drew roughly $10bn of interest and was upsized by about $1bn, so demand exists — at a price. That gap is the story. Regulators are making it cheaper to structure this debt with less equity underneath; bond investors are demanding several hundred extra basis points regardless. When the rule-writers and the buyers disagree, believe the buyers.

Crypto and TradFi contagionhigh

Strategy is now a preferred-stock retirement vehicle

The largest corporate bitcoin holder has stopped being a buyer and started being a liability manager, and it is selling coin to do it.

Strategy sold $333.7mn of common stock last week and used it to buy back preferred shares and build cash. It bought no bitcoin. It has not bought any since mid-June. Over five weeks it has sold about $2.1bn of common, repurchased about $347mn of its STRC preferred, and sold $213.3mn of bitcoin. The cash reserve is now $4.8bn against roughly $58bn of coin. Understand what has inverted. The original machine worked because the shares traded above the value of the bitcoin behind them: issue equity at a premium, buy coin, bitcoin-per-share rises, repeat. Every part of that requires the premium. Without it, issuing stock to buy coin destroys value for existing holders — so the issuance is now being pointed at the liabilities instead. STRC preferred carries a cash dividend; bitcoin pays nothing. Strategy earlier retired $1.5bn of its 0% 2029 convertibles for about $1.38bn cash, cutting converts from $8.21bn to $6.71bn. This is deleveraging, executed with equity and, when necessary, with coin. The prices say the same thing from the other side. Bitcoin is up 1.2% over twenty days at $64,659. MSTR is down 7.8% over the same period at $94, and down about 73% over a year. The coin is flat; the wrapper is not. That is the market pricing the financing structure, not the asset. So far this is orderly — cash raised, claims retired, no forced seller. It is also the clearest live example of a leveraged holder unwinding without a crisis, which is what the good version of this looks like.

Household creditmedium

AI borrowing arrives in the housing market

The AI capex cycle has become a household interest-rate shock, transmitted through the long end of the Treasury curve.

US pending home sales fell 2.3% in July to an index of 71.2 — the lowest since January and, per Bloomberg, matching the second-worst reading in data back to 2001. Economists had expected no change. All four regions fell; the West dropped 4.7%. Housing starts also slowed. The cause is not mysterious: the average 30-year mortgage rate is just below the one-year high of 6.81% it hit in late July. The transmission chain is the thing to hold onto. Mortgage rates track the long end of the Treasury curve. The long end is being set by two forces at once: a federal debt pile approaching $40tn, and a corporate borrowing wave from the AI buildout. Investment-grade companies have sold nearly $1.5tn of bonds this year, up 36%. Nomura estimates 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. "Whoever's issuing, be it a government or a hyperscaler," Nuveen's Tony Rodriguez told Bloomberg, "is now competing with more borrowers. And therefore yields have to be higher." So the data-centre capex boom is now a cost of capital for people buying houses. Meanwhile household credit is already carrying stress: NY Fed Q2 data show overall delinquency at 4.7%, seriously delinquent student loans at 10.6%, auto at 5.5%, and 12.9% of credit-card balances 90+ days late. Consumer credit names are not confirming distress — Capital One up 8.1% in twenty days, Synchrony up 13% — which is a genuine disagreement between prices and delinquency data.

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. "If Anthropic is growing 800 per cent a year, you'd think at the incredibly low end they would trade at 30 times revenue," one investor told the FT. "That would make them a $3tn company."

Financial Times

Taiwan will hand every eligible resident NT$10,000 ($314) next year, funded by an AI-chip windfall: the economy grew 12.9% in the second quarter, the stock market is up almost 60% this year, and the government is on course for its fastest annual growth in four decades. It is a semiconductor dividend paid to 23 million people because many of them feel they are missing out.

Bloomberg

An NBER survey of bond investors, ordinary voters and people with economics or finance degrees found that all three groups put the average probability of a US debt crisis within ten years at close to 50%. Among concerned respondents, 72% of investors reported making no portfolio change at all.

FT Alphaville

Venezuela's government and its opposition have united on one thing: demanding $4bn of gold reserves back from the Bank of England. Gold is at $4,413 an ounce, up 8.4% in twenty days, which is roughly why everyone suddenly wants theirs in hand.

Financial Times

The rumour mill

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

Confirmed

“A business financially tied to the Trump family works with a platform distributing Chinese AI models, including from companies under US restriction, and Trump family crypto products earn revenue from it.”

Reuters reports World Liberty Financial — 38% owned by the Trump family — is collaborating with Hong Kong-based WorldClaw, which offers about 43 of 90 models from Chinese firms flagged by the Pentagon and accepts World Liberty's USD1 stablecoin as payment. Separately, the OCC granted World Liberty conditional approval for a national bank trust charter.

Claimed by Gregory Mannarino

Partly true

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

Directionally right, not uniformly. LSEG consensus has Microsoft, Alphabet, Amazon, Meta and Oracle spending more on capex than they generate in free cash flow by 2027, and BofA models an eight-company aggregate going from +$180bn in 2025 to -$144bn in 2027. But FactSet has Alphabet and Microsoft still positive, so "already negative for several of them" overstates it.

Claimed by Meet Kevin

Partly true

“The Fed's FIMA repo facility has a $60bn cap, was drawn to the cap in the 2023 regional bank crisis, has read zero for eight straight weeks, and Bessent is urging it be upsized.”

The $60bn per-counterparty cap, the March 2023 full draw and Bessent's public interest in expanding the facility all check out; the eight-consecutive-zero-weeks detail does not — the last meaningful use reported was about $3bn in early February 2026. Worth watching, because a foreign central bank tapping this facility rather than selling Treasuries is one of the cleaner early signals of dollar funding strain.

Claimed by Mark Moss

Confirmed

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

Confirmed by Japan's Ministry of Finance, Reuters and the WSJ; Goldman estimates roughly $85–88bn deployed across 30–31 July. Nikkei reports Japan has also repurposed a COVID-era Fed repo facility allowing the BoJ to borrow dollars against its $1tn-plus Treasury holdings — meaning it can defend the yen without dumping Treasuries.

Claimed by Mark Moss

Earlier developments

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

The AI capex bubblehigh

Nvidia now guarantees the rent

The creditworthiness of the AI buildout increasingly rests on one chipmaker's balance sheet rather than on the tenants actually using the buildings.

Nvidia has agreed to guarantee up to $105bn of OpenAI's lease payments on a data-centre campus being built by SB Energy in Pike County, Ohio, and to invest $1.5bn in SB Energy itself. OpenAI has contracted for roughly 8 gigawatts of capacity at the site, the first 800 megawatts due in 2028. One gigawatt powers as many as 750,000 US homes. The $105bn covers only the first 4.25GW phase; Nvidia may take on more for the second (Bloomberg). Here is the plumbing. Whoever lends against that campus is not underwriting OpenAI, a five-year-old company with enormous revenue growth and enormous cash burn. They are underwriting Nvidia, the most valuable company in the world. The guarantee is what converts an unrateable tenant's rent into something a bond fund can buy. That is the whole trick, and it is the same trick as the Microsoft-linked QTS bond and the $500bn of compute-financing platforms Nvidia has agreed with Apollo, Blackstone, KKR, BlackRock, Brookfield and Goldman, where Jensen Huang has said Nvidia could backstop up to $125bn. What has to stay true: OpenAI keeps paying. If it doesn't, Nvidia writes cheques for twenty years on buildings full of its own chips. Nvidia says this is not circular financing. What we do not know is how the guarantee is accounted for — whether it is recognised at fair value as a financial guarantee or disclosed in a footnote, and what triggers a draw. Until that is in a filing, the largest single credit enhancement in the AI buildout is a press release. Nvidia shares fell 2.5% on the day.

Private credit and BDCshigh

The loans are going bad; the funds are not marked down

Credit stress in private lending is now measurable, but it is showing up in manager earnings expectations rather than in the value of the loans themselves.

Loans placed on non-accrual by the twenty largest listed business development companies rose to a median 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. Fitch says private credit defaults hit a record in July. FS KKR reported 7.1% of its loan 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" (FT). Non-accrual means the borrower has stopped paying, or the fund thinks it is about to. It is the one number in private credit that is hard to argue with, because everything else — the fair value of a loan nobody trades — is an estimate made by the person being paid to manage it. Now look at the prices. Ares Capital is 2.5% off its recent high. Blackstone Secured Lending is 3.0% off. Golub's own BDC is 2.2% off. High-yield spreads are 2.67% and tightening. The equity market is, in effect, accepting the marks. Where it is not accepting them is one level up. Blackstone is down 5.2% over five days, Apollo 5.4%, Blue Owl 4.1% — all still up sharply over twenty days, but sold harder than the funds they run. That ordering makes sense: the funds' books shrank again last quarter, with repayments outpacing new lending at KKR, Blue Owl and Apollo's MidCap. Fewer loans means fewer fees. The market is marking down the fee stream before it marks down the loans.

Bond market dysfunctionhigh

Multi-decade high yields, and the calmest bond market in a month

The long end is absorbing government and AI borrowing simultaneously, and the price of that absorption is a permanently higher discount rate for everything else.

The 30-year Treasury yield touched 5.33% on Tuesday, the highest since 2007. Last week's $25bn 30-year auction cleared at 5.216%, the worst since 2001; the $42bn 10-year the day before cleared at 4.683%, the worst since 2007. This is not a US story: 30-year Bunds hit 3.78%, the highest since 2011, French 30s 4.9%, the highest since 2008, and 10-year JGBs reached 2.93%, a level last seen in 1996 (FT). Three things are pushing at once. Brent above $90 revives inflation fear. Deficits keep growing, with US debt near $40tn. And a new borrower has arrived at the long end: investment-grade companies have sold nearly $1.5tn of bonds this year, up 36%, and Nomura estimates the roughly $200bn from the biggest tech firms alone equals about 25% of the Treasury's net issuance of notes and bonds to private investors — five times the 2025 share (Bloomberg). That is crowding out, running the wrong way. Governments are supposed to squeeze out companies. Here the hyperscalers are competing with the Treasury for the same finite pool of people willing to lend for thirty years — and there are not many, because a Bank of America survey shows fund managers hold their lowest bond allocation relative to benchmark, and their highest equity allocation since 2021. One number cuts against the drama. The MOVE index of Treasury volatility sits at 10.2, down 14% in five days and 28% in twenty, and 43% below its recent high. Yields at 2007 levels with volatility at a one-month low is an orderly repricing of the price of duration, not a market breaking. It is the combination to watch: if MOVE turns up while auctions keep tailing, that changes.

Hidden leverage and shadow bankinghigh

$1.65tn of debt that is not on the balance sheet

The most-analysed companies on earth carry more obligation in footnotes than on their balance sheets, which is exactly the pattern Crash Lab tracks: risk moved somewhere it does not have to be marked.

Nikkei went through the filings of Alphabet, Microsoft, Amazon, Meta and Oracle and totted up the obligations that sit outside the balance sheet: $1.65tn, against roughly $1.35tn of debt actually booked. Four years ago the off-balance-sheet figure was an eighth of that. Meta's is about $420bn, nearly triple its recorded debt. Oracle's reached $273.3bn as of the end of May, a more-than-thirtyfold rise in four years (Nikkei). None of this is improper. Under lease accounting, a lease liability lands on the balance sheet when the asset becomes available for use — not when the contract is signed. GPUs and servers ordered under long-term agreements but not yet delivered are commitments, disclosed in footnotes. So a company can sign a twenty-year obligation on a building that does not exist and report no additional debt. The consequence is timing. These liabilities do not stay hidden; they arrive, in a lump, as each campus is handed over. A company that looks conservatively financed today can look heavily levered in 2028 without signing anything new. Moody's flagged ballooning not-yet-commenced lease commitments in February; Morgan Stanley has written it up in detail. The equity market has started to sort the group. Meta is down 14.9% over twenty days and 19.6% from its recent high, the worst of the five, while Microsoft is up 20.8% over the same window. That is not a verdict on AI. It is investors beginning to price the difference between spending cash you have and promising cash you will have.

Household creditmedium

Record car loans, record-ish car losses

Consumer credit stress is concentrated and priced, not systemic — which is a reason the household beat is not the fuse, at least not yet.

Americans took out $211bn of auto loans in the second quarter, a record, according to the New York Fed. Over the same period the share of auto debt rolling into serious delinquency rose to 3.00%, from 2.93% a year earlier. KBRA's July servicing data puts net losses on non-prime auto securitisations at 9.46% and prime at 0.66% — a fourteen-fold gap between the two halves of the same market. The honest counterpoint: 60-plus-day delinquencies in subprime auto ABS were 5.67% in June, down 64 basis points year on year, and overall household delinquency slipped to 4.7% of balances from 4.8%. This is not a consumer falling over. It is a consumer splitting in two. The mechanism is in the funding. A lender with a prime book borrows cheaply in the ABS market, can price aggressively, and takes share — which is what record originations at improving average credit quality looks like. A lender concentrated in deep subprime faces losses near 9.5%, needs a wider spread to cover them, and finds fewer borrowers who can pay it. That shows up cleanly in the share prices: Credit Acceptance, the deep-subprime specialist, is down 7.4% over twenty days and 10.3% off its high, while Capital One is up 8.7% and Synchrony up 13.5% and sitting at its high. Credit card balances, meanwhile, reached $1.263tn, up $54bn on the year, with 12.92% of outstanding balances 90 or more days delinquent. That is a large number that has been a large number for a while. The thing to watch is whether the prime book starts to converge on the subprime one, not whether subprime gets worse.

Crypto and TradFi contagionmedium

A Shanghai IPO is being priced on a crypto exchange

Unregulated 24/7 derivatives are now generating the reference prices that regulated markets react to, with no delivery mechanism underneath them.

Unitree Robotics lists on the Shanghai Star Market on Wednesday at a $9bn valuation. Perpetual futures on the Hyperliquid blockchain were trading Tuesday at about $100, implying $40.5bn — more than four times the IPO price. The offering was 5,526 times subscribed (Bloomberg). A perpetual future — a "perp" — is a contract with no expiry date that trades around the clock. Nothing is ever delivered. It is kept roughly in line with its reference price by a funding rate: when longs outnumber shorts, longs pay shorts a periodic fee, which is supposed to drag the contract back to spot. That mechanism works when there is a spot price to be dragged towards. For a company that has not yet listed, there isn't one. There is no arbitrage available — you cannot buy the shares and short the perp — so the funding rate disciplines nothing. The $40.5bn is not a price. It is a poll of leveraged opinion, taken among people who cannot obtain the asset. What makes it worth your attention is that these polls have started to be right about direction. The same venue foreshadowed the surge in CXMT, the Chinese memory-chip maker, last month, and perps on SpaceX drew broad interest earlier this year. An offshore crypto exchange has become the pre-listing price discovery venue for mainland Chinese state-adjacent semiconductor and robotics IPOs. Nobody designed that. The rest of the crypto complex is not sharing in the enthusiasm: bitcoin is flat at $64,227 over twenty days, Coinbase is down 16.2% and Galaxy Digital 17.6%.

The AI capex bubblehigh

Rated investment grade, priced like single-B

When a bond's rating and its yield disagree by two rating bands, the disagreement is the information — and it is about what an AI data centre is worth once the lease ends.

A bond sale codenamed "Project Odyssey" is being upsized to about $3.9bn after demand topped $8bn. It is five-year paper issued by two special-purpose subsidiaries of QualityTech LP, a unit of Blackstone-backed QTS Realty Trust, to bankroll a data centre venture in Georgia leased by Microsoft. Citigroup, Goldman Sachs, JPMorgan and Morgan Stanley ran the calls. Here is the interesting part. The bonds are expected to be rated investment grade. They are being pitched with a coupon around high-6% and a discount price that pushes the yield to low-to-mid 7% — which is where single-B rated junk bonds currently trade (Bloomberg). Ratings agencies and bond buyers are looking at the same instrument and reaching different conclusions. The agencies are rating the lease: Microsoft's covenant, over a defined term, on a building. The market is pricing something else — the residual. What is a purpose-built AI data centre worth in year six if the tenant walks, or if the chips inside it are two generations obsolete, or if the power contract reprices? Nobody has tested that, because no such asset has ever reached the end of its first lease. The two-notch gap between rating and yield is not a scandal. It is a price the market is charging for a risk the rating does not capture. It also tells you why the deal got done: at 7%, with an investment-grade stamp, this paper fits inside mandates that would never buy junk. That is precisely how risk migrates from where it is measured to where it is not. US high-grade issuance hit a record $145.2bn in August, taking the year to $1.46tn.

Private credit and BDCshigh

The SEC quietly took sponsors off the hook

The single most important post-2008 rule about who eats securitisation losses has been read not to apply to the fastest-growing securitisation asset class.

Last month the SEC's staff agreed with the law firm Latham & Watkins that some data centre debt is not an asset-backed security, and therefore is not subject to the risk-retention rules written into Dodd-Frank after 2008 (CNBC). Risk retention is the "skin in the game" rule. It exists because in 2007 the people packaging mortgages into securities kept none of the losses, and so stopped caring what was in the package. The rule forces a sponsor to hold roughly 5% of the credit risk of what it sells. "It gives them the opportunity over time to push down the required equity in the deal," Orion Mountainspring, a securitisation lawyer at Orrick, told CNBC. Seth Messner of Katten put it more bluntly: Latham asked the SEC to say data centre securitisations aren't ABS, and the SEC basically agreed. The timing matters. Days later Nvidia signed memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build "compute financing platforms" aimed at mobilising more than $500bn of third-party capital, with Nvidia itself prepared to backstop up to $125bn — about a quarter. Two caveats worth stating plainly. This is a staff opinion, not a rule, and it can be revisited. And it is not yet clear whether the Nvidia platforms will use securitisation at all; the lawyers CNBC spoke to think the guidance sounds applicable, which is not the same as knowing. But the direction is unambiguous. The largest new pool of collateral in the market is being structured so that the people assembling it are legally permitted to keep less of the downside.

Bond market dysfunctionmedium

Japan's life insurers are $194bn underwater

The marginal buyer of long-dated global debt is sitting on losses that could force it to become a seller instead.

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 (Nikkei). That figure now exceeds their unrealised gains on domestic stocks of ¥30.03tn. The standard reassurance is that held-to-maturity bonds pay out at par and the loss evaporates. Mostly true. Two things break it. First, impairment. If a bond's market value falls 50% below acquisition cost, insurers may have to recognise the loss immediately. Bonds bought in the ultra-low-rate late 2010s have now fallen that far. Nippon Life booked ¥44bn of impairments in April–June; Meiji Yasuda ¥25.3bn. Small numbers against ¥30tn — but they establish that the threshold is live, not theoretical. Second, duration. Life insurers buy 30- and 40-year JGBs to match liabilities. If assets are longer than liabilities, rising rates cut net assets, and the fix is to shorten duration — which means selling long bonds and crystallising losses. Thirty-year JGB yields were around 3.9% at end-June and have kept rising. The ten-year is reported near 2.93%, the highest since 1996. The transmission to everyone else runs through the same balance sheets. Japanese life insurers are among the world's largest cross-border buyers of duration. An institution that needs to shorten its own book is not adding to a $25bn US 30-year auction that already cleared at 5.216%, the highest since 2001. The global long end is now short of exactly the buyer it used to rely on.

Hidden leverage and shadow bankingmedium

The market-maker that was a hedge fund

The largest liquidity providers in US markets have quietly become directional risk-takers, and the disclosure regime has not followed them.

Jane Street lost $15bn in July — its first losing month in a decade — on directional bets, Bloomberg reports. The firm has still generated more than $40bn of net trading revenue this year, ahead of its 2025 record. In its latest debt offering it described its strategy as evolving to include longer-term positions, in the manner of a hedge fund (Bloomberg). A note on the numbers. The FT, ranking the loss against history's biggest, attributes a $35bn July trading loss to Leopold Aschenbrenner on leveraged AI stock bets and ties it to Jane Street's record (FT). We cannot reconcile $35bn with $15bn from what has been published, and we are not going to pretend otherwise. Either could be gross versus net, position versus P&L, or something else. Treat the exact figure as unsettled; treat the order of magnitude as real. What is not in doubt is the structural point. A market-maker earns a spread thousands of times a day and holds inventory briefly. Its capital requirements, its funding, and everyone else's assumptions about it are built on that. A firm taking multi-billion directional positions in AI equities is a hedge fund with a market-maker's balance sheet and a market-maker's regulatory treatment. That matters because banks, clearing houses and counterparties size their exposure to these firms partly on what they believe the firms are doing. A loss of this size at a bank would be a career-ending scandal and a supervisory event. Here it was a bad month, disclosed after the fact, at a firm nobody supervises the way they supervise a bank.

Household creditmedium

Ten quarters above the 2008 benchmark

Consumer credit stress that persists for ten quarters without breaking is a cost of doing business, not a crash trigger — and knowing the difference matters.

The New York Fed's Q2 data is out, and it says two things at once. The aggregate is improving. Total household debt in some stage of delinquency fell to 4.7% from 4.8%. Subprime auto 60-day-plus delinquencies were 5.67% in June, down 64 basis points year on year. The tail is not. The share of credit card balances flowing into 90-plus-day delinquency was 6.97% in Q2 — the tenth consecutive quarter above the roughly 6.75% level that marked the onset of the Great Recession. Card balances hit $1.263tn, up $21bn on the quarter. Auto originations set a record at $211bn while the flow into serious auto delinquency ticked up to 3.00% from 2.93%. KBRA's July data has non-prime auto net losses at 9.46% against 0.66% for prime. The gap between 0.66% and 9.46% is the whole story. This is not a household sector under strain; it is two household sectors, one of which is fine and one of which has been running at crisis-era loss rates for two and a half years without producing a crisis. The reason is that the losses are absorbed in ABS tranches and lender provisions, quarter by quarter, rather than arriving all at once. The equities disagree with the gloom. Capital One is up 7.5% over twenty days, Synchrony 10.4%, Sallie Mae 7.6% — though all four consumer lenders we track fell on Monday, Affirm by 4.85%. The market's read is that this is a chronic condition being priced correctly, not an accident waiting to happen. On the evidence, that read is holding.

Crypto and TradFi contagionhigh

Saylor is running the machine backwards

The digital-asset treasury model was a premium-to-NAV perpetual motion machine, and the first large one to lose its premium is showing what the reverse gear looks like.

Strategy sold $333.7m of common stock last week. It bought no bitcoin. It has bought none since mid-June, and over the last five weeks has sold about $2.1bn of stock, repurchased roughly $347m of its STRC preferred shares, and sold $213.3m of bitcoin (Bloomberg). Its cash reserve is now $4.8bn against about $58bn of bitcoin. The shares are down around 73% over a year. Understand what has been inverted. The original machine: issue equity above the value of the coins per share, buy coins with the proceeds, coins-per-share rises, the premium justifies itself, repeat. It only works while the stock trades above net asset value. When the premium goes, the same mechanism runs in reverse — issuing stock now dilutes holders rather than enriching them, and the proceeds go not into bitcoin but into servicing the preferred stack that was built during the good years. That preferred stack is the part to watch. It pays cash dividends. Bitcoin does not. So a company whose only real asset is a non-yielding one has a permanent cash outflow, and three ways to meet it: sell stock, sell coins, or refinance. It is currently doing the first two. The market's short-term read is odd. MSTR rose 4.95% on Monday while bitcoin fell 0.77%, and over twenty days the stock is flat against a flat coin. Investors appear to be pricing the shift to "capital management" as prudence rather than distress, which is defensible: buying back preferred below par retires an expensive liability. It is also what a company does when it can no longer grow.

Private credit and BDCshigh

Non-accruals went from 2% to 2.8% in one quarter

Non-accruals are the one number in private credit that isn't a matter of opinion, and they moved sharply in a single quarter.

Loans placed on non-accrual status by the twenty largest publicly traded business development companies climbed to a median 2.8% of cost in the second quarter, up from 2% at the end of March. That is the highest since 2017, when the industry was digesting the oil crash. Fitch says private credit defaults hit a record in July. (FT) What non-accrual means matters here. A BDC lends to a mid-sized company at a floating rate and books the interest as income whether or not cash arrives. Putting a loan on non-accrual is the fund conceding that the interest is not coming — either the borrower has stopped paying or default looks imminent. It is one of the very few things in private credit that is not a matter of judgement. The valuation of the loan itself is a judgement. The non-accrual flag is close to a fact. So a 40% relative jump in one quarter is the clearest signal yet that the credit cycle the industry spent two years denying has arrived. Golub Capital's David Golub said as much to 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." Two further details. FS KKR reported 7.1% of its loan book troubled — far above the median. And funds run by KKR, Blue Owl and Apollo's MidCap Financial shrank in the quarter, with repayments outrunning new lending. A shrinking loan book is a lender choosing not to lend. The important caveat: these are the listed vehicles, which must disclose. The much larger non-traded and evergreen funds report on their own schedule, to their own marks.

The AI capex bubblehigh

Nvidia agrees to stand behind OpenAI's rent

The AI buildout is being financed by a chipmaker's credit rather than its customers' cash flows, and the rules requiring lenders to retain risk have just been waived for it.

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 will also invest $1.5bn in SB Energy. OpenAI is contracting for roughly 8 gigawatts of capacity — a gigawatt powers as many as 750,000 US homes — with the first 800MW due in 2028. Nvidia's backing phases in from 2028 to 2030 and runs for twenty years. The $105bn covers only the first 4.25GW; it may take on the rest. (Bloomberg) The plumbing: SB Energy needs to borrow to build. Lenders will lend against the lease. The lease is only as good as OpenAI, a company that does not make money. So Nvidia — which will sell the chips that go inside — puts its own balance sheet behind the rent, and the lenders are effectively lending to Nvidia. The same purchases could be worth about $600bn of Nvidia revenue through 2030. Jensen Huang says this is not circular financing. Read that literally: Nvidia is not handing OpenAI cash to buy chips. It is doing something subtler — converting its equity-market credibility into third-party debt capacity for its own customer. This is the week's second such structure. Nvidia has also signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR to mobilise over $500bn of compute financing. And last month SEC staff agreed with Latham & Watkins that some data-centre securitisations are not asset-backed securities, and so escape Dodd-Frank's requirement that sponsors keep skin in the game. (CNBC) It is a staff opinion, not a rule. It is enough.

Hidden leverage and shadow bankingmedium

The debt that isn't on the balance sheet is bigger

The most heavily analysed companies on earth have more debt in their footnotes than on their balance sheets, and the footnotes are where risk goes to be unmeasured.

Nikkei went through the filings of Alphabet, Microsoft, Amazon, Meta and Oracle and totted up the obligations that sit in the footnotes rather than the balance sheet. The figure is $1.65tn — more than the $1.35tn of debt these companies actually report. Four years ago it was roughly an eighth of that. Meta's off-balance-sheet total is about $420bn, nearly triple its recorded debt. Oracle's is $273.3bn as of end-May, a thirty-fold rise in four years. (Nikkei) None of this is improper. Accounting rules say that a lease on a data centre that is not yet operating, and a purchase commitment for GPUs and servers not yet delivered, are disclosed in the notes and not capitalised on the balance sheet. The obligation lands the day the lease commences. Until then it is real money owed and invisible to anyone reading the summary statements — which is most people. The reason it has exploded is structural. Hyperscalers keep initial cost down by having an operator supply the land, the building and the power, and then leasing it for twenty years. Meta's $27bn Hyperion financing reportedly sat in an SPV, backed by that twenty-year lease promise, and so did not appear on Meta's balance sheet. The counterweight, which is genuine: this group will earn around $1tn in operating profit this year. Bloomberg's tally puts total commitments through debt and contracts as high as $4.2tn, against $1.6tn of reported liabilities as of June — nearly double two years ago. (Bloomberg) Solvent companies with enormous cash flows and hidden fixed obligations are exactly how a corporate credit cycle starts quietly.

Bond market dysfunctionhigh

Thirty-year yields at 5.31% and the AI bid for capital

The cost of long money is being set jointly by the US deficit and five companies' capex plans, and neither is slowing.

The 30-year Treasury yield hit 5.31% on Monday, the highest since 2007. Last Thursday's $25bn auction of the same bond cleared at 5.22%, the highest auction yield since 2001; Wednesday's 10-year cleared at 4.683%, the highest since 2007. (Bloomberg, FT) The usual explanation — near-$2tn deficits, inflation above target for five years — is most of it. The new part is on the other side of the ledger. US investment-grade issuance hit $145.2bn in August, a record for the month with two weeks to run, and $1.46tn for the year, running 8.5% ahead of the 2020 record pace. Nomura reckons the roughly $200bn borrowed by the biggest tech companies alone equals about a quarter of the Treasury's net note and bond issuance to private investors — five times the 2025 share. That is crowding-out running backwards. The classic theory says government borrowing squeezes companies out. Here a handful of AI borrowers are competing with the sovereign for the same finite pool of duration demand, and everyone's yield goes up. "Whoever's issuing... is now competing with more borrowers," as Nuveen's Tony Rodriguez puts it. (Bloomberg) It is now visible abroad too. Alphabet raised C$8.5bn in Canada, then Amazon raised C$14bn. Canada's AA corporate index — the top rung, where both sit — has widened past the lower-quality A index. "They basically come in and reprice all of the high-quality curve," says AllianceBernstein's Souheir Asba. Alphabet is now marketing up to A$5bn in Australia, after a 100-year sterling bond in March. (FT) Small credit markets are being repriced by borrowers who do not live in them.

The AI capex bubblehigh

A $14bn data centre with $450m of property cover

When the insurance market won't take a risk at any sensible price, the risk doesn't disappear — it ends up with the lenders, unpriced.

Meta and BlackRock's Sopaipilla campus in El Paso costs $14bn and will draw one gigawatt. On the advice of Marsh, the world's largest insurance broker, it has bought all-risk property cover of up to $427mn during construction and $450mn once operating, for a premium of about $5mn a year. Terrorism cover is capped at $645mn, delay-of-completion cover at $218mn, and general liability at $50mn per event and $50mn in aggregate. The project is not insured against total loss. (FT) Run the ratio. Roughly 3% of the build cost is covered against property damage. Everything above that sits with the venture and, if the venture cannot pay, with its lenders. The reason is not carelessness. It is capacity. Insurers will write data-centre risk, but few want a concentrated exposure to a single site worth tens of billions, especially one packed with chips whose replacement cost and lead time nobody can pin down. Full cover has become, in the FT's phrase, prohibitively expensive. For smaller data centres lenders demand it. At this size they have learned to accept a fraction. This is worth noticing because insurance is where a hazard gets priced by somebody with money at stake. When the insurance market declines to write the risk, the risk does not vanish; it migrates to whoever holds the debt. Those holders are increasingly bond funds, private credit vehicles and insurance company general accounts — the same balance sheets absorbing the data-centre ABS the SEC has just exempted from risk retention. We do not know what covenants the lenders extracted in exchange, and that is the part that would tell you how bad this is.

The dollar, gold and reserve statusmedium

Gold up 9% in a month while nothing is wrong

Gold rising 9% in a month with equity vol at 16 means someone is buying it for a reason that has nothing to do with the business cycle.

Gold is at $4,448, up 1.5% on the week and 9.3% over twenty days. Over the same period the dollar index fell 1.35% to 99.63 and the Bloomberg Dollar Spot Index hit its weakest since 15 May. Traders have cut the odds of a Fed hike next month to one-in-three from about 75% in late July, after soft employment, inflation and retail sales data. Only about 36bp of further tightening is priced for the cycle. (Bloomberg) A weaker dollar and lower rate expectations explain part of a gold rally. They do not comfortably explain 9% in a month while the S&P sits 0.7% off its high and VIX is at 16. Something else is in the bid. The candidate is official demand. An OMFIF survey of reserve managers reports 74% expect the dollar's share of global reserves to fall over five years and 89% expect central bank gold holdings to rise over the next year — described as the first such survey in which more managers plan to cut dollar exposure than add to it. That is a survey of intentions, not flows, and should be read as such. The texture item is better than the survey. Venezuela's government and its opposition — who agree on nothing — have jointly demanded that the Bank of England return $4bn of Venezuelan gold. (FT) The reserve-status question is not being settled by anyone selling Treasuries. It is being settled by a slow preference for the one asset that is nobody's liability, and by sovereigns deciding they would rather have the metal at home.

What would change our mind

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

High-yield OAS breaking above 4% (from 2.70%) or IG above 1.20% (from 0.81%) while AI-linked issuance continues — that would mark the point where credit stops absorbing the buildout at a price.

Would move the number

A data-centre bond or securitisation pulled, or priced wider than about 8.5%, or a neocloud disclosing it cannot draw on committed financing.

Would move the number

Evidence of a forced unwind in funding markets: SOFR spiking above the Fed's target range, the FIMA repo facility being drawn, or reserve-management purchases restarted before 14 September.

Would move the number

The other direction: 30-year Treasuries back below 5% with a well-covered auction, and median BDC non-accruals falling back toward 2% in Q3 — both would cut ignition materially.

Would move the number