Banks are paying record sums to move loan risk elsewhere

Archived reading, published Tue, 25 Aug 2026 14:25:50 UTC (27 hours ago). This is not the current state of the meter.

See the live reading →

CRASH-O-METER

Cracking

how close are we

-1 since the last reading

Fragility89

how much tinder is stacked up — moves slowly

Ignition41

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 23 sources and rewrites this page.

Why it moved: Down one to 65. Nothing new broke in this window and the tape eased: the crypto complex ripped (bitcoin +8% in five days, Strategy +33%, Coinbase +25%, all three treasury companies at period highs), the rate-volatility gauge we track is at a 20-day low, high yield is still 270bp and VIX 15.8. Fragility holds at 89 — the record SRT year and the BDC discount data are sharper measurement of risk migration we already carry, not new tinder.

Reporting from 24 Aug to 25 Aug

Crash points

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

What’s moving the needle

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

Hidden leverage and shadow banking25 Aughigh

The trade that takes risk off bank balance sheets

Capital relief only works if the protection seller can pay when the loans go bad, and nobody outside the deal knows who that is or how they are funded.

AIB Group is working with Banco Santander and the insurance broker Howden on a significant risk transfer covering about €2.5bn ($2.9bn) of project finance loans, due to complete later this year, Bloomberg reports. Sales of SRTs are on course for a record $45bn in 2026, up from $41bn in 2025, on Crescent Capital's estimates. The mechanics matter. In an SRT the bank keeps the loans. It buys protection on the first slice of losses from outside investors, who are paid coupons that Bloomberg describes as reaching double digits. The bank's regulatory capital requirement falls — AIB told investors in July that this deal should be worth 25 to 30 basis points of core tier 1 — and that freed capital can be lent again or paid out to shareholders. So the loan is still measured, still supervised, still on a published balance sheet. The loss is not. It sits with whoever sold the protection, in a fund or an insurance wrapper that marks its own book and discloses on its own schedule. The presence of an insurance broker on the deal suggests part of the protection is insurance-structured rather than a cash-collateralised note, though the report does not say, and it does not name the buyers. This is Crash Lab's thesis in its purest form: not a hidden loan, but a hidden loss position, growing at roughly 10% a year in a market that only began publishing size estimates recently. AIB's first SRT was in November 2024. Its third is nearly twice the size of its first.

Private credit and BDCs25 Augmedium

Saba offers 65 cents for a dollar of Blue Owl NAV

A discounted bid from a professional buyer is the only test of a non-traded fund's marks that does not depend on the manager's own arithmetic.

Saba has offered to buy into Blue Owl's non-traded BDC at roughly a 35% discount to the fund's marked value, according to reporting gathered in our sweep of the week's private credit coverage. Alongside it: redemption requests at the ten largest non-traded BDCs averaged 13% of assets in Q1 2026 and 14% in Q2, enough to force most managers to activate gates. Work through what those two facts are together. A non-traded BDC's net asset value is set by the manager, quarterly, on loans that do not trade. Investors are promised limited liquidity — typically 5% of assets a quarter. When 14% ask to leave, the gate closes and the queue becomes the price. Saba's bid is the first hard number attached to that queue: a real buyer, with capital, saying a dollar of stated NAV is worth about 65 cents to someone who has to hold it. Underneath, the credit data moves the same way. Non-accruals across the 20 largest listed BDCs rose 0.8 points in Q2 to a median 2.8% of cost, the highest since 2017. BCRED's went from 0.6% to 2.4%, driven by Medallia — marked at 60 cents on a $1.1bn cost basis — and Affordable Care International at 70. And yet: Blackstone Secured Lending is 0.9% off its period high, Ares Capital 1.2%, Prospect 0.9%. The listed vehicles are priced as though none of this is happening. These figures reach us through secondary aggregation rather than primary filings, which is why we hold this at medium confidence — but the direction has been consistent for three quarters.

Bond market dysfunction25 Augmedium

The long end repriced, then stopped panicking

A market that needs official buying to clear, and whose marginal buyer is levered overnight, is calm in a way that can end in one session.

The 30-year Treasury touched about 5.34% on 18 August, the highest since 2007. The Treasury doubled its liquidity-support buybacks to at least $4bn per operation from 9 September to 4 November. Yields fell roughly 10 basis points, then gave nearly all of it back within two days and finished the week near 5.27%. Every outlet in our window reads that as dysfunction, and the intervention-that-didn't-stick is a fair description. But the rest of the tape says something else. The rate-volatility gauge we track is down 15% in five days and 25% over twenty. Long Treasuries are up 1.7% in five days. Investment grade is at 81bp, high yield at 270bp. The level of long yields moved a lot this month; the volatility of them has been falling. Markets can reprice duration and then be calm about the new price. That is repricing, not a funding accident. What is worth understanding is the funding side of the buyback. CNBC reports Treasury officials consider the $935bn General Account available to pay for the purchases rather than issuing bills. Paying out of the TGA moves cash from the government's account at the Fed into private hands — it adds bank reserves. That is an easing of financial conditions, which is the most plausible reason a crypto short squeeze of about $2.7bn showed up in the same week. The fragility is not today's yield. It is that the OFR now counts about $2tn of hedge-fund cash Treasuries, much of it a repo-financed basis trade, sitting where the marginal buyer of duration should be.

The AI capex bubble25 Augmedium

Nebius borrows $5bn, and its shareholders pay the coupon

Convertibles let a company book cheap interest while handing over the equity upside, which makes debt-funded capex look more affordable than it is.

Nebius upsized its convertible bond to $5bn on 20 August: $3bn due 2030 at a 0.5% coupon and $2bn due 2034 at 4.5%, per Reuters-sourced detail in our sweep. It is one of the largest converts on record. The company spent $5.7bn on capital projects in the quarter. Two numbers to hold together. The raise funds slightly less than one quarter of spending at the current run rate. And a 0.5% coupon on seven-year money is not generosity — it is the price of the equity call embedded in the bond. Nebius is not borrowing cheaply; it is selling its own share-price upside and receiving a low interest rate in return. If the stock triples, the 2030 holders convert and existing shareholders are diluted at a price fixed in August 2026. The near-term mechanism is more direct. Convertible buyers are frequently arbitrage funds who hedge by shorting the underlying stock. Nebius is down 12.6% in five days and sits 21.8% below its period high, while still up 27.9% over twenty days. We cannot separate hedge-related selling from ordinary repricing of a neocloud, and nobody can from the outside — but the pattern is what you expect after a large convert. Set against the sector: the same reporting has global AI-related debt issuance at $489bn by mid-2026, US AI infrastructure issuance at $220bn as of 10 August, and incremental debt funding roughly 32% of hyperscaler capex on a trailing basis, up from 9% in fiscal 2024. The buildout has moved from cash flow to capital markets. Nebius is what that looks like for a company without the cash flow.

Crypto and TradFi contagion24 Aughigh

Strategy builds a second cash pile, and this one can buy bitcoin

A treasury company with fixed cash obligations against a non-yielding asset is solvent exactly as long as its own share price stays expensive.

Strategy has created a new $1.59bn pool it calls USD Cash, which can be used to buy bitcoin, pay preferred dividends and debt service, repurchase securities, or anything else the company needs. It sits alongside the original $5.1bn reserve, which — without board authorisation — can only pay preferred dividends and interest, Bloomberg reports. The plumbing: bitcoin pays no coupon. The preferred stock and the converts do. So the cash to service them has to come from somewhere else — selling shares, or selling coin. Strategy sold about $2bn of common stock in the previous week and bought back $136.4m of its Stretch preferreds, which have traded below par. Nansen's Nicolai Sondergaard put it exactly right: the new pool "gives Strategy more time and optionality, but it does not remove those underlying obligations." Selling shares only works at a premium to the value of the coins held. Bloomberg says that premium remains "well below earlier-cycle levels" and the flywheel is "impaired" — and this is the part the prices are arguing with. Strategy's stock is up 33.4% in five days against bitcoin's 8.1%, and it is sitting at its period high. Bitmine is up 35% and SharpLink 32%, both also at highs. The equity is outrunning the coin four to one, which is precisely what rebuilds issuance capacity. So the machine is repairing itself, for now, by the only mechanism available to it: enthusiasm. That is genuinely better than last month. It is also the same dependency that broke in the spring.

Household credit24 Augmedium

Record auto lending into a record delinquency rate

Consumer credit stress is now concentrated in the loans that were sold on, so the lenders' share prices may be the last place it appears.

Two figures from the New York Fed's Q2 2026 household credit report, gathered in this window's sweep. Auto loan originations hit a nominal record $211bn. And the share of auto debt transitioning into serious delinquency reached 3.00%, the highest since 2010. Fitch's July index puts 60-day subprime auto delinquency at 6.13%, against 0.49% for prime — more than twelve times higher. Student loans 90 or more days delinquent rose to 10.6% from 10.3%. The credit card numbers circulating in the same reporting cannot be reconciled: one figure has 12.8% of balances 90+ days past due, another 6.97%. We are not using either until we can trace it to the primary release. What the auto pair means mechanically: lenders are writing the largest volume of new loans on record into a borrower pool whose worst cohort is deteriorating. That is not necessarily reckless — deep subprime lenders price for high losses, and higher car prices mechanically lift origination dollars without lifting unit volumes. But it does mean the 2026 vintage is large, and vintages written late in a deteriorating cycle are the ones that hurt. The prices disagree with the alarm. Credit Acceptance, the deepest subprime auto lender we track, is up 8% in twenty days. Capital One is up 2.5%, Synchrony 3.8%, Oportun 31.5%. Only Ally is down, 3.1% over the month and 8.7% off its high. Equity holders sit behind ABS bondholders in these structures, so if the market were pricing a loss cycle, these are the shares that would show it first. They are not showing it.

Signs of the times

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

Twelve people, half a trillion in volume

Trade.xyz, a startup of about a dozen people built on top of the crypto exchange Hyperliquid, has generated some $500bn of trading volume since launching in October, in perpetual futures on crude oil, precious metals and companies about to go public including SpaceX. When the US and Israel struck Iran on a Saturday in February, with conventional oil markets shut, its blockchain oil derivative kept pricing the war.

market structurecrypto24/7

24 AugBloomberg

A data centre borrows at 0.5%

Nebius's $5bn convertible includes $3bn due 2030 at a 0.5% coupon — for four-year money, the US Treasury currently pays 4.24% on two-year notes. The coupon is low because the bond comes with a claim on the shares, which is a cost that shows up later and somewhere else.

ai capexconvertiblesdilution

24 AugReuters

Two record fundraisings, one Monday

Alibaba raised the equivalent of $10bn in Hong Kong's largest-ever secondary share sale for chips and data centres on the same day SoftBank filed for Japan's biggest-ever retail bond, ¥1tn at 4.3–4.9%, to fund its OpenAI commitments. Alphabet, Meta, Microsoft and Amazon have between them pledged nearly $2.4tn of AI-related spending.

ai capexissuancescale

24 AugBloomberg

Twenty-two new banks in nineteen months

The OCC has approved 22 bank charters in the first nineteen months of this administration — more than in the previous five years combined — with about 40 applications filed since 2025, as many as in the preceding thirteen years. The newest preliminary approval, dated 14 August, is for World Liberty Trust Company, part-owned by a Trump-family entity, to custody the reserves behind the USD1 stablecoin, which has more than $4bn outstanding. Reported by Il Messaggero; we have not seen the charter details corroborated elsewhere.

stablecoinscharteringsingle source

25 AugIl Messaggero

The rumour mill

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

Confirmed

“Asked about further bond market intervention after yields rose back up, President Trump said 'the ultimate intervention is our military' and 'if we have to use that, we will.'”

Said on camera at Joint Base Andrews on 21 August, in direct response to a question about whether Bessent was considering another intervention; carried by Yahoo News, HuffPost, NDTV Profit and others. We note it as a fact about how this administration frames bond market policy, not as a forecast of anything.

Claimed by Heresy Financial

Confirmed

“Nvidia's stock has fallen after each of its last four earnings reports despite beating expectations.”

Confirmed: -5.5% after Q4 FY26, -3.2% after Q3 FY26, -0.8% after Q2 FY26 and -1.8% after Q1 FY27, on beats each time. Nvidia reports again on 26 August, with its shares 5.3% off their period high and up 2.4% on the day.

Claimed by Meet Kevin

Partly true

“Oracle's credit default swaps keep climbing on data centre debt, and the stock is down nearly 60% from $351 to $143.”

The CDS half checks out — five-year protection reached 219bp in late July, the widest in at least six years, versus Nvidia at 78bp and Meta near 93bp. The equity half is overstated: Oracle is at $144.90, roughly 50–55% below its peak, not 60% below $351.

Claimed by Meet Kevin

Partly true

“A UBS analyst forecast Micron could generate as much as $400bn in free cash flow through 2028, enough to retire up to 40% of its shares.”

UBS's Timothy Arcuri did project roughly $380–450bn of cumulative free cash flow through end-2028 and a buyback capable of retiring about 40% of the shares; outlets differ on whether the figure is $400bn or above $450bn. Worth keeping as a marker of what sell-side models now assume about memory pricing.

Claimed by Meet Kevin

Earlier developments

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

Hidden leverage and shadow banking25 Aughigh

The SEC probes Wall Street's AI leverage

The episode is a live test of whether bank risk controls have improved since Archegos, or if prime brokers are still extending massive leverage to single-strategy funds.

The Securities and Exchange Commission has subpoenaed Goldman Sachs, JPMorgan, Citigroup, and Bank of America over the leverage they provided to Situational Awareness, the AI-focused hedge fund run by 24-year-old Leopold Aschenbrenner. The fund amassed $20 billion in assets in under two years before nearly collapsing in July and dumping its book to Citadel in a 24-hour fire sale. The plumbing here is prime brokerage. Hedge funds do not have $20 billion in cash; they post equity and borrow the rest from Wall Street banks to amplify their bets. When a concentrated, levered portfolio drops, the banks issue margin calls. If the fund cannot meet them, the banks seize and liquidate the collateral, which can drive prices down further and trigger margin calls at other funds. Aschenbrenner has said he will no longer borrow from banks to fund his bets, but the SEC's "information-gathering" probe suggests regulators want to know exactly how much rope the prime brokers handed out, and whether their risk models accounted for the concentration of the collateral.

Bond market dysfunction24 Aughigh

The Treasury may drain its checking account

Using the TGA to suppress long-term borrowing costs is a direct intervention in the price of money, signaling that the administration is unwilling to let the market clear at 5.3%.

The Treasury Department is considering using its $935 billion General Account (TGA) to fund its recently expanded bond buyback program, according to CNBC. Last week, Treasury Secretary Scott Bessent doubled the size of long-end buybacks to at least $4 billion per operation after the 30-year yield hit 5.34%, its highest level since 2007. The TGA is the government's checking account at the Federal Reserve, funded by tax collections. Usually, when the Treasury buys back long-dated bonds, it funds the purchase by issuing short-term bills—a "Treasury Twist." But issuing bills still adds to the total debt supply. Draining the TGA instead injects cash directly into the financial system without issuing new debt, effectively acting as a stealth form of quantitative easing. Citadel Securities called the intervention "financial repression," arguing that preventing Treasuries from clearing at lower prices merely shifts the pressure into the currency market. The dollar index has slipped, and gold is up 4.5% in five days.

The AI capex bubble25 Aughigh

The AI buildout is running on borrowed money

The AI boom has migrated from equity valuations into credit markets, with $410 billion borrowed for data centers this year alone.

The scale of debt funding the AI infrastructure boom is becoming clearer. Broadcom is reportedly in talks to raise up to $100 billion in debt—including a $30 billion junior tranche—to finance AI chips for Anthropic and others. Nebius Group just upsized a convertible bond to $5 billion to fund data centers. And Nvidia is providing up to $105 billion to finance a new OpenAI data center in Ohio. This is not cash-flow financing; it is project finance and vendor guarantees. When a chipmaker guarantees the debt of the data center buying its chips, the revenue is booked today, but the risk stays on the chipmaker's balance sheet. If the AI models do not generate enough cash to service the leases, the data centers default, and the chipmakers are left holding the bag. The market is starting to price this in: while the S&P 500 is near its highs, the debt-funded AI complex is selling off. CoreWeave is down 18% in five days, and Applied Digital is down 13%.

Private credit and BDCs24 Aughigh

A federal probe tests the insurance loop

If regulators force a re-evaluation of the assets backing life insurance policies, the entire private credit funding model could face a capital shock.

A federal investigation into Mark Walter's business empire is spotlighting the mechanics of private-equity-owned insurers. Walter's Guggenheim Investments acquired a string of insurers after 2008, and billions of dollars of premiums were subsequently directed into assets like sports teams and real estate debt. Now, a whistleblower report regarding accounting at a subsidiary has helped spark a wider probe into whether assets were improperly characterized. The core mechanism of the private credit boom is regulatory arbitrage. Life insurers take in premiums and are required to hold safe assets to ensure they can pay out policies. Private equity firms buy these insurers and replace the publicly traded bonds with private credit originated by their own affiliated asset managers. Because private loans do not have daily market prices, they can be marked at par, smoothing out volatility. But if the underlying assets are riskier than their ratings suggest, the capital buffer is an illusion. The only continuously traded instrument on Walter's empire, a $1.18 billion Guggenheim loan, fell to a record low of 72.5 cents on the dollar this week.

Household credit24 Aughigh

Subprime auto delinquencies pass their 2008 peak

The consumer is not universally strong; the bottom quartile is exhibiting stress levels unseen since the Great Financial Crisis.

The 60-day delinquency rate on U.S. subprime auto loans reached 6.13% in July, the highest July on record dating back to 1994, according to Fitch Ratings. Another cut of the data puts the figure at 6.9%, blowing past the 2008 crisis peak of 5.0%. Meanwhile, prime auto delinquencies sit at just 0.49%. This is a story about the bifurcation of the U.S. consumer. High-income households are insulated by fixed-rate mortgages and rising asset prices. Lower-income borrowers are being crushed by inflation and higher interest rates. Subprime auto loans are often packaged into asset-backed securities (ABS) and sold to institutional investors. While the senior tranches of these securities are protected by overcollateralization, the equity and mezzanine tranches absorb the first losses. Ally Financial, a major subprime auto lender, is down 2.6% over the last five days, and the broader regional bank ETF (KRE) is down 3.4%.

The AI capex bubble24 Aughigh

Japanese households take the OpenAI duration risk

When bank credit committees decline AI-related duration and retail investors take it instead, the loss-absorbing layer of the boom moves to households who cannot mark it or sell it.

SoftBank plans a ¥1 trillion ($6.3bn) retail bond sale — the largest by any issuer in Japan — to fund its commitments to OpenAI. Seven-year paper, indicative coupon 4.3% to 4.9%, pricing 4 September, expected to be rated A by Japan Credit Rating Agency. It is SoftBank's third retail bond this year, after ¥418bn in April and ¥260bn in June. The shares fell 5.3% in Tokyo on the news (Bloomberg). The interesting part is why it is a retail deal. "Banks are finding it difficult to take on the risk given weak deposit growth, the credit rating and the seven-year duration, leaving the deal more reliant on retail investors," Yuuki Fukumoto of NLI Research Institute told Bloomberg. Read that plainly. A credit committee looked at seven years of unsecured holdco exposure to a company whose principal asset is a $60bn-plus commitment to OpenAI, and declined. The same risk was then repackaged at a headline yield a Japanese saver has not been offered in a generation and sold to people who will hold it to maturity because there is nowhere liquid to sell it. This is the migration thesis in a single transaction. Risk does not disappear when a bank passes on it; it relocates to a holder with no capital requirement, no mark-to-market discipline and no risk committee. SoftBank's OpenAI stake is illiquid and unlisted. The bondholder's recovery depends on SoftBank's ability to monetise or refinance it in 2033. The filing does not say what happens if it cannot. Context for the coupon: Alphabet sold $25bn of bonds this month against a $205bn 2026 capex plan, and Bloomberg data cited this week put total borrowing for data centres and AI at more than $410bn in 2026 so far.

Private credit and BDCs24 Aughigh

The only price on a private empire keeps falling

Private-equity-owned insurers hold assets nobody has to sell, so the single traded loan becomes the market's only referendum on whether the marks are real.

A $1.18bn first-lien loan issued by GIH Borrower LLC, the financing entity of Guggenheim Investments, was indicated down more than five cents on Monday at 72.5 cents on the dollar — the lowest since it was priced in November 2024 (Bloomberg). The sequence matters. Guggenheim reported second-quarter revenue down 38% year-on-year, with one measure of earnings collapsing 77%. The loan slid. The firm convened a call with lenders to explain that the drop was largely a delay in recognising advisory fees at Guggenheim Private Investments — the subsidiary whose accounting was the subject of a whistleblower report that helped trigger a federal probe into Mark Walter's empire. The loan briefly recovered. On Monday, Anne Walsh, the asset-management arm's chief investment officer, said on Bloomberg Television that the firm had "engaged with our auditors and we feel that the accounting treatment was appropriate." The loan then made a new low. None of Walter's businesses trade on a public exchange. That is the whole point of this dispatch. When an empire is private — insurers, asset management, sports teams, real-estate debt, loans to Carvana and Wendy's — the only continuous, adversarial price anyone can observe is a syndicated loan quoted by dealers. So the loan is doing all the price discovery for a structure that also sits behind insurance policies. 72.5 is not a default price. It is a price that says lenders do not believe they can verify the numbers. The New York Times reports the federal investigation is examining whether some of Walter's companies improperly characterised tens of billions of dollars of assets. That is a question about marks, and marks are the one thing this business model requires you to take on trust.

Hidden leverage and shadow banking24 Augmedium

Gates are working. That is not the same as fine.

A gate converts a liquidity problem into a duration problem for the investor and postpones the moment anyone finds out what the loans are worth.

Two numbers from this week's private-credit reporting sit oddly together. Redemption requests at the ten largest non-traded BDCs averaged 13% of assets in Q1 2026 and 14% in Q2, against typical quarterly gates of 5%, forcing most managers to restrict withdrawals. And listed BDC shares are at or near their period highs: Blackstone Secured Lending is 0.0% off its high, Ares Capital 0.74% off and up 6% in twenty days, FS KKR up 11%. One widely circulated commentary item claims listed BDCs trade "around 60 cents on the dollar" with quality names at 30% discounts to net asset value. We cannot reconcile that with the tape in front of us. Discounts to NAV exist across the sector, but the listed vehicles are not being sold; they are being bought. The mechanism worth understanding is the gate itself. A non-traded BDC promises quarterly liquidity against loans that take five years to repay. When requests run at 13–14% and the gate is 5%, the manager is not forced to dump loans to raise cash — which is genuinely good news, and the reason this is not a 2008-style run. But three things follow. The queue lengthens, so an investor who asked to leave in March is still in. The marks are never tested, because no loans change hands. And the portfolio left behind is progressively less liquid, since managers meet redemptions with the easiest assets first. Separately, payment-in-kind is now 11% of BDC portfolios, $15.7bn at cost, and the median non-accrual rate across the twenty largest listed BDCs rose to 2.8% of cost in Q2, the highest since 2017. PIK income is interest a borrower could not pay in cash. Both numbers describe the same thing: deterioration that shows up in accruals long before it shows up in a price.

The AI capex bubble24 Augmedium

Now they are selling the electricity

Power contracts were the most concrete evidence that AI demand was real; the market is now discounting them faster than it is discounting the chipmakers.

Last week the split in the AI complex ran along the financing line: debt-funded names fell, cash-funded ones did not. This week the selling moved upstream to power. Talen Energy is down 14.4% in five days and 19.5% off its high. Vistra is down 7.2% in five days and 13.6% in twenty. Core Scientific fell 14.2%, Applied Digital 13.7%, IREN 11.4%, CoreWeave 18.6%, Nebius 21.6% — the last of those after selling a $5bn convertible, one of the largest on record. Over the same five days Microsoft rose 1.6%, Alphabet 1.2%, Amazon 0.3%. High-yield spreads sat at 270bp and VIX at 15.85. The merchant power names were re-rated over the past two years on a straightforward premise: data centres would take every spare megawatt at prices set by scarcity, on contracts long enough to underwrite new generation. That premise is a derivative of the tenants' financing. If a neocloud's cost of capital rises, the megawatt it signs for in 2029 gets signed later, smaller, or not at all. Nvidia's reported decision to raise server prices by more than 15% on systems shipping early next year, driven by memory costs, pushes the same way. A neocloud buys GPUs with debt and rents them by the hour. If the equipment cost rises 15% and the rental rate is already contracted, the extra cost comes out of the equity — or, eventually, out of the lender. What we cannot tell from the tape is how much of the power selloff is AI demand doubt versus power prices or rates. It is not rates: long Treasuries rose 1.5% over the same five days. Nvidia reports on 26 August, which will resolve some of it.

Crypto and TradFi contagion24 Aughigh

Twelve people, half a trillion dollars, SpaceX perps

If a twelve-person offshore venue becomes the reference price for private companies, illiquid marks across the fund industry inherit its liquidity.

Trade.xyz, a market built on the crypto exchange Hyperliquid and run by a team of about a dozen, has generated roughly $500bn of trading volume since launching in October and accounts for more than 99% of activity on Hyperliquid's third-party market system, HIP-3. It lists perpetual futures on crude oil, precious metals, stock indexes and companies that have not yet gone public, including SpaceX (Bloomberg). When the US and Israel struck Iran on a Saturday in February and the oil futures markets were shut, traders there priced the conflict for hours before conventional venues opened. A perpetual future has no expiry. Instead, a funding rate paid periodically between longs and shorts pulls the contract's price toward a reference price. For crude oil that reference is a real, deliverable market, and the mechanism is disciplined by arbitrage. For a pre-IPO company there is no such thing. There is no continuous spot price for SpaceX shares, and no arbitrageur can deliver one against the contract. The price is whatever the positioning of the participants says it is, anchored to an index the venue itself defines. That is not a criticism of the design — it is the design. But it means the instrument is a betting market wearing the clothes of a futures market. The reason this matters beyond crypto is what happens if those quotes start being used. Venture funds, holding companies and secondary marketplaces all mark unlisted stakes with considerable discretion. A visible, continuous, 24/7 price is enormously attractive to anyone who has to justify a mark. Trade.xyz and Hyperliquid are lobbying US regulators to allow pre-IPO perps onshore. Hyperliquid does not currently serve US traders.

Household credit23 Augmedium

Student loans quietly went to 10.6%

Household stress is now concentrated in the two loan books — student and subprime auto — where the losses land outside the banking system entirely.

The number that moved this week and nobody wrote about: the share of student-loan balances 90 or more days delinquent rose to 10.6% in Q2 2026, from 10.3% in Q1, on New York Fed data. Alongside it, auto debt transitioning into serious delinquency hit 3.00%, the highest since 2010, and Fitch's July subprime auto 60-plus-day rate came in at 6.13%, a record for the month, against 0.49% for prime — a ratio of more than twelve to one. A warning about the credit-card figures circulating this week. Some summaries report card delinquency at 6.97%, others at 12.8% or 12.9% of balances. These are not the same measure: one is the rate at which balances transition into serious delinquency over a year, the other a stock of balances 90-plus days past due. Sources have been mixing them. Treat any single card number you see this month with suspicion unless the definition is stated. On the loss channel, the structure is what matters. Subprime auto paper mostly sits in asset-backed securities held by insurers, funds and, increasingly, private-credit vehicles — not on bank balance sheets. Deep-subprime lenders write loans at rates wide enough to absorb 6% delinquency provided recoveries hold, and recoveries depend on used-car values, which none of these releases measure. That is the variable we would want and do not have. The equity market has made its call. Credit Acceptance is up 8.8% over twenty days, Oportun 32.9%, Synchrony 6.8%. Ally is down 2.1%. Either the losses are already in the price at generous assumptions, or the recovery rate is about to matter a great deal.

The dollar, gold and reserve status24 Aughigh

The pressure has to come out somewhere

When a government suppresses the price of its own debt, the adjustment does not disappear — it relocates to the currency, and gold is where you watch it happen.

Citadel Securities has put a name to what the Treasury is doing. Nohshad Shah, its head of EMEA fixed-income sales, wrote to clients that Scott Bessent's expanded buyback programme "amounts to financial repression at the margin," and that "preventing Treasuries from clearing at lower prices does not eliminate that pressure. It merely shifts it elsewhere." Elsewhere, in Citadel's argument, means the currency: lower yields reduce the appeal of the dollar, a weaker dollar raises import prices, and the inflation that pushed yields up in the first place gets worse (Bloomberg). The funding question is the interesting plumbing. Dealers assumed Treasury would pay for the doubled buybacks — $4bn per operation, 10- to 30-year maturities, 9 September to 4 November — by issuing bills. That is a duration swap: the government retires long debt and takes on rollover risk instead. Two senior officials have now told CNBC the near-$1tn Treasury General Account is available for it. The TGA stood at $935bn on 20 August, against the previous administration's stated target of $550–600bn (CNBC). That is a different animal. The TGA is cash sitting at the Fed. Spending it down moves money from the government's account into private hands and adds bank reserves. It is not quite money-printing, but it is the fiscal authority adding liquidity while buying its own long bonds, and our reading is that the market has noticed. The evidence is in the prices. The 30-year hit 5.34% on 21 August, its highest since 2007, having erased the intervention rally within a day. Gold is up 16.9% in twenty days to a record, with the broad dollar index down 2.19%. Bonds rejected the policy. Bullion accepted it.

Crypto and TradFi contagion24 Aughigh

A bitcoin company that now needs dollars

Digital-asset treasury companies convert a volatile asset into fixed dollar obligations, and the reserve tells you which side of that trade is currently under strain.

Strategy has created a second cash pool. The new one, called USD Cash, holds $1.59bn and can be spent on bitcoin, preferred dividends, debt payments, securities buybacks or general corporate needs. It sits alongside the original reserve — now $5.1bn — which without board authorisation may be used only to pay preferred dividends and interest on debt (Bloomberg). The mechanism worth understanding is why a company whose entire thesis is holding a volatile asset forever needs $6.7bn of ringfenced dollars. Strategy's preferred stock and converts are hard obligations: fixed cash out, on fixed dates, denominated in a currency it does not hold. The original flywheel solved this by issuing common stock above the value of the bitcoin behind it and using the proceeds to buy more bitcoin, so the coin count per share rose and the dividends were funded out of an expanding capital base. That only works while the premium exists. Bloomberg reports the valuation premium is still well below earlier-cycle levels and the financing model "remains impaired." So the company sold roughly $2bn of common stock in a week and repurchased $136.4m of its Stretch preferreds. Nansen's Nicolai Sondergaard put the trade-off plainly: "dilution in exchange for flexibility" — the equity raise strengthened the balance sheet but "did not immediately increase Bitcoin-per-share exposure." The share price disagrees with all of this. MSTR is up 25.6% in five days and sits at its period high, against bitcoin's 9.25%. The market is paying roughly three times the coin's move for a company that has just built a dividend reserve. One of those two readings is wrong, and the filing is the one with numbers in it.

Hidden leverage and shadow banking24 Augmedium

Two trillion of Treasuries, bought with borrowed money

The largest marginal holder of US government debt is a levered relative-value trade financed overnight, and the plumbing meant to make it observable has stalled.

The Office of Financial Research published a number on 19 August that deserves more attention than it got: hedge funds now hold about $2.0tn of cash Treasuries against $1.4tn of short Treasury futures. A significant share of that is the basis trade (OFR). Estimates of the trade's size from the CFTC, BIS, Fed and sell-side research span $350bn to $1.5tn — a range wide enough to tell you nobody actually knows. The trade itself is simple. Cash Treasuries trade slightly cheaper than the futures that promise delivery of them. Buy the bond, sell the future, collect the difference. The difference is tiny — a few basis points — so you scale it. You finance the bond in the repo market at a haircut of a fraction of a percent, the lender rehypothecates the collateral into another repo, and the short futures position needs only initial margin against a large notional. The OFR describes exactly this: chains of borrow-and-buy transactions that let funds "expand their positions dramatically." What breaks it is not credit. It is repo rates spiking or margin requirements rising abruptly, at which point the position is unwound by selling cash Treasuries — into the long end that the Treasury is currently spending money to support. Separate Form PF-based analysis puts large hedge funds at roughly 8.5% of privately held Treasuries and $4.0tn of gross exposure. One detail cuts against comfort. Central clearing of Treasury repo plateaued in the first quarter of 2026, and since May the SEC has not required affiliate repo transactions to be centrally cleared provided the affiliate clears everything else. The reform designed to make this leverage visible has stopped advancing.

Household credit23 Augmedium

Subprime auto is past 2008. The lenders are up.

The consumer data and the consumer-lender share prices are telling opposite stories, and only one of them can be right about how much loss is already provisioned.

Fitch's subprime auto 60-day delinquency index printed 6.13% for July, the highest July reading in a series that starts in 1994. One syndicated cut of the Fitch and Equifax data puts the rate at 6.9%, against a 2008 peak of about 5.0%. Prime auto is at 0.49%. The New York Fed's Q2 2026 household report has the share of auto debt transitioning into serious delinquency at 3.00%, the highest since 2010, with 5.49% of auto balances 90+ days late — and record originations of $211bn in the quarter. Credit-card balances 90+ days delinquent are around 12.8%, student loans 10.6%, and 4.7% of all household debt is in some stage of delinquency. Now look at the lenders. Credit Acceptance, which sits at the deepest end of subprime auto, is up 5.4% in five days and 8.8% in twenty, 5.4% off its period high. Oportun is up 32.9% in twenty days. Synchrony is 1.2% off its high. Only Ally, down 2.7% in five days and 8.7% off its high, looks like a stock reading the same data. The two facts are reconcilable, and it is worth being precise about how. Deep subprime lenders do not price loans expecting them to perform; they price a forecast of collections at origination and earn a spread over it. What hurts them is losses coming in worse than forecast, not losses being high. Record originations also flatter the ratio: a fast-growing loan book puts new, not-yet-delinquent paper in the denominator. So the equity market's view is that the loss rate is already in the price. That may be right. It is a bet on the accuracy of forecasts made two and three years ago, at a delinquency level nobody in the series has seen.

The AI capex bubble24 Augmedium

One quarter of capex, raised in one convertible

The financing for AI infrastructure is still available at scale — the price is being paid in the equity of the companies raising it, not in the coupon.

Nebius upsized a convertible bond sale to $5bn on 20 August, described by Reuters as one of the largest on record, to fund data centres and its AI platform. The company spent $5.7bn on capex in the quarter (Reuters). A record convertible therefore buys about eleven weeks of building. Since the deal, the stock is down 21.6% in five days and sits 24.1% below its recent high, still up 12.2% over twenty days. Convertibles are cheap for the issuer because the buyer is paid in optionality rather than coupon, and a large part of the buyer base hedges that optionality by shorting the underlying equity. The prospectus is not in front of us and the reporting does not identify who bought, so we cannot say how much of the fall is hedge flow and how much is sentiment — but the direction is what you would expect, and it is the cost that does not appear in the coupon. The wider number: companies have borrowed more than $410bn this year for data centres and AI investments, on Bloomberg's compilation. Hyperscaler debt issuance alone was $220bn by 10 August, per BNP Paribas. Carson Group puts incremental debt at roughly 32% of hyperscaler capex by mid-2026, against 9% in fiscal 2024. The split in the tape is the same one we flagged last week and it has widened. Microsoft is up 1.6% in five days, Alphabet 1.2%, Amazon 0.3%. CoreWeave is down 18.6%, Core Scientific 14.2%, Talen 14.4%, Applied Digital 13.7%, Nvidia 7.4%. Capital funded from operating cash flow is holding. Capital funded from someone else's balance sheet is not.

The AI capex bubble24 Aughigh

The cost of the asset just went up 15%

Every loan written against AI hardware was sized against a cost per unit that has just moved against the borrower, not the lender.

Bloomberg reported on 22 August, and Reuters and CNBC confirmed, that some of Nvidia's biggest customers have been told the price of servers built around its AI chips will rise more than 15% in many cases, driven by soaring memory costs. The increases apply to Vera Rubin and Grace Blackwell systems shipped early next year. This is not a headline about sentiment. It is a change in the unit cost of the collateral underneath the entire AI financing complex. Consider how a neocloud works. It signs a multi-year contract to supply compute at an agreed price, then borrows against that contract to buy the hardware. CoreWeave closed a $2.6bn delayed-draw term loan at SOFR plus 5.5% on 10 August; Lambda raised $926m of term loan B at S+300. Both were sized against an assumed cost per rack. If the hardware costs 15% more and the revenue contract does not move, the equity cushion in each project thins — and the equity cushion is what the lender is relying on. For the hyperscalers it is simpler and less dangerous: Alphabet's $205bn 2026 capex plan buys less than it did when the plan was written, and Alphabet has the cash flow to absorb it. The tape has already sorted the two groups. Over five days Nebius is down 21.6%, CoreWeave 18.6%, Core Scientific 14.2%, Applied Digital 13.7%, IREN 11.4%. Over the same week Microsoft rose 1.6%, Alphabet 1.2%, Amazon 0.3%. What we do not know is who absorbs the increase — whether neoclouds can pass it through, eat it, or delay deliveries. Nvidia reports on Wednesday with revenue expected above $92bn, and the gross-margin line will be the first place the answer shows up.

The AI capex bubble24 Aughigh

Broadcom's guarantees are now visible in its own spread

A guarantee that never appears as debt is still a claim on the guarantor, and the CDS market is the only place currently pricing it.

The price of five-year credit protection on Broadcom has climbed 28 basis points so far in August — more than either Oracle or SpaceX — and the yield on its 5.15% bonds maturing 2031 is up about 14bp over the same period, Bloomberg reports. The reason is not Broadcom's operating business. It is what Broadcom has promised. The company is in talks to raise more than $60bn of debt for an AI chip financing package expected to benefit Anthropic, and could end up guaranteeing part of a senior-secured tranche. That would follow an agreement earlier this year to backstop most of a $35bn debt package, in which Apollo and Blackstone financed the purchase of custom AI chips to lease to Anthropic. Here is the plumbing. Broadcom sells the chips. Private-capital investors put up the money to buy them. The chips are leased to a customer that is loss-making at scale. And Broadcom promises the investors that if the lease payments do not arrive, it will make them whole. The debt sits in the vehicle, not on Broadcom's balance sheet, so it does not appear in its leverage ratios. The obligation is real all the same. "The rise in Broadcom's CDS seems to me is more specific to their balance sheet than overall angst over AI investment," Impax portfolio manager Tony Trzcinka told Bloomberg, adding it was probably about expectations of further guarantees. This is the running argument of this site in its purest form: risk moved into a structure where nobody has to mark it, and the first place it becomes visible is a derivative market that only professionals watch. Bloomberg counted roughly $70bn of such shadow credit backstops across the sector on 15 August. The equity market has not repriced Broadcom. The credit market has.

What would change our mind

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

High-yield spreads breaking above 350bp, or investment grade above 110bp, from today's 270bp and 81bp — the index level that would tell us credit has stopped believing the marks.

Would move the number

A non-traded BDC marking a portfolio down toward Saba's 35% discount, or suspending redemptions outright rather than gating them.

Would move the number

An SRT tranche taking a first loss, or a protection seller failing to post — the test of whether bank capital relief was real.

Would move the number

Strategy's premium to its bitcoin holdings collapsing while preferred dividends remain due, forcing coin sales rather than share sales.

Would move the number

A Treasury buyback operation or coupon auction clearing badly, which would mean official demand is no longer enough to hold the long end.

Would move the number