The Instrument That Did Not Move

Two numbers describe this week and they disagree. The Federal Reserve’s target range rose for the first time in more than three years. The gauge this letter publishes to measure how much crash risk is assembling did not budge, and reads twenty-five out of a hundred for the second week running. Both are accurate, and holding them together is the work of the next two sections.

The bond market did the moving, and it moved at one end only. The two-year Treasury yield, the maturity that prices what the central bank is expected to do next, rose sharply. The thirty-year did not move at all. A market that reprices the next two years and leaves the next thirty alone has changed its mind about policy and not about the long-run price of lending money, and that is not what a straightforward inflation scare looks like. The third section has the figures and the second has the argument.

Our own instrument read the same number this week as last, and it did so because one of its dials improved by exactly as much as another deteriorated. A reading that has not changed is not the same thing as a world that has not changed, and this week it is emphatically not.

The gauge below reads what it read last Saturday, and it is the most misleading number in this edition. Two of its eight dials changed band this week, in opposite directions, by exactly the same amount.

Every yield in this edition is a Friday close on the US Treasury constant-maturity series for 18 September 2026, read from this letter’s own rate record and not from any figure printed elsewhere on this page.

The hike, and what it was actually aimed at

The Federal Reserve did not raise rates this week because inflation is high. The working reading in this letter is that it raised them because it has decided this particular inflation will not pass through and disappear on its own, which is a judgement about next summer rather than a reading of this month. That is an inference about the committee’s reasoning and not a quotation of it: the statement itself says only that inflation remains elevated and that the move supports a timelier return to the two per cent goal. Everything else in this section is the working behind that sentence.

The decision, from the statement of 16 September: the target range for the federal funds rate goes to 3.75 to 4.00 per cent, a rise of a quarter of a percentage point, on a vote of twelve to nothing. It is the first increase since July 2023. The accompanying projections show a median expectation for the end of 2026 above the new range, and sixteen of the eighteen participants expect at least one further rise this year.

Now set that against what the committee can see. Core inflation, built to exclude food and fuel, has been easing on the annual comparison. The measure the Federal Reserve actually targets is PCE, the Personal Consumption Expenditures index, and the August reading is not published until around 26 September, so the committee moved on the earlier consumer-price figures and on its own forecasts. Meanwhile the price it does not set went the other way over the month.

So the tightening is not aimed at the current number. It is aimed at what economists call second-round effects, which is the point at which a one-off rise in the price of fuel starts to show up in the price of everything that is moved by fuel, and then in what people ask to be paid. A central bank cannot produce a barrel and cannot reopen a strait. What it can do is make it more expensive to fund the wage and price increases that a fuel shock invites, and that is a real mechanism, not theatre. It is also a mechanism whose whole cost lands on borrowers and whose benefit, if it works, is an inflation rate a year from now that nobody will be able to attribute to it.

The number that did not move, and why that is the week’s most instructive fact

The energy dial improved from red to amber, for the calendar reason set out in the scorecard above. And the breadth dial deteriorated from green to amber: the share of companies in the American index trading above their two-hundred-day average price fell to roughly fifty-seven per cent from sixty-four. Breadth is the count of how many members of an index are participating in its direction, and it is the difference between a market being carried by its whole membership and a market being carried by a handful of very large companies. Five points off for one, five points on for the other, and a composite that looks untouched.

One caveat belongs here rather than in a footnote, because it is the number the week turns on. The breadth reading is an estimate spanning roughly 55 to 57 per cent, and the amber band runs from 40 to 60. At 61 per cent the dial would be green, the composite would read 20.0 and the story in this section would be a gauge that fell five points. The reading is low-precision and the conclusion is therefore held with less confidence than the arithmetic implies.

A reader who saw only the composite would conclude nothing happened. The more accurate reading is that the risk in this market rotated: away from the price of fuel, which is easing on a four-week view even at a hundred dollars, and into the narrowness of the equity market itself. That is a worse trade than it sounds, because a fuel shock is visible and self-limiting and a narrowing market is neither.

What the rate model says, and where the published view sits against it

This letter’s internal read of the next move is built on ten indicators in three blocks, anchored to a Taylor Rule prescription, which is simply an arithmetic rule for what a policy rate ought to be given inflation and unemployment. It reads hike-leaning, and that is identical to last week.

The hike itself did not move the model, which is worth explaining rather than glossing. The increase lifted the midpoint of the funds range to 3.875 per cent and so narrowed the gap between the actual rate and the rule’s prescription of 4.90 to 1.03 percentage points, from 1.28. But a gap that is both positive and narrowing scores the same either way on this panel, so the anchor did not move; the two-year at 4.76 holds the rate signal where it already was. The model is also running on three-quarters of its panel weight, because several inputs had no fresh release this week, which makes the composite provisional and is stated and not buried.

The published view moves only when two of the three blocks agree, and this week only one did. So the view is a hawkish hold on the path: the front end has stopped pricing cuts, and the next argument is about the long end rather than the next meeting. A positive Taylor gap still means a cut is not yet justified by the rule, which is the guard against calling the turn early.

One competing explanation has to be named before that stands. Friday was a triple-witching session, when large tranches of options and futures expire together, and it fell in a week of heavy Treasury supply. Either could produce a nine-hundredth jump in the two-year that has nothing to do with anybody’s view of policy. The two readings are distinguishable and the test is cheap: a flow effect unwinds inside two or three sessions, so if the two-year is back below 4.70 by Wednesday this was a flow effect. If it holds above 4.70 into the following week, it was a revision of the path.

The American index moved eight hundredths of one per cent this week. Its largest members rose almost a per cent and its smallest fell a percentage and a half, so the flat headline is an average of a disagreement rather than a description of calm.

Minus 0.08, plus 0.94, minus 1.50. Three numbers from the same market over the same five sessions: the broad American index, the hundred largest non-financial companies on the Nasdaq, and the index of smaller American companies. Nearly two and a half percentage points of daylight between the top and the bottom, and none of it visible in the figure that gets reported.

What that spread buys is a different market from the one the headline describes, because the index held its level on the work of a small number of very large companies. This letter’s breadth dial, the share of index members trading above their average price of the last two hundred days, fell from roughly sixty-four per cent to fifty-seven in the same week. Both facts describe one condition: the market is being carried by fewer names than it was a fortnight ago. That condition can persist for a long time and it has done so before. It is also the condition in which an index looks safest and is least so, because the average conceals how few things are holding it up.

Government bonds sold off, and they did it at one end of the curve and not along the whole of it. The two-year Treasury yield rose 13 basis points to 4.76 per cent, the five-year 8 to 4.86 and the ten-year 5 to 5.01. The twenty-year did not move. The thirty-year fell a hundredth, to 5.34, and was the only maturity on the table to do so. The gap between the ten-year and the two-year narrowed from 0.33 to 0.25.

That pattern has a meaning, and it is not the obvious one. A curve that flattens because its short end is rising is a market saying the policy rate will be higher for longer than it previously believed, while simultaneously declining to ask for any extra compensation for lending over thirty years. Those two positions sit awkwardly together. If rates really are going to stay high, a thirty-year lender at 5.34 is being underpaid; if they are not, the two-year at 4.76 is too high. Those two positions can be reconciled, but only by assuming the long-run price of lending money is unchanged and that lenders are content to be paid very little for the extra years. The two ends are expressing a tension the curve has not resolved, and this week’s tell is a bet on whether that assumption survives.

Digital assets had their third consecutive strong week while equities went nowhere. Bitcoin rose 4.75 per cent and Ethereum 3.94. Both remain well down for the year, Bitcoin by 7.94 per cent and Ethereum by 11.99, so this is a recovery from a poor start rather than a new advance. What makes it worth a line is the company it keeps: gold up 1.34 per cent, silver up 3.10, copper up 2.25. Monetary metals, industrial metals and digital assets all bid in a week when equities were flat and bonds fell. That combination is hard to reduce to a simple flight from risk. One reading is that money is moving out of instruments whose value depends on a discount rate and into instruments whose value does not. The simpler competing reading is that the industrial metals are following the same physical tightness the freight and crude lines have shown all year, and copper has a large industrial-demand component either way. The two separate on copper: on a discount-rate rotation it should lag gold as the front end reprices, and on a physical story it should lead.

American retail sales rose 1.2 per cent in August in the Census Bureau’s release on Wednesday, the strongest monthly figure in five months and well ahead of the 0.8 per cent expected. July was revised to minus 0.5 per cent from minus 0.6. Total sales were 773.9 billion dollars, six per cent higher than a year earlier. This matters because it is the opposite of the print a central bank would want before tightening into a fuel shock, and it removes the negative-consumer flag this letter has carried since July. The American consumer absorbed a hundred-dollar barrel and went shopping.

Dry bulk freight fell 3.91 per cent on the week, from 3,507 to 3,370 on the Baltic Dry Index, which measures the cost of shipping raw materials by sea. It comes after a fortnight of daily gains of four and five per cent, so the interesting question is whether the freight spike has rolled over. If it has, it belongs beside the two distillate builds in the previous section as evidence pointing the same way. The index measures dry-bulk shipping rates and nothing wider, so it is one input to that question and not an answer to it. The index is still 79.06 per cent higher than it began the year, which is the largest year-to-date move on this letter’s Scoreboard and has been for most of the year.

Europe was the weak spot, with the euro-area benchmark down 1.41 per cent and Germany down 1.03, both worse than America on the week and both a long way behind it for the year. Japan was the exception among the large markets, up 1.57 per cent and now 25.45 per cent higher for the year, the strongest equity index on this letter’s Scoreboard.

Two dials changed band this week, in opposite directions and by the same amount, so the composite is unchanged at twenty-five. The risk did not go away. It moved from the price of fuel into the narrowness of the equity market.

The thing a risk scan should have been looking at

This framework measures the machinery of finance, and the machinery is calm. What it does not measure is whether a large, capital-hungry building programme can still fund itself, and that is where the week’s most concrete risk sat.

Consider one company as the instance, not as the argument. Kodiak Gas Services rents out equipment that compresses natural gas, and it has spent this year turning itself into a supplier of electricity to data centres, buying a fleet of engines and turbines in April and signing an equipment framework with Baker Hughes in July for roughly a gigawatt of generation by 2030. It paid for that with borrowed money and new shares. The deconstruction column later in this edition takes the financing apart properly.

Now the number that is missing. On its own second-quarter disclosures and earnings call of early August, the contracted offtake on that new generation fleet, meaning megawatts a customer has actually signed to buy, is zero. The most advanced item is a limited notice to proceed on engineering for a project under a hundred megawatts, with the long-term contract still being negotiated. By contrast its old compression business is described as already half contracted for its 2027 deliveries. So the company has committed to build, has raised the money to build, and has not yet sold the output.

This is not a prediction that Kodiak fails, and this letter is not making a call on it. It is an illustration of where the financing risk in the artificial-intelligence build-out actually sits, which is not in the chips and not in the models but in the ordinary corporate balance sheets that are buying long-lived equipment on borrowed money against demand that has been described rather than contracted. Its shares fell 23 per cent between late June and this Friday, through a raised profit forecast and through that Baker Hughes announcement. The market has already begun to mark this. The dials above cannot see it, and saying so is more useful than pretending the composite covers everything.

Four of the eight readings are carried from earlier dates and not measured this week, and each is marked as such on its card and in the appendix: bond volatility, factory new orders, insider selling clusters, and the breadth figure, which is a mid-September estimate rather than a daily print and is marked low-precision. Nothing carried is presented as fresh.

The cost of doing a unit of artificial-intelligence work is falling far faster than the cost of the electricity to do it, which moves the binding constraint from the chip to the building and the permit.

The research house SemiAnalysis published a modelled comparison on 14 September of the newest data-centre computing rack against the generation it replaces, and put the improvement at roughly sixty-seven times the work per dollar of total ownership cost on a like-for-like basis. Two cautions belong in the same breath. That is a model instead of a measurement, built on assumed utilisation and assumed power prices, and the honest way to read any such figure is as a direction with a very wide error bar. And the comparison is against a single prior generation, so it is not an annual rate of improvement.

Take it as directional and it still matters, because it cuts against the argument this letter has weighed all year: that the cost of computing would become the natural ceiling on how far automation could go. If the work per dollar improves at anything like that pace, the ceiling is not the chip. It moves to whatever cannot be improved at the same speed, which is electricity, land, transformers and permission.

The same house argued on 13 September that the memory shortage now constraining the whole complex has a route out that does not require a single new factory: a change in how memory is stacked, described as four-high rather than the current arrangement, which raises output from existing capacity. Their own timeline puts the relief in 2027, and they say plainly that near-term prices go the other way. That matters for the section below on the compute stack, where memory contract prices printing down month on month is one of only two conditions this letter watches for evidence that the scarcity is genuinely breaking. A credible route to relief in 2027 is not the same as relief, and the difference is worth eighteen months of positioning.

Apollo published a comparison on 17 September that is more useful than most things written about European competitiveness. American data-centre construction has been financed in part through a specific market: bundling the lease payments from finished data centres into bonds and selling them, which lets a builder recycle capital instead of holding every project on its own balance sheet. Their figures put that market at roughly 81.4 billion dollars in the United States since 2018, with about 18 billion in the first half of this year alone. The comparable European figure is about 1.7 billion, and the British figure about 2.3 billion.

If those numbers are right, and they are their numbers, not ours, then one important constraint in Europe is the underdevelopment of the machinery that turns a finished building into a tradeable security, which is a regulatory and legal question with an identifiable cause and therefore a fixable one. That is not the same as saying Europe has no shortage of capital or of engineers, which these figures do not test. That is a more hopeful diagnosis than the usual one, and a more specific one.

Saudi Arabia built a pipeline across the peninsula specifically so that its oil would not have to pass through the Strait of Hormuz. With the pipeline shut, the oil is going back through Hormuz, under American naval escort.

The crews doing it are on short-haul tankers. The route, as reported this week, runs from Saudi terminals inside the Gulf, out through the strait with a naval escort, and then transfers the cargo to long-haul vessels standing off Sohar on the Omani coast. Each barrel is now handled twice, in the one stretch of water everybody had spent a decade trying to avoid. Aramco is reported to have sold about twenty million barrels to Asian refiners this week for collection just outside Hormuz.

The reason is that the alternative failed. Drone strikes on 10 and 11 September, which the office of the Iraqi prime minister has said originated in the south-eastern province of Maysan, hit a pumping station on the East-West Crude Oil Pipeline, the 1,200-kilometre line built in 1981 to carry crude from the eastern fields near Abqaiq to Yanbu on the Red Sea. Satellite imagery published on 14 September showed fire damage around the station. Riyadh shut the line as a precaution. No group has claimed responsibility and neither Saudi Arabia nor Iraq has publicly named one.

The scale is what makes it a market event and not a news item. The kingdom had been routing roughly five million barrels a day through that pipeline precisely because the strait was already effectively closed. So the redundancy and the thing it was redundant against both failed inside a fortnight, which is the specific risk that redundancy is supposed to remove and almost never does: two routes that are independent on a map are not independent if one adversary can reach both.

The new development this week, and it is the reason crude fell back rather than kept climbing, is that the line is expected to reopen. On 16 September American officials said the damaged pipeline would restart within days, and the oil price fell on the statement. Independent assessments reported this week expect a restart within a couple of weeks at something like forty to sixty per cent of capacity, which on a line carrying roughly five million barrels a day is two to three million. Crude peaked near a hundred and six dollars on Tuesday and gave most of that back over the following three sessions. Riyadh has published neither a damage assessment nor a timetable of its own, so the restart expectation currently rests on third-party statements instead of on the operator.

What a reader should take from the price action is narrow but real: the market is pricing a repair, not a resolution. Hormuz flows are reported below two million barrels a day against a normal eight to nine million. A hundred dollars with the bypass down and the strait constricted is not the same hundred dollars as a hundred dollars with both working, and the difference is entirely in what happens if the repair slips.

Away from the Gulf, and this is the flashpoint worth watching precisely because it is not about oil transit. Albertans vote on 19 October on whether to hold a future binding referendum on separating from Canada. The provincial government launched a four-million-dollar advertising campaign on 10 September urging a vote to remain, which tells you how seriously it is being taken by the people closest to it. Polling through early September suggests a broad majority want to stay, while separatist organisers say publicly that they expect to win regardless of what the polls show.

The financial interest is not the constitutional question, which on the polling is unlikely to resolve in separation. It is that Alberta is a very large gas producer with export capacity well below its production, so it sits on fuel that is cheap because it is difficult to move. Research this letter reads makes the argument that this is exactly the combination data-centre developers are hunting, and that the political question determines who gets to negotiate access. That argument is a source claim rather than a verified one and is carried as such. What is verifiable is the date, the campaign and the money behind it, and a constitutional vote in a major energy-producing province is a risk that does not appear in any of the eight dials in this edition.

Bom Kim cancelled his own stock market listing on the weekend it was due to go to the printers, when the company was already profitable, and then spent seven years and four billion dollars of accumulated losses proving he had been right to.

The architect

Most founders in this series inherited something. A broken industry, a mispriced asset, a technology somebody else invented and nobody had applied. Bom Kim is the other kind, the one who looks at an empty site and decides to build the whole thing himself, including the parts a sensible person would rent. Call him the blank-canvas architect. It is a description of how he operates, not a compliment, and by the end of this profile it will have cost a great deal of money and, as the record now shows, rather more than money.

He was born in South Korea and moved to the United States as a child. He went to Deerfield Academy, then Harvard, where he read political science and founded a student magazine that he sold to Newsweek in 2001. A second magazine, for Harvard alumni, raised roughly four million dollars and folded in the wreckage of 2009, though he had sold it before it did, which protected the people who had backed him. He enrolled at Harvard Business School and left six months in. His own explanation, given to CNBC in December 2019, is worth quoting because it is the whole man in one sentence: he had a belief in graduate school that he had “a very short window to really make something that had an impact”.

He founded Coupang in Seoul in 2010, with Sun-joo Yoon and Jaewoo Koh and about three billion won. It was a daily-deals site, one of more than thirty competing clones of the same American idea, and it had no way of being different from any of them.

The problem, in his words rather than mine

What he decided to fix was not shopping. It was delivery. Korean e-commerce in the early 2010s ran on third-party couriers, which meant that the thing a customer actually judged you on, whether the parcel turned up, was the one thing no retailer controlled. Kim’s description of the target is blunter than any analyst would write: he set out to improve what he called South Korea’s “disjunctured postal system”.

Three facts about Korea made the idea rational and not romantic, and he has named all three: extreme urban density, extreme urbanisation, and unusually good telecommunications. A country where most of the population lives stacked within a short drive of a warehouse is a country where owning the last mile can pay. Almost nowhere else in the world was it obviously true in 2014.

The weekend before the printers

By around 2013 the daily-deals business had passed a billion dollars of sales inside three years and had turned profitable for the first time. His investors told him to take it public. He ran what he has described as an arduous six-month listing process, with the underwriting in place and the prospectus ready.

Then he stopped it. His stated reason is what makes this a case study and not a profile: “once you go public, it’s much harder, at least in the near-term, to pivot or to really change your direction.” The test he then applied to his own company was a simple one. Was the platform he had built “creating a 5 per cent difference”, or was it the kind of thing that would make the customers he cared about wish they had found it sooner? His answer, on the record, was no.

He pulled out, in his words, “right at the eleventh hour, literally the weekend before we were meant to go to the printers”. Then he took a profitable company apart: “We had to change our entire technology stack, the way we did business, our business model.” In 2014 the rebuilt company launched Rocket Delivery, ordered by midnight and delivered by seven the following morning, run on its own warehouses, its own vans and its own employed drivers.

The personal stake here needs no embroidery, and there is none available: he walked away from his own liquidity event, at the one moment it was securely available, on the grounds that the company was not yet good enough to be locked into its current shape. He has since called it “the most difficult, but the choice that I’m most proud of”.

What it cost to be right

The bill is in the audited filings and it is worth reading slowly. In 2018 Coupang did 4.05 billion dollars of revenue and lost 1.098 billion, a net margin of minus twenty-seven per cent. Gross profit that year was 189 million dollars on that 4.05 billion of sales, a gross margin of 4.7 per cent. Read plainly: it was selling goods at very close to cost and paying for a national logistics network out of shareholders’ equity.

The losses then narrowed every year, to 699 million in 2019 and 475 million in 2020. By the end of 2020 the accumulated deficit stood at 4.118 billion dollars. That is the price of the decision he took at the printers, and he paid it over seven years in public.

What made it survivable was somebody else’s conviction arriving at the worst moment. Sequoia put in a hundred million dollars in May 2014 at a one-billion valuation. SoftBank Group put in a billion in June 2015 at five billion. And then, in November 2018, inside the calendar year of what was then the largest annual loss it had reported, the SoftBank Vision Fund put in two billion dollars at a nine-billion valuation. By that point Coupang listed more than 120 million items, offered four million of them on guaranteed one-day delivery, and was moving more than a million parcels a day.

The company listed on the New York Stock Exchange on 11 March 2021, priced the night before at 35 dollars a share, raising 4.6 billion dollars at a valuation of roughly 60 billion. It closed its first day at 49.25, up 41 per cent, worth about 84.5 billion dollars. Seven years after he cancelled the first attempt, the second one was the largest American listing of that year to date.

Then the arc completed. Coupang recorded its first full year of operating profit in 2023, at 473 million dollars of operating income. Its reported net income that year was 1.36 billion, but the scope matters and most write-ups drop it: 895 million of that was a one-off, non-cash release of a tax valuation allowance instead of money earned, which is why the company’s own adjusted figure for the year is 465 million. The cleaner numbers are the years since, 154 million of net income in 2024 and 208 million in 2025 on revenue of 34.5 billion dollars. Revenue compounded at about thirty-six per cent a year from 2018 to 2025.

Whether it matters, which is the fair question

Size it before believing it. Korean online shopping transactions ran at 272.3 trillion won in 2025, about 188 billion dollars, on the national statistics office’s figures. Coupang’s reported revenue of 34.5 billion dollars is about eighteen per cent of that, and the two bases are not comparable, because Coupang books its own sales gross and third-party sales net, so its share of actual transactions is meaningfully higher than eighteen per cent. Published estimates of that share run from about twenty-eight to forty per cent depending on who is measuring and what they count, which is why no single figure appears in this paragraph.

The cleanest measure of scale is the customer one. Coupang had 24.7 million active customers in its retail business in the second quarter of this year, in a country of about 51.7 million people. Close to half the population bought something from it in three months. And the competitive record is not an estimate: Tmon and WeMakePrice, two of its rivals, collapsed in 2024, and Gmarket, 11st and Lotte On have all been losing money for years. This is the company that outlasted the field.

What it is not is a rapid-scaling story in the present tense, and this letter is not going to write it as one. Revenue grew four per cent in the second quarter as reported, ten per cent stripping out currency. Guidance for the third quarter is eight to nine per cent. Adjusted profitability has gone backwards, from a 4.9 per cent margin to 2.5 on a trailing-year basis, and free cash flow has fallen 87 per cent. The correct frame is a company that scaled explosively, won its market, and is now being tested on something else entirely.

The bill that arrived twelve years later

On 18 November 2025 Coupang disclosed a data breach affecting about 4,500 accounts. Twelve days later it confirmed the real number: 33.7 million customer accounts, names, email addresses, telephone numbers, delivery addresses and partial order histories. No card details and no passwords. Korea’s privacy regulator, counting non-members in delivery records as well, put the total at 37.5 million, which is more than half the population of the country.

The cause is the part that connects back to the architecture. According to the regulator’s June 2026 findings, the breach was not a sophisticated intrusion. A former employee who had helped build Coupang’s alternative authentication system left with the signing key, and the access ran for roughly five to six months undetected, because keys were viewable by staff who did not need them and were not rotated when people departed.

On 11 June 2026 the regulator fined Coupang 624.7 billion won, split between the breach itself and the unauthorised collection of user activity data. That is about 409 million dollars at the exchange rate when it was levied, and on the regulator’s own account it is the largest privacy penalty it has ever imposed, against a previous record of 134.8 billion won levied on SK Telecom. The shares fell about five per cent on the day. Coupang has booked the charge, is appealing through the courts, and has disclosed that the fines are not tax-deductible in Korea.

Where this sits

The shares ended the week at about fourteen and a half dollars, a market value of about 26 billion dollars. That is roughly 57 per cent below where they stood a year ago and about 59 per cent below the 35 dollars the company listed at in 2021. The second quarter of this year was the worst in its listed history, on its own second-quarter report: a net loss of 570 million dollars, of which 410 million is the fine.

The lesson worth carrying away has two halves, and only the first is the one the company would write. Vertical integration that looks like a cost disease while you are building it becomes a genuine moat once density arrives, and the only route there is a willingness to stay unprofitable for longer than your competitors can survive. Kim was right about that, he was right about it before anybody funded him, and the audited numbers settle the argument.

The second half is what 2026 has added, and it is the reason this profile is not a tribute. The same decision that produced the moat concentrated every operational risk onto one balance sheet with nobody to share it. A company that hires its own drivers owns the deaths on its own night shifts. A company that owns the customer relationship holds 33.7 million customer records, and therefore owns the consequence of one departing employee and an unrotated key. There was no third-party courier to absorb the first and no platform partner to absorb the second, because the whole strategy was that there would not be. That is the bill for building it all yourself, it arrived twelve years after the decision, and it is the part of the case that no investor in 2014 was modelling.

Nought of the two conditions that would say compute scarcity is genuinely breaking are met. The one that can currently be measured was checked on 16 September and it moved further away from being met, not closer.

This gauge answers one question: is the scarcity that the whole artificial-intelligence trade is priced on actually dissolving? It reports two break conditions instead of a score, and it does not average them with anything.

Break condition one, memory contract prices printing down month on month. Not met, verified 16 September 2026. Down, not merely decelerating. A fortnight ago there was a case that the direction had turned, on the back of a spot-price dip around 10 and 11 September. The primary read on 16 September settled it the other way: that dip was a two-day wobble, spot prices for the commodity grade printed a new high at 45.79 dollars, roughly 7.7 per cent above their level on 10 August, contract prices for server memory are up thirteen to eighteen per cent quarter on quarter for the current quarter, and high-bandwidth memory contracts are still rising. Counter-signals exist and are recorded rather than scored: the producer price index for semiconductors has fallen for four consecutive months, and buying interest is described as subdued with discounting on some branded parts, which together suggest prices are being held up by supplier quotes and not by demand. That is worth watching. It is not the condition.

Break condition two, counterparty funding stress. Not yet built. It is not met and it is not unmet, because the instrument does not exist yet. It requires a basket of the companies financing this build-out and a matched-duration comparison, and until those are named and shown to be genuinely liquid it would be a number without a meaning. Stated rather than left to look like a clean reading.

The book, which is a separate fact and is not merged with the thesis

Four of the seventeen companies this gauge follows are more than twenty per cent below their highest close of the last ninety days. The worst is down 36.4 per cent. The concentration is worth naming. Three of the four sit in the power complex, where the independent generators and the data-centre equipment makers are down between twenty-three and thirty-three per cent from their peaks. The worst single name is a memory producer, and it has been falling since July for reasons none of the conditions above captures. Meanwhile memory and semiconductor shares rallied hard into Friday, which is one sector going two ways at once.

A strategically correct thesis and a badly performing position are different facts and this letter reports them separately. The point of separating them is that a single verdict would have read the same every week through a stretch in which the worst name in the book fell by more than half.

Reported, not scored

Interruptible rental for the previous-generation data-centre processor printed two dollars an hour on Thursday and Friday, against 1.467 on the previous Friday, 11 September, and 1.487 on Saturday morning across 28 live offers. Guaranteed capacity was 2.856 dollars an hour on Friday. The fact worth recording is a streak breaking: interruptible rates had sat continuously below the 1.65 level for thirty-four days, from 14 August to 16 September, and that run ended on the 17th and 18th before the price fell back below it again.

That 1.65 figure is a widely used estimate of what it costs an operator to run the machine. It is not scored here, and the reason is worth stating: this letter has not built the cost analysis that would defend the number, so using it to draw a conclusion about operator margins would be borrowing somebody else’s arithmetic. Reported because it is interesting. Not scored because it is not ours.

Two other series that this gauge would like to use are not observed: pricing for the newest processor generations is populated in only two weeks out of the last eight, and the frontier token-price tier printed an identical figure five weeks out of six and then moved fivefold on a change in how it was sampled. Neither is used.

The binding constraint

Memory, and it has not moved. The same constraint has been binding for the last several readings, and the closest thing to a challenger is Chinese domestic memory capacity, now at roughly a tenth of global supply and still expanding. Capacity arriving into a market its own producers still describe as tight has not yet arrived in the sense that matters here.

Appendix A11 carries the state of each condition with the date it was last verified. The definitions and why scarcity inverts are on the standing method page.

On one Wednesday in September a central bank raised its policy rate twice in a single day to defend a position, and by the evening the position was gone.

At about eleven in the morning on 16 September 1992, the Bank of England raised its base lending rate from 10 to 12 per cent. Sterling kept falling. In the early afternoon it announced a further rise, to 15 per cent, to take effect the next day. Sterling kept falling. At about twenty to eight that evening the Chancellor, Norman Lamont, stood outside the Treasury and announced that Britain was suspending its membership of the European Exchange Rate Mechanism, the arrangement that had committed the pound to trade within a fixed band against the German mark since October 1990. The 15 per cent rise was never implemented. The Treasury subsequently put the cost of the day’s currency defence at 3.3 billion pounds.

What the market did was decline to be impressed. Two rate rises inside seven hours moved the pound not at all in the direction intended, because the people selling it had worked out that the rate the Bank was quoting was not a rate Britain’s economy could live with, and that a commitment which cannot be lived with will not be kept.

What rhymes, and what does not. The rhyme is the use of the policy rate against a price the central bank does not set. In 1992 it was an exchange rate; this week, in three jurisdictions, it is the price of fuel and the price of long-dated government debt, and British ten-year borrowing costs reached a five-year high in the same week the American Federal Reserve tightened. The difference is the one that should trouble a reader and not reassure them. In 1992 there was a peg, so there was a single evening on which the policy visibly failed and everybody could see it. In 2026 there is no peg and therefore no breaking point: a central bank raising rates against an oil price cannot be publicly defeated on a Wednesday, which means there is no obvious moment at which it is told to stop.

What happened next is the reason this is not a cautionary tale. Freed from the band, Britain cut its base rate to 6 per cent within five months, inflation did not return, and the economy entered an expansion that ran for fifteen years. Some people took to calling it White Wednesday. The lesson is narrower than the mythology: the market delivers its verdict on a rate decision within hours, and the economy delivers its verdict over years, and the two are frequently opposite. Anyone reading this week’s bond market as a judgement on the Federal Reserve’s decision should hold both of those in mind.

A measure that excludes the cost it is being used to justify is not a measure. It is a permission slip.

“We can’t tie discretionary cash flow directly to the dividend anymore.” That is the chief financial officer of Kodiak Gas Services, on the company’s second-quarter earnings call of 7 August 2026, in the same session in which the dividend had been described as covered more than three times over. Both statements were made on the same call, about the same company, on the same morning. This column takes one phrase each time it runs and asks what it is doing, and this month the phrase is discretionary cash flow.

What it is, mechanically

Start with cash from operations, which is the money the business actually generated. Subtract maintenance capital expenditure, the spending required to keep the existing assets working. What is left is called discretionary cash flow, and the name is accurate as far as it goes: it is the cash the management has discretion over.

Now the part the phrase does not say out loud. It is struck before growth capital expenditure, the spending on new assets. So a company in the middle of a large expansion programme can report a healthy discretionary cash flow and a large negative free cash flow in the same quarter, because free cash flow subtracts all the capital spending and discretionary cash flow subtracts only some of it. Neither number is wrong. They answer different questions, and only one of them answers “can this company pay its dividend out of what it earned?”

The number

Kodiak’s own filings supply it. In the first half of 2026 the company generated 170.6 million dollars of cash from operations and spent 318.6 million dollars on capital expenditure, so free cash flow was negative 148.0 million dollars, and it paid 92.6 million dollars of dividends. On a trailing twelve-month basis to 30 June 2026 free cash flow was 5.0 million dollars, against a dividend running at roughly 198 million a year. Five million dollars covers about two and a half per cent of a hundred and ninety-eight million.

So the dividend is genuinely covered three times over on the measure the company used, and is covered about one fortieth of once on the measure that counts all the cash going out of the door. The difference between those two statements is the growth capital expenditure, which for this year is guided at four hundred to four hundred and fifty million dollars on power generation alone. Nobody is being misled by a false number. The selection of the number is doing the work.

Where the risk went

This is the part worth carrying away. The phrase does not hide a loss; it relocates a decision. By defining coverage before growth spending, it converts a question about solvency into a question about intention: the dividend is covered provided the growth programme can be financed some other way, and the other way is debt or new shares. Kodiak raised a billion dollars of notes in March and about 836 million dollars of new equity in May. The dividend was, in a real sense, covered by the equity markets rather than by the business, and the metric is constructed so that this does not appear anywhere in the coverage ratio.

And note what happened to the shares: they fell 23 per cent between late June and this Friday, through a raised profit forecast. A flat dividend declared in a record quarter is the tell that management can read the same arithmetic.

The question to reuse

This is not a debunking column and some phrases survive the surgery intact. This one does not, but the fault is in the use instead of in the definition: discretionary cash flow is a legitimate measure of managerial flexibility and an illegitimate measure of dividend safety, and it is almost always deployed as the second.

So the transferable question is this: what does this metric exclude, and who is paid out of the excluded part? Apply it to adjusted earnings, to underlying profit, to net debt excluding leases, to like-for-like sales. In every case the adjective is the instruction. Find the thing it is adjusting away from, ask who bears that cost, and you will usually find the answer is either a lender or a future shareholder. It works on any pitch and it takes about a minute.

Four pillars, four separate observations, and a correction to how this letter read the compute pillar a week ago.

Adoption speed, the Augmenter. Europe has the capital and the engineers and, on Apollo’s figures published 17 September, roughly one fiftieth of the American market for financing finished data centres, which is friction rather than demand.

Labour, the Augmenter. No large employer named artificial intelligence as the cause of a headcount reduction in a filing this week, and because an absence is weak evidence it carries a positive fact beside it: American retail employment sits inside a month in which retail sales rose 1.2 per cent.

The type of shock, the Augmenter. Households have not marked up what they expect prices to do overall: one-year inflation expectations in the New York Federal Reserve’s survey are unchanged at 3.6 per cent. Two limits on that, because they cut against the point and it is the weakest of the four this week. The survey is the August round, published on 8 September, two days before crude first closed above a hundred dollars, so nobody has been asked anything since the barrel went through. And in that same release the expected rise in petrol prices went UP by 1.7 points, to 4.6 per cent, which is the component a fuel shock reaches first. A fuel shock that has not reached the household’s own general forecast is not yet destroying real income, which is what the Displacer needs on this pillar.

Compute constraints, the Displacer, and this is the correction. A week ago this letter treated the cost of computing as the natural ceiling on how far automation could run. The modelled improvement reported earlier in this edition, roughly sixty-seven times the work per dollar of ownership cost against the previous generation, says that ceiling is receding fast if the figure is even directionally right. A point to the Displacer, and a point against the reasoning this section used last week.

Week 36: the Augmenter 3, the Displacer 1. Running total, the Augmenter 46, the Displacer 22.

What would flip it: a named employer attributing a headcount cut specifically to artificial intelligence in a filing takes labour outright, and a memory contract price printing down month on month would take the compute pillar back.

The framework and its scoring rules: standing method.

For most of the period this chart covers, most countries most of the time had inflation in the high single figures or worse. For the last thirty years, most countries most of the time have not.

The line below is the median country’s consumer price inflation, as the World Bank compiles it from International Monetary Fund data: in any year, half the countries it tracks sat above the line and half below it. One hyperinflation cannot drag a median, so the shape is about the typical country changing rather than the largest ones. It starts in 1981, which is as far back as the world series goes, and it starts at 12.4 per cent. What follows is not a smooth improvement. It falls to 5.7 by 1987, climbs back into double figures by 1994, and only then steps down, reaching 3.0 in 1999 and staying since in a band between about one and a half and five per cent. That band has been broken twice, in 2008 and again in 2022 and 2023, and both times by the price of energy and food.

The mechanism behind that step is the institution being complained about this week. New Zealand gave its central bank an explicit, published inflation target in 1990. Within a decade most of the developed world and a great deal of the developing world had copied the design, usually alongside statutory independence for the bank. A committee raising rates into a fuel shock because it is worried about next summer, which is precisely what happened on Wednesday and is precisely what this edition has spent several sections questioning, is that design working as intended. It is not obvious that the design is right in this instance. It is reasonably clear that the design is one important reason the chart changed shape, alongside globalisation, demographics, the commodity cycle and China’s integration into world trade, none of which this chart separates out.

Two limits, because this section is a counterweight and not a comfort. A median tells you about the middle and says nothing about how much of world output sits in either tail, and the tail is still there: Turkey sits on this letter’s own Scoreboard with the dollar 37.74 per cent stronger against its currency this year, which is the lira 27.4 per cent weaker read from the other side of the same quote. The line also begins in 1981, so the 1970s are not on it; they are the decade its opening level is the aftermath of. And the series ends in 2025, so it does not contain the fuel shock this edition is about. The claim here is not that inflation has been solved. It is that the thing which made the 1970s the 1970s, a world in which high and unpredictable inflation was simply the normal condition, was not a natural state that has been temporarily suspended. It was dismantled, deliberately, by people doing an unpopular job.

They raised it a quarter, and why?
For a barrel no rate can supply.
The short end complied,
the long end denied,
and the gauge did not move. Nor did I.

The next evidence arrives in three places. The petroleum inventory report on Wednesday 23 September carries the third distillate reading, and with it the answer to whether this letter has been explaining the oil price correctly for three weeks. Friday’s close settles the 5.15 line on the thirty-year. And around the 26th, core PCE gives the first look at the inflation measure the committee actually targets since it raised rates.

Taken together, the uncomfortable combination is a third build, a thirty-year holding above 5.15 and a core reading at three-tenths of a per cent or more: a refining bottleneck rather than a shortage of crude, a bond market that has accepted higher rates for longer, and fuel genuinely passing into everything else. That is the case in which the central bank is right about the inflation and wrong about the cause, and tightens anyway. The likelier split is a third build alongside a thirty-year that breaks lower, which would say the market has decided the tightening will do the work on demand that the barrel was never going to do on supply. Either way the levels were named in advance, and the result goes in the same place next Saturday.

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

Crude holds second place on the board for a third week, at plus 58.7 per cent for the year, and freight holds first, so the two lines at the top of a twenty-six-line table are both physical goods rather than financial assets. The gap between best and worst narrowed for a second week, from 105.8 points to 96.2, and both ends contributed: freight gave back seven points at the top and natural gas recovered two at the bottom.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 18 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 18 September 2026. MSCI EM is derived from the verified EEM fund close of 67.03 on 18 September multiplied by the ratio 28.367, which is locked for the year and never re-derived, giving the 1,901.44 in the table above. The fund is a proxy for the index and not the index itself, and the conversion happens at publication rather than in the price record, so the two scales are two views of one series