Growth And The Bill For It

Three things happened this week that cannot all be the same story. Long-dated government bonds sold off hard enough to be the worst line on the board. Shares rose. Gold fell 2.34 per cent, below where it started the year for the first time in 2026.

The third is the useful one, because it closes off the easy explanation. When long yields rise and gold rises with them, the reading is usually debasement: investors demanding more to hold paper promises and buying metal instead. Gold went down. What is being demanded at the long end is not insurance against the currency. It is compensation for lending into an economy that has just reported the fastest output growth in five years and the hottest input costs in four.

Underneath the index level sits a second divergence that has been building for a month. The proportion of the index trading above its own two-hundred-day average fell to 46.5 per cent, from a range of 55 to 57 a week ago and 64 at the start of September. An index a whisker off its record with fewer than half its members trending up is not calm. It is concentrated.

Every yield here is a Friday close on the US Treasury constant-maturity series for 25 September 2026, and any intra-week level is named with its day, taken from the rate series this letter keeps and not from any figure printed elsewhere on this page.

The front end is a forecast. The long end is an invoice.

On Wednesday morning S&P Global published its flash reading of American business activity for September. The composite output index came in at 58.4 on the release’s own figures, the strongest in sixty-two months. The same release had input prices rising at the fastest rate since October 2022, backlogs building at the sharpest rate since May 2022 with severe supplier delays, and hiring at its quickest since June 2022. One survey, two findings usually reported as opposites.

The bond market read it the way a lender would. The ten-year went from 4.94 on Tuesday to 5.18 on Thursday, closing at 5.17; the thirty-year reached 5.49, its highest close of the year; the two-year, which prices the central bank, hit 4.87 on Thursday and fell back to 4.81. That is a bear steepening: the whole curve sold off and the long end three times harder than the front, which is what happens when the market revises up not what the Federal Reserve will do but what it will cost to lend into the world it operates in.

That distinction is the substance of this edition. A tightening is decided by twelve people in a room and the two-year is where it gets priced. The term premium (the extra yield a lender demands for accepting thirty years of uncertainty about inflation, supply and politics) is decided by nobody. It moved this week while the policy expectation did not.

The rate framing, and where it comes from. This letter’s internal rate model reads hike-leaning, unchanged, with the rule-based prescription at 4.90 per cent against an actual midpoint of 3.875 and a gap of 1.03 points. The gap is unchanged in level and its direction has gone from narrowing to stable, which is small and real: a gap that stops closing means the rule is no longer being caught up with. The model is also running on three-quarters of its panel, because August core PCE (the Personal Consumption Expenditures index, the inflation measure the Federal Reserve actually targets) and the September factory survey are unpublished, so the verdict is provisional. The 30 September figure is the first thing that can move it.

Note where the model and the market disagree. The model leans towards another increase; the two-year fell back six hundredths from Thursday. If it keeps drifting down while the thirty-year climbs, the market is saying the tightening is nearly done and the cost of the aftermath is not.

One thing did not register, and it changes how to read this week’s number. Treasury volatility rose roughly thirty per cent, its largest weekly move since April 2025, and closed above 100 for the first time in three months of readings. The crash gauge did not move, because the band that scores it begins at 110, so it reads 25.0 for a third week. The instrument registers crossings, and this week’s event was a large move inside a band.

The competing explanation, and the tell that separates it. The reading above puts the long-end move down to term premium repricing on a growth-and-cost survey. The strongest alternative is supply: the Treasury funds a large deficit at the far end, and late September brings quarter-end balance-sheet management that pushes dealers out of inventory. That produces the same shape. The separator is the twenty-year. If this were about the economics of lending for decades, the thirty-year should yield more than the twenty, since it carries ten more years of the same risk. It does not: 5.54 against 5.49, an inversion that widened from four basis points to five. That is a fingerprint of who buys which maturity, so the flows explanation owns part of the move. If the gap closes while long yields stay high, the economics are doing the work. If it stays open, the mechanics are.

Where I could be wrong

The survey could be one print. A flash purchasing-managers reading is a diffusion index from a sample, it gets revised, and 58.4 after a run in the mid-fifties sometimes does not survive its own final estimate. If the final comes back near 55 and the October flash follows, the growth-plus-costs sequence was an artefact and a bond market that moved on it moved on noise. The tell is the final September reading and the 1 October factory survey.

The inversion argument could prove too much. Five basis points between two very long maturities is close to the noise in the daily par curve, and one pension rebalancing can produce it. Concluding from that that the whole week was mechanical goes further than the evidence: a sixteen-basis-point move in the ten-year is not a rebalancing.

And the gold argument could be a currency story wearing other clothes. Gold fell in a week the dollar firmed against both currency pairs on this letter’s own board, so part of what reads as an absence of debasement demand may just be the currency the metal is quoted in. If the dollar turns and gold rises while long yields hold, that reading comes straight back and will deserve more than a sentence.

On Monday the market believed a rate rise had settled the inflation question. By Wednesday a survey said the opposite and the bond market agreed with the survey.

Capacity. The survey came out on Wednesday morning. American business activity grew at its fastest rate in sixty-two months on S&P Global’s own figures. Order backlogs built at their sharpest rate since May 2022. Suppliers were late. Firms hired at the quickest pace since June 2022. And input costs rose faster than at any time since October 2022. Growth and a cost shock, in one release. That is not a contradiction. It is what a supply-constrained expansion looks like, and it is the least comfortable combination for a central bank that has just raised rates.

Who pays. The bond market answered by Thursday. The ten-year went from 4.94 to 5.18 in two sessions; the thirty-year closed at 5.49, the twenty-year at 5.54. Long-dated government bonds are now the worst line on the board, down 15.86 per cent for the year. Investment-grade corporates are down 5.33 per cent and the aggregate bond market 6.88.

Forced flows. Gold fell 2.34 per cent to 4,321.20 dollars an ounce, taking it to minus 0.46 per cent for the year after being positive every week since January. Silver fell 3.47 per cent. The dollar firmed against both pairs on our board, 1.05 per cent on the rand and 0.33 on the lira. This closes off the lazy reading: rising long yields plus falling gold is not a debasement trade but a real-yield trade, and higher real yields are what makes a metal that pays nothing less attractive.

Cost structure. Oil went the other way. WTI fell 7.87 per cent to 92.41 dollars. A senior Iranian official told a wire service on Thursday that the workable bargain is navigation through the Strait of Hormuz for an end to the American naval blockade, and both sides are said to be working through the sequencing. The market bought it, as it bought a version in June that expired in August.

Capacity, again, and this one is domestic. Natural gas rose 9.75 per cent, the largest weekly move on the commodity board, and it was not the weather. Thursday’s storage report showed an injection of 53 billion cubic feet against a five-year average build of 76 for the week, with forecasters looking for about 50. Ordinary against expectations, light against the season, and that distinction is the story. Production fell to an eleven-week low near 108.4 billion a day, and gas touched a thirteen-week high on Thursday before giving back 3 per cent on Friday.

Where value is captured. Shares rose. The American index gained 1.21 per cent to 7,743.41 and the Nasdaq 100 gained 3.25 per cent. Neither is a record: the standing high is 7,798.99, set on 13 August, which the index approached and did not pass. Bitcoin rose 3.98 per cent and climbed six places up the board, the largest single move in the ranking. Emerging markets gained. Japan gained. Almost everything with equity in the name gained, in the week long money got dearer.

Enforcement. A state visit to Washington, 23 to 25 September, the first in eleven years, extended the trade truce by two months to 10 January. No purchase commitments. No artificial-intelligence framework. Nothing on Iran. An extension is not a settlement, and the new expiry sits in the same week as much of next year’s positioning.

Commoditisation. A frontier language model was released on Tuesday at four dollars per million input tokens and twenty per million output, about a fifth below the version it replaced, with cached reads down some sixty per cent. Published prices, not projections. The cost of machine cognition fell again in a week when the cost of thirty-year money rose.

Three named decisions carried the week and not one was a rate decision. A survey published a number. An Iranian official floated a bargain. A storage report missed a consensus. Everything above is downstream of those.

Three dials deteriorated this week and five held, and the reading that moved furthest of all still scores green.

Three dials deteriorated, five held, none improved, and the composite came out at 25.0 for a third week: a fair description of the machinery of finance and a poor one of the week.

The pair of volatility rows is the thing to watch. Bond volatility rose about thirty per cent and equity volatility did not move, and those two indices measure the same emotion in two markets that have travelled together for most of two years. When they separate it is usually because one market is repriced by people who have to hold it and the other by people who choose to. Pension funds, insurers and banks own government bonds under rules that tell them how much they must hold; shareholders can leave. A forced holder reprices by demanding a higher yield. A voluntary holder reprices by selling, and nobody sold.

Where value is captured, and a note on frontier claims. Mazama Energy, a private American company, announced 135 million dollars this month for superhot-rock geothermal drilling, and its release of 17 September projects up to ten times the power of a conventional well. Its own paper, given at the fifty-first Stanford Geothermal Workshop in February 2026, reports a flow test at its Newberry well delivering an electrical equivalent of about 0.7 megawatts. The marketed figure per well is fifteen. Both numbers come from the same company and neither is false: a result, a design target and a resource estimate are three different objects. When a company hands you a multiple, ask what it is a multiple of, and where it has actually been observed. This one moved from six-to-eight times to ten in eleven months without a new well test.

The price of machine thinking fell about a fifth this week and the price of thirty-year money rose. Those two facts will spend the next decade arguing with each other.

Commoditisation. On Tuesday a frontier language model was published at four dollars per million tokens of input, a token being roughly three-quarters of a word, and twenty dollars per million tokens of output. That is roughly a fifth below the version it replaced, with cached reads cut about sixty per cent to twenty cents. Those are list prices on a public page, so they are facts and not forecasts. The claim attached to them, that a typical workload gets about forty per cent cheaper, is the company’s own and depends on the workload. Take the prices; leave the percentage.

Capacity. The same week, the physical side of the industry hit three walls. The governor of Texas ordered a halt to data-centre permits, weeks after a moratorium, in the state absorbing more of the build-out than any other. California signed a package of bills on data-centre energy and water, and a federal energy bill for the sector stalled in the Senate. Local objections blocked about 68 billion dollars of American projects in the second quarter alone, according to Data Center Watch via Bloomberg on 21 September, and unverified here; if it is even roughly right, community resistance has stopped being a sentiment survey and become a capital-expenditure line.

Who pays. Six American states now have statutes governing the use of artificial intelligence in hiring, and one published a framework this week built around regulation and retraining. Recruitment is the most automatable corporate function and the one acquiring the most law, which is a pattern worth keeping: the easiest tasks to automate are often the ones where being wrong is most legally expensive, and that asymmetry appears in no capability benchmark.

Put the three together and the shape of the next few years appears. The marginal cost of a unit of cognition is collapsing and the marginal cost of the electricity, the buildings and the legal cover to deploy it is not. A technology whose software cheapens every quarter while its infrastructure dearens does not diffuse evenly. It diffuses fastest where it needs least new steel and slowest where it needs a substation.

An energy shock that arrives as a queue is a price problem. One that arrives as an election result is a fiscal problem, and it lasts a great deal longer.

At half past ten on Sunday morning, 20 September, French diesel touched 2.4097 euros a litre on the government’s own portal, which covers more than eight thousand seven hundred stations. A record, about 10.40 dollars a US gallon, with nine per cent of French stations short of a grade. The cause is the week this letter did not carry: about fifty fishermen blockaded the oil terminal at Fos-sur-Mer, the depot at Frontignan and the port of Nice between the 15th and the 18th, wanting a meeting with the fisheries minister. They got it, and on 22 September a 450 million euro package raising their fuel aid from 25 to 35 cents a litre. Last week’s edition made no mention of any of it.

The same Sunday, nine hundred miles away, the Christian Democrats took 4.9 per cent of the vote in Mecklenburg-Western Pomerania. It is the worst result the party has recorded in any German state election, and it puts it under the five per cent threshold and out of the state parliament. Alternative for Germany won the state with 38.2 per cent, 2.7 points ahead of the Social Democrats and short of a majority. It took 43.8 in Saxony-Anhalt a fortnight earlier. The narrowness is the point: a governing party wiped out and an insurgent one winning without a majority is a parliament that must buy its way to agreement. The Chancellor called the night a disaster.

The mechanism is why those two paragraphs belong together. Europe imports well over half the energy it uses and has the least capacity of any large developed bloc to substitute quickly. When that price rises the cost does not stop at the forecourt. It arrives as a demand on the state, the only actor with a balance sheet big enough to absorb it, and it arrives twice: through subsidies and price caps, and through parties promising to make the problem stop. France has spent four years learning the first. Germany spent Sunday learning the second.

The tradeable consequence is not the oil price, which fell nearly eight per cent this week. It is European sovereign credit and the regulated utilities. A government facing blockaded ports and an insurgent party at 38 per cent does not hold the line on fuel taxation. It buys its way out, and every instrument it uses is a claim on a budget already carrying defence commitments it did not have in 2021.

The Gulf, and it does not lead this section for the first time in five editions. A senior Iranian official told a wire service on Thursday that the realistic bargain is navigation through the Strait of Hormuz for an end to the American naval blockade, and both sides are said to be weighing how a reopening would be staged. Crude fell 7.87 per cent. The June memorandum expired in mid-August, which is reason to hold this one loosely. The rule here is that a flashpoint may not lead more than three consecutive editions without a new named event; the Gulf has had four, so a fifth needed something genuinely new. This qualifies and still does not lead, because the more consequential energy story happened in a French port and a German state.

And the meeting that was mostly a photograph. A Chinese state visit ran from Wednesday to Friday, 23 to 25 September, the first in eleven years. It bought two more months of the existing trade truce, to 10 January, and a panda. No purchase commitments, no framework on artificial-intelligence risk, nothing on Iran. The market-relevant part is the date, not the spectacle: a truce expiring on 10 January puts a dated policy risk inside the fortnight when allocators set next year’s positioning, and two months is a shorter fuse than the one it replaced.

She spent thirty-six years at Shell, latterly running its unconventional wells with about a billion dollars a year of spend. She now sells electricity from the technology Shell looked at and did not fund.

Cindy Taff had the key and her employer did not want the door opened. For most of four decades at Shell she drilled and completed wells, ending as vice-president of unconventional wells and logistics, with more than three hundred and fifty staff, twelve hundred contractors across five countries and roughly a billion dollars of annual spend. Geothermal came up inside the company from time to time. Her own account, and it is hers alone because no Shell statement exists, is that the projects never advanced because the economics did not hold up, and no wells were drilled. Then her daughter started asking why she was still in oil and gas and whether there was a future in it. Taff described the effect in her own words to the Associated Press, in an interview published on 1 October 2024: as a parent you want to give direction, and was she giving the right one. She left. Her daughter interned at the company she left for, was asked by its executives to help recruit her own mother, and is now one of its engineers.

Call her a safe-cracker. Not the inventor of a category, but the holder of one rare skill that is the exact skill a guarded industry could not hire. Enhanced geothermal is not a subsurface science problem, because the heat is there at depth everywhere. It is a drilling-cost problem, which is what the American shale industry spent fifteen years and a great deal of other people’s money learning to solve. Sage is that playbook pointed at rock instead of shale, run by somebody who wrote parts of it.

The hard part was never the technology. It was whether an incumbent would let her prove it, and that turned on 22 September. Sage announced it had selected Blue Mountain, at Winnemucca in Nevada, for Project Vector: a two-well system drilling into hot rock and feeding the heat into a geothermal plant that already exists and is already connected to the grid. First well this quarter, first electricity in 2027, full production in 2028. Blue Mountain belongs to Ormat Technologies, which co-led Sage’s January round with twenty-five million dollars of a ninety-seven-million total alongside Carbon Direct Capital. The plant that will prove the technology to outsiders is owned by one of the insiders who funded it.

That bargain cuts both ways. Drilling two wells into an interconnected plant is the cheapest route to a commercial reference: no interconnection queue, no new turbine, no greenfield permitting, just the subsurface work the company is good at. It is also, as the strongest argument against it puts it, a sale to an insider. Ormat is co-lead investor, site owner and customer at once. A later investor will discount that revenue, and should.

Around a hundred and fourteen million dollars is disclosed across its rounds, and every financial metric that matters in a normal analysis is not disclosed at all: no revenue, no gross margin, no burn, no runway, no headcount. It sold its first electricity in the second quarter of this year, from a three-megawatt commercial plant at Christine, Texas. Three megawatts. The United States has roughly four gigawatts of geothermal capacity, so this company holds about 0.07 per cent of one of the smallest generating categories in the country. The leader in its own niche, Fervo Energy, has raised close to two billion dollars, has five hundred megawatts under construction and listed in May at about ten billion. Sage has one order of magnitude less capital and two orders of magnitude less plant. A 150-megawatt agreement with Meta for data-centre power, signed in 2024 with a first phase due in 2027, is contracted demand far ahead of anything it can currently generate.

This letter put Sage through its full private-company analysis before writing any of the above, and it did not clear the bar that would make this a recommendation, chiefly because the financials are undisclosed and the Ormat relationship is unresolved. The story is excellent; the investment case is not yet made, and those are separate questions.

The second product came out of a mistake. During pressure testing in south Texas the team realised the engineered fracture that stores heat also stores pressure, and so energy: pump water down when power is cheap, let it push back through a turbine when power is dear. A geothermal company found a battery nobody set out to build.

None of this is really about geothermal. Taff’s account is that Shell declined the projects because the payback took too long, which is not a failure of imagination but a large company applying a hurdle rate correctly to a long-dated asset. The consequence repeats across every capital-intensive industry: incumbents decline technologies whose payback exceeds their hurdle, and the people who know most about building those technologies are on the incumbents’ payroll, so the expertise and the capital sit in one building and the willingness does not. The investable version is to watch where senior operating people go when they leave, because a thirty-six-year drilling executive resigning to drill something else says more about a category than any amount of venture enthusiasm. It is also free, because it is published on a professional network.

One borrower paid five hundred and sixty-six million dollars in cash this week for the privilege of not issuing shares.

Thesis: nought of two break conditions met, and only one of the two is yet measurable. This instrument watches two things only, because they are the two that would say the scarcity underneath the compute trade is genuinely breaking and not merely being talked about.

Condition one, memory prices falling, is not met, and the reading is carried. When it was last verified, on 16 September, commodity-grade spot had printed a new high, server contract prices were up 13 to 18 per cent quarter on quarter and high-bandwidth contracts were still rising. That is nine days old, labelled carried and not as this week’s.

Condition two, counterparty funding stress, is not yet built. Neither met nor unmet. The basket of financiers has not been named and shown to be liquid and genuinely dependent on these contracts, and a number without that work behind it would be worse than none.

And this week’s reading was not produced at all, for a second week running. The instrument that generates the weekly observation did not run for the Friday, and nothing above has been rebuilt by hand, because a hand-built reading destroys the only thing a refusal is good for.

Which brings us to the week’s actual event, and it needs handling carefully. On 22 September CoreWeave borrowed 4.2 billion dollars, having raised the size twice from an initial three billion. It borrowed as a convertible bond, which is a loan the lender may later swap for shares, here at a price 22.5 per cent above the share price on the day. Alongside it the company spent about 566 million dollars on an option contract whose only job is to cancel that future swap up to a ceiling, so existing shareholders are not diluted if the shares rise as far as it. A capped call is routine and cheap debt is rational, so the structure says what the company preferred and not what it could not have had. What it preferred, at a cost of more than half a billion dollars, was to borrow at 2.875 per cent rather than sell shares.

Last week this section published, as one of its three watch conditions, that either of the two large financiers of the build-out raising primary capital with structure attached where plain equity would be expected is the cleanest early warning available, and that it is public and free. This is not that condition firing. CoreWeave is an operator and not one of the two named, so the condition stands unfired and is not scored. What happened is the borrower one link down the same chain doing exactly the thing the condition was written to detect. That is adjacent evidence, reported as adjacent evidence, and the distinction is why the condition names obligors instead of a mood.

Book: four of the seventeen names sit more than twenty per cent below their ninety-day peak, and the worst is down 29.7 per cent. Reported beside the thesis and never merged into it, because a correct thesis and a bad position are two different facts. The worst is a Korean memory producer, improved from 36.4 last week; two power names sit at 24 per cent and a Chinese foundry at 25. Two of the seventeen carry a close one session old, stated rather than smoothed.

Reported, not scored. Open-market hourly rent for the previous generation of accelerator closed Friday at 1.067 dollars interruptible and 2.301 guaranteed, across twenty-three live offers. That is the lowest interruptible price this register has recorded, and it went back under 1.30 dollars on Wednesday for the first time since 4 September. It sits well below the 1.65 dollars this letter treats as an operator’s break-even, and that figure is not scored, because no cost analysis has been published here to defend it. Watch the offer count beside the price: twenty-three quotes this week against thirty-three last and sixty-four in July, so a thirty per cent weekly fall in quotes is a market emptying as much as a price falling. Pricing for the newest generations appears in two of the last eight weeks and is not observed rather than estimated.

The binding constraint is memory, and it has not moved for seven weeks. Not power, not networking, not lithography. What decides how fast capacity arrives is whether there is enough high-bandwidth memory to sit beside the processors, and every fresh reading this month says there is not. One caution, because a constraint that never moves is either genuinely unmoved or unexamined: the largest American memory producer reports on 30 September, and that is the best available test of which.

Three watch conditions. A monthly memory bulletin showing contract prices flat or down rather than revised up. Either of the two named financiers raising primary capital with structure attached, which has still not happened. And interruptible rent holding below 1.30 dollars for four consecutive weeks instead of dipping and recovering, which is what it did a fortnight ago.

What would re-tighten the scarcity: a memory producer announcing a delay or cancellation of planned capacity, which pushes the 2027 relief out in one announcement. Nothing of the kind happened; two Chinese producers are expanding.

Positions changed by this gauge since inception: nought. Printed weekly until it is not.

The break-condition definitions, and why scarcity in a supply chain inverts before it breaks: standing method.

Fewer than half the index now trades above its own two-hundred-day average, and it finished within one per cent of its record anyway.

The number was 64 at the start of September, a range of roughly 55 to 57 last Friday, and 46.52 on this one. Over the same four weeks the index itself rose. Fewer than half the five hundred largest American companies are now trading above their own average price of the last two hundred days, while the aggregate of those five hundred sits just under its August high.

File that as a paradox and you have filed it wrongly. It is arithmetic with a distribution behind it. An index weighted by market value is not a survey of five hundred companies; it is a survey of about ten with a long tail attached. When the ten rise and the four hundred and ninety drift, the index rises and the breadth measure falls, and both instruments are telling the truth about different populations.

The distributional question is who benefits and who does not, and this week it has an unusually clean answer, because the mechanism is the same one driving the bond market. Long-dated yields rose fifteen basis points at the thirty-year. Every company that finances itself with long-dated debt got fractionally poorer this week. Every company sitting on net cash got fractionally richer, because the cash now earns more and the competitor with the debt now pays more. The largest American companies are, almost by definition, the ones with net cash. The four hundred and ninety are, on average, the ones with the debt.

So this week’s divergence is neither sentiment nor momentum. It is a balance-sheet sort. A rise in long real yields transfers from borrowers to savers, and at the corporate level the savers sit at the top of the index and the borrowers through the middle. The index rose because the transfer favoured its heaviest members.

Two things follow. A breadth reading of 46 per cent in a rising market describes a rate regime rather than warning about one, and it will mean-revert when long yields stop rising and not before. The less comfortable one: a market whose gains concentrate in the companies least sensitive to the discount rate has become, without anybody deciding it, a bet on that discount rate. The companies in the middle are already carrying the cost of higher long rates. The large ones have not yet been asked to.

And a measurement caveat that cuts against the drama. Last week’s figure was an estimated range of roughly 55 to 57 per cent and this week’s is a precise close from one source, so part of the ten-point fall is the measurement sharpening and not the market narrowing. The four-week direction is not in doubt, because 64 was a precise print too. The size of the single-week step is.

A Lincolnshire carpenter spent forty years making a clock, because the Board of Longitude would not accept that a clock was the answer.

Longitude, Dava Sobel (1995). Sailors could measure latitude and could not measure longitude, so ships were lost, routinely, with everybody aboard. Parliament offered a fortune for a solution. The astronomers on the adjudicating board were certain the answer lay in the stars. John Harrison, a self-taught joiner with no Latin and no patron, was certain it lay in a clock accurate enough to keep London time at sea. He was right. He spent four decades proving it against a board of examiners who included the men whose astronomical method his clock was displacing, and who marked his work.

A hundred and eighty pages, it reads like a thriller, and it is the best short book there is on what happens when the people judging an idea hold a stake in the alternative. Chosen for that.

“We’re seeing a pivotal shift in the US economy: investment is shifting away from residential investment and towards computers.” Adam Shapiro, vice-president, Federal Reserve Bank of San Francisco.

He was writing about a figure in the national accounts published this month. American corporate spending on data centres and information-processing equipment ran at an annual rate of about 752 billion dollars in the second quarter, against 748 billion for residential investment, on the national accounts. For the first time the computers figure is above the houses figure. The gap is half of one per cent, inside the range these series are routinely revised by, so it is a crossover to mark and not a measurement to lean on.

The type of shock, the Displacer, and the strongest single item of the week. That crossover is capital formation moving out of the household-facing economy into the capital-owning one. Houses employ tradesmen, furnish themselves and generate local property taxes. Data centres employ a few dozen people once running and send their returns to shareholders. The same dollar has a materially lower wage share on the second use, and this is the quarter the second use became the larger.

Adoption speed, the Augmenter. On 21 September the governor of Texas ordered a halt to new data-centre permits, weeks after a moratorium, in the state that had been absorbing more of the build-out than any other. Adoption is not a software curve when the software needs a substation, and the constraint here is political rather than technical.

Labour, the Augmenter, awarded partly on an absence. No named employer attributed a headcount reduction specifically to artificial intelligence in a filing this week. Because an absence is weak evidence, a positive fact runs beside it: a study of chief human-resources officers published this week found four in five saying these tools create invisible extra work, and nearly half of employees surveyed saying they have added to their workload. Tools that generate new tasks are complements.

Compute constraints, the Augmenter. Prices for twelve-inch silicon wafers under long-term agreement are reported rising 15 to 25 per cent for 2027, and more than 40 per cent on some grades. That is input-cost inflation one layer below the chip, in the substrate the chip is made on, and it is the clearest evidence this week that the ceiling on how far this can run is physical and is going up rather than down.

Week 37: the Augmenter 3, the Displacer 1. Running total, the Augmenter 49, the Displacer 23.

What would flip it: a named employer attributing a headcount cut specifically to artificial intelligence in a filing takes the labour pillar outright. A memory contract price printing down month on month takes the compute pillar. And permitting halts reversed rather than extended takes adoption speed straight back.

The framework and its scoring rules: standing method.

“We actually think it’s a game changer.” A rural electricity co-operative chief executive in Oregon, on a power station that does not exist yet.

He is not wrong to be excited and he has not signed anything, and that gap is where most of the money in frontier energy sits. A good week for enthusiasm. A panda changed hands in Washington. And a survey of purchasing managers shifted the price of thirty-year money more than a unanimous central bank decision had managed the week before.

Spare a thought, finally, for the French fishermen. They wanted a meeting with a minister. They blockaded three ports, put nine per cent of their country’s filling stations on the wrong grade of fuel, and moved French fuel policy further in four days than two years of communiqués had. The government found four hundred and fifty million euros by the Tuesday.

A survey of factories said
“We’re busy, and costs are ahead.”
  The thirty-year read it,
  Repriced the cost of credit,
And nobody voted. Just said.

The next evidence arrives in three places and all three land inside eight days. August core inflation on 30 September, the measure the Federal Reserve actually targets and the first published since it raised rates, which is also the input this letter’s rate model is currently missing. September payrolls around 2 October, where a figure below a hundred thousand alongside a two-year back under 4.45 per cent is the stated pair that flips the rate view dovish. And the gap between the twenty-year and the thirty-year, which is this edition’s own separator between an economic explanation of the week and a mechanical one.

Suppose inflation comes in at 0.3 per cent or above, payrolls hold up, and the twenty-to-thirty gap closes. Then the term-premium reading above is right and the repricing is on the economics. Long money stays dear, borrowers keep paying, and the breadth problem gets worse before it gets better. If inflation is soft, payrolls are weak and the gap stays open, the week was a survey and a quarter-end squeeze, the long end gives some of it back, and the front end was the honest one all along. The uncomfortable combination is a hot inflation print with weak payrolls, because that is the one in which the bond market is right about the cost and wrong about the growth, and nothing in this edition has a good answer for it.

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

The two ends of this table pulled apart again. Freight holds first place and long-dated American government bonds took last off natural gas, and the distance between the two ends widened from 96.2 points to 97.9 after two weeks of narrowing. The line that moved furthest up the ranking is Bitcoin, six places, on a week in which shares rose and gold fell.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 25 September 2026. The basket was fixed on 1 January and cannot change: twelve equity indices, four bond and credit funds, six commodities, two currencies and two cryptocurrencies. Five rows are abbreviations: MSCI ACWI (All Country World Index) is the broadest global share index, AGG the US aggregate bond market, LQD investment-grade corporate bonds, HYG high-yield, and TLT long-dated US government bonds. The single-company calls in Portfolio Watch are tracked separately. The masthead figure scores this year’s thesis on five pre-registered lines, not all twenty-six: shares, bonds, gold, crude and high-yield credit, two up and three down across a 53.1-point range. A wide range with split signs is the thesis working. The five priced commodity rows are front-month futures, not spot, so gold reads tens of dollars above a spot quote; the Baltic Dry is an index. Annualised volatility is how much each line typically swings in a year, judged from its last eight weeks: higher means bumpier, not worse. The lira reads under one per cent because it crawls by the same small amount each week, which is a managed rate and not a quiet one.

Every close is read from the same price record the table draws on, so the figures here cannot drift from the figures we hold. All closes are Friday 25 September 2026. Emerging markets is the one derived line: the verified fund close of 67.98 multiplied by the ratio 28.367, locked at the start of the year and never recalculated, gives 1,928.39. The fund tracks the index and is not the index.