Treasury announced on Wednesday morning that from 9 September it will lift the ceiling on its long-dated buybacks from $2 billion per operation to at least $4 billion, and the long end rallied about nine basis points while the dollar fell almost one percent. That was the loudest number of the day. The more consequential pattern was quieter, and it was about electricity. Mining lost money at the gross line inside Core Scientific while renting the same buildings to AI tenants threw off more gross profit than the whole company earned. Analog Devices guided nine percent above consensus at a seventy-four percent margin on the power-conversion layer nobody benchmarks. Pennsylvania can now stop a data centre because roughly a billion dollars goes into a hundred-megawatt site before a permit matters. And a silicon carbide industry built for electric cars is being kept alive by eight-hundred-volt server racks. The scarce input in artificial intelligence has moved from the chip to the delivered watt and the permission to deliver it, which is why the assets now deciding the buildout are mostly assets nobody built for it. Watch Analog Devices' fiscal fourth quarter in late November, the first print that tests whether the layer between the racks holds its pricing while the layer above it reprices.
Korea's KOSPI closed up 5.89 percent at 6,852.58, its sharpest session of the year, and the Korea Exchange suspended program trading for five minutes around 10 a.m. local time when the KOSPI 200 futures index cleared a 5 percent gain and triggered a buy-side sidecar. SK hynix rose 12.73 percent after announcing a 40 trillion won buyback, about $28.7 billion, and Samsung Electronics rose 9.49 percent into an expected shareholder-return announcement later this month. A halt mechanism firing on the way up is still a halt mechanism, and it is the same index that broke circuit breakers downward on 29 July.
Asia: Nikkei 225 up 1.21 percent to 66,118, with Asian equity generally rebounding as long-end yields backed away from multi-year highs. Europe opened against Wednesday's flat Stoxx 600 at 652 and a German ten-year Bund at 3.22 percent, its highest since May 2011.
US futures are modestly higher into the open: S&P 500 futures up 0.16 percent, Nasdaq 100 futures up 0.42 percent, Dow futures up 0.06 percent. Bitcoin trades near $68,200 after a high of $69,749, gold near $4,480 an ounce and the ten-year Treasury yield near 4.64 percent. The buyback rally in Markets and Macro is what Asia traded overnight.
Crypto data provided by CoinGecko
Treasury will buy at least twice as much long-dated debt per operation from 9 September, and it will pay for the purchases at the front end, which raises its own interest bill while shortening the maturity of what it owes. Treasury's release takes the maximum size of liquidity-support buybacks in the ten-to-twenty and twenty-to-thirty-year nominal sectors from $2 billion per operation to "at least $4 billion," effective 9 September through 4 November. "At least" is a floor, and the tape has priced the minimum. The arithmetic underneath is less flattering than the reaction: retiring low-coupon bonds issued when money was free, funded by leaning harder on bills, swaps a cheap thirty-year liability for an expensive three-month one. The yield falls and the interest bill rises. Joseph Wang and Damped Spring both expect the next step to be cutting long-end auction sizes. We expect Treasury to print above its own floor on 9 September, because "at least" is what an issuer writes when it wants room to do more without having announced it.
The Federal Reserve's July minutes moved nothing, because the meeting they record happened before every soft number that has arrived since. The committee held at 3.50 to 3.75 percent on a nine-to-three vote, with Beth Hammack, Neel Kashkari and Lorie Logan each preferring a quarter-point hike. Three dissents in the same direction is the first such split since September 2016, when George, Mester and Rosengren dissented for a hike and the Fed delivered one that December. Since that meeting, payrolls turned negative, retail sales fell and Michigan's preliminary sentiment dropped sharply, none of which the minutes could have known. Futures-implied odds of a September hike sat in a 31-to-34 percent band before the release and inside the same band after; the day's move happened at half past eight, hours before this document existed. We expect September to be a genuinely contested meeting rather than the settled hold the pricing implies, because the three votes that wanted to move are all still on the committee.
The typical American household expects a bit over four percent inflation next year, the average household expects roughly twice that, and the number that decides spending is the average. Michigan's own preliminary August release, published 16 August, puts the median one-year expectation at 4.3 percent, up from 4.2 in July, with the long-run expectation steady at 3.3 for a third month. Noelle Acheson, charting the same survey's distribution in "A Tale of Two Markets," reads a mean running near twice the median. Michigan does not headline that mean, so treat it as her read of the microdata rather than a published figure. A median that low is the evidence usually cited for anchored expectations; a mean that far above it says a large minority is bracing for something the median cannot see. The personal savings rate ended June at 2.7 percent, against 4.6 percent a year earlier. We expect consumption to deteriorate faster than the aggregate implies this autumn, because a household with no buffer and a high inflation expectation does not behave like the median household, and the median is what the models are fit to.
Target's quarterly profit doubled, and forty percent of the earnings per share was a refund of tariffs the Supreme Court had already ruled the government could not charge. The company booked $994 million pretax on duties paid under the International Emergency Economic Powers Act, worth $1.65 of the $4.11 it reported against $2.05 a year ago. Strip it and the figure is $2.46. Gross margin printed 33.7 percent and Target says 3.7 points of that is the refund, so the merchandising rate is 30.0. TJX, which runs T.J. Maxx and Marshalls, booked $331 million the same morning against $112 million of incentive compensation accrued in the quarter. The honest counterweight is that Target raised its underlying outlook too, by 75 cents ex-refund, on 3.6 percent traffic growth, the hardest retail number to manufacture. But when the 2017 tax act threw off one-time windfalls they landed in the tax line, where every model strips them automatically. A credit to cost of goods sold gets read as merchandising skill. Where a windfall is booked decides whether anyone sees it.
Bitcoin mining lost money at the gross line inside Core Scientific last quarter, before a dollar of overhead, in a company whose revenue doubled. Self-mining brought in $21.5 million against $33.7 million of cost of revenue, a $12.2 million segment loss. Secondary coverage puts mining revenue at $27.5 million, most likely total mining including hosted; the segment note says $21.5 million and both belong on the page. Colocation, renting the same buildings to AI tenants, did $136.7 million of the company's $164.2 million and threw off $80 million of gross profit, more than the consolidated total. TeraWulf, a miner turned data-centre landlord, ran the identical split: $31.9 million of lease revenue against $12.8 million of crypto, down from $47.6 million a year earlier. Read as an AI story this is a margin upgrade. Read as a balance sheet it reprices what these firms ever owned, which was energised interconnect, built when bitcoin was the only bidder for stranded power. The asset never changed. The tenant did, and the tenant was always the whole business.
Merck and Moderna's melanoma therapy worked, and the interesting thing about it is that there is no such thing as a dose of it. Moderna shares more than doubled on the session, closing up about 177 percent. Intismeran autogene, given with Keytruda, met its primary endpoint of recurrence-free survival and a key secondary in 1,137 patients with completely resected stage IIB to IV melanoma, at a pre-specified interim analysis, the first Phase 3 win for an individualised neoantigen therapy. Each course is built from that patient's own tumour, encoding up to 34 mutations unique to it. Every pharmaceutical model since the 1950s rests on spending enormously once and then producing identical units at almost no marginal cost, which is what Keytruda is: $31.7 billion of sales last year, about half of Merck's revenue, on a compound whose US patent expires in December 2028. Its replacement has no identical units. The variable cost is a sequencing run, a design step and a synthesis, per person, forever, which makes the binding constraint plant throughput rather than patent life. The competitive question stops being chemical and becomes industrial.
OpenAI has put its largest frontier reinforcement-learning run on hold indefinitely and now spends about a fifth of its monitored inference compute watching its own models. Jakub Pachocki and Sam Altman each posted the detail on the evening of 18 August: Astra's training paused for two weeks, the larger frontier run held with no restart date, monitoring mandatory for every Astra inference involving tools since 7 August, and a thirty-minute service level from a flagged trajectory to a human being paged. Zvi Mowshowitz treats this as the first time a frontier lab has priced alignment work in the currency that constrains it. Twenty percent of monitored serving compute is not a research budget. It is a tax on the thing that earns, and it landed the week the Wall Street Journal reported the quarterly loss widening $3 billion to $12.3 billion on revenue of $6.7 billion. We expect the halted run to be the more durable signal, because a compute tax gets optimised down and a lost training schedule does not.
Analog Devices posted a record $4.02 billion quarter and guided the next one about nine percent above consensus, on the layer of the AI buildout nobody benchmarks. Revenue rose 40 percent from a year earlier and adjusted earnings of $3.45 a share rose 68 percent, with the company naming data centre and industrial as the drivers rather than either alone. The fiscal fourth-quarter guide is $3.86 a share plus or minus fifteen cents against roughly $3.55 expected, with gross margin guided near 74 percent. That margin is the whole point. Analog content is power management, signal conditioning and precision timing, the parts between the racks, and no one runs an evaluation on a voltage regulator or switches supplier over a published score: qualification cycles run years and designs get locked. It printed on a session when technology led the decliners. A layer that competes on qualification rather than on benchmarks keeps its pricing through exactly the repricing that takes the layer above it apart.
Pennsylvania now requires an AI data centre to win local community approval before it is built, and has pulled every data centre out of the state's fast-track permitting programme. Governor Josh Shapiro signed the order on the evening of 18 August, effective immediately, adding environmental and transparency conditions and barring agencies under his jurisdiction from signing non-disclosure agreements with developers. His office calls them the strictest standards in the nation, which is his characterisation rather than a comparison anyone has run. What makes it bite is the capital that goes in before a permit matters: published 2026 construction benchmarks cluster around $11 million per megawatt of IT load, and Brookfield's 2026 infrastructure outlook uses roughly $10 million, so a hundred-megawatt site carries a billion dollars or more of spend before it earns anything. A permitting queue is financeable because it has a distribution. A community vote does not, which is the difference between a delay you can price and one you cannot. We expect siting rather than silicon to bind 2027 capacity, and expect developers to crowd into the states that have not moved.
A Malaysian prime minister called Taiwan a province of China on camera, and the argument about what else he said is the more instructive part. In an Al Jazeera interview aired 16 August, Anwar Ibrahim described Taiwan as a province of China. Pressed on Beijing's potential use of force, he answered by analogy: if a Malaysian province tried to secede, protecting the unity of the nation would be his responsibility, by force if needed. Two readings are live and both are in print. Chinese state media and several outside analysts read the analogy as backing force for reunification. Malay Mail reads the answer as being about Malaysia's own territory. Taipei's foreign ministry condemned the remarks on 18 August; Beijing's spokesman Lin Jian said on 19 August that China appreciates them. Xinhua's own headline says Anwar reaffirmed the one-China principle, the formulation Beijing prefers to the one-China policy Kuala Lumpur has used since 1974, though no Malaysian document has changed. What is at stake sits in the back end of the chip supply chain, where the Malaysian Investment Development Authority puts the country at 13 percent of global assembly, testing and packaging. The word is free to say today. It prices on the day those plants sit on one side of a perimeter. We expect further ASEAN realignment to arrive as vocabulary rather than as treaties, because vocabulary costs nothing to issue and is hard to withdraw.
Iraq is selling crude at as much as $29.80 below its own official selling price and its oil revenue has halved, while eighty-five percent of spending is a payroll it cannot cut. Bachar El-Halabi's accounting: southern loadings ran 1.8 to 2 million barrels a day in the first half of August, well below Iraq's recent norm, while SOMO widened August differentials on Basrah Medium and Heavy to between $25 and $29.80 under the official selling price, from $14 to $18.80 in July. First-half oil revenue came in at 28.5 trillion dinars, about $21.8 billion, against 57 trillion a year earlier, on spending of 57.18 trillion. From 1 September a cabinet-approved mechanism lets buyers take Iraqi crude at a 30 percent discount. We expect Iraq to keep clearing volume at almost any price into 2027, because a state whose budget is a wage bill cannot wait, and a forced seller of two million barrels a day caps the Gulf premium from the supply side, the direction nobody watches.
The record for the most atoms ever held in a quantum array is 6,100, and not one of them has run a calculation. Quanta's Ben Brubaker surveyed the field on 19 August, and the figure that stops you is the gap between two records usually quoted as one. A Caltech group trapped 6,100 individually controlled cesium atoms in an optical tweezer array about a millimetre across, which is the number the field cites for how far the hardware has come. The count actually used to execute an algorithm remains far lower, and nobody has bridged the two. Holding a thing in place and getting work out of it are separate achievements with years between them, and a surprising share of the capacity numbers quoted in any industry are quietly the first kind.
A sixty-year-old textbook explanation for why insects evolved societies has just failed the first proper test anyone ran on it. Sachin Suresh and Timothy Linksvayer at Arizona State mapped about 69,000 insect species onto two large phylogenies and checked whether haplodiploidy, the genetic system making sisters more related to each other than to their own daughters, predicts eusociality. In Current Biology on 15 August they report that almost the entire statistical signal comes from one clade, the aculeate Hymenoptera, and that haplodiploid insects outside it go social at roughly the same rate as ordinary diploids. By the authors' account the hypothesis had almost never been tested comparatively in sixty years of textbooks. It survived on the elegance of the arithmetic rather than the evidence, which is a specific and repeatable way for a field to be wrong.
A single parasite took six and a half million American households out of the packaged-salad aisle in one month. Numerator's July purchase data, relayed by restaurant reporter Jonathan Maze, shows 6.5 million households stopping salad-mix purchases over cyclospora and fast-food salad trips down 16 percent, and Salad and Go named the outbreak as the last straw before its bankruptcy. The shape matters more than the pathogen: the risk was invisible, the substitute was trivially available, and the category emptied in weeks rather than quarters. Categories built on convenience have no loyalty to fall back on when trust goes, which is why consumer collapses in them look nothing like the gradual erosion the models assume.
Context signal: Overall delinquency improved in the same quarter that the loans already sick got much sicker
The Mortgage Bankers Association's National Delinquency Survey, released 13 August, put overall mortgage delinquency at 4.37 percent of one-to-four-unit loans in the second quarter of 2026, down seven basis points on the quarter. One line down, the book splits in two. Serious delinquency, ninety days past due or in foreclosure, rose for the fourth consecutive quarter: up 227 basis points from a year ago on FHA loans, up six on conventional. Same quarter, same labour market, same rates, and a thirty-eight-fold difference in how two books are ageing.
What separates them is the part of a fixed-rate mortgage that was never fixed. Escrow, meaning property tax and homeowner's insurance, collected monthly and reset once a year, is up 45 percent nationally since 2019 on Cotality's March 2026 figures, and up 77 percent in Colorado and 70 percent in Florida. In some states escrow is now in the mid-forties as a share of the monthly payment. The borrower's debt-to-income was underwritten against the escrow at closing, and the reset arrives annually, roughly a year behind insurance renewals that were priced long ago. FHA's borrower is the thin-reserve borrower, concentrated in exactly those states.
We expect FHA serious delinquency to keep climbing through 2027 while the headline number looks unremarkable, because a loan this deep does not cure at a 6.9 percent mortgage rate with little equity behind it. The cost lands on the servicer before it lands on credit: Ginnie Mae issuers must advance principal, interest and the escrow itself on delinquent loans, which is working capital scaling with the sick part of the book at PennyMac Financial (PFSI) and Rithm Capital (RITM). It reaches the entry-level builders too, D.R. Horton (DHI) and LGI Homes (LGIH), whose FHA buyer now faces a payment that rises without a rate ever moving.
Context signal: Chipmakers built far too much silicon carbide for cars, and the collapse in its price is what makes the 2027 AI data centre electrically affordable
Silicon carbide was scaled for the electric vehicle. Then EV growth disappointed and the capacity did not. Yole's December 2025 assessment put 2025 utilisation near 50 percent in upstream substrate processing and 70 percent in device fabrication, with Chinese producers already holding roughly 40 percent of wafer and epiwafer capacity as of 2024 and still building. Six-inch substrate prices collapsed and eight-inch fell steeply behind them. Yole's forecast has the downturn running to 2027 or 2028.
The second market was not in that forecast. Nvidia is working with more than twenty infrastructure partners on 800-volt DC distribution, aimed at the Rubin Ultra generation in 2027 and rack densities up to 600 kilowatts. That architecture converts utility medium voltage straight to 800 volts DC, and the front-end stage, the solid-state transformer, is a silicon carbide job, with gallium nitride taking the conversion inside the rack. Enphase (ENPH) showed a working version in July 2026 at 342 modules per 1.25-megawatt rack.
The transferable mechanism: a technology's cost curve is usually crashed by overbuilding for the market it was designed for, and that crash is the subsidy that opens a market nobody forecast.
We expect the silicon carbide downturn to end earlier than the 2027-28 consensus, and to show up first in substrate utilisation rather than in device prices, which means the wrong names re-rate first. The device makers sitting on idle fabs, onsemi (ON), STMicroelectronics (STM) and Infineon (IFX.DE), take the volume; Navitas (NVTS) and Enphase hold the rack-level gallium nitride. Exposed are the vendors whose content is the AC power chain 800 volts DC deletes.
More than thirty AI models, from twelve different developers, sit above the European AI Act's training-compute line of 10^25 floating-point operations, on an Epoch AI count dated June 2025, which is its own small piece of evidence. Above the line a model is presumed to carry systemic risk and its provider owes the Commission notification within two weeks. Every serious proposal for governing AI runs through a number like that one, and the historical record on such numbers is worse than the debate assumes.
In early 1963 a single number separated Washington and Moscow on a comprehensive nuclear test ban. Khrushchev offered three on-site inspections a year; Washington asked for seven. Neither figure came from a seismologist. Each was a political judgment about how much residual uncertainty a government would accept, and the talks died on it. What was signed that August banned testing in the atmosphere, in space and underwater, and excluded underground tests, the one category requiring anyone to be let in.
Call the mechanism verification exposure: the access that produces assurance is the same access that produces intelligence, so the monitored party prices an inspection as a disclosure, and a technically stronger monitor can have a smaller agreement space than a weaker one. Andrew Coe and Jane Vaynman supply the formal version, that assurance and vulnerability travel together, which is why arms control is rare rather than merely difficult. The corollary does the damage. An unverifiability objection and a refusal to be verified produce the same sentence. "That scheme cannot be made to work" is what an honest engineer says and what an unwilling party says, and no improvement in the instrument separates them, because the unwilling party's objection was never about the instrument.
The Biological Weapons Convention has had no verification regime since 1975, and the compliance protocol negotiated from 1995 to 2001 was rejected by the United States on the stated ground that it would not improve verification, with the loudest domestic opposition coming from pharmaceutical firms on commercial confidentiality. Both readings fit every fact in the record, which is why an unwilling party reaches for the technical argument: it is free, it sounds rigorous, and no better sensor answers it.
Read the AI Act against that and it stops looking like a first draft of AI governance and starts looking like the Limited Test Ban Treaty of it: real, binding, and agreed quickly because its standing obligations are declaratory. Providers self-notify, document and report. That is not the regime's weakness; it is why it exists. The statute also concedes its headline number is provisional: Article 51(3) obliges the Commission to amend the threshold by delegated act as technology moves, and Annex XIII holds non-compute criteria in reserve for when compute stops tracking capability, which inference-time scaling is already making happen.
So we expect the split in AI governance to run along access rather than ambition. Declaration-based regimes keep advancing and keep being agreed quickly; anything requiring standing, reciprocal entry to weights, training data or a facility keeps stalling, and a US-China arrangement is least likely of all, because the inspection that creates the assurance is the inspection that reveals the asset. On the European number, we expect the Commission to reach for Annex XIII before it raises 10^25, since raising it concedes the figure was never a measurement.
Where this might be wrong. The strongest objection is a treaty that did the opposite. The Chemical Weapons Convention carries challenge inspections, routine industry inspections and a standing organisation in The Hague, agreed over confidentiality complaints from the chemical industry every bit as loud as the pharmaceutical ones that sank the BWC protocol six years later. The difference was design: the OPCW built managed access and a confidentiality annex into the instrument, bounding exposure rather than leaving it open. If that is the lesson, verification exposure is an engineering problem already solved once, and the AI case is early rather than barred.
The second objection is in the Act itself. Article 92 lets the AI Office demand access to a model through APIs "or further appropriate technical means and tools, including source code," with Article 101 fines for refusal. That is not a declaratory obligation, and it was legislated without the exposure problem stopping it. The honest answer is that it survives by being discretionary and one-directional: a regulator triggers it against firms in its own jurisdiction and none of them inspect back, while the mechanism is about reciprocal verification between parties who are each other's risk. Article 92 has never been exercised against a frontier lab. Held as a threat it costs the Commission nothing; exercised, it becomes the thing this argument says is hard.
The third is that the framework is over-determined. If neither Washington nor Beijing expects durable parity, there is nothing to trade, and ordinary bargaining theory explains the missing regime without any of this. That is probably right about the US-China case, and it is the reading this Take least wants to be true.
"A beginner should not hold two arrows. It will make him rely on the second arrow and be careless with the first."
— Yoshida Kenkō, Essays in Idleness, c. 1330 (trans. Donald Keene)
You would call the archer prudent, not lazy. He has brought a spare because things go wrong, which is what a sensible person does. His teacher tells him the spare is the problem.
Read the sentence again, because the claim is stranger than it looks. It is not that the second arrow makes him fail. It is that holding it changes what happens in his hands during the first shot, before anything has gone wrong, in a way he cannot feel. Kenkō is explicit: the pupil is unaware of any carelessness, and the teacher sees it from across the range.
You are holding one right now, and you know which. The second job application you keep open in another tab. The apartment you have not stopped looking at. The version of the conversation where you do not have to say the hard sentence. Each was acquired for a good reason, and each is quietly deciding how much of you shows up to the thing you say you are doing. The cost is not paid at the moment you fall back on it. It is paid continuously, by the part of you that stays behind to keep it warm.
Kenkō ends with a student who thinks at night that he has tomorrow and in the morning that he has tonight. A reserve option and a later moment are the same place to put the part of yourself you have not committed. Wendell Berry's Muse of Realization, in this section a day ago, returns to say the work is harder than you thought, and you catch that failure by hearing yourself describe a slightly different project. This one you cannot catch that way. Both hold, and the difference is the diagnostic: Berry's failure answers to attention, Kenkō's only to removing the object.
Today's practice: take the piece of work you are making right now and find its second arrow, the alternate draft or the safer version you built in case the real one fails. Remove it before you next sit down, and give the next hour to the first arrow only.
In 1693 William Molyneux wrote to John Locke with a question he could not shake. A man born blind has learned by touch to tell an ivory cube from an ivory sphere. His sight is restored, the two objects sit on a table, and he may not touch them. Can he tell which is which by looking? Locke thought not, and philosophers argued about it for three hundred years.
Then somebody checked. In 2011 Pawan Sinha's group at MIT, working with a charity eye hospital in New Delhi, tested five patients aged eight to seventeen within forty-eight hours of the surgery that gave them sight. Touch-to-vision matching came in at 58 percent, near chance. Molyneux's answer was no. Then two of the five, retested days later on a new set of shapes with no training in between, were scoring above 80 percent.
Mechanism. Molyneux's question looked unanswerable for three centuries and was empirical the whole time. Nobody had built the situation in which it could be asked. That is one kind of hard question: the answer is unknown, the method is fine, and what is missing is an observation. There is a second kind. Write down the best answer your method can produce, in your method's own vocabulary, and check whether the question survives it. If it does, you are not short of data. You are short of vocabulary. David Chalmers named the sharpest example in 1994: you can specify everything a nervous system does, every discrimination, every report, every integration of information, and the question of why any of it is accompanied by experience is still sitting there untouched, because none of those answers were in the business of addressing it.
A second domain. Revealed preference gives a complete account of choice: what was chosen, at what price, against what alternatives. Run it to its limit and you have described every transaction a person ever made and said nothing about whether they were better off, because "better off" was never a term in the vocabulary. Which is why welfare economics exists as a separate argument rather than a corollary. More transaction data was never what was missing.
Sizing. Most questions are the first kind, and the failure mode of this model is treating too many of them as the second. Declaring a question fundamentally beyond your method is satisfying and usually premature. Wöhler's 1828 synthesis of urea from inorganic material is the case everyone cites, and the pattern holds even where the popular telling does not: a question that felt like it needed a new category of explanation turned out to need better chemistry. Assume that first.
Failure mode. The reverse error is quieter and more expensive. You keep gathering data on a question no amount of it can settle, because gathering is legible and reframing is not. A team measuring engagement to find out whether the product is good. A person auditing their week to work out whether the life is right. The instrument returns real numbers forever, the question stays untouched, and the activity feels rigorous because it is producing something.
The decision tool: before you spend on a hard question, write the strongest answer your current method could give, then ask whether the question is still standing. If it collapses, you have an empirical problem and should go get the observation, the way Sinha did after three hundred years of argument. If it is still standing, more of the same will not move it, and your task is to find the vocabulary in which the question becomes answerable, or to decide honestly that you will act without one. Both are respectable. Measuring harder is not.
Give E. coli all the oxygen and glucose it wants and, past a certain growth rate, it does something that looked for decades like incompetence: it stops burning sugar all the way down and ferments it instead, excreting acetate and pouring most of the available energy into the medium. Markus Basan and colleagues measured why, in Nature in December 2015, and the answer was not a flaw in the machinery. It was the machinery. Respiration extracts far more energy per molecule of sugar but takes roughly twice as much protein to run, per unit of energy produced. A cell has a fixed protein budget, and every enzyme built to burn sugar efficiently is an enzyme not built to grow, so when sugar is abundant and protein is the binding constraint, the cell rationally wastes the sugar. It is not being careless with food. It is being careful with apparatus.
What transfers is not the metabolism. It is that efficiency is never free. Being efficient in one resource requires machinery made out of another, and which one you ought to conserve depends entirely on which is currently scarce. Nothing here is hidden; what moves is the exchange rate between the two costs. Most efficiency habits were formed when one specific thing was scarce, and they survive unexamined long after abundance arrives, quietly spending the resource that has become precious to protect the one that no longer is.
The decision tool: when you catch something being obviously wasteful, do not fix it until you have priced the fix in the other currency. Name the resource the waste consumes, then name the resource the efficient method would consume: the hours, the attention, the standing overhead of whatever system keeps the waste from happening. If that apparatus costs more of what is genuinely scarce than the waste costs of what is abundant, the waste is the correct answer, and eliminating it is a downgrade wearing the costume of discipline. Test it inside a week: pick one thing you feel guilty about wasting, write down what conserving it would actually cost you, and notice whether you had ever counted that side at all.