The Wrong Instrument

Almost nothing moved in the markets that price risk, and that is the place to start. American shares fell 0.80 per cent on the week and the technology index 0.59. Expected equity turbulence, on the index that measures it, finished at 15.84, which is low by the standard of any year this decade. The premium riskier borrowers pay over the American government sits at 270 hundredths of a percentage point, less than half the level at which it would start to matter. A hundred-dollar barrel, eight destroyed tankers and a rate rise in Frankfurt produced no reaction whatever in equities or in credit.

What did move, moved in two places. Government bond yields rose at every maturity, hardest at the short end. And crude rose 9.4 per cent. Every equity market on this letter’s twenty-six-line Scoreboard fell, and so did gold, silver, copper, freight and natural gas. Three lines rose alongside oil: the rand and the lira, both of which are the dollar strengthening rather than anything happening in South Africa or Turkey, and Ethereum, which is nobody’s story about a barrel.

The through-line is the quote at the top of this page. A central bank can set the price of money and it cannot set the price of a barrel, and this month three of them are being asked to do the second using the first. The cost of that attempt does not land on the people who produce oil. It lands on everybody who borrows.

Two of the eight signals in the crash framework are now red, the yield curve and the energy dial, and the six that measure credit, volatility and positioning are all green. That is what a shock to costs looks like before it becomes a shock to finance, and the two are not the same event.

Every yield in this edition is a Friday close on the US Treasury constant-maturity series, which is the standardised rate the Treasury publishes for each maturity, unless it says otherwise.

Two prints, one shape

The European Central Bank’s decision of 10 September raises the deposit facility rate to 2.50 per cent, the main refinancing rate to 2.65 and the marginal lending facility to 2.90, all with effect from 16 September. Underneath it sits the August inflation reading the bank published in the same breath: 3.3 per cent in total, 14.3 per cent for energy, and 2.2 per cent for everything except energy. The bank’s own staff now expect the headline rate to average 3.0 per cent this year, 2.5 next and 2.1 in 2028, and it described the outlook as highly uncertain with the risks to inflation on the upside and the risks to growth on the downside.

Twenty-six hours later the Bureau of Labor Statistics published the American August figures and the shape repeated. Consumer prices rose 0.4 per cent in the month, with petrol alone accounting for more than a third of it; core prices, which the Federal Reserve watches precisely because they exclude the things a barrel moves, rose 0.3 per cent in the month and slowed to 2.4 per cent over the year from 2.5. A month ago that second number would have been the story. This week it was not read at all.

Put the two documents beside each other and they say something neither says alone. The European Central Bank and the Federal Reserve are both being pushed to tighten into inflation measures that are falling, because a third measure neither of them controls is rising. Neither has hidden it. Both published it, in their own numbers, in the same week.

What a rate rise can actually do to an oil price

Nothing. A central bank cannot produce a barrel and cannot reopen a shipping route. Raising the cost of borrowing does not add one unit of supply to the thing that is short.

What it can do is narrower, and it is the actual argument for doing it. A one-off rise in a price is not inflation; it becomes inflation when it passes into wages, then into other prices, then into the expectation that this will keep happening. Tightening is a bet that dearer credit now makes that second round less likely later, and the bet is paid for by borrowers, none of whom moved the barrel. So the question is whether there is evidence, this week, that the second round is beginning.

There is, and it points both ways. The New York Federal Reserve’s consumer survey for August, published on 8 September, held one-year inflation expectations at 3.6 per cent and lowered the three-year to 3.2. Inside it, expected petrol-price growth rose 1.7 percentage points to 4.6 per cent, which is a second round starting to form. In the same survey the share expecting higher unemployment in a year rose to 44.4 per cent, the highest reading since April 2020 on the survey’s own series. Households expect fuel to cost more and the labour market to weaken at once, which is what a supply shock looks like from a kitchen table and is exactly what makes a tightening decision uncomfortable.

What the rate model says, and where this letter departs from it

This letter’s internal read of the Federal Reserve’s next move has crossed from Hold into hike-leaning this week. And it crossed on one input: the two-year breaking 4.45, which lifted the market block of the model by a full step. The inflation block is still reading softer on three-month momentum. The labour block did not change its score, and that is a judgement worth showing rather than hiding: August payrolls came in at 162,000 against a twelve-month average around a fifth of that, and one month five times the trend is not yet the trend.

So the published view does not move with the model, and this is the week to say why rather than to let the two quietly diverge. The rule this letter runs under is that the rate view only changes when two of the three blocks agree, and this week one did. The view stays at Hold with a hawkish tilt. What would move it is a second block, and the nearest candidate is the labour block: continuing claims, the number of people still drawing unemployment benefit week after week, breaking above 1.85 million, or a September payroll print that confirms August rather than reverting to the trend. And the Federal Reserve’s own rule book still does not justify a cut. The measure of that is the Taylor gap, the distance between what a standard policy rule prescribes and where rates actually sit: plus 1.28 percentage points, and stable. A positive gap means the rule says rates are too low rather than too high, which is why this letter has not called a cut all year and is not calling a rise now either.

One competing explanation deserves naming before the argument stands. The move may be positioning into a meeting rather than a repricing of the path: a market that has spent a month leaning towards a cut will cover that lean before a decision, and the result looks identical to a change of mind. The two are separable, and the separation is this edition’s tell. If it was positioning, the two-year gives the move back after the 16th whatever the committee does. If it was a repricing, it holds.

Crude oil rose 9.4 per cent and every other commodity on the board fell. A barrel that goes up while copper, gold and freight all go down is not telling you about world demand. It is telling you about one waterway.

West Texas Intermediate, the American crude benchmark, settled on Friday at 100.05 dollars a barrel against 91.48 seven days earlier, a rise of 9.4 per cent, and 21.4 per cent against the 82.40 it closed at on 14 August. The hundred-dollar line was crossed on Thursday, not Friday: crude closed at 102.48 on the 10th, its first close above a hundred this year, and gave back 2.43 of it on the 11th. Second-best line on the Scoreboard either way, up 58.3 per cent since January.

The cause is not a market one. It is eight destroyed tankers and a waterway that is barely open, and the detail belongs with the geopolitics rather than with the tape. What belongs here is the shape of the week. Brent, the European benchmark, touched 99 dollars on Tuesday and traded above 101 on Wednesday, its first time there since May, and the American barrel peaked on Thursday. Thursday is worth marking, because it is the same day the two-year Treasury crossed the line this letter had put its name to. Both did less on the Friday than the coverage suggested. And the move was not shared by anything else: every other commodity on the board fell in the same five days.

The whole American government bond curve sold off and the front end sold off hardest. The two-year yield rose to 4.63 per cent from 4.37, the five-year to 4.78 from 4.54, the ten-year to 4.96 from 4.78, the twenty-year to 5.38 from 5.25 and the thirty-year to 5.35 from 5.24. In basis points, which are hundredths of a percentage point, that is twenty-six at the front and eleven at the back.

The shape is the information. A curve that rises twice as much at two years as at thirty is a market changing its mind about the next eighteen months of policy rather than about the next three decades of inflation. It also narrowed the gap between the ten-year and the two-year to plus 0.33 percentage points from plus 0.41, which flatters nobody: that gap is the standing red on this letter’s crash framework and a narrowing does not clear it.

The Swiss Market Index fell 4.31 per cent on the week, to 13,775.27. On the eight-week volatility this letter prints in the Scoreboard’s own last column, 12 per cent a year, an ordinary week for it is a swing of about 1.7 per cent, so a 4.31 fall is roughly two and a half times that. It is the largest such multiple among the lines with a meaningful printed volatility, and not the only one above two: the aggregate bond index fell 1.05 per cent against a printed 3 per cent a year, which is about two and a half ordinary weeks for it as well, and for the same reason as everything else in this edition. The managed Turkish pair is excluded, because its printed volatility rounds to zero and a denominator of zero produces a number rather than a fact. It fell roughly two and a half times as hard as the German index, which was down 1.83 per cent, and four times as hard as the Euro Stoxx 50, down 1.06.

So this is not European weakness arriving in Zurich. The Swiss index is unusually concentrated in a small number of very large defensive companies, which means a move of this size is normally one or two names rather than a market. What this letter has not established is which ones, or why. No company-specific event has been verified to a primary source at the time of writing, and the honest position is that a reader holding European equities should expect to find a single-stock explanation rather than a macro one, and should check before believing either. It is reported here because it is the week’s largest statistical outlier and because a reader will see the number and want to know whether anybody looked.

Gold fell 1.44 per cent to 4,366.20 dollars an ounce in a week containing a hundred-dollar barrel, eight destroyed tankers and a shooting war in the Gulf. Silver fell 2.27 per cent, copper 1.93 and natural gas 4.84. The Baltic Dry Index, which measures what it costs to hire a ship to move dry bulk cargo, fell 3.34 per cent to 3,507, though it remains the best line on the board at plus 86.3 per cent for the year.

Gold pays no income, so its attraction falls when the yield on government bonds rises, and yields rose at every maturity this week. That explains a one and a half per cent fall without reaching for geopolitics. Copper is the more interesting number. It is the metal the market normally uses to read global industrial activity, and it fell in the week crude rose almost ten per cent. Those two usually move together when the world is growing. They move apart when one of them is being set by a blockade rather than by an economy, and this week they moved apart by eleven percentage points.

The week’s equity tape does not add much to that. American shares fell 0.80 per cent, the technology index 0.59, small companies 2.41. Hong Kong fell 3.30, India 2.09, Japan 1.55 and the British index 1.67. Only three lines on the board rose alongside crude, and two of them are the dollar rather than the country: the rand weakened 1.38 per cent and the lira 0.33. The third was Ethereum, up 2.27 per cent, which is the sort of fact this section reports and declines to explain.

One dial changed band this week and it is the same dial that changed band last week, one step further. Everything measuring credit, volatility and positioning is where it was, which is why a barrel through a hundred dollars lifts the composite by five points rather than fifty.

The other thing a risk scan should be looking at

The gauge measures the machinery of finance and it has nothing to say about what is happening inside the earnings of the market’s largest sector, so this is where that goes. Standard Chartered’s global chief investment office published a decomposition of consensus 2026 earnings growth for American equities on 8 September, drawing on FactSet and Bloomberg data: semiconductors up 123.3 per cent, technology hardware up 42.2 per cent, and software and services up 15.7 per cent. Those are their figures, from client research, and they are reported with the source named rather than presented as this letter’s own work.

Fifteen per cent growth is a good year for almost any industry. It is a rounding error next to a hundred and twenty-three. The concentration everyone worries about in the index is also a concentration inside the profit pool: the money is arriving in the layer that sells the equipment, not the layer that sells the applications. A companion note from the same team on 9 September put the sectors most exposed to being disrupted by the technology, software among them, at roughly six to eight per cent of global equities. The thing the machines are disrupting fastest is the software industry, which is the industry everybody assumed would be doing the disrupting.

Four of the eight readings are carried from earlier dates rather than freshly measured on Friday, and each one is asterisked and dated where it is printed.

Somewhere in a permitting office this month, a clerk is reading an application for a nitrogen-oxide allowance, and the answer will decide whether a data centre gets built this decade or next. That is not a metaphor. It is one account of the current state of the artificial-intelligence build-out, published by SemiAnalysis on 10 September, and it is the most specific one this letter has seen. It is their research and their reading rather than an established fact.

The headline number from that work: roughly seventy-five gigawatts of firm, binding orders tracked in the supply chain for power generated on site rather than drawn from the grid, with about twenty gigawatts of that ordered in the second quarter of 2026 alone. Seventy-five gigawatts is a quantity of generating capacity comparable to the entire electricity system of a mid-sized European country, contracted privately, for one industry. Generating your own power on site used to be what you did when the grid failed. It is now the plan.

Their explanation is the useful part. When the thing you produce with the power is worth far more than the power, you stop optimising the power bill and start optimising the calendar, and you will pay almost anything to be connected two years sooner.

Which moves the queue somewhere unexpected. On their account the binding constraint is no longer turbines, whose delivery slots are sold out past 2030 and now trade in a secondary market of their own, but reciprocating engines, the large piston engines that can be installed faster; and behind that, air permits. A site that would breach a nitrogen-oxide cap, or cross the threshold at which a far heavier review is triggered, is being redesigned mid-build to use different equipment. The technology choice is being made by the regulator rather than by the engineer. This is client research, it is credited to SemiAnalysis, and the figures are theirs.

The demand side of the same trade reported on Thursday. Oracle’s first-quarter results for its 2027 financial year, released after the close on 10 September, put revenue at 19.3 billion dollars, up 30 per cent over the year, with cloud infrastructure revenue up 121 per cent to 7.4 billion. The number that matters is the backlog: remaining performance obligations of 664 billion dollars on the company’s own release, up 209 billion over the year and 26 billion on the quarter before.

Remaining performance obligations are contracted revenue not yet delivered, a measure of commitments made rather than work done, and eight and a half years of the company’s current revenue signed in advance. Not all of that is incremental demand for the kind of computing the first half of this section is about, and the company says the new obligations are not expected to touch its revenue until 2028 or later. What it does establish is that the commitments are being made years ahead of the buildings, and the buildings need the gigawatts above, which need the engines, which need the permits. The contracts are signed long before the capacity exists, and on the research cited above the thing standing between the two is increasingly an emissions permit.

The crews of eight Iranian tankers had a bad week. American forces struck three of them on Sunday 7 September and destroyed five more on Tuesday the 9th, in stated retaliation for Iranian missile fire directed at American naval vessels, reported by CNBC, CBS News and NBC News across the 8th and 9th of the month. This is the third consecutive edition in which the Gulf leads this section, and the rule this letter runs under permits that only with a new, named event inside seven days. Eight destroyed vessels qualifies.

What is new in the analysis rather than the tape is the second chokepoint. The oil market has spent seven months watching the Strait of Hormuz, where crude flows are reported below two million barrels a day against eight to nine million before the war, on trade estimates this letter has not verified against a primary shipping source. MacroEdge Oil and Gas Research argued on 10 September that the binding constraint has moved to refined product rather than crude, and that Houthi positions along the Red Sea now give effective leverage over Bab-el-Mandeb, the strait at the southern end of the Red Sea through which traffic between Europe and Asia passes. Their case, credited to them.

If it holds, it changes what a settlement is worth. A Hormuz de-escalation reopens the barrel. It does not reopen the route that carries the diesel refined from it, and diesel is what moves lorries, tractors and ships. That is a reason to be careful about treating good news from the Gulf as automatically good news for the cost of moving goods, and it is the second time this letter has had to make that point about this war.

Japan, and the barrier that was lowered on 31 July

The flashpoint furthest from oil is a currency, and it is a week away. The Bank of Japan decides on 18 September. A Reuters poll conducted between 1 and 8 September found that all but two of sixty-eight economists expected a rise to 1.25 per cent, against 57 per cent of them in the previous round, and that twenty-four of sixty-six expected a further move to 1.50 in October or December.

The mechanism inside that poll is the part worth carrying. Eighty-two per cent of those economists said that the joint American and Japanese intervention to buy yen on 31 July, undertaken as the currency reached forty-year lows, together with remarks on Japanese policy by the American Treasury Secretary, had significantly or somewhat lowered the political barriers to raising rates. A co-ordinated currency intervention by two governments is a rare instrument, and this is the clearest statement anyone has produced of what it bought: not a level for the yen, but permission for a central bank to do something its domestic politics had been resisting.

That matters beyond Tokyo, and it matters directly to a position this letter holds. A Japanese bank earns more as domestic rates normalise, and the thesis in Portfolio Watch rests on exactly that. It also cuts the other way: a currency that strengthens on intervention and then on a rise makes Japanese exporters less competitive, and those exporters are the borrowers. Two limbs of one trade, pulling in opposite directions, and the meeting is on Friday.

One further item, reported rather than adopted. PIMCO’s United Kingdom team published a note on 10 September arguing that the rise in British government bond yields has tracked global factors rather than domestic political risk, and that deferred spending decisions are the thing most likely to dominate volatility at the long end into 2027. The British index fell 1.67 per cent this week, which is a large move for it. Their argument, named and dated because it is client research and because it cuts against the more comfortable domestic explanation.

In 1918 two men named Israel Cohen and Abraham Siegel opened a wholesale grocery business in Worcester, Massachusetts. A century later their firm, C&S Wholesale Grocers, was one of the largest privately held companies in America and still controlled by the Cohen family, and it was in the business that almost nobody wants: buying food in bulk, storing it in enormous sheds, and delivering it to supermarkets on a margin that rounds to nothing. Grocery distribution is a trade where a full percentage point of operating margin is a triumph. Everything in it is volume, and almost all of the controllable cost is the labour of people walking around a warehouse picking cases off pallets.

Richard Cohen, the founder’s grandson, ran that company. The interesting fact about him is not that he inherited a wholesaler. It is what he did with the problem he inherited, which was that the economics of his own industry were being squeezed flat and there was no product he could sell his way out of. He could not raise prices, because the supermarkets he served were themselves fighting on price. He could not cut the labour, because the picking had to happen. So he set out to rebuild the shed.

The renovator

There is a type of founder who arrives at an industry from outside, sees that it is absurd, and builds the alternative. Cohen is the other type. He had spent his working life inside the thing he set out to fix and went at the wiring of an old building rather than sketching a new one. The robots were not built to create a market. They were built to take cost out of his own warehouses, and tested there, on his own goods, with his family’s name on the building. Only afterwards did it occur to anybody that the machine was the product.

That is the whole origin of Symbotic, which was carved out of the family business, taken public in 2022, and is now a listed company in its own right with Cohen as chairman and chief executive. The system it sells is a fleet of autonomous robots that move cases inside a dense storage structure and assemble outbound pallets in the order a shop wants to unload them. That last detail is small and its consequence is not: a pallet built in aisle order means a shop worker stocks shelves in one pass instead of three.

Whether it matters, which is a fair question to ask of a robot company

It matters. In the third quarter of its 2026 financial year, reported on 5 August 2026, Symbotic recorded revenue of 721 million dollars, up 22 per cent on the same quarter a year earlier. It recorded net income of 55 million dollars against a net loss of 21 million in the comparable quarter. And adjusted earnings before interest, tax, depreciation and amortisation, a rough measure of what the operating business earns before financing and accounting choices, came in at 95 million dollars against 45 million. It guided the following quarter to revenue of 760 to 780 million. Those are the company’s own figures from its own quarterly release, and the scope is the quarter rather than the year.

Two further numbers give the scale. The backlog reported in the same quarterly materials stands at 22.5 billion dollars, contracted rather than hoped for, against revenue running near 2.9 billion a year. And on the reporting of that filing cycle, about 85 per cent of revenue in its 2025 financial year came from a single customer, Walmart, under an agreement running to 2037 to automate the American regional distribution centres of the largest retailer on earth. Its own annual filing for the 2023 financial year put that single-customer share at 88.4 per cent, against 94.4 the year before and 66.9 the year before that.

So this is not a company doing something interesting at the edge of an industry. It is the firm rebuilding the physical distribution system of the largest retailer in the world, and doing it profitably for the first time. It is also, on those same numbers, a company whose fortunes are eighty-five per cent one relationship.

The number that could not be billed

On 25 November 2024 Symbotic told the market it had found errors in its own accounts.

The company announced that it had identified errors in revenue recognition relating to cost overruns on certain deployments that would not be billable, and separately that goods and services relating to specific milestones had been expensed before those milestones were achieved. It estimated the total effect of correcting the errors at a reduction of 30 to 40 million dollars to system revenue, system gross profit, income before tax and adjusted operating earnings for the 2024 financial year. Those figures are from the company’s own restatement announcement, and the scope is the full 2024 financial year rather than a single quarter. A securities class action followed.

The mechanism underneath the accounting language is the thing worth carrying away. Symbotic’s large contracts had been widely understood as cost-plus: the customer pays what the job costs, plus a fee, so if steel goes up the customer absorbs it. What the restatement exposed is that those contracts also carry cost caps. Above the cap, the overrun stops being the customer’s problem and becomes the contractor’s. When installations ran late or ran inefficiently, Symbotic was not passing the cost on. It was eating it, and it had been booking revenue as though it were not.

Read that beside the business it is in and the irony does the work without help. The company that exists to take the uncertainty out of other people’s warehouse operations, whose entire proposition is that a robot does a job in a predictable number of minutes at a predictable cost, could not predict the cost of installing its own robots. Selling precision and performing precision are different trades. The first can be priced in a contract. The second happens on a building site in Arkansas in the rain, with a schedule, and it is where the money actually goes.

The transferable part is smaller and more useful than a moral about hubris. A contract described as cost-plus is only cost-plus up to its cap, and the cap is where the risk lives. Anybody reading a supplier agreement, a construction contract or a fund’s expense arrangement can ask the same question the restatement answered the hard way: at what number does this stop being somebody else’s problem? It is a question with an answer, it is written down somewhere, and almost nobody asks for it until the overrun has already happened.

Where this sits

This letter runs a full analytical screen on any company before it is named, and this one was run before the profile was written. It agrees with the three claims above and adds a fourth. The quality is genuine and improving: profitable for the first time, a contracted backlog, and switching costs about as high as they get, because a system poured into a building’s concrete is not displaced by a cheaper bid but by a rebuild. The concentration is the dominant risk rather than one risk among several, and the backlog does not dilute that so much as term it out. And the restatement is direct evidence that those contracts allocate execution risk to the supplier in a way the market had not priced.

The fourth thing the screen says is about price, and it is worth printing because it cuts against the tone of everything above. Take the company’s own guidance for the current quarter, 100 to 105 million dollars of adjusted operating earnings, annualise it, and set it against a market value a little under 25 billion. That is a multiple close to sixty, on a derived figure rather than a published one, and it is stated as derived. The shares are not cheap on the concentration risk. They are expensive in spite of it, at a price that requires this year’s margin inflection to hold across the remaining decade of deployment. Which is the strategic uncertainty in one sentence: the question is not whether Walmart leaves, it is what the price of staying turns out to be.

None of that is a recommendation. Symbotic is not an On the Radar call, it carries no entry price and no horizon, and nothing here is a view on where the shares go. What earns it the chair this week is the shape of the story: a man who spent his career inside a one per cent margin business, built the machine that fixed it, sold the machine to the industry, and then found that the hardest thing to automate was his own installation schedule.

Nought of the two conditions that would say the shortage in computing hardware is genuinely breaking have been met. Six of the seventeen shares that live off that shortage are more than a fifth below their own three-month high. Those are different facts and they are not averaged.

The two conditions

Memory prices falling: not met, last verified on 5 September. The condition is contract prices for conventional memory chips printing down month on month, not merely rising more slowly. They are still rising, at roughly plus 13 to 18 per cent in the most recent printed quarter against plus 58 to 63 in the one before and plus 90 to 95 before that, with plus 8 to 13 guided for the current quarter, on the TrendForce series. The rate of increase has collapsed across three quarters and the direction has not changed, which is exactly what the condition was written to distinguish. The reading is a week old, carried and dated.

Funding stress at the two large financiers of the build-out: not yet built. Fourth in the build order, and unbuilt until the five instruments involved are named and shown to be both genuinely tradeable and genuinely dependent on the contracts in question. It is neither met nor unmet, and saying so is the only accurate thing available.

The book, reported beside the thesis and never merged into it

Six of the seventeen shares tracked against this thesis sit more than 20 per cent below their own ninety-day high, measured from our own price file on Friday’s closes. The worst, on our own price record, is the Korean memory maker SK Hynix at 37.9 per cent below its own ninety-day peak, an improvement on the 43.6 per cent recorded a week ago. Behind it: the power producer Talen at 28.6 per cent, the data-centre equipment maker Vertiv at 28.2, the Chinese foundry Hua Hong at 24.1, the fuel-cell maker Bloom Energy at 20.3 and the compute landlord CoreWeave at 20.0.

The composition moved even though the count did not: the broad semiconductor fund left the list at 19.5 per cent below its peak and CoreWeave entered it. A thesis about where a bottleneck sits and a description of what a group of shares has done answer different questions, and they are printed separately because for a stretch of the summer this instrument read the same verdict every week while the worst name in the book was down more than half.

Reported, and not scored

The open-market rent for an H100 graphics processor, the chip most artificial-intelligence work is trained and run on, closed Friday at 1.467 dollars an hour for capacity that can be taken away from you, against 1.267 a week earlier, across forty live offers. Guaranteed capacity, which cannot be reclaimed, closed at 2.761 dollars against 2.896.

That answers a condition this letter set last week, and it answers it more bluntly than the question was asked. The test was whether the interruptible rent would stay below 1.30 dollars for two consecutive weeks or bounce back above it. It never came close. On our own daily record the rent was below 1.30 on exactly one day, Friday 4 September, and was back above it the next day; every reading since has been higher. The two-week test is dead rather than pending, and the reading that the rental market had entered a durable glut lasted a single session. The rent remains below the estimated all-in hourly cost of owning and running the machine, 1.64 dollars, which is a locked first-publication figure rather than a fresh one. It is not defended by a published cost build, so it is shown and it earns no verdict.

Three further price series are recorded as not observed rather than merely missing: the H200, B200 and B300 columns, populated in two weeks out of eight, and the frontier token tier, identical five weeks out of six and then fivefold on a change of sampling.

The constraint we can actually observe

Memory, unmoved, for the fifth consecutive week. This is the constraint the evidence in front of us identifies, which is not the same claim as its being the binding constraint across the whole of the compute stack, and the difference is worth keeping. The component market is tight and the rental market is loose, and this week the rental market got tighter while the components stayed where they were. Nothing in the evidence suggests the bottleneck has moved somewhere the two conditions do not look, and that remains the failure this instrument is most exposed to.

Three things to watch, and they are not three of the same thing. Two of them can break the thesis and the third is a price worth watching, and they are kept apart deliberately. Break condition: whether the memory contract series prints a down month rather than a smaller up month at its next quarterly reading. Watched price: whether the interruptible rent makes a second visit below 1.30 dollars at all. One touch in one session is noise; a second, inside a month of the first, would be the start of a series. Break condition: whether either of the two large obligors financing the build-out raises primary capital carrying structure rather than plain equity, meaning preferred terms, conversion rights or guarantees attached to a raise that used to be ordinary shares. That is the cheapest available signal of funding stress and it is public the day it happens. The rent is the watched price, and a watched price is not a break condition however interesting it gets.

What would re-tighten the scarcity: a memory supplier withdrawing or delaying announced capacity, or a new binding constraint appearing somewhere neither condition currently looks.

Positions changed by this gauge since it began: zero. That line stays on the page every week until it is not zero. How this is built.

A million humanoid robots in 2030 would consume roughly ten thousand silicon wafers a year, which in semiconductor terms is nothing at all. What they would consume is memory: the chip at the centre of one carried 32 gigabytes four years ago and carries 128 today. And the thinking cannot be done on board. SemiAnalysis reported on 9 September that running one of the current world-action models on the newest data-centre processor tops out at seven robots per chip at just over a second of worst-case response time, so a single eight-chip server serves fifty-six robots and the machine in your warehouse is a terminal. Robotics does not arrive as a boom in edge silicon. It arrives as more demand for data centres, and as more pressure on a memory market that this edition has just reported is still getting tighter.

The thing most people believed on Monday is that the American share buyback is the strongest floor under this market in years, and the aggregate supports them. Neuberger Berman’s Market Signals, on data to 30 June, puts repurchases over the trailing twelve months at a record 1.10 trillion dollars. That is their figure, from client research, and it is the number everybody quotes.

The composition inside it goes the other way. On the same work, within the artificial-intelligence complex the companies spending most heavily on capital expenditure cut their buybacks by around a third. Which is to say that the firms with the largest weights in the index, the ones whose share count actually moves the index-level earnings figure, are the ones withdrawing from the bid while the total sets a record.

A buyback does two things. It puts a buyer in the market, which supports the price, and it shrinks the share count, which raises earnings per share without the business earning a penny more. The second effect is the one that has quietly flattered the largest American companies for a decade. Take a third of it away from precisely the names carrying the index and you have removed a tailwind from reported earnings growth at the moment those same companies are spending record sums on buildings and chips. Nothing about the headline trillion tells you that. The headline is the aggregate, and the aggregate is being held up by everybody else.

The transferable question is the one to take away, because it generalises well beyond buybacks: when a total reaches a record, ask which of its parts is doing more and which is doing less. A record can be set by the same people doing more or by different people arriving. Those are opposite facts and they look identical from the outside.

And a second place where the price and the fundamentals disagree

High-yield corporate bonds, the borrowings of the riskier companies, had a poor week in a market that did not. The exchange-traded fund that tracks them fell 0.71 per cent, which sounds like nothing until you know that the Scoreboard’s own volatility column puts it at 3 per cent a year, which is a typical week of about four-tenths of one per cent. Measured on that, a 0.71 fall is roughly one and three-quarter ordinary weeks in a single one, and it happened while the American index fell only eight-tenths of a per cent and while the credit spread this letter tracks widened by five basis points, which are hundredths of a percentage point. Something small moved in the price of risky credit.

Cutting directly against it, Apollo’s asset-backed finance note of 10 September argues that the prevailing story about a deteriorating American consumer has the distribution wrong: on their work, net-worth growth for the bottom half of the income distribution has outpaced other groups since the end of 2019, with wage growth holding up. Their argument, credited to them, and it is a genuine challenge to a narrative this letter has itself carried.

Both cannot be describing the same thing. A market price saying credit is softening and a set of household balance sheets saying it is not can be reconciled in two ways: either the price is early and the balance sheets will follow, or the price is about something else entirely, most plausibly the rate move rather than the borrower. The second is the duller explanation and this week it is also the better one, because the whole bond curve sold off and high-yield bonds carry interest-rate sensitivity like everything else. Which is worth saying plainly: the more interesting reading is not always the right one, and a fall of seven-tenths of a per cent does not need a story about the American consumer to explain it.

Volkswagen told its workforce this week that a further fifty thousand jobs will go, taking the total programme to a hundred thousand. It is the largest European redundancy programme of this cycle and it lands in the middle of an argument about machines replacing people. The company’s own stated reasons are energy costs and structural competition. Not automation.

This week: the Augmenter 2, the Displacer 2. A draw. Running total from the reset at Week 20: the Augmenter 43, the Displacer 21. Sixteen weeks at four points a week is sixty-four, and 43 plus 21 is 64, so the ledger reconciles.

Adoption speed, to the Displacer. Oracle’s order book, reported earlier in this edition, is many years of its own current sales already signed. Whatever friction is supposed to be slowing enterprise deployment, a book of that shape is not the evidence for it.

Labour, to the Augmenter, and the Volkswagen number is the evidence rather than the exception. A hundred thousand jobs are going in European manufacturing and the company attributes them to the price of energy and to competitors, which is the attribution this framework is obliged to use rather than the one that would be more dramatic. No large employer named artificial intelligence as the cause of a headcount reduction in a filing this week. That is an absence, so a positive fact is carried beside it: the Economist puts the number of American jobs created by the technology at around a million, and work by the Ramp Economics Lab finds that firms adopting it hire rather than replace. The visible job destruction of 2026 is being done by an energy price. The technology is not yet doing it.

The type of shock, to the Displacer. The earnings split reported in the risk scan above is close to eight to one between the layer that sells the equipment and the layer that sells the applications. Gains concentrating in the owners of the machinery, rather than spreading through the businesses that use it, is the concentration of income in capital that this pillar exists to measure.

Compute constraints, to the Augmenter. The on-site power orders described earlier in this edition, and the fact that the thing standing in their way is now an emissions permit rather than a price. That is a physical ceiling forming in front of substitution rather than a cost curve falling through it.

What would flip the score. A named large employer attributing a headcount cut specifically to this technology in a filing hands the Displacer the labour pillar outright. A memory contract price that finally turns downward takes the compute pillar. And a quarter in which the software earnings line grows faster than the semiconductor line takes the shock pillar straight back to the Augmenter.

How this is built.

Consider the position of a central banker this week, holding an instrument that sets the price of money and facing a problem that is the price of a barrel. The European Central Bank did the only thing available to it, published the figures showing why it was not the right tool, and raised rates anyway. There is something almost admirable in the candour, and something faintly absurd about a document that contains both the decision and the evidence against it.

Elsewhere in the week’s small print: the crash framework on this page turned one of its dials red on a four-week comparison with a barrel that traded on 14 August, a date now four Fridays behind us and receding at the rate of one Friday a week. Hold crude exactly where it is for seven days and the dial goes back to amber without a single trade. The measurement moves even when the world does not, which is worth knowing about any indicator that compares today with a fixed point in the recent past, including this one.

The bank, with a barrel to fight,
reached first for the rate overnight.
Though the core had come down,
and the tanker had drowned,
they tightened, and called it a fight.

Three things to watch next week, and two of them land on the same Friday. The Federal Reserve’s decision on Wednesday 16 September, with the target range at 3.50 to 3.75 per cent and a committee that held 9 to 3 in July with three members already preferring a rise. The two-year Treasury at Friday’s close, against the 4.63 per cent this letter has put its name to. And the Bank of Japan on Friday 18 September, where the economists polled this week were all but unanimous. If all three move the same way, meaning a hawkish Federal Reserve, a two-year that holds above 4.63 and a Japanese rise, then the last two large central banks still supplying cheap money have both stopped, inside seventy-two hours, because of an energy price neither of them can influence. That is the configuration the whole of this edition has been describing, and it would be the first week since the war began in which the policy response, rather than the shock, was the story. If they diverge, and the likeliest divergence is a Federal Reserve that holds and talks the front end down while Tokyo goes ahead, then the hawkish read here was a reaction to one inflation print and this letter will say so next Saturday in the same place it said this.

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

Crude holds second place on the board for a second week, at plus 58.3 per cent for the year, behind freight and ahead of every equity market on it. The gap between best and worst across the twenty-six lines narrowed from 110.0 points to 105.8, and every bit of that came from freight falling back at the top. The bottom of the board got worse rather than better: natural gas widened it by two points.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 11 September 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 11 September 2026. MSCI EM is derived from the EEM exchange-traded fund close of 67.84 multiplied by the locked index ratio of 28.367, which gives the 1,924.42 published above; the fund price and the index level are two scales of one series, and the ratio is locked and never re-derived. Baltic Dry is the Baltic Exchange Dry Index as published by Hellenic Shipping News, 11 September, 3,507. 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.