An announcement before we begin. Supply Chains ships as its own post this week.

The Commodity Disruption Index was always meant to sit inside Supply Chains rather than be the whole of it. Alongside it now runs the Sector Register — all 16 industries under watch, walked in turn, whether or not anything in them has broken. A chain does not have to be in crisis to be moving. It runs alphabetically. Read it through, or jump to the one you came for.

The framing the last issue was built on has been superseded, by data that arrived within a fortnight of publication.

The claim was a genuine two mandate bind — a labour market cracking without the disinflation that would license a cut, so a Fed that could neither ease into the weakness nor tighten against it. Three supports under that reading have since failed. Waller rebutted the participation datum technically: the fall sat in the noisy 25 to 34 cohort, broader underutilisation measures including the marginally attached actually declined, and wage growth near 3.5% is consistent with the target rather than against it. Morgan Stanley then found that 0.4pp of the 0.8pp annual participation drop is a one time population control adjustment — half the figure I leaned on hardest is an artefact of how the survey is benchmarked, not people leaving work. And claims printed 208K: no layoff impulse in a market that had supposedly cracked. Meanwhile the inflation side resolved the other way, June core coming in at 0.0% on the month against a consensus that was universally sticky. A July hold is now better than 90% priced.

What is left is not a bind but an inflation question, over a labour market soft in hiring and firm in firing.

The distinction is between facts and weight. The 57K payroll level, the 74K of downward revisions, participation at 61.5% — all still true, none shown otherwise. What was overstated is the structural load I put on them. The method failure underneath is the more useful admission: I named the correct trigger, the 14 July print, 48 hours in advance, then leaned past it on the strength of a single speech the day before. The market made the identical error on the identical timetable — July hike odds ran from under 10% to roughly half on Waller's Monday remarks, then collapsed on Tuesday's print. When the analysis and the consensus are wrong for the same reason at the same time, the analysis is adding nothing.

One thing I decline to take as corroboration, and it cuts against me rather than for me. The headline accompanying that cold core — down 0.4%, petrol off 9.7% — is a photograph of an economy that no longer exists, priced off a crude curve near $70 that reflected the June memorandum. That instrument is now void in practice: Iran declared yesterday that it has suspended every obligation under it and is implementing none of them. But the print was never sound even while the memorandum stood — a memorandum does not clear a strait, does not reverse damaged infrastructure, and physical transit lags any announcement by weeks. Those energy components were never true, and I won't bank the half of the print that flatters the disinflation case while rejecting the half that doesn't. Core excludes energy, which is why the shelter led deceleration survives the objection and the headline does not. Brent is already back above $85.

One widely circulated reading appears to restore the finding and does not. The number of Americans outside the labour force hit a record 105.8M in June, rising 832K on the month. But a record level in a growing, ageing population is close to mechanical, the monthly jump runs into the same population control adjustment that hollowed out the participation figure, and the measures that count people who actually want work fell in June rather than rising. It is the same datum from the other side of the identity, and I am not going to reach for it a fortnight before the print I said would settle this.

The tests named last time are reporting mixed: a fund has permanently gated and begun liquidating, the gold floor absorbed its drawdown, and Korean guidance is going against me — each scored properly below. The restoration path is dated. A genuinely weak 1 August payroll — soft headline, the 25 to 34 cohort still falling, revisions extending down rather than reversing — restores the finding, because the rebuttal that unseated it is itself a prediction that June reverses in July.

The deceleration that shut the July hike door came from a single component, and it is the one carrying more than a third of the index. Shelter rose 0.1% on the month, the smallest increase since January 2021, and that line is most of what took core to zero. It is also the mechanism Williams had named publicly 5 days earlier — moderating rents feeding through a series that lags them — which matters, because it means the softness was forecastable rather than a surprise.

What it is not is a broad disinflation, and the Bureau's own breakdown of core services says so in a way that cuts both directions. Shelter's 12 month rate is 3.3%, but its 3 month annualised pace is 4.07% — the component that produced the cold headline is running faster over 3 months than over 12, which makes June the outlier rather than the turn. Other personal services are annualising above 13%. Against that, transportation services are annualising at −2.4% and medical care at 1.6%, both well beneath their own yearly rates, and that is real disinflation in two large categories rather than a rounding artefact. Half the 6 categories the Bureau publishes sit at or above 3% on the year. Waller's own breadth argument, on a finer cut than the published 6, puts close to 70% of core services categories above 3% on both horizons — worth citing precisely because it comes from the committee's most consistent dove, and worth flagging that his category set is his rather than one I can reproduce from the release. The picture is dispersion, not a turn.

The staleness problem runs deeper than a single print, and it runs in two directions at once. Backward, the official forecasts now in circulation were built on an assumption that has since broken: the Fund's July outlook was finalised on 10 June, and its entire baseline assumes the Strait begins reopening in mid July and normalises by early 2027. The regional line is where the assumption is visible as a number — Middle East growth projected to collapse to 0.7% this year and rebound to 6.5% next. That rebound is the reopening, expressed as arithmetic, and it is the most exposed figure in the document. Published growth of 3.0% and inflation of 4.7% are therefore a floor rather than a centre.

Forward, the same distortion runs the opposite way. The tariff cliff on 24 July is pulling import volume into July — record port throughput now, an air pocket in August — so July's trade, inventory and possibly output data arrive pre-loaded and August's arrive hollowed. One set of numbers describes a June that has ended; another describes a July that has been borrowed from August. Neither measures what it appears to measure, and they are running simultaneously in opposite directions.

Underneath both, the pass through is queued rather than spent. Among firms that actually paid tariffs over the past year, 47% of service companies and 44% of manufacturers plan further price increases, roughly a third of them within 6 months — and those surveys predate the reinstatement measures. Small firms are further along than planning: 38% raised their average selling prices in June, the highest share since January 2023 and a fourth consecutive monthly increase, against a long run average near 14%. In the same survey, 21% named inflation as their single most important problem, the most since October 2024 and 10 points higher than a year ago. That reading is from the month petrol was cheapest. Import prices run 7.1% on the year.

Which leaves the committee genuinely split, and widening at both ends rather than converging. The June minutes put a hike faction in the record: a few participants saw a case for raising immediately before joining the unanimous hold, and the dissent that once ran toward cuts has reappeared pointing the other way. The dot split was 9 to 8 with no lean. In the same window the Vice Chair drifted audibly dovish, seeing a positive near term inflation outlook and then looking through the war's renewal entirely. Staff raised their inflation forecasts and cut growth, naming two causes — tariffs, and the AI buildout. That second one is the notable line, because it is the Fed's own staff writing that strong demand for AI infrastructure would likely sustain upward pressure on prices for technology products and electricity.

The Chair's behaviour in that setting is the part I would not have predicted. Warsh has a documented preference for trimmed inflation measures, and the trimmed gauge is currently running a full point below core; a statistical revision at the end of September will shave roughly another 0.2pp off the conventional number for technically sound reasons. He testified to both chambers within 48 hours of the coldest core print of his tenure, holding a cold producer print as well — the clearest dovish opening he has had — and declined all of it. He acknowledged the energy driven improvement and then discounted it explicitly: strip out energy, and inflation is still too high, and has been for more than 5 years. He named the data centre buildout as an inflationary force in his own report to Congress and signalled no path.

There is an inversion sitting inside that is worth holding on to. Warsh's stated personal view is that AI is eventually disinflationary through productivity. His committee, his staff and his own report all say the opposite for now. If a dovish turn arrives from this Chair, it is likelier to come through that belief than through any softening in the data — and his own report notes that one of the Fed's numerical policy rules currently calls for a higher rate than the band he is holding.

The meeting lands on 28 and 29 July, 9 days from now.

The release most likely to break this argument arrived in mid July and, on its face, broke it. Total foreign holdings of US Treasuries rose $18bn in May to $9.37tn, the second highest on record and $549bn above where they stood a year earlier. The seven largest financial centres added $33bn between them, to an all time high of $3.24tn. China, the sovereign whose retreat gets quoted most, bought $8bn. Any argument resting on foreign demand retreating in aggregate is simply wrong, and I retire that phrasing where I have used it.

That the aggregate hides a recomposition is not new here. The Composite Buyer framework set out in April held that the official retreat was a decade long structural trend masked by private inflows, and that the strength was never demand but substitution. Two things in the May data are new, and one of them corrects a reading of my own.

The first is that the official sector has stopped drifting and started selling. Net official purchases of Treasury notes and bonds ran −$26.8bn across 2024, −$28.3bn across 2025, and −$39.3bn in the 12 months to March, with −$37.9bn of that in March alone. Official bill buying collapsed 86%.

The second is the identity of the replacement, and it is not the one I described. Japan — the largest foreign holder and the most official, most real money of them — sold $67bn in May, taking its position to the lowest since May 2025. The United Kingdom rose to a record $949bn. When this account listed the top 3 foreign holders in the spring, the United Kingdom sat in that list as a sovereign holding, which it is not: it is the custody centre through which London hedge funds run their Treasury basis trades, and the private bid filling the official hole is, on Brad Setser's work at the Council on Foreign Relations, substantially those trades — cash bonds funded in the repo market. The bid that replaced the central banks in the long end is not the domestic cash pool I catalogued; it is offshore and it is borrowed.

That last clause is the argument. A basis trade is not a view on America; it is a spread between a cash bond and a future, earned on borrowed money, and it exists only while the borrowing is cheap. Hedge fund repo borrowing has reached $2.5tn, more than doubling in 2 years, with 73.8% of it done at zero or negative haircuts — meaning the lender is taking no cushion at all. Cayman domiciled funds hold $1.85tn of Treasuries, a trillion more than in 2022, at leverage the Federal Reserve puts above 18 to 1 at the largest of them. Alongside it, banks now hold $1.14tn of loans to nondepository financial institutions, the fastest growing segment of their loan books since the financial crisis at a 21.9% compound rate over 14 years, nearly 3 times the next fastest; the FDIC began disaggregating the category in its own reporting in December 2024 because it had outgrown the line it was sitting in.

So when 2 long end auctions tailed in the same week — the 10 year clearing 4.580% against 4.538% prior, the 30 year through 5% to 5.058% — the tail is not noise. It is the price that book charges to take paper the official sector is dumping.

The obvious objection is that the funding market looks calm. The overnight secured rate ran 8 to 10 basis points below the effective funds rate through both tails, on $3.1tn of daily volume, drifting lower rather than higher as the auctions cleared — no scarcity anywhere in it. That reading is correct and it disposes of one version of this argument: the tails were not a funding event. What it does not dispose of is the dependency. Cheap repo is not neutral here; it is the enabling condition of the bid absorbing the issuance, which makes the calm the mechanism working rather than evidence the mechanism isn't needed. The market has swapped a price insensitive official buyer for a price sensitive, leverage dependent, repo funded one, and if that funding ever tightens the marginal buyer does not simply bid less — it is forced to sell into the market it was holding up. March 2020 is the template, and it took the Federal Reserve buying at scale to stop it.

Against that structure sits the supply, and the honest way to state fiscal pressure is as a rate of change rather than a level. Gross federal debt reached $39.39tn on 6 July, $3.16tn higher than a year earlier — the level is the number that gets quoted, and the extrapolations built on it are not worth much. The coupon is what matters: the average interest rate across all marketable debt is 3.348%, against bills clearing at 3.760%, 10 year at 4.580% and 30 year at 5.058%. Every security that matures refinances above the one it replaces, with $6.59tn of bills rolling continuously. Net interest ran $104bn in the single month of June and $827bn for the fiscal year to date, $78bn more than the same point last year, heading for the largest annual figure on record — second only to social security among all outlays, and above Medicare, health and defence. That is a floor that rises mechanically even if yields never move again. And the revenue side moved the wrong way in the same month: net customs duties turned negative at −$26bn as court ordered refunds flowed, a swing of about $53bn in a line that had been running around +$27bn a month.

Two dates test all of it. The August refunding announcement is where the deferral of coupon increases expires against a curve that has already repriced end to end, and Treasury must either raise coupons or accept whatever the leveraged bid charges. Behind it, the clearing mandate arriving for cash Treasuries late this year and repo in mid 2027 is a scheduled margin tightening aimed precisely at the buyer now absorbing record issuance.

The condition that would break this reading is specific and observable: if the secured rate moves sustainably above the effective funds rate while the tails persist, the funding story reopens in the opposite direction, and I would be reading a liquidity problem where I had insisted there was a structural one.

Two ratios describe the same rotation from opposite ends. Foreign official holdings have fallen to 12.5% of total marketable Treasuries, the lowest this century and roughly 24 points below the 2009 peak. Gold has climbed to 27% of global central bank reserves against 22% for Treasuries — the European Central Bank's June read, up from 20% and 25% respectively a year earlier. One asset is being vacated and the other filled, by institutions with the same mandates and, in several cases, the same balance sheets.

The Treasury side is a story about pace rather than exit. Since 2009 marketable debt has grown by $23tn, or 379%, to $29.1tn; official holdings have grown $1.5tn, or 63%, to about $3.9tn. Nobody had to sell for the share to collapse — issuance simply outran a bid that was never going to keep up with it. What has happened more recently, as the flow data shows, is that the official sector stopped even trying.

The gold side is the reciprocal, and the reporting understates it by construction. The official tallies capture the least: the People's Bank reported around 15 tonnes in June, its heaviest month in at least 2 and a half years and its twentieth consecutive month of accumulation, taking reserves to a record 2,346 tonnes — about 9% of its foreign exchange reserves. Trade and market data capture more. Goldman Sachs puts China's May purchases on the London over the counter market at 48 tonnes, nearly 5 times what the central bank reported for that month and the largest identifiable single buyer month in over a year. That figure is an estimate and I hold it as one; what is not in doubt is the direction of the gap, which runs the same way every time anyone measures it. May sat in the middle of the drawdown. The heaviest accumulation was happening while the price was falling — which is the whole argument about who is buying, made without needing the estimate at all. Breadth sits underneath it: the World Gold Council's survey has 89% of central banks expecting global gold reserves to rise over the coming year and a record 45% intending to add to their own, answers given after the drawdown rather than before it. Set 12.5% against 9% on the same balance sheet and the rotation is legible as arithmetic rather than narrative.

The crossover held through the fall, which is the part that matters for the reading offered last time. Gold sits well below its February high, and the drawdown was real. But gold's share of official reserves rose past Treasuries' while the price was falling, and the accumulation ran through the entire round trip — a buyer that adds at the top and adds again at the bottom is not timing anything. That is what makes a floor hold: it was set by holders for whom the price was never the variable.

What has changed since is the cap. Across three consecutive escalations in July — tanker strikes on the 7th, the ceasefire's collapse on the 8th, the Hormuz declaration on the 13th — gold fell each time, roughly $95 on the last of them, while crude rose more than 5% and the dollar strengthened. That is the inverse of the safe haven reflex, and three instances across three distinct shocks is a pattern rather than an accident. The market is pricing war through the rates channel: oil up, inflation risk up, hike risk up, dollar up, and the asset that pays no coupon marked down against a discount rate that just rose. The cap is not a verdict on the metal; it is the cost of holding it going up.

Which produces a reading that has to run on two clocks, and I want to name the confusion in advance. A dollar rally on an escalation headline is the rates channel working exactly as described — it is not evidence against the erosion, because the two operate on different clocks. Reserve allocation moves over quarters and years; the rally moves over sessions. The clearest illustration landed on a single day earlier this month, when the People's Bank announced a substantial yuan infrastructure package and the dollar rallied hard on the Iran escalation in the same session. Both were true. An analysis that cannot hold both is misreading one of them.

That infrastructure is the leg that gets least attention and may prove the most durable. The package doubled the offshore renminbi business facility to 500bn yuan, raised the Southbound Bond Connect quota to 800bn, put a centralised Hong Kong gold clearing system into trial with a delivery link to the Shanghai exchange, pledged support for yuan denominated commodity futures, and opened cooperation with the London Metal Exchange on adding renminbi elements to commodity pricing. The gold clearing system has gone live with its own benchmark, now carried on Bloomberg and backed by 11 institutions, and Hong Kong's gold imports have surged well above their 2 year average to stock the settlement inventory behind it. A benchmark is a bid for the price formation role London has held for decades. It may not win it. But building one is a different order of commitment from selling a bond.

The honest complication runs the other way, and it belongs in at equal weight. A consortium of more than 140 firms launched a Treasury backed dollar stablecoin at the end of June, and stablecoin reserves are a genuine and growing source of Treasury demand. Settlement rails are being rebuilt in several places at once, and they do not all point the same direction: one of these developments bypasses the dollar and the other deepens demand for it. Both are the dollar's evolving role rather than its decline.

The Hormuz toll episode is instructive for what did and did not survive it. On 13 July Washington declared itself guardian of the strait and demanded 20% of all cargo shipped — roughly $30m a vessel, against the $2m Iran had been reported to be contemplating — then withdrew it 5 hours before it was to take effect, replacing it with trade and investment arrangements with Gulf states. No implementing instrument ever existed, nothing was collected and nothing was priced. What does operate is the Iranian regime, which never depended on the American announcement: transit reportedly charged around $32m per supertanker against up to $2m before, turning the chokepoint into a continuously variable cost instrument.

The claim itself cannot be withdrawn, and that is the part that outlasts the week. Transit passage through an international strait is non suspendable and cannot be conditioned on who is sailing or on payment — the principle the United States has spent decades enforcing by sailing warships through other states' claims, and the precise basis on which Washington rejects Chinese assertions in the South China Sea and rejected the Iranian toll weeks earlier. The United States is not party to the convention that codifies it; it defends the rule as customary law, which is the stronger claim, because customary law binds everyone including the state asserting it. The International Maritime Organization has since restated that there is no legal basis for mandatory straits tolls — an institutional counter that now points at both claimants rather than one.

The reason this reaches a section about reserves is what the escort function actually underwrites. Freedom of navigation is not a service rendered to ships; it is the condition that makes the trading system open to every participant on identical terms, and that universality is one of the legs under the dollar's role as the default currency of physical trade. It is not the only leg — network effects, market depth and pricing convention carry more of the weight — but it is the one that was given a number in public. A commons carries no political condition; a toll is nothing but one. I would not build a forecast on a statement retracted inside a day, but the audience for it was the same set of reserve managers whose diversification is already measurable above, and they were told, briefly and at the highest level, that the terms of access are discretionary rather than principled.

Here is what would prove this wrong. If a transit levy is ever implemented and the observable consequence is a shipping cost and inflation shock with no measurable response in reserve composition over 2 or 3 quarters, then the reserve leg of this argument is wrong and should be retired even if everything else in it holds.

The tension that has sat unresolved in this analysis for 2 months is that the AI price shock is well evidenced on the producer side and absent from consumer prices. The Fed's own staff attribute part of core goods inflation to AI related pressures; contract memory prices are guided up 13 to 18% this quarter; Micron printed an 85% gross margin. And yet core commodities fell 0.1% on the month and rose 0.8% on the year. Something is absorbing the cost between the fab and the shelf.

The missing half is that the price the model layer can charge is collapsing faster than the cost of serving it. Per token inference prices have deflated over roughly 3 years by about as much as personal computer prices did over a decade and a half — a rate of decline something like 5 times the PC cycle — while the inputs those models run on inflate. That is a margin scissor rather than a productivity gain, and it explains the absence: the scarcity rent is captured upstream by whoever owns the constraint, the squeeze lands on the intermediary, and the intermediary has no pass through channel because its own pricing is falling. Nothing reaches the consumer because the middle is eating it.

One caveat limits the claim. The index measuring that deflation is partly quality adjusted — it tracks intelligence per dollar, not revenue per customer — so some of the measured collapse is capability improving rather than pricing failing. The scissor is real; its severity is overstated by how it is built. It closes if per token pricing stabilises while memory contracts keep rising, which is the condition to watch.

The scissor has a direction, and the filings already show it moving. Across the 5 big hyperscalers — Amazon, Alphabet, Meta, Microsoft and Oracle — combined free cash flow peaked near $234bn and has fallen to about $175bn in the most recent reported year, a quarter off the top, with Oracle alone swinging to −$23.7bn. Over the same stretch the 4 semiconductor names they spend with — Nvidia, Micron, Broadcom and Applied Materials — went from roughly $28bn to $131bn, a 4.7 times increase. That is the transfer as realised: not yet an inversion, but a decisive change of direction on both sides at once.

The projection runs further than the record. Bank of America's 12 month forward measure has the hyperscaler basket falling through zero for the first time in a series running back to 2007, while the semiconductor basket climbs past $430bn. Those are consensus estimates rather than reported figures and I hold them as such — but the direction they extend is already in the filings, and the mechanical consequence holds either way: firms whose free cash flow is collapsing cannot fund a capital programme from operations. The buildout has to be borrowed.

It duly is. Off balance sheet data centre lease commitments across the majors now exceed $850bn on roughly 20 year terms, up more than 200% year on year; Oracle has taken on $58bn of new debt and cut 21,000 jobs at a severance cost of $1.84bn; investment grade issuance has run at record volume, with the technology sector's share of it more than tripling to about 20% of the entire market.

Which is where this connects to the Treasury argument above, because the same thing is happening to that paper. 5 year credit default swaps on the largest hyperscalers have widened to around 75bp, close to a 7 year high; strip Oracle out and the remainder sits near 49bp, the highest since at least 2018, with both more than doubling since the start of 2025 and now above their 2022 lows. More telling than the spread is the demand behind the issuance: the new issue cover ratio for hyperscaler bonds — orders received per dollar of paper sold — has collapsed from about 4.7 times in February to roughly 1.7 times in July, against an average of 3.4 times across investment grade as a whole. Roughly half the market's cover, on $194bn issued this year, a record share of total supply. That is the same saturation signature the long end auctions are printing: a demand curve for duration that requires escalating concession as issuance mounts. Two markets and one mechanism, both absorbing record supply from a thinning bid, both charging for it.

The demand side produced its first named casualty in July, and the cause is on thesis rather than incidental. On 13 July, Jim Cramer went on Mad Money and said "Buy some now and add on panic dips" about IBM. It then fell 25.21% the following session, the worst single day in the company's history — worse than Black Monday — erasing about $68bn, after pre announcing a quarterly miss it attributed to clients reprioritising capital in the final weeks of June, redirecting spend away from software and infrastructure and toward servers, storage and memory chips to lock in supply before prices rose further. The memory scarcity that expands Micron's margin contracted IBM's revenue. ASML raised guidance the following day, which is the same event from the other side. I would hold the magnitude lightly until the full release on 22 July, because single session moves of that size often retrace; the cause is what matters.

Meanwhile the trade expressing all of this broke, and the break was not about any of it. SK Hynix completed the largest foreign listing in US history on 10 July, then fell 15.4% in Seoul within days on a single brokerage note about the timing of high bandwidth memory shipments, tripping a market wide circuit breaker as the index fell nearly 9%. Micron dropped 15% in two sessions to more than 30% below its June high. Samsung's record quarter could not lift its own stock. Around $1.5tn of semiconductor value has gone since late June, and none of it was an input event or a demand crack — it was positioning unwinding in the most crowded trade in the market, with Korean retail margin debt at a record and a memory exchange traded fund that launched in April having already passed $25bn in assets, overtaking a Korea fund that has existed for 26 years.

The tape is carrying two different mechanisms at once, and they are easy to conflate. Net foreign outflows from Korean equities were already running heavily by late May, well before July, and are characterised by several houses as mechanical rebalancing driven by index weight effects and single name ownership caps. That is a months long structural rotation. The July move is a leverage unwind. Reading the outflow as a reaction to the crash gets the sequence backwards.

Underneath the tape the physical position is intact but slowing. Server memory remains undersupplied and prices keep rising, but the pace of increase has moderated sharply: contract gains guided at 13 to 18% this quarter against roughly 60% last. The shortage has not eased; only the rate at which it reprices has.

Which brings this to the test the last issue staked itself on, and it is not going our way. The claim was that the disrupted inputs fall on the Korean fabs rather than on Micron, so the asymmetry would express as input driven production cuts in Korea while Micron's guidance stayed clean. Three outcomes were named, only one of which confirmed it. What has happened so far is the third: neither has guided down on inputs. Samsung printed a record quarter and flagged no input driven cut; Micron guided to $50bn with about $100bn contracted; SK Hynix's stumble was about shipment timing, not supply. The constraint did fire — the two Japanese producers of the interconnect gas, together roughly 30% of world capacity, permanently ceased production this month, with contracts locked 70 to 90% higher — but it fired at the price level and has so far been absorbed before reaching output. The honest reading is that the input register was a real risk that has not reached the print. SK Hynix reports later this month, and the second half guidance is the remaining read; if that also comes back clean, the register was noise and I will say so.

The last issue named 4 conditions that would loosen this argument, and the first was gates that keep holding. That one has fired the other way — and the miss has a sharper edge than a simple oversight, because the last issue published on 5 July, more than 4 months after the fact it missed.

The framework had the sponsor right, well before that. In April this account placed Blue Owl at the bottom of a three tier hierarchy: Blackstone meeting redemptions with sponsor capital, the mid tier gating at their caps, and the most concentrated sponsor forced into a partial fire sale into a market that did not want the inventory. What failed in July was a sentence describing the same quarter with the benefit of hindsight, not the framework. Writing on 5 July, I said the gates held across every major sponsor through Q1. Blue Owl's OBDC II vehicle had permanently halted redemptions in late February and begun selling assets to raise cash — a fact that was fully on the record by the time that sentence was written. The claim was too broad by one vehicle, and it was the vehicle the framework had already named as the weakest tier 3 months earlier.

A cap binding for a quarter defers a problem; a permanent halt with asset sales behind it is a fund liquidating into whatever bid exists. Those are different objects, and only the first is the mechanism working.

The second quarter then produced the largest redemption wave the sector has recorded. Investors requested $15.6bn from non traded business development companies, a third consecutive quarterly increase, and 38% of it was met — leaving an unmet backlog of $9.7bn, also a record. Blue Owl held its 5% caps on two further vehicles; Apollo, Ares, Morgan Stanley, HPS, Cliffwater, Monroe and Blackstone all exceeded theirs. Across the first 5 months of the year, $12.9bn was pulled from the wealth channel specifically. The managers themselves have lost about $265bn of market capitalisation since September, with Blue Owl down somewhere between 40 and 67% from its peak depending on where you mark the top.

The honest qualification is that this is still a liquidity event with credit features rather than a credit crisis. Sector non accruals remain low, between 0.6 and 2.6%, and loans are broadly performing in line with the index that tracks them. Goldman's vehicle stayed under its cap at 3.24% while Apollo limited withdrawals after a tender of 16.8%. Some books are being run better than others, and the gap is wide enough that describing the sector as one object would be wrong.

The counter comes from the firm whose data this rests on. Stanger reads the wave as rotation rather than retreat: capital leaving credit for real estate and infrastructure, where fundraising is up 33% on the year and infrastructure alone up 61%, with credit's share of alternative fundraising falling from over half to about a third. On that reading the caps are not a failure but the structure doing its job — investors exiting on defined terms without forcing distressed sales inside the funds. The evidence is real and the dispersion above supports it. What it does not explain is the fund that stopped being a gate altogether, or the $9.7bn that did not get out.

Nor would I read the low non accrual rate as health, because of how the deterioration is being recorded. The private credit default rate reached a record 6.0% for the 12 months to April, and 94% of the downgrades underneath it are distressed exchanges — maturity extensions and payment in kind toggles that defer the loss rather than crystallise it. Interest coverage across the leveraged loan index has fallen to about 4.6 times from 6 in 2022. A book where 9 tenths of the deterioration is resolved by moving the maturity date is not one where credit is fine; it is one where recognition has been postponed and the arithmetic underneath is getting worse.

That is not a private credit idiosyncrasy. Research from Wharton and the NBER puts latent commercial real estate distress at roughly 4 times reported delinquencies, with regional banks lowering standards to roll over distressed loans rather than recognise them — the same concealment in a different book, and both sit on the same regional balance sheets. Deferral is the operating mode of the middle tier credit system, not a quirk of one asset class.

Two transmission mechanisms run through banks, and the first is distribution. UBS advised its wealth clients holding large private credit allocations to diversify; the fund worst hit was Blue Owl's $3bn technology income vehicle, which UBS helped design in 2022 for those same clients and through which roughly 60% of its capital arrived. First quarter requests exceeded 40% of net asset value, and firm wide requests reached $4.7bn in the second. When 60% of a fund's capital comes through a single wirehouse, that bank's change of house view is not advice — it is a synchronised run, and the wealth channel converts a research opinion into a liquidity event.

The second is that the same institutions sit on the other side. The FDIC's risk review this year names lending to nondepository financial institutions as a formal credit risk category — the banks as lender, through warehouse lines, subscription facilities and the risk transfer paper they sell to move data centre exposure off their own books. A bank's change of view can trigger the run and sit exposed to its consequences at once.

Which brings this to the awkward fact. Every major US bank has just reported an extraordinary quarter: JPMorgan's net income rose 41% to $21.2bn, the largest quarterly profit in the history of American banking, with Goldman up 78% to $6.6bn, Citigroup 45% to $5.8bn, Bank of America at $9.1bn and Wells Fargo $6.4bn.

It does not refute any of the above, because the earnings and the exposures are on different statements. The profits are capital markets income. The exposures are positions: $325.1bn of unrealised securities losses in held to maturity portfolios and accumulated other comprehensive income, which touch earnings only if the paper is sold — and that figure is dated 31 March, before the 10 year tailed to 4.580% and the 30 year cleared through 5.058%, so the true mark is worse. The lending book to nonbank financials, the commercial property extensions, the repo financing behind the leveraged Treasury bid: none are profit and loss lines. Trading revenue is precisely what funds the capacity not to recognise them. Record earnings are not the refutation of extend and pretend; they are its fuel.

So I should refine the claim. Capital and liquidity are strong, provisions are up only about 1.8% on the year, and a bank earning $21bn a quarter is not on the edge of anything. The constraint is on the bid, not on the bank — an accounting and incentives problem rather than a solvency one. I am not forecasting a banking crisis and will stop phrasing it as though I were.

One structure ties the two halves together. Net equity issuance in the US has turned positive: corporates are now issuing more stock than they retire, reversing the buyback regime that has been the market's most reliable source of demand for over a decade. The single largest component is the $86bn SpaceX flotation in June, the biggest in history — which is also the single largest driver of those record profits. The fee income is the fee on the equity supply draining the market's own bid. The banks are being paid to remove the buyer.

The connection to the section above is direct. Software lending runs around 26% of the average direct lending book, and it is the exposure the Dallas Fed president singled out in mid July. The risk transfer trades that move data centre debt off bank balance sheets are bought primarily by the same funds now gating redemptions. And UBS is modelling private credit defaults climbing toward 15%, citing that software concentration. An AI derating does not need to reach the banks to reach this book; it only needs to reach the borrowers, and the wiring was built by capital rules that make lending to the intermediary cheaper than lending to the end borrower.

The 5 markets above do not share a theme; they share a mechanism, and it is the same one each time. In each, a holder who did not care much about the price has been replaced by one who does.

In Treasuries the foreign official sector held on a reserve mandate across decades and has been replaced at the long end by basis trades funded overnight in the repo market. In equities the buyback bid — mechanical, indifferent, the most reliable source of demand for over a decade — has gone the other way, with net issuance now positive and its single largest component a flotation the banks were paid to bring. In hyperscaler credit the bid did not leave so much as thin and reprice, cover collapsing from 4.7 times to 1.7 against an investment grade average of 3.4. In private credit the capital that arrived through one wirehouse turned out to be capable of leaving through it as well.

The price each new holder charges is legible in its own market, and it is the same price in each — the auction tail, the cover ratio, the gate. Every one is a demand curve that now requires concession where it used to absorb, and reading them as separate stories about separate markets is the error.

The exception proves the point rather than softening it. The one genuinely price insensitive bid left standing sits under gold, where a buyer added at the top and kept adding through a 4 month drawdown — and that buyer is the same official balance sheet that has been leaving the Treasury market. The money did not become price sensitive. It moved.

What has made the swap survivable so far is a second structure sitting underneath the first, and it is the load bearing one: almost nothing in this system is currently being marked. Repo is cheap, which is what lets the levered bid absorb record issuance rather than what proves it does not need to. Private credit loans are valued by the managers who hold them. 94% of the credit downgrades beneath a record default rate are maturity extensions rather than resolutions. Commercial property distress runs at roughly 4 times what delinquency data reports. The banks' unrealised securities losses touch earnings only if the paper is sold, and the last figure available predates the tails that widened them. A June inflation print priced energy off a memorandum that has since been abandoned outright.

Deferral is not a quirk of any one of these books. It is the operating mode of all of them, and it is what allows a position to sit still while the pressure behind it builds. The swap from a patient holder to a levered one is affordable precisely for as long as nobody has to say what anything is worth.

All of it still prices off the level, which is where this connects to the last issue and where that issue was wrong. I argued the Federal Reserve was caught between two live mandates and could move in neither direction; the first section of this one concedes that reading has been superseded. The level holds anyway, on a narrower basis — an inflation question over a labour market soft in hiring and firm in firing. The destination survived; the mechanism did not, and the mechanism was the part I was confident about.

The tail worth naming is not any single one of these markets breaking. It is that a rate shock reaches the levered Treasury bid and the deferred credit book through the same door, because the deferral is itself priced off the level. A funding tightening does not make the marginal Treasury buyer bid less; it forces it to sell. The same move ends the arithmetic that lets an extension look like a cure. Positions that appear uncorrelated because they sit in different markets are correlated through the one variable underneath all of them.

Which is also the shape of what would prove this wrong, and it is not exotic. If the extensions cure — if borrowers refinance into a curve that eventually comes down, if the levered bid absorbs a genuine funding tightening without becoming a seller, if the gates reopen and the queues clear — then what I have described is a transition rather than a vulnerability, and the concessions above should be read as the leading edge of a longer retreat rather than as corrections at the margin of a durable claim.

There is a pattern in those concessions worth ending on, because it is the same one. Each was a case of treating a published number as though it described the present: a payroll figure carrying more structural weight than a benchmarking artefact could bear, an energy print measuring a world that had already ended, a quarterly summary read as complete when one fund inside it had stopped being a fund 4 months earlier. That is precisely the error the deferral above is built to produce, at scale, in every book described here.

Treasury's quarterly refunding announcement lands in the first week of August, and it turns the third section's argument into a test with a published answer.

Coupon sizes have been held flat for several quarters while the financing leaned on bills. That was defensible when the curve was cheaper than the average coupon; it is not now, with the average rate across marketable debt at 3.348% against bills clearing at 3.760% and the long end repriced to 4.580% and 5.058%. Raise coupons, and the levered bid is asked to absorb more duration at scale. Hold them, and the deficit finances at the same front end rate the marginal buyer of the long end borrows at — two dependencies converging on one variable.

Either way the answer sits not in the announcement but in the auctions behind it, and the observable is the tail. If tails hold as sizes rise, the financed bid is absorbing at price. If they widen, the concession is a function of supply.