The Committee held on 29 July and the curve came apart in two directions in the same session. The 2 year fell about 4bp to 4.236%, the mechanical response to a hike that was priced and not delivered. The 30 year rose roughly 11bp to about 5.193% and the 10 year to 4.657%. Read together the curve said less tightening near term and a higher price for lending to the United States over 30 years. Equities followed the long end rather than the front, and the decline arrived after the press conference rather than the statement, with the Dow closing down some 840 points.

The vote underneath was 9–3, and the composition is the part worth pausing on. The three dissents — Hammack of Cleveland, Kashkari of Minneapolis and Logan of Dallas — all favoured a quarter-point increase, and all three are regional presidents. Not one governor dissented. The inflation objection is arriving from outside the Board, delivered with no projections and a statement that again declined to forecast.

The dispersion is measurable: September hold odds at 41.9%, up from about 24% the day before the meeting, with two hikes priced across 2026 against a committee median of one and year end official projections spanning 3.6% to 4.1%. The market is pricing more tightening than the committee's own median.

What makes this datum different from every previous one in this argument is what was absent from it. Each prior instance came from primary issuance — the May coupon sweep, the 22 July 20 year, every tail in between — and a concession at auction can always be attributed to the mechanics of that day's supply. The 29 July move arrived in the secondary market, on a policy decision, with no auction to which anything could be attributed, and with the front end moving the opposite way. Whatever is repricing the long end is not an artefact of the auction calendar. The condition that would unwind the reading is specific: if the 30 year retraces materially below the pre-meeting level inside the window without a policy change, the session was a reaction rather than a repricing.

The Chair supplied his own account from the podium, and most of his facts were right. He described the repricing accurately, placed the inter-meeting move in the top decile of two decades, flagged AI-related equipment and software running near 20% four-quarter growth — then attributed part of the move to the withdrawal of forward guidance, called it a change for the better, and said the institution is just getting started. A convenient, unverifiable claim of cause says plenty about the person making it, but zero about the event itself. The mechanism inside it survives on its own terms and is small: removing pre-commitment machinery does mechanically raise the uncertainty component of a term premium, whether or not the stated rationale is credible.

The demand side has finally been measured end to end. The price insensitive share of the Treasury market has fallen from roughly 75% in 2007 to about 52%. The bid that replaced it is levered — $2.5tn of hedge fund repo borrowing with 73.8% of it at zero or negative haircuts, and Cayman domiciled funds holding $1.85tn of Treasuries at leverage the Federal Reserve puts above 18 to 1 at the largest of them. Behind both sits the link never previously sized: primary dealer balance sheets have shrunk roughly fourfold relative to Treasuries outstanding since 2007, with holdings at a record $409bn and capacity warnings attached.

That is the whole chain, and it reads as a subtraction. The fallback behind a market whose captive bid has halved is a dealer complex 4 times thinner, relatively, than before the last crisis. The 22 July 20 year is what that looks like in practice — a tail at 5.163% against a 5.158% when issued, graded D+, with domestic demand much weaker than average and dealer takedown higher, one day after the desk manager had publicly warned that money markets may tighten

The strongest objection landed in the same session as that tail and belongs in at full weight. The German 20 year cleared at 3.60% against 3.38%, a 22bp concession, with the 15 year at 3.46% from 3.30%. If the US tail were purely a function of American fiscal supply or dollar demand erosion, a simultaneous German concession of that size is not what one would expect. The US specific composition evidence is not overturned — officials are still net selling duration, the price insensitive share is still 52%, the replacement bid is still repo funded — but the weighting between US specific and global explanations is now open, and July has moved it toward the global component without establishing how far. The test is whether European long end concessions keep tracking American ones through August.

Which leaves the funding leg, where this argument has always been most vulnerable and where the series moved this month.

The front end calm beneath the tails is manufactured at a known rate: reserve management purchases at $10bn a month, held for a third cycle through 13 August, plus roughly $17.6bn of agency to bill reinvestments — about $27.6bn a month of standing official bill demand against reserves at $3.14tn and a balance sheet grown to $6.747tn. Quantitative tightening has effectively ended, whatever it is being called. And the desk manager warned on 21 July that money markets must absorb a large volume of net bill issuance across this month and next, that conditions may tighten, and that the reserves demand curve could shift back up.

Which raises the question of what would count. A funding indicator moving a basis point or two is meaningless on any given morning; the series drifts on settlement mechanics, and month end routinely pushes secured rates up for reasons unrelated to scarcity. So the bar has to be numeric and set before the data is read, because a qualitative word like sustained is how a watch quietly never fires. Mine: the funding leg reopens if SOFR closes above the effective funds rate on 5 consecutive business days, or 5bp or more above on any 3 days within a rolling 10, in either case while long end tails persist. Month end effects are excluded by construction, so anything confined to 30 and 31 July counts toward neither leg.

They did tighten, on his timetable. For 5 consecutive sessions from 23 July, secured overnight funding cleared above the rate the Federal Reserve actually targets — 3.64% against 3.63% for the first three, 3.65% against 3.63% for the last two — alongside the 20 year tail and a 30 year at 5.193%. The gap is 1 to 2 basis points and I will come to what that does and does not mean. What matters first is the sign. This argument has rested throughout on repo running below policy, 8 to 10 basis points below it as recently as the second week of July, because that discount is the enabling condition for a levered bid absorbing the duration the official sector is dumping. The series travelled 10 basis points, changed sides in the window the desk manager had named, and stayed there. That is the threshold met exactly, and the funding leg is reopened at it.

Now the size. The spread is 1 to 2 basis points on a series published to two decimal places. SOFR at 3.65% sits at interest on reserves rather than above it, well inside a 3.50–3.75% target range, on roughly $3.0tn of daily volume with no scarcity signature in the distribution. The second leg never came close: 0 of the last 10 sessions reach 5bp, against a series that touched 6bp in mid-June with none of this true. A market in funding stress does not look like this. What has happened is that the cheapest condition in the system has stopped being cheap at the margin, in the window the desk said it would, on the series that gates whether the levered bid can keep absorbing what the official sector no longer wants.

So the reading refines rather than escalates. If that funding tightens, the marginal buyer does not bid less; it is forced to sell into the market it was holding up. What the last two weeks establish is that the condition is no longer fixed — it is a variable now, moving in the direction that matters, at a magnitude that does not yet matter. The observable is whether the spread widens beyond the noise band while the tails persist, or reverts once August's bill supply is absorbed. Both are legible from a page that updates every morning.

The US Treasury bought yen on 31 July, and the New York Fed executed it by selling euros through two primary dealers — the first direct American intervention in support of that currency since 2011.

Start with what the operation cannot be. Reported at $5–10bn, it sits against Japan's own roughly $73.6bn operation earlier this year that did not reverse the trend, and against a stabilisation fund whose net position is a fraction of its gross assets. Nothing at that size squeezes a complex holding tens of billions of one way exposure. It is not a scale operation and cannot be argued into one.

What it is instead is legible from the sequence. A senior official called the currency very undervalued and substantially overshot; banks were told to stand ready for future action; a note was photographed at a cabinet meeting in a state legible enough to be read; then the purchase executed. Four stages, three of which cost nothing, and the currency moved from roughly 163 to 157.96 on news of potential intervention before a single reserve was spent.

The funding is the part worth holding on to, and almost nobody priced it. The issuer of the world's reserve asset spent reserves to strengthen a foreign currency — which is to say, to weaken its own — while running record issuance into an absorption chain whose residual link is the thinnest in the post-war market. It funded that by selling euros rather than dollars or Treasuries, which left its own duration market untouched and exported the balance sheet adjustment onto European paper. Whether that reflects deliberate protection of the Treasury market or simply the composition of the fund is not established and I am not going to guess it. The posture stands either way: an issuer acting to reduce the value of the currency its foreign creditors are paid in.

The reason this reaches a section about the Treasury market is what the yen funds. Yen funded carry is one of the largest private bid structures in US duration, and the positioning is record crowded and two sided — roughly $39.7bn of institutional long dollar exposure against a $17.2bn Japanese retail dollar short — which makes a squeeze the base case in either direction of surprise. The authority now bidding deliberately against that funding currency is the issuer of the debt the complex helps absorb.

The Japanese side was already destabilising before anyone intervened. The 10 year JGB sits near a 30 year high, domestic inflation has accelerated, and Japanese investors turned net sellers of foreign bonds at −¥714.4bn, reversing a large prior inflow — the first repatriation flow of this cycle. The Bank of Japan then held at 1.00% on an 8–1 vote and named rising chip prices from AI demand as an upside inflation risk, projecting core clearly above 2% from the second half of the fiscal year. Chip prices now sit inside two reaction functions separated by 250–275 basis points, and when a common input shock meets a spread that wide it is the low rate jurisdiction's currency that does the adjusting rather than its policy rate.

The reading I had been carrying held that this vector fires on one of two paths — a compression of the US side rate differential, or a disorderly capitulation. It did not contain what happened: an official bid for the funding currency arriving while the differential is unchanged or still widening. Forced appreciation without differential compression is a third path and it belongs in the set. I am adding the path and explicitly not forecasting that it fires; a currency defence is not a currency floor, and an operation this size against a complex that size is a statement rather than a mechanism.

Two conditions on all of it. The sourcing is people familiar grade, and nothing about the magnitude is settled until the stabilisation fund's monthly statement and the Treasury's reserve position publish. If that data shows no material euro sale, the operation was smaller or differently funded than reported, the announcement channel finding survives, and the funding mechanism does not.

Aramco shut the Jazan refinery on 27 July after a strike 2 days earlier, with damage to the integrated gasification block and the tank farm and a restart targeted for 15 August. Abqaiq and the East-West Pumping Station were hit the same day. Yanbu had sat inside a Houthi claim on the 25th, and the exit from the Red Sea has been under a declared blockade since around the 20th.

Crude fell on 27 and 28 July.

That is what this section is about, and the price is not the subject — the price is the thing that failed. Brent touched $101.01 intraday on 23 July, shed roughly $16 across three sessions on peace-talk reporting to about $84 by the 28th, and settled at $90.74 on the 29th. A benchmark that round-trips $17 in six sessions and declines into the 2 days that removed 400,000 bpd of refining and struck the largest crude stabilisation facility on earth is not lagging the physical picture; it is measuring something else. So the instrument changes. For everything below, the series are inventories, loadings, cracks, transit counts and outage schedules. Flat price is withdrawn.

Read that way the picture is worse than the tape. Commercial crude drew 7.2M barrels to 404.5M for the week ending 24 July, a cycle low roughly 6–7% under the 5 year average — while refineries ran at 97.2% of operable capacity. The framing I carried a fortnight ago was a migration of stress from crude to products with crude comfortable. There is no migration, because there are not two balances. Maximum refinery effort converts tank inventory into product faster than imports at 5.7 mb/d, running 6.9% below the year ago 4 week average, can replace it.

The bypass is the structural finding, and it is not that the routes around the strait are under attack, though they are. It is that there is one route. Kpler puts Yanbu at roughly 92% of Saudi seaborne crude exports in June, against about 973 kb/d a year earlier: the Hormuz closure did not diversify Saudi routing, it collapsed it from two channels into one. Every model of a strait shock, mine included, has carried an implicit price ceiling because some volume can go around — through the East-West line to Yanbu, through Fujairah, through the northern routes. That ceiling was never a law. It was a piece of physical capacity, and this month all three links of the single remaining chain were struck inside a week: the pumping station that fills it, the field infrastructure behind it, and the sea lane it exits into.

This logic breaks under one specific, near term scenario: terminal loadings bounce back to trend and the war-risk premium disappears within weeks. If so, it was a raid, not a reframing. I won't reduce the finding to a single metric; calling for $120 is price tagging, not analysis, and it relies on the exact mechanism I just rejected.

A third gate opened in the early hours of 30 July, outside both named theatres. The Caspian Pipeline Consortium halted loadings at Novorossiysk after a tanker was struck on the cargo deck while loading and a second was attacked on approach — the fifth attack on CPC facilities, three days after loadings had resumed from a week long halt. The consequence is physical rather than a routing cost: Tengiz output more than halved to about 406 kb/d against a July average of 925, and Kazakh national output fell to 1.63 mb/d from 2.07. CPC carries roughly 80% of Kazakh exports and close to 2% of world oil. No party has claimed either attack and I offer no attribution.

Three gates now, and they do not select the same way. Hormuz is administered — a declared permit regime overlaid by a naval blockade, running 20 to 34 crossings a day against 120 to 147 before the war. Bab el-Mandeb is affiliational, a blockade the Houthis describe as targeting Saudi Arabia rather than closing the strait, which has let unaffiliated tonnage through while striking Saudi hulls. Novorossiysk is neither: attritional interdiction with no discrimination, where the terminal itself is the target. The insurance market has priced that distinction to a thirtyfold spread on one sea — roughly 0.1% of hull value for Jeddah and Yanbu calls against up to 3% for Saudi linked tonnage — and Lloyd's underwriters have begun excluding vessels with any Saudi touchpoint from Red Sea cover. A shipper's exposure has stopped being a property of its route and become a property of its counterparties, and no routing decision manages that.

The buffer that has absorbed all of this for five months is running out on a calendar, and the American leg is worse than the level suggests. The strategic reserve stands at 307.7M barrels, its lowest since March 1983 — but the binding constraint is deliverability. Drawdown capability sits at 61% of design, 2.700 against 4.415 mb/d rated; fill at 56%; distribution at 53%; more than a quarter of inventory unavailable on outages, and $230M of repairs unappropriated. The 10 June solicitation offered up to 40M barrels and awarded 500,000. The operational floor near 300M that I have described does not exist. A reserve that cannot be withdrawn at rate is not a buffer; it is a number.

Commercial cover sits at 43 days forward, a 45 year low, and the adjustment mechanism sits on the demand side. China's June crude imports collapsed 41.3% YoY to the lowest since 2016, with refinery runs at a 10 year low. The largest importer on earth took 41% fewer barrels — not because it could not source them, but because it curbed runs, drew inventory and substituted. The world did not replace the lost barrels; it stopped consuming them and spent reserves it will have to rebuild.

And the tightness that reaches the consumer has a deferral attached with a date. Refinery maintenance outages ran an average of 470,000 bpd of capacity between January and May, against 700,000 in 2025 and 900,000 in 2024 — maintenance not performed, quantified in barrels, while the 3-2-1 crack set records. That work accumulates into an autumn turnaround window overlapping hurricane season, in a fleet at 97.2% with no spare to absorb an unplanned outage.

One thing runs the other way and belongs in at full weight. Novak indicated on 25–27 July that the Russian producer level diesel export ban will be lifted as the market recovers, expressly to stop refineries facing a glut and cutting runs, while the petrol ban was extended to end 2026. The diagnosis sits in that pair: the strike campaign destroyed secondary conversion capacity, which makes petrol, not primary distillation, which makes diesel. A European margin at $60.17 a barrel has lost its principal single support on no stated timetable.

So the call on the 12 August print, with that caveat attached. The headline goes up. The June energy relief was priced off a curve reflecting a memorandum rather than a cleared strait, and that base breaking is arithmetic; the live part is larger. The pump ran from $3.84 on 9 July to $4.09 on the 23rd and held there, most states above $4 and the national average nearly a dollar above a year ago. Energy ran 15.7% on the year in June with petrol up 26.7% even in the month energy fell 5.7%. Airline fares are up 26.5%.

Core is the genuinely open question and I hold it as one. It printed 2.6% on a shelter line easing to 3.3%, and shelter is a third of the index and lags rents still moderating, so that disinflationary mechanism is real and continues. Against it runs a change on a separate clock from the war: the emergency tariff authority expired on 24 July and forced labour tariffs of 10 to 12.5% took effect the same minute under permanent statutory authority, across roughly 60 economies covering about 99.4% of US imports. Roughly 13% of receipts backfill to roughly 7% — a partial durable replacement rather than a cliff, close to a wash for many importers. For prices the effect is queued rather than relieved, because a duty with no statutory expiry and no rate cap converts a contested cost into a settled one and lands it in core goods. Whether the shelter deceleration outruns the goods pass-through is what the print settles, and I would not pretend to know.

What would make me wrong on the headline: a refining crack that collapses through August as the Russian ban lifts and deferred capacity returns, taking retail fuel down with it.

The private credit argument has been that this is a liquidity event with credit features rather than a credit crisis, and the qualifier has been doing real work: redemption requests running 2 to 8 times the structural design, gates binding across every major sponsor, one vehicle permanently halted and liquidating, roughly $265bn of manager market capitalisation erased — all of it about who can get out, none of it yet about what the loans are worth.

That distinction now has a datum on the other side, and it arrives in the instrument closest to the underlying marks. CLO equity funds fell roughly 15% in the first quarter, their worst since 2020, with managers attributing the loss directly to software sector loan prices.

The identification is the analytical content rather than the number. CLO equity is the first loss tranche: it absorbs deterioration in the pool before any other holder, and before manager marked fund values move at all. The gap between what a manager says a loan is worth and what it fetches on exit is the unrecognised loss this whole thread has been about, and the first loss tranche is where that gap shows first because it has nowhere to hide. The anchor moves from liquidity plus to liquidity plus with valuation impairment, on a valuation reading rather than another redemption headline. Two caveats limit it: CLO equity is levered, so a 15% move corresponds to a much smaller move in the pool beneath and cannot be read across as a 15% credit loss, and the reporting runs with roughly a 75 day lag. Second quarter performance extending the decline establishes a trend; reversing it suggests the first quarter was the software fear spike the counter reading describes.

Software is now the common thread across three independent sources, which is why the instrument matters more than the magnitude. A major bank modelling private credit defaults toward 15% cited software concentration. The Dallas Fed president flagged software sector vulnerabilities publicly in mid July. And the first loss tranche has now marked down on software loan prices. Software lending runs around 26% of the average direct lending book.

Underneath sits a factor I had been describing as static, and it is not. Bank lending to non depository financial institutions stands at $1,419.5bn as at March, on a series growing at a 21.9% compound rate since the financial crisis — the fastest growing segment of bank loan books over that period by nearly three times, with roughly 23% lent to private credit intermediaries. A common factor connecting several books is one kind of object if it sits still and a different one if it compounds at roughly a quarter per year, because every timeline built on it shortens as it grows.

And the bank sits on four sides of the same loop. Lender, through warehouse lines and subscription facilities. Distributor, where a change of house view at a single wirehouse functioned as a synchronised run on a fund through which roughly 60% of the capital had arrived. Risk transfer counterparty, through the paper by which data centre exposure is moved off bank balance sheets and into the same funds now gating redemptions. And repo financier of the levered Treasury bid previously described. Duration is not one of the banking system's risks; it is the common factor across the securities book, the loan book and the financing book at the same time.

The deferral machinery those books run on is familiar: latent commercial property distress at roughly 4 times reported delinquencies, with regional banks lowering standards to roll over distressed loans rather than recognise them; and 94% of the downgrades beneath a record 6.0% private credit default rate consisting of distressed exchanges — maturity extensions and payment in kind toggles that defer the loss rather than crystallise it.

There is now a third book, and it sits outside the banking perimeter entirely. Official research documents life insurers holding roughly $849bn of private credit, standing behind a state guaranty system whose member assessments are creditable against state premium taxes. Follow that to its end: a substantial portion of the cost of an insurer failure is ultimately borne by state fiscal capacity rather than by the industry that funded the guaranty. The legal mechanics are not in dispute; the systemic reading is mine and I mark it as such — a public backstop for deferred credit risk that is fiscal rather than federal, sits outside the prudential perimeter, and has never been sized against the exposure it now stands behind. This premise breaks down if insurer holdings turn out to be confined to senior tranches whose historical loss experience remains insulated from general direct lending defaults.

Three books, three different supervisors — a banking regulator, a securities regulator, and fifty state insurance commissioners — and no one of them sees the whole position. That is not a coordination failure. It is what happens when the same credit risk is moved often enough that each move is legitimate under the rules of the venue it lands in.

The floor under the machine is now visible. There were 372 large corporate bankruptcies in the first half, the highest first half count in 16 years, concentrated in industrials, consumer discretionary and healthcare. Extension is what creditors do to avoid recognition; the bankruptcy count is what happens where the extension option has been exhausted. The two numbers belong together and are almost never read that way — a record default rate resolved 94% by extension, against a bankruptcy count at a 16 year high. The book is not holding because the credit is fine; it is holding because most of it can still be moved.

None of which is contradicted by the bank earnings, because the earnings and the exposures sit on different statements. Every major US bank posted an extraordinary quarter, capital and liquidity are strong, and none of that is in dispute. The constraint here is on the bid rather than on the bank — an accounting and incentives problem, not a solvency one. Trading revenue is precisely what funds the capacity not to recognise.

The claim has been that value in this complex moves from the firms doing the spending to the firms selling them the constraint. Until this month that rested on one measurement. It now rests on three.

The cash flow axis is the original. Twelve month forward free cash flow across the largest technology buyers has gone from roughly $270bn to below zero on a series running back to 2007, while semiconductor free cash flow climbs toward $430bn. Those are consensus estimates and I hold them as such; the realised figures point the same way, the hyperscaler basket down from a peak near $245.6bn to about $198.3bn while the semiconductor names went from roughly $18.6bn to $95.1bn. Either measure gives the same consequence: firms whose cash generation is collapsing cannot fund a capital programme from operations, so the buildout is borrowed.

The earnings axis is new. The largest buyers at roughly 34%, plus semiconductors at roughly 31%, together account for about 65% of index earnings growth, up from 52% — two segments on opposite sides of one transaction, jointly carrying two thirds of what the index calls growth.

The correlation axis is the most telling. The relationship between capital expenditure announcements and semiconductor performance has collapsed to roughly 0.0165, from +0.78 in April: the market itself ceasing to read capital expenditure as good news for the spender, inside 4 months, on unchanged inputs. One refinement sharpens it — the market is not punishing capital expenditure, it is punishing unfunded capital expenditure. Spending that breaks the cash flow statement was penalised while spending carried by operating cash flow was not. That converts a spender versus recipient dichotomy into a funding structure test, and it fires downward if a cash flow positive name is punished for raising capex, or a cash flow negative one rewarded for it.

The borrowing is larger than reported debt shows. A review of securities filings estimates $1.65 to $1.8tn of AI related obligations outside headline balance sheet debt — roughly $1tn of purchase commitments and more than $800bn of leases not yet commenced — with sale leasebacks and joint ventures carrying loss guarantees as the mechanisms, and sector leverage roughly doubling from 0.9 to 1.8 times. Two figures are confirmed directly from filings: approximately $420bn at one firm, about three times its on balance sheet debt, and roughly thirty times at another. I hold those at face value and the aggregate as an estimate, and I do not adopt the accounting fraud framing circulating alongside them. What the number supports is narrower: true leverage is materially larger than reported debt suggests, which makes the credit market's repricing rational rather than an overshoot.

That repricing is legible in two places. Five-year credit default swaps on the major buyers have widened to roughly 75bp, near a 7 year high, with the basket excluding the most levered name at its highest since at least 2018 — sector wide rather than one issuer's problem. And the new issue cover ratio has collapsed from about 4.7 times in February to roughly 1.7 in July, against a 3.4 times investment grade average, on record issuance at roughly 9% of total supply. That is the same saturation mechanism the first section described in Treasuries, appearing in corporate credit.

Now the part that changes the shape of the argument, and it came from the constraint owners themselves. Both reported in the fortnight, both printed records, both disappointed. SK hynix posted the highest operating margin in the industry — 76%, on revenue of ₩79.3tn and operating profit of ₩60.5tn, up 557% — and the shares fell about 11% to a record low in their American listing on a revenue miss against a consensus near ₩84tn. TSMC guided its third quarter roughly $2bn above the street and its shares reset too.

The explanation is neither positioning nor a demand crack. SK hynix has signed multi year agreements with around ten customers. A spot priced shortage is being sold forward under contract, which caps the price capture the equity market was extrapolating while securing the volume the buildout needs. Record margins beside a falling stock is exactly what that looks like, and it is a better account than any unwind story, including mine. Contractualisation does not falsify the constraint owner thesis; it dates it. The rent is being fixed in advance, at a level the owner accepted, for a period neither side discloses.

The other half of the ceiling is being bought down. TSMC raised 2026 capital expenditure from $52–56bn to $60–64bn against a $58bn consensus and added $100bn of Arizona commitment. Advanced packaging has run from roughly 35,000 wafers a month at end 2024 to about 75,000 at end 2025, with a target of 125,000–130,000 by year end and the supply demand gap projected to narrow from as wide as 20% toward roughly 10%.

Which forces a correction to the most directional claim I hold. I have described advanced packaging as the binding physical constraint on this trade, indifferent to anything the market prices as an all clear. That was stated too broadly. Packaging is immune to the war, not immune to change — the variable that moves it is capital, and capital has arrived. Two qualifications keep the relief from being an all clear: Nvidia is reported to have booked around 60% of capacity through 2026 and more than half the 2026–27 expansion, so it is largely pre-allocated before it exists; and the capital doing the relieving is being raised into precisely the credit repricing described above. The bottleneck is being bought down with money that is getting more expensive.

The constraints without that path are the gases and the grid. Helium: roughly a third of global supply behind a chokepoint exposed jurisdiction, damaged trains under a repair estimated at up to five years and gated by turbine availability rather than by any transit or political change, semiconductor manufacturing consuming about a quarter of global use, Northeast Asian spot roughly doubling. There is an ownership asymmetry in the export controls worth noticing: the restricting state produces only about 1.6% of global supply and imports over 80% of its own use — chokepoint control without resource ownership, a more replicable instrument than controls over materials whose production is genuinely concentrated. Then electricity, with data centre demand projected toward a fifth of US national consumption and the longest lead time of the three.

The sequencing is the content: chips, then gases, then power, with the constraint that binds last the one capital can do least about. And one distinction keeps them honest rather than lumped — packaging and power bind regardless of the conflict; gases bind because of it. A market pricing de-escalation as an all clear is mispricing two of the three and correctly pricing one.

Which settles a test named in print last month. The claim was that disrupted inputs fall on the Korean fabs rather than on Micron, so the asymmetry should express as input driven production cuts in Korea while Micron's guidance stayed clean. Samsung printed a record and flagged no input driven cut, Micron guided to $50bn with about $100bn contracted, and SK hynix's miss was shipment timing and product mix, stated by the company — not supply. Three for three clean. The input register was a real risk that has not reached the print, and I said I would say so.

I owe a sharper admission alongside it. Last month I described the mid July Seoul crash as positioning, triggered by a single brokerage note about the timing of high bandwidth memory shipments. The dismissiveness in that phrase was the error: the note said shipments would slip, and management has now confirmed exactly that. It was carrying a real physical datum about a shipment schedule, and I filed it as noise because of who published it — treating the provenance of a number as though it settled its content.

The tape then did something the fundamentals do not explain, and I closed a watch too early. Having judged that the memory complex had extended its drawdown rather than round tripping, the Korean index closed up 17.91%, the largest single day advance in its history, with the two memory makers up 29.95% and 26.81%. The closure was premature and I reverse it; the error is adjudicating a two outcome watch inside a volatility cluster, where three sessions is a sample of one event. The scoreboard does not restore the position, and running only the rebound would be the flattering half: the index still finished the month down more than 20%, having drawn down 43.9% from its June peak against 31.1% in the 2020 crash. Underneath, leveraged retail positioning financed the advance through single stock vehicles introduced in May 2026 and liquidated mechanically on the way down, with new deposit requirements taking effect on the day of the rebound — a product 3 months old amplifying both directions of a 44% drawdown and a record recovery.

One related claim of mine has broken cleanly and the break is instructive. The broader thesis has been that efficiency gains cap what the constraint owners can extract. At the hardware input layer that is now false: the Chinese entrant, which raised $8.6bn in Shanghai and rose 466% on debut, is pricing above the incumbents rather than undercutting them, at cost per bit more than 30% higher. In a genuinely supply constrained market a second source does not need to undercut — it can capture the scarcity rent instead of competing it away. The ceiling may well persist at the model inference layer, where marginal cost is software like and capacity is not physically rationed. Two layers, opposite outcomes, one thesis stated too broadly — and the consequence runs against comfort, because the input floor is higher than assumed while the output ceiling holds.

I am not forecasting a capex pause. What I bank is that the economy has been configured so that one would transmit to growth, earnings, credit and household wealth simultaneously.

Every book described above has the same property, and it is not leverage.

Private credit loans are valued by the managers who hold them, and the first genuinely independent mark — the first loss tranche — came in 15% down. 94% of the downgrades beneath a record default rate are extensions rather than resolutions, which is a decision not to say what something is worth. Commercial property distress runs at roughly 4 times what delinquency data reports. Banks' unrealised securities losses touch earnings only if the paper is sold. Between $1.65tn and $1.8tn of AI related obligations sit outside headline balance sheet debt, which is not concealment but is certainly not disclosure. Reserve releases flatter an inventory picture while leaving the series that sets the pump price alone. And a memory shortage has been sold forward at a price neither party will name.

None of that is fraud. Each is legitimate under the rules of the venue it sits in. The aggregate is a system in which the price of most things is a matter of assertion rather than of transaction, and where the number that would settle it is either not published, not required, or not yet due.

What is being priced, then, is not the assets. It is the compensation for not knowing what they are worth, and that compensation is visible everywhere at once, in exactly the places described above. A 30 year real yield at its highest on record — dating from when the Treasury resumed 30 year TIPS issuance in 2010 — which is a statement about duration risk rather than about inflation. A new issue cover ratio at 1.7 times against a market average of 3.4. An auction tail on a D+ grade with dealer takedown above average. A gate. A first loss tranche marking down before the fund does.

Read as separate stories about separate markets, those are five unrelated pieces of bad luck. Read as one price, they are the same demand curve appearing 5 times: a buyer who used to absorb and now requires concession, charging for the part of the position it cannot see.

The same shape runs outside the balance sheets. A transit levy on the strait was declared by social post, moved markets, and was withdrawn 5 hours before taking effect with no implementing instrument ever in existence; policy has been repeatedly moved by anonymous leak placed with a breaking outlet, carrying market moving framing with no path to named channel verification; and an intervention in the yen was signalled by a note photographed at a cabinet meeting, with the currency moving on the possibility before a single reserve was spent. Three instances all of the same shape: a market repriced by something that was not an instrument, with the announcement channel cheaper every time than the action it described. Whether any of it was deliberate is unfalsifiable from outside and I decline to speculate; the effect is identical under either reading, which is what makes the effect the thing worth grading.

Put the two together, because they are the same problem from opposite ends. The scheduled, verifiable, pre-committed channels are being narrowed, and the unscheduled, unverifiable, uncommitted ones are doing the work instead. A market that cannot price the path from the institution will price it from whatever else moves — and what else moves is a post, a leak, a legible piece of stationery.

The last issue argued that the swap from a patient holder to a levered one was survivable for as long as nobody had to say what anything was worth. That still holds. What has changed in a fortnight is that the market has begun charging for the difference in a venue where the concession cannot be blamed on an auction.

The condition that would unwind this is not exotic and requires nobody to admit anything. If the long end retraces while the opacity stands — if the 30 year comes back through its pre-meeting level without a policy change, if cover ratios recover as issuance continues, if the gates reopen and the queues clear — then what I have described is a repricing that has already happened rather than one that is under way, and everything above should be read as the end of a move rather than the start of one.

The framework is set aside here. What follows is opinion.

Four things have happened at the Federal Reserve in 3 months. Forward guidance was removed outright at the first meeting under the new Chair, alongside a statement cut to 130 words from 341, he declined to submit his own projection in June while announcing five task forces, one of them reviewing the institution's data sources, and at the July meeting he raised reducing the number of regularly scheduled policy meetings from eight, with a stated willingness to scale back post decision press conferences and reporting that a calendar decision could arrive before September.

The counter deserves stating properly before I make the point: forward guidance was a crisis era instrument that arguably overpromised, and a chair who refuses to pre-commit while the inflation question is genuinely unresolved is being candid rather than opaque. Eight meetings is a convention, not a law; plenty of central banks meet less often and manage. And he has been more hawkish than anyone expected of him — he had the clearest dovish opening of his tenure in mid July with a cold core print in hand, and refused it twice.

None of which survives the thing he will not say.

He describes inflation as a choice, has said it has been too high for more than 5 years, and he is right about that, he then held, from the dovish side of a 9–3 vote in which every dissent wanted to tighten, and when asked what would move him, he offered nothing — no threshold, no condition, no number. Candour about uncertainty would mean naming what would change his mind and admitting he does not know which way it will break. He has done the opposite: removed the machinery, declined the criteria, and told the podium that the resulting yield move is a change for the better.

Take the meeting calendar seriously as a mechanism rather than a housekeeping item. Fewer scheduled decisions means more accumulated data arriving at each one, which means policy adjusts in larger and less frequent increments, and it means the uncertainty carried between those points is greater because there is no intervening meeting at which the path could be corrected cheaply. That lands on a committee which has already removed the pre-commitment machinery a market would otherwise use to price the path in advance. More violent adjustment at fewer points, from the institution and from the market both.

The refunding resolves this week and the observable is unchanged: the tail behind the announcement rather than the announcement itself, against coupon sizes held flat for nine quarters while the average rate on marketable debt sits at 3.348% and bills clear near 3.760%.

The other thing is a test I set out in public in June, and it is about to be leaned on harder than it has been.

On 1 August, Fars News Agency published a list of energy installations it described as within range of Iranian precision missiles should Washington act: Ghawar, Abqaiq and Khurais in Saudi Arabia; Ruwais and Zakum in the UAE; the North Field and Ras Laffan in Qatar; Burgan in Kuwait; Sitra and Ma'ameer in Bahrain; Leviathan and Tamar in Israel.

The Consequence Ceiling framework holds that a class of target exists which is threatened constantly and struck rarely — not because it cannot be reached, but because reaching it destroys the thing that makes the threat worth having. Every name on that list is in the class. Every name has been in range since February. And the framework's proposition is that the list is more useful published than executed, which is precisely why it was published.

That produces an asymmetry worth stating carefully, because the two branches are not symmetrical in probability or in consequence.

If the ceiling holds, this recurs. Threat, counter threat, a wire list, a premium that builds over a weekend and decays over a fortnight, and no strike at scale. That is not a prediction that nothing happens — Abqaiq was hit on 27 July and Ras Laffan was hit in March, with roughly 17% of Qatari LNG capacity damaged and repairs put at 3 to 5 years. It is a prediction that the difference between a facility being damaged and a capacity being removed keeps being the difference nobody crosses. The threats become an instrument in their own right, costing nothing, working every time, and requiring no execution.

If it breaks, the arithmetic is not gradual and nothing in the sections above is standing by to absorb it. The strategic reserve cannot be withdrawn at rate — 61% of design capability, 2.700 against 4.415 mb/d rated; commercial cover is 43 days forward, a 45 year low; the refining fleet is at 97.2% with a maintenance backlog falling due in autumn; the demand side adjustment that absorbed the last 5 months has already been used: China curbed runs 41.3% and drew down inventory it will have to rebuild; and for the item that carries roughly a fifth of global LNG trade there is no reserve at all, no bypass, and no buffer beyond what is already on the water.

Which is the part that is easy to get backwards. The ceiling holding has not been evidence of resilience. It is the reason every buffer described above was permitted to run down — 5 months of absorption, financed by inventory, on the assumption that the class of target which has never been hit at scale would continue not to be. If it breaks, it breaks into a system that spent its entire cushion on the version where it did not.