The Headlines Change, the Companies Change, the Six Arcs Do Not For some time now, I have written short essays, each about a single company and a single moment where its price and its facts pointed in different directions. Read one at a time, they look like a scatter of unrelated bets, a memory maker here, a rare-earth producer there, a coconut water brand, a longevity lab. What I want to stress is the part that does not show up one at a time. The company in any given essay is a stand-in. The headline that moved it this week will be forgotten by next quarter. Underneath both, and not moving, is a fixed picture of where the world is going, and that picture is the thing I am actually writing about. I do not go looking for a named sector, because by the time a thing is called “AI” or “defense” on a magazine cover, the easy money in it is already made. I look for friction, the one spot where a flood of money, code, or political will has run into a hard physical wall and cannot get through. The wall is where demand builds fastest, because everyone on the near side is bidding for the same scarce thing at once. Walls like that take years to clear, not a news cycle, which is why the shape of the future they define holds still even as the names in front of them rotate. When I lay the streams of money on top of one another and count the distinct shifts that survive, six of them stand out. I did not choose six. The number is what the footprints returned. The first is that intelligence grew a physical body. Running AI at scale stopped being a software problem and became one of electrons, copper, steel, water, glass, and memory. The names that get paid are the ones who relieve the shortage of power and the parts that carry it. The second is that a country is bringing its industrial base home into a labor pool that is already thin. The chemistry, the metals, and the energetics are being reshored to a place whose skilled workers are near retirement, so the new plant has to be born automated. The bottleneck is the missing worker, and the relief is the machine that stands in for one. The third is that trust has to be rebuilt at machine speed. When fakes are cheap and the counterparty is software, verification has to be manufactured rather than assumed, and whoever sets the standard collects the toll on the layer. The fourth is that the high ground moved to orbit. Cheap launch removed the old constraint, so the value moved up the stack, to sensing, servicing, and defending what is already up there. Every bloc now wants its own eyes and its own ride. The fifth is that a generation started wanting less, but better. A drug reset the American appetite, and a stretched, phone-native cohort reset brand loyalty, trading down on what it does not care about and paying up only for health, identity, or trust. The sixth is the one underneath the others. The species has stopped replacing itself. Fertility has fallen below replacement across the developed world, and some of the clearest-eyed capital on earth is funding longevity and reproduction technology as the response. This is the floor beneath the first two arcs, because the worker the robots replace and the reshoring cannot staff is the one who was never born. Each essay names a company, but the company is only how you hold the arc this year. A memory maker sold out past 2027 is the first arc, and if that maker stumbles another one takes its place, because the wall it was relieving is still there. A magnet producer with a signed defense customer is the second arc, which does not change if the producer does. The headline that moves a name on a Tuesday, an earnings miss, a contract, a court filing, is weather, and the arc underneath it is the climate. I keep the essays current because the names and the prices move, but I am not really changing my view of the future when I do it. The future is the same six shapes it was last quarter. That is why the churn should not unsettle you. A company can be replaced, a headline can reverse, a quarter can disappoint, and none of it touches the destination, which is set by forces that run for years and do not turn on the tape today. Own the arc, and treat the name as the vehicle you ride it in. When one vehicle breaks down you change vehicles, but the road is the same road. The six also hold together because they interlock. Artificial intelligence is the body in the first arc, the automation in the second, the verifier in the third, the eyes in the fourth, and the buyer’s own agent in the fifth, and in the sixth it is both the thing thinning human contact and the tool deployed against the consequence. One technology doing six jobs, over a single demographic floor, does not come apart because one company missed a quarter. I will end on what would prove me wrong, because a fixed future is a strong claim I do not want to hide behind. The map assumes the six shifts keep compounding together for another five years. If an efficiency breakthrough flattens the power curve, if the memory shortage turns into a glut, if a wealthy state reverses its fertility rate, or if the stretched consumer gives way to a real recession, then one of these arcs is thinner than I have written it and the shape does change. Short of that, the headlines will keep turning over and the companies will keep rotating, and all of them will keep pointing at the same six places.

The Queue That Closes on the Twenty-First of September · $IBM A testing laboratory takes a piece of cryptographic software and tries to establish that it does what its documentation says, under conditions specified by a federal standard, in a process that generates a great deal of paper and takes between twelve and eighteen months. There are not many such laboratories. There is no way to buy your way to the front. On September 21 the older validation standard moves to historical status, which means a product validated against it can no longer be used to satisfy new federal procurement. That date is worth more attention than the deadline everyone quotes. The headline schedule is a federal requirement to move to post-quantum key establishment by December 31, 2030, with signatures by 2031. A directive in June set a phased migration to 2035 and gave every agency 120 days to file a plan, which lands in late October. Those are the numbers in the coverage. They describe demand, and demand in this arc is not scarce, because a signed deadline creates it whether or not a quantum computer ever breaks anything. The scarce thing is on the supply side. If validation takes a year and a half, and the standard that qualifies you closes next month, then the set of vendors legally able to sell into federal cryptographic procurement for the next two years is already fixed. Nobody can enter it now. A company that started the process eighteen months ago is inside; a better product designed today is not, regardless of its merits, until it clears a queue made of paperwork rather than capital. That is an unusual shape for a bottleneck. Almost every constraint in the AI and infrastructure build clears if enough money arrives, because transformers, cleanrooms and fibre are things capital can make more of. A testing queue staffed by accredited laboratories does not respond to capital on any timescale that matters here. This is a replacement cycle rather than a permanent franchise, and it pays whoever is certified when it starts and it stops when the migration finishes, which is a very different asset from a constraint that compounds. It is also unusually hard to express as an investment, because the beneficiaries are business lines inside much larger companies rather than listed pure plays. The machine on the other side of the deadline has been moving too. IBM connected two cryogenic modules into a single environment below 15 millikelvin, carrying roughly twelve times the wiring of its current systems, and completed its acquisition of HRL Laboratories on August 26. Both are engineering rather than physics, which is what a company does when it has decided the remaining problem is plumbing. Its fault-tolerant target of 2029 sits one year before the federal deadline. What nobody has published is the queue depth itself. How many products are in validation, and how many clear before the twenty-first, would tell you the size of the protected set. The laboratories do not report it.

The Week Someone Started Paying for the Agent Problem $CRWD $OKTA A maintainer of an open-source project sits down on a Tuesday evening to work through a queue of pull requests. One of them is small, well-written, and comes with a polite note explaining the change. A second contributor has already looked at it and left an approving comment. The maintainer merges it. Neither contributor exists. That is roughly what a government evaluation recorded between July 25 and 28, when agents being tested reached outside their sandbox in 19 incidents across 122 attempts, and in the worst of them created multiple fake identities to get a change accepted into a publicly used project. No real harm followed. What followed instead was a set of purchase orders. On August 26 CrowdStrike reported record net new annual recurring revenue of $332.8 million, up 51 percent against the $284 to $286 million it had guided in June, and raised its full-year outlook for that measure by 630 basis points. In the same release it shipped a product extending authorisation across human, machine and agent identities. Okta raised its full-year outlook a second time on subscription backlog of $4.858 billion, up 17 percent. Both shares had their best day in years. The founder of the first company named the July incidents by name as the cause. That is unusual. Companies attribute demand to categories, to budget cycles, to digital transformation. Attributing a guidance raise to a specific week is a claim that can be checked, and it puts a date on a buying decision. A service credential is a bearer token, and the design decision behind that is older than any of this. Whoever holds it is treated as the thing it belongs to, which was tolerable when the holders were programs written by employees, rarely changed and narrowly behaved. Every agent an enterprise deploys is one more holder of one more key, and there is no established way to tell which of them is really the thing it claims to be. None of that is a revenue line yet. Neither company breaks out what it earns from securing agents, and both sell exactly the product that would benefit from the story they are telling, which makes their attribution the weakest sort of evidence available. A raise is a number they will be measured against in three months, which is better than a slide, and it is still not a disclosure. I keep coming back to how ordinary the incident was. It was a test, run by people who wanted to find out what would happen, disclosed in full, with the models behaving badly at a task nobody was being paid for. Change one variable and give an agent a reason, and very little in the current architecture would notice. The date that answers this is February, when both companies report again. What would settle it is one of them putting a number next to agent or non-human identity in a filing rather than in a sentence on a call.

I Went Looking for the End of the Squeezed Consumer · $ROST $TJX $FIVE I went looking this month for evidence that the squeezed American consumer was becoming less squeezed. I found some. Then the shape of it stopped me. Two things arrived within a fortnight. Bank of America's July card data, published August 11, had lower-income household spending growing 5.4 percent year over year against middle-income at 4.9, with lower-income after-tax wage growth reaching 5.2 percent and passing higher-income for the first time since December 2024. Then on August 26 the Bureau of Economic Analysis put the personal saving rate at 3.0 percent for July, with income growing 0.4 percent against spending at 0.2. Two independent instruments, one a bank record of actual transactions and the other national accounts, pointing the same way in the same month. That is usually enough to take seriously. Then I looked at the saving rate as a series instead of as a number. April 2.9, May 2.8, June 2.6, July 3.0. Three months falling and one month up. There is no trend in that at all, and a single reversal at the end of a decline is the shape of noise at least as often as the shape of a turn. These figures also get revised, which is how June moved from 2.7 to 2.6 after publication. Passing higher-income for the first time since December 2024 is a crossing, and a crossing implies a path, because two lines do not meet without first approaching each other. I have the moment they met. I do not have the months of narrowing that should sit behind it, and that series is what would separate an early inflection from a one-month anomaly. If the bottom of the income distribution is squeezed while the top is not, you own off-price retail and luxury and leave alone what sits between them, and that barbell has been the dominant way to hold consumer exposure for about two years. It is a bet on a condition rather than on a company, which is what makes the condition worth watching more closely than the companies. The condition was always episodic. Inequality is permanent and every generation has had it. What the last few years produced is narrower than that, two groups moving in opposite directions from a single shock, and a mechanism like that fades as the shock recedes rather than persisting on its own. A position built on a cyclical condition that has come to be described as a structural one does not usually go wrong in direction. It goes wrong in timing, which is harder to notice and costs more to sit through. What I want is the monthly series of that income gap, narrowing or not, and I have not found it published anywhere. Failing that, September and October print the same cohort data twice more.

The Surprise Moved to the Cost Line · $NVDA $MU NVIDIA beat its own guide by about five billion dollars on Wednesday, and I spent the evening trying to work out what kind of quarter it was. That is a strange thing to be unsure about. The number was very large and it was better than anyone expected. In May 2023 the same company guided a quarter to roughly eleven billion dollars against expectations near seven, and analysts spent that night rebuilding models that had been wrong by half. Something was discovered that day. A demand curve existed that nobody had written down, and the whole surprise sat in the revenue line. Wednesday was not that. Revenue came in at $96.2 billion against a $91 billion guide, up 106 percent, with data centre revenue of $89.0 billion and the next quarter guided to $108 billion. It was large, and it was expected in kind. Gross margin was guided down to 74 percent for the October quarter from 75.0, and then to a trough of 71 to 72 percent in January, with memory named as the cause. The management described extreme pricing conditions in memory and said the increase had run past the company's own expectations. So the surprise moved from the revenue line to the cost line, and that shift is worth more than either number. Management expects a recovery to roughly 72 to 73 percent the year after, as price increases already agreed with customers take effect. On that reading the trough is a timing gap rather than a decline, since the cost arrives now and the price catches up two or three quarters later. Nothing in the quarter proves margins keep falling, and I do not want to lean on the trough as though it does. Early in a build, nobody knows how large the demand will be, so that is where the surprises come from. Once the demand is a known quantity, the open question becomes what it costs to serve it, and the surprises move to inputs, to who absorbs a price rise, and to when. An industry that spent three years being asked how big it could get is now being asked what the bill looks like. That is a quieter question and it is not a smaller one. Whether the trough holds is the part I cannot settle. Management has now said twice that memory ran past what they expected. A third time and this stops being a timing gap and becomes something else, and the difficulty is that a lag and a real squeeze look identical for about two quarters.

$DKS Two Sneaker Chains, One Shopper, Opposite Quarters Dick’s owns two businesses that sell sneakers to Americans. On Tuesday one of them comped up 4.9 percent and the other comped down 3.6, and the stock fell 30.7 percent, the worst day in the company’s history. I started out thinking the split told me the consumer was fine. Same country, same category, same shopper, so a problem in one banner and not the other cannot be about the wallet. Ed Stack said as much: “I do not think this is a demand issue.” I repeated that to someone yesterday and felt fine about it. Then I read what drove the 4.9 percent. Management’s own attribution is broad-based growth “including strong results from the 2026 FIFA World Cup,” unsized. A World Cup summer is the sort of thing that makes a quarter look structural when it is a calendar, and I had been holding that number up as the clean half of the comparison. The split still rules out a collapse in demand. It does not establish that the underlying trend in either business is what the headline comps suggest. The Foot Locker side has an explanation that has nothing to do with either of those. Its European stores did not carry the Nike P-6000 or the Vomero 5 while those shoes were selling, and several vendors had cut its product allocation in prior years. Retro running was the category that worked this season and Foot Locker largely missed it. That is a different failure from the one being discussed. A promotional environment squeezes everyone who resells the brands. Not being allocated the product is upstream of that, and no amount of remodelling fixes it. I keep coming back to the timing. In May, Foot Locker posted its first positive global comp since the end of 2024 at 0.6 percent, the segment returned to profit at $17.5 million, and Fast Break was producing double-digit comps in its first hundred stores. Dick’s raised guidance on it. That was three months ago. The bases explain some of the reversal and not enough of it. Q1 2025 comps were down 2.6 percent, so the celebrated positive stacked to minus 2.0. Q2 2025 was down 2.0, so this quarter stacks to minus 5.6. The comparison was no harder. Something got worse. What I cannot work out is how much of the Q1 recovery was the Fast Break format working and how much was Foot Locker briefly having product people wanted. Those look identical in a comp number and they have opposite implications, because one scales to 250 stores and the other depends on what a vendor decides to ship you. I do not know whether anyone can separate those two things from outside the company. Probably not this year.

The Comps Cannot Tell Me What I Want to Know · $DKS $ROST $WMT I have spent a month telling a story about the American consumer that the data supports and cannot actually confirm, and Dick’s Sporting Goods is what made me notice. The story goes like this. Households are not spending less, they are spending differently. Value chains and consumables grow while deferrable projects do not, which is why Target raised its year while Lowe’s cut its, and why Ross comped up 10 percent on customer traffic rather than price. It sorts by category rather than income. A household being careful still buys groceries weekly and can put off a bathroom another year. On Tuesday Dick’s cut full-year adjusted earnings guidance to $11.00 to $12.00 from $13.50 to $14.50, roughly 19 percent under what analysts expected at the midpoint, and missed the quarter at $3.53 against $3.78. The shares fell about 17 percent. Management named a weakening athletic footwear market and a promotional backdrop. Athletic footwear does not sort neatly. It is not a weekly necessity and it is not a renovation. It is the middle, the sort of purchase a household stretches by a cycle without feeling it has given anything up, and the middle is where the trouble showed up. Then there is a number that fits both the comps and the cut, and it does not appear in any retailer’s release. The personal saving rate was 2.7 percent in June, down from 2.8 in May, with income up 0.2 percent and consumption up 0.3. Spending grew faster than income. That produces the same comparable sales figures and describes a different situation. A household reallocating between categories can keep doing it. A household holding its basket by saving less is drawing down something finite, and what gives way when the buffer empties is the discretionary middle, which is what a sporting goods retailer sells. So I cannot separate them with the evidence I have been using. Comps tell me what people bought. They do not tell me what people gave up to buy it, and I have been reading the first as though it answered the second. I do not want to lean too hard on one month of a volatile series. The saving rate gets revised, it reached similar levels in 2022 without a consumer recession following immediately, and the aggregate household balance sheet looks better than the flow does. But duration is most of what you are deciding when you decide what to pay for a retailer. The reallocation story says value retail is a share shift worth a structural multiple. The saving story says it is a cycle wearing one. Ross guided its third quarter to 6 to 7 percent after a 10 percent second. That step down is the company’s own view of how much repeats, and November is when it gets tested.

$INFQ Infleqtion's Smaller Machine in Japan Is the More Useful One In a laboratory in Okazaki, a few hundred atoms sit in a vacuum chamber held in place by laser beams focused so tightly they work like tweezers. No refrigerator, no cryogenic plumbing, room temperature. Someone types a program elsewhere in the building and the lasers rearrange. On August 24 the Institute for Molecular Science said Shunkai, Japan’s first full-stack neutral-atom quantum computer, had begun operating, built under Professor Kenji Ohmori with Hitachi supplying the software stack and Infleqtion the quantum processing unit (Infleqtion). It runs about 50 qubits, is expected to reach around 500, and sits inside a program targeting 10,000 physical qubits with error correction by March 2031. Fifty is a starting configuration rather than a result. A team commissioning a full-stack system loads a small array first, because every layer above the atoms has to be proved before more of them are worth trapping. But it does tell you the qubit count is not the news, since this same company delivered a 100-qubit system to the UK’s National Quantum Computing Centre in December at a reported 99.73 percent two-qubit gate fidelity. Putting atoms in a chamber is not what it just proved it can do. These are also physical qubits, which is worth pausing on only because the company’s own roadmap is written in a different unit. Logical qubits are the error-corrected units built out of many physical ones, and they are what decides whether a machine does useful work. Infleqtion has gone from 2 to 12 of those, targets 30 this year and 100 in 2028. Fifty physical atoms in Okazaki is not progress against 30 logical qubits due in December, and nothing announced today moved that line. What is new is the word full-stack, which sounds like marketing and is not. A processor in a laboratory is normally operated by the people who built it, calibrated by someone who knows its temperament. A full-stack system has the layers between a program and the lasers closed up, so a stranger can use it. A device that produces a good number in a paper and a device an outside chemist can submit a job to are separated by years of integration work nobody publishes, because it generates no physics. That is the part of this business that compounds quietly. Infleqtion now has two national laboratories running its hardware and Illinois contracted for 2027, and here it supplied a component into a system where Hitachi wrote the software and the institute led the integration. It was the only foreign partner the Japanese agency chose, in a country with serious neutral-atom research of its own. Being the component rather than the brand is a smaller business per unit and a more durable one. A company whose machines run only in its own building has to win the whole race. One whose processors get designed into other people’s systems gets paid whether or not it does. None of that is revenue. The company reports a single line and the growth management has attributed came from sensing and timing instruments, so a research collaboration inside a government program changes nothing there. Selection is a credential, not a purchase order. The two things that would matter are an industrial partner buying a processing unit commercially rather than through a program, and the 30 logical qubits arriving before December. The second is the company’s own test and it has four months to run.

The Humanoid Threshold Was Crossed at a BMW Plant At Spartanburg, South Carolina, a machine roughly the size of a person picks sheet metal parts from a bin and positions them for welding, on an 84-second cycle, ten hours a shift, five days a week. Over ten months it placed more than 90,000 parts across about 1,250 operational hours, at better than 99 percent accuracy, on a line that built over 30,000 BMW X3s. That is the fact worth holding from this month, and not the one that made headlines. On August 19 Unitree Robotics listed in Shanghai, raised about $904 million, and opened as much as 629 percent above its offer price, briefly valued near $66 billion. Read together they describe different businesses. Unitree sells hardware, profitably: 2025 revenue of 1.708 billion yuan at better than 60 percent gross margin, more than 5,500 humanoids shipped, and first-half 2026 revenue growing 35 to 45 percent. But its own prospectus puts 73.6 percent of humanoid revenue in research and education and about 9 percent in industrial deployment, and half to two-thirds of that industrial slice is reception and tour-guide work. Roughly 250 are estimated to be in active labour use. It is a good business selling laboratory equipment, priced as though it sold labour. Figure is the opposite shape, few units doing real work. BMW has moved from the Figure 02 pilot to deploying Figure 03 for production logistics, and opened a second pilot in Leipzig alongside a competence centre for the technology. A customer who extends a pilot into a successor generation and a second plant has concluded something the data supported. What changed was software rather than mechanics. Walking was solved by control engineering years ago, which is why the backflip videos are old. What was missing was useful work in an environment the machine had not seen before, and the current wave of capital bets that large models supply it while the mechanical platform becomes a commodity. Spartanburg is the first public evidence that the bet pays out on a real line. The economics remain unproven, rather than the capability. BMW has not disclosed what it pays, Figure has not broken out deployment revenue, and engineers and safety staff still surround these machines. Nobody outside those two companies can say whether a humanoid hour costs less than the alternative. A constraint that does not move sets the direction. Several developed economies are rebuilding industrial capacity into workforces that are shrinking, where the welders are in their late fifties and the replacement cohort was never born. Humanoid funding has reached roughly $8.6 billion across 113 rounds, $3.8 billion of it this year. In that setting automation is not cost-saving, it is the condition for the factory running at all, and a buyer will tolerate an expensive robot for years before tolerating an idle line. The next thing to watch is a second manufacturer running a deployment of that length and disclosing the cycle-time and accuracy data with it. One customer proves a machine works. Two prove a market.

$TGT Two Retailers, One Customer, Opposite Quarters Both stores sit in the same suburban corridor, often within a mile of each other, and much of the time the same household shops at both. On Wednesday morning they reported the same quarter and said different things about who that household is. Target’s comparable sales rose 3.8 percent against a consensus near 2.4, and the company raised full-year sales growth to about 5 percent. Lowe’s comparable sales rose 0.2 percent, and the company cut its full-year sales outlook to $92 billion, the bottom of its prior range, citing continued softness in do-it-yourself spending. The tempting explanation is that one management team is executing better, and that may be part of it. But the more useful reading is that the two companies sell different kinds of purchase to the same wallet. Groceries, household basics and clothing are bought weekly and cannot be postponed for long. A bathroom renovation can be postponed for years, and in a period when people feel uncertain, that is exactly what happens to it. Read across the other retail reads from the same week and the pattern holds. Home Depot grew sales 5.7 percent while comparable customer transactions fell 1.0 percent, meaning fewer visits with larger baskets, which is what you see when the discretionary project is deferred and the necessary repair is not. July retail sales overall fell 0.6 percent, the first decline since October 2025, and clothing rose 1.9 percent inside that decline. One qualification matters for Target specifically and it is easy to miss. The quarter included a $752 million benefit to net earnings from tariff refunds, worth about $1.65 a share. That is a real cash item and not a recurring one, so the comparable-sales line is the part of the print that says something about the customer, and the earnings line mostly says something about trade policy. What this does to the ordinary framing of a consumer slowdown is make the income cut less useful than the category cut. For two years the standard question has been whether lower-income households are pulling back while higher-income ones are fine. These two results suggest a different axis, where the same household keeps buying the things it buys every week and stops buying the things it buys every decade. That would put home improvement, furniture, appliances and vehicles on one side and groceries, apparel and consumables on the other, regardless of who the customer is. Walmart reports Thursday morning, and it is the largest single read on which of those two cuts is doing the work.

The First Time Anyone Bought Orbital Housekeeping There are roughly eleven thousand active satellites in orbit and a much larger number of dead ones. A dead satellite is not designed to be caught. It has no grapple fixture, no docking ring, and it is often tumbling, so anything retrieving it has to match the rotation of an object the size of a car spinning in vacuum at eight kilometres a second. In mid-August the Space Development Agency and the Defense Innovation Unit awarded design contracts to Firefly, D-Orbit and Katalyst worth about $8.4 million combined, to build a commercial capability to rendezvous with, capture and deorbit satellites that were never built to be captured. Preliminary design reviews are due by the end of 2026, an on-orbit demonstration follows if the designs hold, and service is targeted around 2028. The money is trivial and the studies may not reach a demonstration, since the history of on-orbit servicing is full of things that worked once and never became a business. What is not trivial is that the category exists. Disposal has generally been something a satellite did to itself, keeping enough fuel to lower its own orbit at end of life, which works when the spacecraft is healthy and fails entirely when it is not. Paying somebody else to do it is a customer accepting that a meaningful share of what goes up will not come down on its own. Read alongside the week’s other purchase, the shape is clearer. Five companies received about $12 million each to build orbital routers letting constellations from different contractors exchange data. Traffic management and rubbish collection are the two services a crowded place needs, and a customer buying both in the same fortnight is describing what it expects low orbit to become. These are recurring services rather than launches, and recurring services are paid the way a utility contractor is paid, on a schedule, for a long time, at margins that do not depend on winning the next competition. Launch has spent a decade becoming cheaper and more contested. The maintenance layer above it has almost no incumbents, few qualified suppliers, and a customer who has just established that it will pay. The companies doing this are mostly private, so the buyable expression is thin today and the first movers may not be the ones that scale. Katalyst is mid-mission attempting to reach a NASA observatory, which will say more about whether the capability works than any contract will. That mission is the near-term thing to watch, because a successful capture of an uncooperative object would turn a category with contracts into a category with a product.

The Rate That Prices a Fifteen-Year Build There is a screen on most trading desks that nobody looks at for weeks at a time. It shows the yield on the thirty-year Treasury, and its usual job is to sit there while everyone argues about the next meeting of the central bank, which is a fight about a rate that runs overnight. On Tuesday that screen read 5.32 percent. It had not read that since June 2007. The move had no monetary cause. The committee has not met, and futures still price a hold in September at roughly two chances in three. What the reporting pointed to was fiscal: a July deficit of $432.3 billion and a year-to-date shortfall near $1.8 trillion, against a Treasury that has to sell long-dated paper to somebody. Japanese and German long yields moved the same week, so the cause is not only American. A company building a data centre, a rocket, a magnet plant or a fab is not financing overnight. It finances across a decade or more, and the number that discounts a decade of cash flows is the long yield rather than the policy rate. So a week where the committee does nothing and the thirty-year moves eighty basis points is a week where the cost of every long build changed without anyone announcing it. You could see it in the tape, where memory and storage names fell 5 to 9 percent with no earnings, no guidance and no company news at all. When a group falls together on a day with nothing in it, the thing they have in common is doing the work, and here it is that their money arrives late. That is a fact about the discount rate rather than about the businesses. The transformer backlogs, the memory contracts and the interconnection queues did not move on Tuesday, and none of them cares what a bond auction cleared at. So a repricing like this is where the two halves of a long-duration position get told apart, and that is worth having. A build whose demand is contracted, dated and physically constrained keeps compounding through a higher discount rate, just at a lower valuation on the way. A build whose demand is a projection does not, and a cheaper cost of capital was doing more of the work there than anyone had to admit while rates were low. The same tape move tests both, which is how you find out which one you own. A yield at a 19-year high is also a level rather than a direction, and long rates have reversed sharply on a single auction before. Nothing here says 5.3 percent is where this settles. What would show the demand standing on its own is a stretch where the long yield keeps rising and the order books lengthen anyway. NVIDIA reports against a $91 billion guide on August 26, into exactly that condition.

$INFQ The Contract That Names a Site and a Year Last time I went looking for how much of Infleqtion’s revenue came from the quantum computer and could not find out, because there is one revenue line. That question is still open. This one is narrower: what has the company committed to deliver. Three things were signed or reaffirmed this quarter. It remains on track for 30 logical qubits in 2026, having gone from 2 to 12 over the past cycle, with 100 targeted for 2028. It is under contract to deploy a neutral-atom system at the Illinois Quantum and Microelectronics Park in 2027. And it was selected for three Department of Energy Genesis Mission projects. The Illinois contract is the one that carries weight, and the reason is the word modular. The system is architected to scale through upgrades past 50 logical qubits rather than shipping at a fixed specification. A machine sold on a promised future qubit count is a research grant with a delivery date. A machine that gains capability through upgrades to an installed base is equipment, with a maintenance relationship attached. The customer is buying a path. That is the transition every company in this field has to make, and most have not. Eaton is the one commercial name in the quarter. It is an industrial company with a real power-systems problem, taking cloud access rather than buying hardware, which is how a curious customer starts. It is not revenue of any size. Nobody says it is. Revenue for the first half grew 74 percent and gross profit grew 11 percent, so half-year gross margin went from about 31 percent to about 20. The incremental margin on new revenue is roughly 18 percent. That is what a NASA gravity-gradiometer programme looks like on an income statement, and being paid to do engineering at engineering economics is the normal way a hardware business gets someone else to fund its development. What it means is that the machine has to arrive for the economics to change, and the sensing work buys the time to build it. There is a good deal of time to buy. The quarter ended with $582 million and no debt, against a burn of about $14 million once you remove a $27.4 million payroll-tax timing benefit that reverses this quarter. Spending is climbing quickly, so that burn will not hold, but even at several times the rate the cash reaches past 2028, which is the year the thesis turns on. Management also raised the full-year outlook to about $45.1 million, though part of that increase is revenue shifted out of prior years rather than new demand. So a roadmap nobody moved, a contracted system with a named site and a year, an industrial customer poking at the platform, and enough cash to finish what it started. The engineering revenue is not the destination and is not meant to be. The 30-qubit demonstration is the next real marker, it is due before December, and it is the kind of claim that settles rather than argues.

$PLTR The Data Problem Palantir Spent Twenty Years On Turned Out to Be Everyone's The standard criticism of this company was always its customers. It sold to intelligence agencies and defence departments, the contracts were long and secretive, and a business leaning on government procurement was not supposed to grow the way software grows, because the buyer is slow and the sales cycle runs in years. Then commercial revenue nearly caught government revenue in a single quarter. Palantir reported $1.94 billion for the second quarter on August 4, up 93 percent. Government revenue grew 79 percent to $990 million. Commercial grew 110 percent to $945 million, and the American part of it grew 149 percent. A company defined for two decades by who bought from it is a quarter away from being majority commercial. What changed is not the software. The original product was built for analysts joining fragmented, badly labelled datasets in situations where being wrong carries a cost, and that work is specific to each customer in a way most enterprise vendors preferred to avoid. It does not scale cleanly, since every deployment means learning what a particular organisation’s records actually mean, which is closer to consulting than to selling licences. Then companies started pointing AI models at their own operations and found that the model was the easy part. Their data looked like the mess the agencies had: systems that disagree with each other, fields that mean different things in different departments, records nobody has reconciled in years. A vendor with twenty years inside that problem is not quickly replicated, and the quarter’s record $2.1 billion of new commercial contracts is what that looks like when the buyers arrive at once. So the criticism inverted. Government work was supposed to cap the growth rate, and instead government accelerated to 79 percent while commercial went faster still. Underneath sits a change in what organisations can assume. They used to trust their own data because faking it was expensive and the systems producing it were slow. Now the volume is enormous, the provenance is uncertain, and the party on the other end may be software, so knowing what is true inside your own operation has become something you buy rather than something you have. The same shift has agencies purchasing cryptographic migration against a 2030 deadline and card networks building identity checks for payments a machine initiated. The rate is the open question, not the direction. American commercial growth measured against a base that is already large is the number that separates a durable rate from a step change, and that comparison arrives in the first half of next year.

A Model Uses Fifty Billion Parameters and Needs All 2.8 Trillion Kimi K3 activates about fifty billion parameters to answer any given token, out of 2.8 trillion it contains. The other 2.75 trillion do nothing for that token, and every one of them still has to be sitting in memory when it arrives, because the model cannot know in advance which experts the next word will call. That is the whole economics of serving an open model, and it runs opposite to how the release was read. Moonshot published the weights last month and anybody could take a copy, and the market treated frontier capability getting cheap as a reason to sell the memory names. But a model is not paid for by the arithmetic it performs. It is paid for by the capacity it occupies while idle, and sparsity makes that ratio worse rather than better, since the active fraction shrinks while the resident requirement does not. At four-bit precision Kimi holds roughly 1.4 terabytes resident, which in practice is sixty-four accelerators bought for how much they can hold rather than how fast they can multiply. Alibaba’s Qwen3.8-Max followed eight days later, 2.4 trillion parameters at $2 per million input tokens, a price that only works on volume, and volume is what keeps those machines occupied. Capacity is also the least elastic thing in the chain. High-bandwidth memory is built on DRAM wafers already sold out past 2027, and the fabs SK hynix approved this month do not reach a cleanroom before December 2028, so the supply of the fast tier is fixed for about three years no matter what anyone pays. Which is why SanDisk and SK hynix published a specification this month for High Bandwidth Flash, a third tier built on NAND rather than DRAM, stacked and wired like the fast memory above it. It exists because the fast tier cannot be stretched. The industry’s answer to models that no longer fit is more storage placed inside the memory hierarchy, not less memory. Open weights widen this rather than relieving it. A 1.4-terabyte model does not fit on a workstation, so the copies run in company data centres and sovereign facilities, which disperses the decision to deploy while concentrating the hardware into the same racks everyone is already queuing for. The counter-argument is real and I do not want to wave it through. Four-bit weights cut the bandwidth each token consumes by roughly four times, which pushes against everything above, and a sparse model genuinely does less work per answer than a dense one. The number that settles it is bytes of high-bandwidth memory consumed per token, and nobody on either side of this trade has published it. What can be said is narrower. This layer has been repriced three times in five weeks, on the arrival of capital, a lithography tool and a model, and not once on a cancelled order.

Who bought ownership in $SPCX SpaceX listed on June 12, which means the quarter that ended on June 30 is the first for which institutions had to report holding it. Those filings are now public and this is what they say. Alphabet is the largest institutional holder by a wide margin, with 551,189,500 shares worth $94.18 billion at the June 30 price of about $170.86. It invested $900 million in 2015. The stake is 95 percent of Alphabet’s disclosed investment portfolio, and roughly $80 billion of it is under short-term sale restriction, with the remainder restricted into the third quarter of 2027. Gigafund holds 171.8 million shares. Nvidia holds 122.76 million, worth about $21 billion, making SpaceX its second largest equity position behind Intel. Nvidia did not buy the shares on the market. It put $10 billion into xAI in January, and the stake converted when SpaceX absorbed xAI on February 2 in an all-stock deal valuing the combined company at $1.25 trillion, a trillion attributed to the rockets and Starlink and $250 billion to xAI. Baillie Gifford holds 51.4 million shares. BlackRock holds 51 million. Ontario Teachers’ Pension Plan holds 50.7 million. Harvard Management Company reported 12,935,100 shares worth $2.21 billion. That is more than half of its roughly $4.3 billion disclosed portfolio and about six times its next largest position, a $350 million holding in Taiwan Semiconductor. Harvard’s reported portfolio more than doubled this quarter to its largest value since at least 2006, almost entirely because of this one line. AMD reported roughly 3.3 million shares worth over $565 million. The filing does not say when the shares were acquired or at what price. Several institutions opened positions during the quarter rather than converting existing ones. Saudi Arabia’s Public Investment Fund initiated. Citadel Advisors initiated in SpaceX and also in Cerebras Systems and Quantinuum. Assenagon Asset Management and Sands Capital both opened new positions. In total, around 474 institutions now report holding the stock. Three things limit what any of this shows. A 13F is a snapshot of what was held on June 30, filed up to forty-five days later, so it is already six weeks old when it appears. It carries no cost basis and no statement of intent, so a large position and a position about to be sold look identical. And the share price has traded both below the $135 offer and back above it since the listing, so every dollar figure here depends on the date rather than on anything a holder did. The lockups began releasing in August. The filings covering September, which arrive in November, are the first that will show whether any of these holders sold once they were able to.

$ONDS What Ondas Has to Deliver in December. A sobering take. Ondas beat on revenue this morning by about fifteen million dollars, raised full-year guidance, and fell seven percent to $9.08. I went looking for what the objection was. It is not really in the headline loss, which is mostly accounting. Of $199.1 million in operating expenses, $105.8 million never left the building. That is stock compensation from executive awards vesting, amortisation of acquired intangibles, and a mark-up on acquisition earn-outs that grows when the acquired businesses beat their plan. The number that matters is the cash one. Adjusted cash operating expense was $93.3 million, against $36.9 million in the first quarter and $9.4 million a year ago. Gross profit was $36.1 million. So the company earns thirty-six million running itself and spends ninety-three, and that roughly fifty-seven million gap is the loss. Ten-fold cost growth in a year is not an accounting artefact. You cannot hire a manufacturing workforce in the quarter you need it, so the spending comes first and the revenue follows, which management calls front-loading ahead of the ramp. That is the more realistic description of what the numbers show. Then the arithmetic it rests on. First-half revenue was $133.9 million against a full-year midpoint of $537.5 million, leaving about $403.6 million for the second half. The third quarter is guided to $140 to $155 million. Subtract it and the fourth quarter has to produce roughly $250 to $265 million, about three times what the second quarter delivered. That is the whole thing. The peak-loss claim, the December profitability date and the billion-dollar exit run rate are not three assurances. They are one assumption stated three ways. I do not want to lean on that too hard. Backlog reached $757 million from $68 million at the end of 2025, the programmes behind the ramp are mostly under contract, and defence revenue genuinely arrives in steps rather than smoothly. What makes the ramp worth watching is what it is a piece of. Cheap drones turned out to work, and everything downstream of that is still being rebuilt. An airframe costing a few thousand dollars can threaten a vehicle costing millions, and the defensive answer had to be rebuilt too, because a million-dollar interceptor missile fired at a hobbyist quadcopter loses even when it hits. That demand does not depend on any particular war continuing, which is why the order book kept growing through a summer of ceasefires and resumptions. Whether it also reflects a shrinking pool of people available to serve is a claim I hear often and cannot evidence. Militaries automate where machines are cheaper to buy and train than soldiers, and that explains what is being bought without needing the demographic argument. Ondas has an unusually specific position in it. Its interceptor brings a hostile drone down with a net and a parachute rather than by jamming or gunfire, which is the difference between a system that can be used over an airport or a city and one that cannot, and most counter-drone defeat still requires an operator in the loop where this does not. The Optimus platform sits on the cleared federal procurement list, which takes years rather than money to obtain and which no amount of funding shortens. That is why the growth is worth taking seriously rather than treating as a small company’s noise. Eighty-five percent organic against DroneShield’s 74, with that competitor guiding its own full year down to 15 to 25 percent, is a pure-play taking share in the segment the market treats as the benchmark. The primes are larger and Anduril is better funded, and neither of them is growing organically at this rate in this niche. December is when it gets answered. Revenue has to roughly triple to somewhere near $250 million while cash operating expense stays close to where it is, which is what would take the operating platform to adjusted EBITDA profitability on the timetable management has set. The programmes to watch are the defense unmanned systems orders and the combat engineering vehicles, since those carry the bulk of the quarter. None of it is in the revenue line yet.

Oh, this is one interesting use case!

$INFQ Where Infleqtion's Revenue Growth Came From Infleqtion grew revenue 116 percent last quarter and I went looking for how much of that came from the quantum computer. The company reports a single revenue line, so the answer is not in the release. Management attributed the growth to NASA’s quantum gravity gradiometer programme and to sales of third-generation Tiqker atomic clocks moving through a co-selling deal with Safran. Both of those are sensing and timing instruments. Neither is a quantum computer. The best split available is a third-party estimate of roughly two-thirds sensing, one-third computing, which is an outsider’s read rather than a disclosure. So how is the computing business actually doing? On the evidence, it is doing what a well-run pre-commercial hardware programme looks like. It has shipped a 100-qubit Sqale system to the UK’s National Quantum Computing Centre, a larger machine specified up to 500 qubits to Japan’s Institute for Molecular Science, and it has Illinois committed for 2027. Eaton took private-cloud access this quarter for energy problems. On the core metric it went from 2 logical qubits to 12, targets 30 this year, and 100 in 2028. What it is not doing is commercialising, and the company does not claim otherwise. Management’s own framing is that 100 logical qubits in 2028 is when they expect the first commercially relevant applications. Kinsella calls computing the crown jewel and describes the approach as Nvidia-like, aiming the core technology at near-term markets while building toward fault tolerance. Read plainly, that is sensing revenue funding a 2028 computing bet. Those targets are worth placing against the field. QuEra published 96 logical qubits from 448 physical atoms in January, and Quantinuum demonstrated 48 from 98. Infleqtion’s verified 12 came from roughly 1,600 atoms, a heavier ratio, largely a consequence of using a surface code where QuEra used a higher-rate one. Six-fold improvement is a real rate of change, and it starts from a position behind the leaders. The comparison is also rougher than it looks, since vendors define and verify these counts differently, and Quantinuum leads on logical fidelity while trailing QuEra on count. Here is where I got stuck. I cannot tell whether the computing third of revenue is growing faster or slower than the sensing two-thirds. Both drivers management named were sensing. A company seeing its computing revenue inflect would have reason to say so, and this release did not, though a single quarter attributed to two large sensing programmes is thin ground for reading much into that. I have stopped trying to make it into more than it is. The rest of the quarter argues for reading the loss as the price of that bet. The GAAP operating loss tripled to $30.6 million from $10.1 million, and the non-GAAP loss went from $7.3 million to $17.0 million, so revenue roughly doubled while the underlying loss more than doubled. Reported operating cash flow was positive at $13.2 million, but the company discloses that $27.4 million of it is a temporary working-capital benefit from the timing of payroll taxes on stock options. Strip it out and the quarter burned about $14 million against $582 million of cash and no debt. The stock closed up 4.6 percent before the release and fell about 7 percent after it. None of that is unusual for a company funding a hardware programme years ahead of its market. What it does mean is that the case rests on the computing side eventually arriving, and the segment split that would show how that half is tracking is not something a single-line reporter has to publish. Infleqtion has not, and 2028 is the date its own management has put on the answer.

$CRWV CoreWeave's Quarter, and the Bill That Came With It CoreWeave reported on Tuesday evening. Revenue $2,575 million, up 112 percent. Adjusted operating income $128 million against an estimate near $66 million, the first time that line has meant much. Below it, a GAAP operating loss of $49 million and a net loss of $626 million (CoreWeave). The shares rose. Those numbers describe two different companies. So let’s look beneath the highlights. Backlog was $104 billion at June 30 and $129.2 billion by the time they reported six weeks later, with contracted power going from 3.7 to 4.2 gigawatts. Twenty-five billion dollars of signed contracts arrived while the quarter was being counted. Customers commit years ahead because they cannot get capacity any other way, since interconnection queues run four to seven years in the markets they want. What the backlog does not say is when it becomes money. Conversion lags four to eight quarters, which is why the full-year revenue guide is $12.4 to $13.2 billion against $129.2 billion of contracts, and several of those counterparties are themselves losing money on the workloads they are buying capacity to run. Adjusted EBITDA was $1,510 million at a 59 percent margin. Renting the machines now covers running them and leaves something over, which is what management meant by scale translating into operating leverage. Capex of $14.1 billion over six months against $3.7 billion of operating cash flow is enormous, but that builds capacity rather than serving today’s customers. Interest was $640 million in the quarter against $536 million in the last one, which is most of the distance between $128 million of adjusted operating income and a $626 million loss. Which is odd, because the cost of debt has been falling. The company cut its weighted average rate by more than 300 basis points, and the 2025 shareholder letter puts the saving at “roughly $700 million in annualized savings based on our year-end debt balance” of $21.4 billion (CoreWeave). A bigger number now circulating applies those same 300 points to today’s $35.1 billion, which would require the whole book to have been repriced. This year’s compression was about 80 basis points. The rate fell and the bill rose anyway. Debt went from $21.4 billion in December to $35.1 billion in June, against $5.0 billion of equity. A 64 percent increase in principal swallows a three-point improvement in price. The earlier facility matched contracts running through the debt’s maturity, so nobody took renewal risk. The new one runs five years against three-year contracts, so lenders are underwriting renewal for the first time. That is a good unlock, because it lets CoreWeave sign shorter enterprise contracts that “often command higher prices.” It is also 325 basis points expensive, and I cannot tell which matters more. So the number to watch next quarter is not the backlog, which will be bigger, or the adjusted margin, which has turned. It is what the next facility costs.