The Third Door

Two kinds of asset went in opposite directions this week and the pairing is the whole edition. Everything you can hold rose: the metals, the barrel, the coins. Everything that is a claim on a future dollar fell or stalled, including the long government bond, on a week in which its own issuer became a bigger buyer of it. The Scoreboard has all twenty-six lines.

Underneath that sat the two events of the week, one loudly reported and one barely: the debt passing forty trillion dollars, and the Treasury quietly enlarging the operation through which it repurchases its own long bonds. Contrarian Corner takes apart what a government in that position can actually do about it, and why the option it reached for is the one nobody has to vote on.

Elsewhere: the economics of hiring out an artificial-intelligence chip got worse for a second week running; a second oil chokepoint opened in a place almost nobody is watching, and two American oil majors own barrels going through it; and the running argument this publication scores every week about whether artificial intelligence adds to human work or replaces it went, for the first time since we reset the scoreboard in May, to the side that says it replaces.

The crash gauge holds at twenty for a second week: the yield curve is still the only red on the board, the energy dial is still amber on one week’s move in crude, and nothing else shifted at all.

Every yield in this edition is a Friday close on the US Treasury constant-maturity series, unless it says otherwise.

Scoring last week’s tell

The test we set was narrow and it has an answer. The thirty-year Treasury had to avoid closing below 5.00 per cent on any session before Friday 21 August. It closed the five sessions at 5.31, 5.28, 5.19, 5.23 and 5.27. The lowest reading came on Wednesday, and Wednesday was the single day on which the United States government intervened in that exact part of the market. It still did not get within nineteen hundredths of a percentage point of the line. The tell held, and it held with room.

What that buys is one week of standing for the claim underneath it: that the long end of the American bond market has stopped taking its instructions from the inflation data and started taking them from the quantity of borrowing. It is one week. The tell below is the next one.

The full working, every weight, every threshold and the arithmetic that reaches twenty, is in Appendix A6. A subscriber with only this page can recompute it.

The argument

“Treasury Announces Increased Sizes of Nominal Long-End Liquidity Support Buybacks.”

That is the headline the United States Department of the Treasury put on a press release on Wednesday 19 August. Nine words, three of them doing all the work. Read them in plain English and they say: the government is going to buy more of its own long-dated debt, and it would like that purchase understood as support for the market rather than support for the price.

The mechanics are small and precise. The maximum size of each buyback operation in the ten-to-twenty-year and twenty-to-thirty-year parts of the curve at least doubles, from two billion dollars to at least four billion. The change begins on 9 September and runs to 4 November. Treasury’s stated reason is the volume of high-quality offers it has been receiving in those operations, which is the polite way of saying that a great many holders would like to sell.

Set the size against the stock before deciding what it means. Four billion dollars an operation, against a stock of debt four figures larger, is about one ten-thousandth. This is not a government suppressing its own yields, and anyone who tells you it is has not done the division. The information is not in the size. It is in the direction, and in the calendar: the day after the debt crossed forty trillion, with the thirty-year at its highest since 2007, the largest borrower in the world enlarged the machinery through which it buys back what it has sold.

The market’s answer took about forty-eight hours to arrive, and it was no. On the announcement the thirty-year fell nine hundredths of a percentage point to 5.196 and the ten-year fell six to 4.647. By Friday’s close the thirty-year was back at 5.27 and the ten-year at 4.74, both higher than the previous Friday’s close, though the thirty-year finished below where it opened the week. An intervention that is absorbed inside two sessions has told you the size of the flow it was fighting.

A government that finds long money too dear has three ways out of it, and Contrarian Corner takes all three apart below. What matters here is the mechanical one: this week the United States reached for the exit that does not require a vote in Congress, and the exit did not open.

What the rate model says

Our internal reading of the Federal Reserve’s next move is Hold, and it moved this week towards the hawkish end of that band rather than out of it, on a panel of which ninety-two per cent was scored on fresh data. One cell did the work: the two-year Treasury, which sold off seven hundredths of a percentage point over the week to 4.24. Unlike the two gauges in this edition, that model does not publish its rubric, so treat the direction as ours and the two-year as yours.

Sitting behind that is the Taylor gap, which is worth one sentence of explanation because it recurs. The Taylor rule is an arithmetic recipe for where a central bank’s policy rate ought to sit given inflation and unemployment, and the gap is the distance between the recipe’s answer and the rate actually set. That gap stands at plus 1.28 percentage points and is stable, meaning the rate is still above what the recipe prescribes, meaning a cut is not yet rule-justified. Stable is the operative word. A narrowing gap is a cut coming into view; this one is not narrowing.

The rate market and the long bond are therefore now arguing about different things. The front end is discussing employment and prices, and it has a central bank that answers to it. The long end is discussing how much paper there will be, and it does not.

Forced FlowsBitcoin rose 24.8 per cent over the week and Ethereum 34.0 per cent, the largest moves on the Scoreboard by a distance. Three things happened at once, and none of them is an opinion about money. The Treasury’s buyback announcement on Wednesday released liquidity into the long end. The American president met crypto executives at the White House the same day and pressed Congress to advance a market-structure bill. And roughly 2.7 billion dollars of bearish positions were forcibly closed inside twenty-four hours, more than a billion of it in a single hour, as prices rose through the levels at which those bets are automatically shut. A short seller who is closed out is a buyer. He is simply not a volunteer.

Who PaysToll Brothers reported its third quarter on Tuesday 18 August, and the two halves of it point in opposite directions. Revenue from home sales fell to 2.65 billion dollars from 2.88 billion a year earlier, deliveries fell to 2,662 homes from 2,959, and the gross margin on those sales fell to 23.9 per cent from 25.6. Yet the value of newly signed contracts rose, to 2.52 billion dollars from 2.41 billion, on 2,508 homes against 2,388. (Company release, 18 August 2026; the figures are fiscal 2026 third quarter against fiscal 2025 third quarter.) Toll sells to the least mortgage-dependent buyer in American housing. When the price of thirty-year money is the binding constraint, the customer who needs a loan disappears and the customer who does not stays in the showroom. That is what a long bond at 5.27 per cent looks like from inside a business.

CapacitySemiconductors fell about twice as fast as the index around them. The main American chip exchange-traded fund fell 4.66 per cent, Nvidia 4.64 and Arista 5.12, against a Nasdaq 100 down 2.45. This is happening while the companies buying those chips spend faster than ever, and while professional money quietly leaves the sector. The Stack Inversion section below takes that apart, because it moves one of the gauge’s eight dials. Money is going into the build-out and coming off the table in the shares of the people building it. Both can be rational. Only one of them can be right about the return.

Cost StructureAmerican retail reported and the answer was the same from four different tills. Walmart’s American comparable sales excluding fuel grew 2.6 per cent, its weakest in more than six years, dragged by its pharmacy business (company release, 20 August 2026). Target, Home Depot and Lowe’s all described a shopper who is still spending and increasingly unwilling to spend more than necessary. The consumer is intact and selective, which is the state in which retailers make the sale and lose the margin.

Where Value Is CapturedSpaceX completed a 60 billion dollar all-share acquisition of Anysphere, the company behind the Cursor coding tool, issuing roughly 391 million of its own Class A shares to do it. It is, on the available record, the largest acquisition of a venture-backed company ever completed. A rocket manufacturer has bought a text editor for the price of a mid-sized listed company, and paid in a currency it prints itself.

Japan was sold three separate ways and it is worth naming as one trade. The Nikkei 225 fell 3.93 per cent, more than any other equity index on our own Scoreboard of verified 21 August closes; the Japan ETF fell 3.09; and Mitsubishi UFJ, an open call in this publication, fell 4.42. The yen barely moved against the dollar, so this was not a currency effect being read as an equity one. It was Japanese equity risk being sold on its own account, in a week when long yields rose everywhere.

Not one of the eight crash-gauge dials changed band this week, which in a week that included a forty-trillion-dollar milestone and a five per cent move in gold is a finding rather than a formality.

Three of the eight readings are carried from earlier dates rather than freshly measured this week: the percentage of the index above its two-hundred-day average, measured 7 August; insider selling clusters, measured 17 July; and the high-yield spread, which is an August level rather than a precise Friday print. Every carry is asterisked in Appendix A6. Insider selling at five weeks old is the weakest of the three, and a green from a five-week-old observation is a weak green.

The most useful piece of evidence about artificial intelligence and work is not a forecast. It is a payroll record. Researchers at Stanford, working through the payroll files of the largest American payroll processor, found that since late 2022 employment of workers aged twenty-two to twenty-five in the occupations most exposed to artificial intelligence has fallen about thirteen per cent relative to older workers doing the same jobs, who held steady or grew. Nobody was replaced. The junior post was simply not created. That is what displacement looks like when it arrives through a hiring plan rather than a redundancy notice, and it is why a headline unemployment rate can stay reassuring for years while a profession quietly stops taking beginners.

The Strait of Hormuz is where everyone is looking. The American negotiating deadline with Iran lapsed on Monday 17 August, fresh sanctions on Hezbollah followed on the 20th, the strait remains effectively closed, and crude finished the week 5.7 per cent higher at 87.06 dollars. That is the story with the headlines, and it is doing the work the headlines say it is doing.

The one to add is a thousand miles north. The Caspian Pipeline Consortium runs roughly 1,510 kilometres from Kazakhstan’s Caspian oilfields to the Russian Black Sea port of Novorossiysk, and carries around 1.4 million barrels a day. Almost none of it is Russian oil. It is Kazakh crude, and the pipeline’s owners include Chevron and Exxon Mobil, whose stake in the Tengiz field dates to a joint venture signed in 1993. A Western-owned energy asset sits inside the Russian export chain, and has done for thirty-three years without anyone needing to think about it.

They are thinking about it now. Ukrainian drone and unmanned-boat attacks have hit tankers loading at the consortium’s terminal repeatedly through July and August, shutting the route three times in a single month and prompting Kazakhstan to cut production. After a request from the American vice-president in late July, Ukraine agreed not to target consortium infrastructure or non-Russian vessels. The strikes at Novorossiysk continued into August regardless.

Here is the mechanism, and it is the reason this belongs in a market letter rather than a foreign-affairs one. Hormuz is a chokepoint where the barrels belong to producers the West is already sanctioning or arguing with. Novorossiysk is a chokepoint where the barrels belong, in part, to two American companies, shipped from a country Washington counts as a partner, through the territory of a country it does not. There is no clean lever. Pressing Kyiv protects an oil flow that runs through Russian infrastructure and pays Russian transit. Not pressing Kyiv puts American corporate assets in a drone corridor. That is an escalation node with no good exit, and it is currently being managed by private request rather than by policy.

For the reader the consequence is simple enough. The oil market has spent this year pricing one chokepoint. There are two, and the second one has no obvious diplomatic off-switch. A gauge dial that reads oil on a four-week window, as ours does, will register the first shock and can be slow to the second.

In 1959 a Norwegian of twenty-four was living in New York with no money and no ship. He was reading a copy of Life magazine, and the article in front of him was about submersible pumps: machines that could sit at the bottom of a deep well and push water up from beneath the American West.

The problem he had been turning over had nothing to do with wells. Liquid chemicals moved by sea in whole-ship lots, because a tanker was one large tank, and a customer who wanted to ship a small quantity of one chemical had to charter the entire vessel or wait for somebody who wanted the same thing. Ships had been built with internal divisions before. They leaked. The pipework ran horizontally along the bottom of the hull and had to pass through every bulkhead to reach the pumps at the stern, and every one of those crossings was a place where one customer’s cargo could contaminate another’s.

The article suggested the fix. Put a pump inside each tank. If every compartment lifted its own cargo straight up and out, no pipe needed to cross a bulkhead, and the divisions would hold.

The bet

Jacob Stolt-Nielsen took a thirteen-and-a-half-thousand-tonne ship called the Freddy and had it fitted with sixteen deep-well submersible pumps, one to each tank. He did it with no plans, no naval architect and no shipyard contract. He pointed, the yard built, and two partners put up the money. This is the moment of genuine uncertainty in the story and it is worth sitting with, because it is not the romantic kind. He was not risking a reputation he had. He was persuading other people to spend their money on a modification that nobody had drawings for, in an industry where a mistake meant a contaminated cargo and a claim larger than the ship.

It worked, and the market called him Jackpot Nielsen. On his first owned vessel he took the larger end of a sixty-forty split despite not having the funds to sit comfortably behind it, which tells you what he thought the idea was worth.

The archetype is the Mad Scientist Chef. Not the visionary who saw the future of chemical shipping, and not the operator who ran it better than anyone. The man who read about one thing, cooked it into another, and could not really explain the recipe until it was on the plate. Cross-domain theft, executed before anyone had approved the method.

What he actually invented

Here is the part that transfers, and it is not the ship.

A parcel tanker is only useful if a buyer will believe that the cargo arriving is the cargo that was loaded. Segregated tanks make that physically possible. What makes it commercially possible is everything that had to grow up around them: tank cleaning to a specified state, documented cargo histories, coating registers, inspection and certification. Stolt-Nielsen’s real invention was a standard. Once a small parcel of a named chemical could be priced, insured and financed on its own, a market existed where before there had been a queue.

That pattern recurs everywhere in this publication. An opaque market becomes a real one at the moment a credible standard arrives, and the moment it does, capital that could never previously participate turns up. It is why graded trading cards became an asset class and ungraded ones did not. It is, in a smaller and more awkward way, the question hanging over the American long bond this week: what happens to a market when the party that most needs it to function starts underwriting it itself.

Sixty-seven years on

Stolt Tankers now runs 167 ships, the largest fleet of chemical parcel tankers in the world, carrying about twenty-six million tonnes a year. Around it sit fifteen storage terminals, nearly sixty-seven thousand tank containers, and a fish-farming business that raises turbot and sole in fourteen land-based farms that recycle their own water. In the quarter to 31 May 2026 the group reported revenue of 750.3 million dollars and operating profit of 93.8 million, against 712.9 million and 113.7 million a year earlier: revenue up, profit down. The family vehicle still holds 51.24 per cent. The chief executive, since 2023, is the first person to hold the job who is not a Stolt-Nielsen.

And the standard he invented is no longer his. Segregation, cleaning and certification are now codified across the industry in the chemical carrier code and the shared inspection regimes that go with it. Everybody has to meet them. So the company that created the moat now competes on scale and network density inside it, which is a considerably less comfortable place to stand: our own analysis of the shares stops a long way short of enthusiasm, with roughly fifty-five per cent of group operating profit entering a multi-year down-leg in freight rates, and 2.36 billion dollars of net debt against a market value near 1.7 billion.

That is the lesson, and it is the uncomfortable half of the first one. Inventing the standard creates the market. It does not entitle you to keep it. The founder who builds the road usually ends up paying the same toll as everybody else, and the returns migrate from the invention to whoever operates at the largest scale on top of it.

The gauge falls to 30.0 out of 100. Still first cracks, and now at the very bottom of that band. One dial moved and it is worth being precise about which, because the direction flatters the trade and the reason does not.

Positioning goes from amber to green. That dial asks whether the scarcity premium is crowded, and it is no longer crowded. Professional investors’ net allocation to semiconductors fell from 23.5 per cent in late June to 15.2 per cent in early August, and July saw the second-largest monthly fall in their overall borrowed exposure on record (Data: Goldman Sachs Global Investment Research, August 2026). This week added a third data point of its own: semiconductor shares fell about twice as fast as the index around them.

Read that carefully, because a falling score usually means supply arrived. It does not mean that here. Nothing upstream has loosened. What has happened is that the people who were crowded into the trade have got out of the way, which lowers the risk of a positioning accident and tells you nothing whatever about whether the bottlenecks are dissolving. A gauge that falls because the crowd left is not the same instrument as a gauge that falls because the shortage ended, and this edition would rather say so than let the number carry an implication it has not earned.

The load-bearing dial stays green, and its second derivative has turned. Memory pricing measures whether the price of the working memory inside every computer is still rising, and it is. But the shape of the rise has changed sharply. Contract prices for DRAM rose about 95 per cent in the first quarter of 2026 and about 63 per cent in the second, and the forecaster the industry watches most closely projects 13 to 18 per cent for the third. Still up, and up by a great deal. Rising at less than a fifth of the pace of six months ago. A shortage does not announce its end with a fall. It announces it with a smaller rise.

Where that lands is on a shop floor, and a third of the cost of a laptop is now a component in shortage. The Displacer section carries that figure, because who pays for a shortage is an argument about labour and capital rather than about chips. No tariff explains any of it.

Four dials are held deliberately and the reasons are printed rather than assumed. The extreme-ultraviolet lithography monopoly stays green: the prototype report that raised the question is dated 27 July, and moving on it now would be a re-reading rather than news. Memory capacity holds amber, because the Chinese capacity story is three editions old and the company at its centre has still refused to cut prices, which is scarcity intact. Foundry access and model-layer commoditisation are unchanged.

The business test: second Friday under the line

This section runs one test that the eight dials do not: does renting out the workhorse artificial-intelligence chip still make money? Last week it dropped back below its cost line, where it had also been on the 17 July baseline, and we set out the condition that would separate a round trip from an inversion: four consecutive Fridays under the line.

Friday was the second. The open-market interruptible rent for an H100, the price at the margin where independent operators compete for work, was $1.33 an hour, taken as the daily median across forty-seven live offers. Against an all-in cost of $1.64 to own and run the chip, that is a margin of minus thirty-one cents an hour, against minus seventeen a week ago.

What changed is not the level, it is the shape. The last four Friday readings run $1.87, $1.73, $1.47 and $1.33: down twenty-nine per cent, in a straight line, without a single week going the other way. Last week the swing inside a single week was wider than the margin being measured, and we said so. This week the daily readings ran $1.60, $1.47, $1.50, $1.33 and $1.33, a narrower range that has stopped crossing the line in both directions and settled underneath it. Two more Fridays and the test this section set for itself is met.

The cost stack. Carried from last week’s recomputation, and the only input that moved was the funding rate, by one hundredth of a percentage point, which does not change the total at two decimal places.

Depreciation remains the line carrying most of the weight, and it is the one an operator can argue with: writing the chip off over five years instead of four saves about twenty-four cents an hour, which narrows the loss from thirty-one cents to seven rather than removing it. A week ago that same adjustment would have restored the margin outright. That it no longer does is the clearest single measure of how far the rent has fallen. That is not a technicality. It is the single assumption on which the entire question turns, and an industry under pressure to show returns has an obvious incentive about which way to resolve it.

The interpretation, and why the two are kept apart

The rental test sits deliberately outside the composite. It measures the economics of one ageing chip; the gauge measures upstream bottlenecks across the whole stack. Mixing them would let obsolescence masquerade as an easing of scarcity, and the H100 is a 2022 part with a successor already shipping.

That competing explanation is still the strong one, and the test that separates them is still unbuilt: if the newest silicon rents at a widening premium while the H100 sinks, this is displacement; if the premium compresses too, it is the tide going out. We do not yet have a clean current-generation rental series. Naming the missing instrument is worth more than guessing at what it would say.

The score falls further if memory capacity is delayed, a demand surge empties inventories again, or the Chinese tool programme slips a year on quality, because any of those would re-tighten scarcity. It rises if the positioning dial that just went green turns back, which would mean the crowd returned rather than that anything real had changed.

The full eight-signal working, the reader explainer and the four bands are in Appendix A11. The price register is in A12, where the memory contract price row remains absent and says so.

In January 2001 the Congressional Budget Office published its regular ten-year outlook for the American federal budget. Its central projection was a cumulative surplus of 5.6 trillion dollars over the decade to 2011.

The actual outcome for those years was a cumulative deficit of 6.1 trillion. The forecast missed by 11.7 trillion dollars, which at the time was roughly a year of American national income, and it missed in one direction.

The point is not that the Congressional Budget Office is bad at its job. It is among the most careful public forecasters anywhere. The point is what its job actually is. A fiscal projection is a conditional statement: this is what happens if current law does not change. Current law is the single input a legislature exists to change, and it changes it in one direction far more often than the other. Tax cuts pass. Spending programmes are extended. Emergencies are funded. The projection is therefore not a forecast of the future so much as a description of a future nobody intends to allow.

Which gives the reader a usable rule. When an official projection says the debt reaches a hundred and twenty per cent of national income, read the number as a floor rather than a path.

Data: Congressional Budget Office. Framing: BCA Research, Martin Barnes, “Stranger Things”, August 2026. Analysis: Weekly Market Pulse.

Three doors, and only three

A government whose debt exceeds its annual income has exactly three ways out, and every historical episode is one of them or a blend.

The first door is growth. Out-earn the debt so the ratio falls without anybody having to lose anything. This is the American case and it is not a fantasy: America still runs the world’s most productive large economy, still adds working-age people through immigration in a way that Europe, China and Japan do not, and has a political class that agrees on almost nothing except that growth should not be interrupted. It is the only door that costs nobody anything, which is why everyone in office prefers it and why it is the one written into the projections.

The second door is default. A correspondent of mine, a Danish writer whose argument deserves a hearing precisely because it is not the house view, puts it starkly: America’s 1919 has already happened, a Truss moment is closer to thirty months away than thirty years, and the American ten-year yield belongs nearer seven per cent than five. Set against a ten-year at 4.74 per cent this Friday, that is not a small claim.

The 1919 parallel is worth taking apart rather than repeating, because half of it holds and half does not. Britain came out of the First World War owing about a hundred and thirty per cent of national income, chose deflationary orthodoxy, and in 1925 went back onto gold at the pre-war parity. The result was a decade of strangulation: to hold a fixed exchange rate the country had to squeeze wages and prices, and it did, through the General Strike and out the other side. What does not carry across is the currency. Britain had borrowed in something it had voluntarily tied to gold, so the only adjustment left was the domestic economy. The United States borrows in a currency it issues and allows to float. The failure mode therefore shifts from strangulation to debasement. Different door, same corridor.

The third door is repression, and it is the one history usually picks. Arrange for the debt to be held at a yield below what a free market would demand. It is done by regulation, when banks, insurers and pension funds are required to hold government bonds regardless of what they yield. It is done by a central bank balance sheet. It is done by inflation that quietly erodes the real value of what is owed while the coupon is paid in full and on time. Britain and America both did it for roughly three decades after 1945, and neither ever announced it.

Here is the contrarian claim, and it is the reason this section exists. The two loud positions, the optimist who says growth handles it and the pessimist who says default is coming, are each half right about the arithmetic and both wrong about the politics. Growth is real and insufficient. Default is possible and unnecessary. The third door is the one that requires no vote, produces no headline, and can be walked through in increments so small that each one is defensible on its own terms.

Which brings it back to Wednesday. A four billion dollar buyback operation against forty trillion dollars of debt is not financial repression. It is roughly one ten-thousandth of it, and anyone who calls it repression is not doing the division. But it is the correct shape. The state buying its own long paper, described as support for market liquidity, announced in a press release, effective in three weeks. Nobody rings a bell for the third door. It arrives as an amendment to an operations schedule.

Almost every method of valuing anything begins in the same place. You take the return available on an asset that cannot fail, and you build everything else on top of it as a series of premiums for the ways in which other assets can. The foundation stone has a name that has been repeated so often it has stopped being a claim and become a category. The risk-free rate.

What it actually says

The mechanism underneath the phrase is narrower than the phrase. A government that issues debt in its own currency will pay you back, because in the last resort it can create the currency in which the promise is denominated. That is default risk, and on that dimension a Treasury bond genuinely is about as close to free as finance gets.

It says nothing whatever about what your claim is worth between now and the day it matures. A thirty-year bond bought today at a yield of 5.27 per cent will pay 5.27 per cent for thirty years. If next year the same bond is issued at six, nobody wants yours at the old price, and the market marks it down until the arithmetic agrees. That is price risk, and the longer the promise, the more of it there is.

The number

Strip the adjective and put a figure under it. The exchange-traded fund that holds long-dated American government bonds, the most widely used way for an ordinary investor to own the risk-free asset at the long end, is down 12.96 per cent so far this year, measured against its locked opening price on 1 January and its verified close last Friday. It is the third-worst line on this publication’s twenty-six-asset Scoreboard, below everything except Ethereum and natural gas, which is unusual company for the asset the textbooks call riskless.

Over the same eight months gold, which promises nothing, pays nothing and can default on no one, is up 6.52 per cent. An investor who put money into the riskless asset and money into the useless one, and did nothing else, is nineteen and a half percentage points apart on the two.

Where the risk went

It did not disappear when the phrase was coined. It moved into duration, which is simply a measure of how much a bond’s price moves when interest rates move, and which rises with the length of the promise. “Risk-free” is shorthand for “free of risk if you hold it to maturity and are indifferent to its value for thirty years”, and there are very few investors of whom the second half is true. Pension funds with liabilities of matching length, and almost nobody else. Everyone else is holding an instrument whose name describes a property they will never get to use.

The phrase survives, in its own house

This is not a debunking, and the term is not marketing. At the short end it is close to exactly right: a three-month Treasury bill held for three months is free of both default risk and, for practical purposes, price risk, because there is not enough time for rates to move the value of a claim that short. The phrase was born there and it is honest there.

The failure is a transfer. A term that is true of a three-month bill was carried up the curve to a thirty-year bond, where only half of it survives the journey, and then placed at the foundation of every valuation model in use.

This week: the Displacer 3, the Augmenter 1. Running total from the Week 20 reset, the Augmenter 36, the Displacer 16. Thirteen weeks, four pillars a week, fifty-two points awarded, and the ledger reconciles.

It is the first week the Displacer has led since the reset. Its best previous result was a draw, three times. This is not a turning point and one week never is, but a run of twelve without a single outright loss ending is worth naming rather than burying in the arithmetic.

What the two sides are, and how the four pillars work, is explained in Appendix A13.

Pillar one, adoption speed: the Displacer

The Week That Was carries the transaction. What it says about adoption is the point here: a manufacturer of rockets has bought an artificial-intelligence coding tool for a sum larger than most listed companies are worth, in its own shares, and completed it. Whatever one thinks of the price, nobody pays it for a technology they believe is still climbing the slow part of an adoption curve.

Pillar two, labour: the Displacer

Here is the shape of it, and it is not the one the word "layoffs" puts in your head. American employers announced 33,429 job cuts in July, the lowest monthly total in two years, and the single most-cited reason for those cuts was artificial intelligence, at 10,970 of them. That is the fifth consecutive month in which it has led the table (Challenger, Gray & Christmas job cut announcement report, July 2026, published 6 August). Total announced cuts for the first seven months of the year were 477,033, down 41 per cent on the same period of 2025.

So the pool is shrinking and the machine is the biggest single thing in it. The caveat travels with the number and matters: these are reasons companies give in their own announcements, not an official statistic, and a firm cutting costs has an obvious incentive to name a technology rather than a mistake. Discount it as far as you like. Displacement that arrives as the leading stated cause inside a falling total is harder to see than a wave of redundancies, and it is not less real for that.

Pillar three, the type of shock: the Displacer

The cost of the build-out has arrived in a household purchase. One of the largest personal computer manufacturers has disclosed that memory is now around thirty-five per cent of the bill of materials for a machine, against fifteen to eighteen per cent before the shortage. A capital investment cycle that raises the price of the consumer’s own computer is transferring income from consumption to capital in the most literal way available, which is the Displacer’s demand-destruction argument arriving as a price tag rather than as a theory.

Pillar four, compute constraints: the Augmenter

The Augmenter’s case has always been that something binds before mass substitution does, and this week that something is the cost of money rather than the cost of silicon. The Analytical Takeaway sets out the mechanism. The consequence for this pillar is direct: a build-out financed at a thirty-year rate of 5.27 per cent, in a market where the borrower of last resort enlarged its own bid and could not move the price, meets an economic boundary well before it meets a technological one.

Four pillars, four different observations, each used once. The check is run explicitly every week because using one week’s fact twice is the error that required a correction in Week 30.

Nothing in this section is advice. When the table returns, the action labels on it will describe what this letter is doing with its own attention rather than what anybody else should do, and several of these strategies are not accessible to a private investor at all.

No new call this week, for the third week running. Nothing cleared the three-step screen this letter requires before a name is published, and the honest version of that is a short section rather than a name it would not stand behind. The horizon minimum is now twelve months, and a twelve-month call is a considerably higher bar to clear than a six-month one was.

Arista Networks closes next Friday

The call was published on 29 May at $147.00. It closed on Friday at $188.65, a return of 28.33 per cent. Its declared thesis horizon is 28 August, which is next Friday, and on that date it is scored in public against the condition pre-registered at entry and against its benchmark, and it is never reopened.

There is no new catalyst this week and it would be easy to pretend otherwise. The raised outlook and the switch demand story both belong to the results published in early August, and a continuation of an old catalyst is not a new one. What there is instead is a live risk that the company named itself: management has flagged industry-wide component shortages constraining shipments and pressing on gross margins.

Which is worth pausing on, because it is the same shortage the Stack Inversion section spends this whole edition measuring. The company whose shares this letter called on the strength of artificial-intelligence networking demand is now telling the market that the binding constraint on it is not demand but the availability of parts. A thesis can be right about the direction and then be overtaken by the mechanism it identified. Six days from now that becomes a number rather than an observation.

The shares fell 5.12 per cent this week, with the rest of the semiconductor complex.

Mitsubishi UFJ and YPF have no new company-specific development this week, so they sit in Portfolio Watch below with their prices, benchmarks and score dates, which is exactly what that appendix is for.

The debt passed forty trillion dollars on Tuesday and the interesting thing about the coverage was how much of it went looking for a comparison. It is enough to buy every listed company in Japan and Germany. It is a stack of banknotes reaching some distance towards the moon. It is roughly a hundred and seventeen thousand dollars for every American, including the ones born this week, who are not consulted.

None of these tells you anything. A debt is not a quantity of money, it is a claim on future taxes, and the only comparison that matters is with the income of the person expected to pay it. But the round number does one useful thing that the ratio never does: it gets reported. The debt passed thirty-nine trillion five months ago and almost nobody mentioned it, because thirty-nine is not a number that fits in a headline.

So we can now say with some confidence that the American fiscal position will next attract public attention at forty-one trillion, which on the current run rate is around the turn of the year, and that between now and then it will be no more or less serious than it was on Monday.

The Treasury, finding buyers thin,
Announced it would purchase within.
    The long end fell nine,
    Then thought better in time,
And ended it higher than seen.

Three things to watch. The thirty-year Treasury against 5.15 per cent, which is this week’s tell and the test of whether the government can move its own long end. Jackson Hole, from Thursday to Saturday, the first symposium chaired by Kevin Warsh, on the theme of financial innovation. And Arista Networks on Friday, when the best open call in this book stops being a thesis and becomes a number.

If all three move the same way, with the thirty-year holding above 5.15, a central bank that treats the long end as somebody else’s problem, and Arista closing well on demand it cannot fully supply, then the picture is coherent and uncomfortable: the AI build-out is intact and the cost of financing it is not coming down. If they diverge, and the thirty-year slips while Jackson Hole talks about the term premium (the extra yield lenders demand purely for lending far into the future), then Wednesday’s operation was the opening move of something larger and I have called it too small.

One of ours closes on Friday. Arista Networks went into the notebook on 29 May at a hundred and forty-seven dollars, benchmarked against the Nasdaq 100, with a scoring date fixed that day and never touched since. It closed this Friday at a hundred and eighty-eight sixty-five. We score it next Saturday against the condition we wrote down at entry, and we print the verdict whichever way it falls.

Until next week. Stay curious and stay hedged.
Anthony Rosenthal

Freight still leads the year at plus fifty-one per cent, but crude has closed to within thirteen points of it and now sits second, and the gap between best and worst across the twenty-six lines is seventy-two points. One commodity is doing most of the widening, which the appendix says plainly rather than dressing it as a broad repricing.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 21 August 2026. The basket is a fixed set, chosen on 1 January and unchangeable during the year: twelve equity indices, four bond and credit funds, six commodities, two currencies and two cryptocurrencies. Five rows are named by an abbreviation: MSCI ACWI (All Country World Index) is the broadest global share index; AGG is the US aggregate bond market; LQD is investment-grade corporate bonds; HYG is high-yield, the riskier corporate borrowers; and TLT is long-dated US government bonds. Nothing else is added, dropped or substituted mid-year, which is the only way a year-to-date table means anything. The single-company calls in Portfolio Watch are tracked separately. Annualised volatility shows how much each line typically swings in a year, judged from its last eight weeks: higher means bumpier, not worse, and a dash means we do not yet have enough weeks to measure it honestly.

Every close here is drawn from the same price record the table itself reads from, so the numbers on this page and the numbers in our record cannot drift apart. All closes are Friday 21 August 2026. MSCI EM is derived from the EEM ETF (exchange-traded fund) close of 67.12 multiplied by the locked index ratio (28.367); Baltic Dry is the Baltic Exchange Dry Index as published by Trading Economics (21 Aug, 2,841). Every year-to-date figure is recomputed from the locked 1 January baselines rather than carried forward, so an error cannot compound week to week.

Every company that has appeared in On the Radar remains tracked here until its thesis horizon. From this month every new call carries a horizon of at least twelve months, because a shorter window judges the weather rather than the climate; the calls entered before that change keep the shorter horizons they were given at entry, because a published date does not move. A company moves from On the Radar to this table when there is no new catalyst that week: the analytical call is intact, but there is nothing fresh to add. Arista Networks, whose thesis horizon closes on 28 August, is written up in full in On the Radar above rather than repeated here. Benchmark returns are recomputed each week from verified closes rather than carried: the Alphabet leg runs from 18 June, the Japan leg from 2 July and the Nasdaq leg behind Arista from 1 June, each being the first verified close in our price series on or after its entry date. Arista’s is three sessions after entry, because the 29 May index close in our price series is a web reading rather than a verified one and we will not publish an accountability figure off an unverified leg. It is being repaired before that call closes. Every return is measured against a benchmark chosen and fixed at entry, because a call that beats nothing you could have held instead is not a win.

Running record, and the denominator named so you can reconcile it: three calls are open, first mentioned in weeks 20, 23 and 25. Arista is up 28.33 per cent against a benchmark down 3.95; Mitsubishi UFJ is up 9.27 against a benchmark up 2.19; YPF is up 1.73 against a benchmark down 6.31. All three are ahead of the benchmark fixed for them at entry, by roughly thirty-two, seven and eight percentage points, and all three are also up in absolute terms, which will not always be true and is not the test. Arista’s own line sits in On the Radar rather than in this table, because it closes on Friday. Six calls have now closed at their thesis horizons since week 13, and the closed book is the one to read hardest: the two most recent were USA Rare Earth on 16 August, which scored plus one, and MP Materials on 9 August, which scored minus two after the condition pre-registered at entry fired on both limbs. The full ledger, both horizons and every loss, is published in the Quarterly Reckoning.