One Basis Point.

The map of the week

• July payrolls fell by twenty-three thousand and May and June were cut by a hundred and three thousand between them, while unemployment printed four point one per cent. This is the week’s hinge and the Analytical Takeaway is where it is argued.
• The fence tell fired, dovish, by a single basis point. The full call, with its levels, is in the 90-Second Read above; it is scored at the top of the Takeaway.
• Everything that likes cheaper money went up at once: shares had their best week of the year, gold rose seven per cent and silver ten, while crude gave back nearly eight. The moves are in The Week That Was; what the pattern means for rates is in the Takeaway.
• Does renting an AI chip still make money? The Stack Inversion asks it and shows the sum, every assumption printed. The answer on Friday was nine cents an hour, on an asset costing thirty-five thousand dollars that will be obsolete in four years.
• Rare earths repriced as a sector, not as two shares. The specialist exchange-traded fund rose almost seventeen per cent, which is what tells you it was policy and not a story about any one company. Named, dated and taken apart in On the Radar.
• Six cars exist, and one of them carries a twenty-five-million-dollar floor next Saturday. The Trophy Asset takes apart what a decade of those prices is actually measuring.

The week’s closes are in the masthead tiles above and, in full, in the Scoreboard; this page does not repeat them.

The gauge falls ten points to fifteen, its lowest reading since the spring, because the oil shock that had been holding the energy dial red for a month finally expired.

The full working, every signal, its weight, its threshold, this week’s reading and the score it earned, is printed in Appendix A6, so you can recompute this number yourself.

Last week’s tell, scored first, and it fired. We left a two-line call on the two-year Treasury yield, the bond that tracks what the central bank is expected to do over the next couple of years. Above four point four five per cent and the market had decided inflation beats the rate cut. Below four point two and the cut was alive. On Friday it closed at four point one nine per cent. It fired on the dovish side by one basis point, a single hundredth of a percentage point, and it did so on the morning the July jobs report landed. That is the honest picture of the week: not a market drifting, but a market that sat on a fence for a fortnight and was pushed off it by one number. A week ago the same yield was four point two eight. The lesson is one this publication keeps relearning and will keep saying: a market does not change its mind gradually and then suddenly. It waits for permission.

The labour market blinked

Start with the number that moved everything. On Friday the Bureau of Labor Statistics reported that American employers cut twenty-three thousand jobs in July. Forecasters had expected a gain. Worse than the headline were the revisions: May was cut from a hundred and twenty-nine thousand to sixty-three, and June from fifty-seven thousand to twenty, a hundred and three thousand jobs that were counted once and have now been uncounted. The unemployment rate printed four point one per cent, and the participation rate, the share of adults working or looking for work, has fallen four tenths of a point since January. Wages are growing three point two per cent a year.

Here is why that matters to somebody who does not follow the jobs report. A central bank raises interest rates to cool an economy that is running too hot, and it holds them high while it is still worried. The single thing most likely to calm it is the labour market. Inflation figures are noisy and backward-looking. When employers stop hiring, the pressure on wages and prices eases with them. A negative payroll number is the clearest signal a central bank gets that the tightening has bitten. So the moment that number printed, the market’s judgement about the autumn changed, and the two-year Treasury broke through a line it had been sitting above for a fortnight.

The caution, and it is a real one. A single month of minus twenty-three thousand is inside the noise of a survey that covers a hundred and sixty million jobs, and the Bureau itself describes both the payroll change and the unemployment rate as having “changed little”. What makes this print heavier than one month is the revisions. Two consecutive downgrades of that size mean the labour market has been weaker than reported for a quarter, not that it deteriorated in July. That is a level shift, not a wobble, and it is why the bond market treated it as new information rather than as noise.

Hold, with the bias moving to a cut: the labour block has now turned decisively dovish, and only the arithmetic of the Taylor Rule is arguing the other way.

Reconciliation: the market moved hard toward a cut within hours of the jobs print, and futures pricing swung with it. We treat that pricing as sentiment to be tested, not an input to follow; our own read comes from the model set out below, which is still at Hold and moved only gently.

The rate read, governed by the model

Our internal rate model, a panel anchored on the Taylor Rule, reads Hold with the bias moving toward a cut, and its composite sits at four on a scale running from minus a hundred to plus a hundred, comfortably inside the Hold band. It is worth explaining what the Taylor Rule is, because it is doing the arguing on the other side. It is a simple formula that says where the policy rate ought to sit given how far inflation is from target and how far unemployment is from normal. On its arithmetic the rate should be about four point nine per cent against an actual three point six, a gap of one point three percentage points. In plain terms: a cut is not yet justified by the rule, because unemployment at four point one per cent is still low by historical standards and inflation is still above target. The gap widened this week rather than narrowing, and it widened for an awkward reason, that unemployment fell.

So the model is being pulled in two directions, and the discipline that keeps it honest is that it moves only when two of its three blocks agree. This week the labour block turned decisively dovish, three of its cells at their most dovish readings, and the inflation block is soft. That is two of three, and it is why the bias has moved to a cut while the level has not. Roughly a fifth of the panel is unscored this week for want of fresh data, so treat the composite as provisional rather than precise; the direction is what carries.

The competing explanation, named. We have read this week as the labour market cracking and the bond market repricing the cut. The strongest alternative is that this was a positioning event. Going into Friday, a great many investors were deliberately holding fewer long-dated bonds than usual, betting that yields would rise. A number that surprised the other way forced them to buy those bonds back in a hurry, and a rush of that kind moves a yield a long way without anybody changing their mind about the economy. On that reading the size of the move tells you how crowded the bet was, not how weak the labour market is. The tell that separates them is persistence. A short-term trading squeeze reverses within a week or two, once everyone who was caught out has finished buying; a genuine repricing holds and extends on the next data point. That is precisely why this week’s tell is set at four point three five rather than at a dramatic level: it is designed to catch the retracement if that is what this was.

Where I could be wrong (the standing self-check): First, I have leaned on the revisions to argue this is a level shift rather than a wobble, and revisions themselves get revised. If the August report restores a chunk of what was taken away, the argument weakens considerably; the tell is a two-year yield back above four point three five. Second, I have treated the fall in crude as removing the inflation risk, and it does remove some, but the tariff pass-through we took apart last month is a separate and slower engine that does not care about the oil price. If July core inflation surprises to the upside on goods rather than energy, the cut goes away with the jobs number intact, and that would be the uncomfortable outcome: a weakening labour market and no room to help it.

One number on Friday morning rearranged the week, and almost every line on the Scoreboard is a consequence of it. Shares had their best week of the year, the metals ran, and crude gave back nearly eight per cent. Below, the week sorted into the categories this publication will use every week from now on, so that a reader who follows for three months ends up with a small set of reusable ways to file the news rather than fourteen forgotten stories.

The barrel handed back the year’s inflation story. West Texas crude fell seven point seven per cent to seventy-eight dollars and eighteen cents, its worst week since the spring, on a threat that was withdrawn rather than carried out; the sequence, and why this publication does not treat it as a resolved risk, is in Geopolitical Watch. That single move is why the energy dial on the crash gauge turned green. Crude is still up twenty-four per cent this year and it is still the fourth-best line on the Scoreboard, but for the first time since June it is no longer the third: it has been passed by the Nikkei. The cost side of the inflation problem is deflating faster than the wage side.

The everything-rally was one trade wearing five costumes. The S&P 500 rose three point six per cent to a record, the Nasdaq 100 five point one, gold seven point two, silver ten, and bitcoin three point three. Assets that normally argue with one another moved together, which is the signature of a rates trade rather than a growth trade: when the market decides money will be cheaper later, everything priced off the discount rate reprices at once. Note what did not join in. High-yield credit rose one sixth of one per cent. If this were a rally about corporate health rather than about the central bank, credit would have been in front, not asleep.

The Baltic Dry Index rose thirteen per cent and is now the best-performing line of 2026, up sixty-four per cent. It is the least glamorous number the WMP tracks and one of the most informative: it measures what it costs to hire a ship to carry dry bulk, iron ore, coal, grain. Nobody charters a bulk carrier as a bet. They charter it because they have something to move. Freight rates rise either because there is more cargo or because there are fewer ships, and the fleet does not change quickly, so a move of this size in a few weeks is usually cargo. It is the fourth consecutive rising session and the index has run from two thousand seven hundred and thirty-two to three thousand and eighty-nine in a week. Set that against a labour market shedding jobs and you have the week’s genuine puzzle: the physical economy is busy while the American jobs numbers say the opposite.

Arista Networks cleared three billion dollars of quarterly revenue for the first time, on total company revenue, per its own results release of 4 August filed with the Securities and Exchange Commission. Revenue reached three point zero four billion dollars, up thirty-seven point seven per cent on the year. Adjusted earnings were a dollar and two cents against seventy-three cents a year ago, and the company raised its full-year guidance for the third time this year, to about twelve point six billion dollars. Two details matter more than the headline. Multi-year purchase commitments stand at nine point seven billion dollars, nearly triple the year-ago figure, which is customers paying to reserve future capacity. And the company told the market it has secured its memory supply for 2026 with visibility “well into 2027”. Hold on to that second one. It is a networking company telling you that the binding constraint on its business is the availability of memory chips, which is the subject of the new gauge two sections down.

Rare earths repriced as a policy asset. On Friday the White House held a mining roundtable announcing over two billion dollars in critical-minerals commitments, with representatives of both companies this publication tracks in the room. The specialist rare-earth exchange-traded fund rose sixteen point eight per cent on the week; the two individual names rose twenty-three and twenty-nine per cent. The fund is the number that matters, because a sector-wide move tells you this was about policy rather than about either company. There is a sharp edge in the detail, and it is taken apart properly in On the Radar.

The yen intervention was confirmed, and the confirmation was the event. Correcting the record on timing, because it is easy to get wrong: the coordinated purchase of yen by Japan’s Ministry of Finance and the United States Treasury took place on Friday the thirty-first of July, the previous week. What happened inside this week was the public acknowledgement. Japan’s finance minister published the statement on Monday the third of August, and the yen extended its gain by a further one point four per cent to a hundred and fifty-five and twenty, a near three-month high. It closed the week at a hundred and fifty-seven and seventy-five, which is weaker than where it started the day the two governments acted. Two treasuries signed it, and the market took four days to sell it back. Who pays for a currency intervention is rarely the currency. It is whoever borrowed in it, and the mechanism, along with what the market did to the intervention by Friday, is in Geopolitical Watch.

Two of the eight readings, the MOVE index and insider selling, are carried from earlier dates rather than freshly measured, and we flag them here rather than let you find them; every carry is asterisked in Appendix A6.

Two months ago the question about artificial intelligence was how fast it could be built. This week the question was who is actually making money from it, and a single chart answered it more brutally than any earnings call.

The profit pool is upside down. Research published this week by Apollo’s chief economist, using second-quarter data, splits the artificial-intelligence economy into layers and measures the margin at each. The upstream layer, the chips, the memory, the networking, the power, is running at roughly plus forty-one per cent margins. The layer everybody talks about, the models and the applications built on them, is running at roughly minus fifty-nine per cent. The people selling the shovels are making forty cents on the dollar; the people digging are losing sixty. That is not a criticism of the diggers, who are buying market share deliberately and can afford to. It is a statement about where, today, a pound invested in this technology actually earns a return, and it is the single most useful frame for reading any AI announcement you see this year.

Every model at the open frontier is now Chinese. The best freely-downloadable artificial-intelligence models in the world, the ones any company can run on its own machines without paying a licence, now come without exception from Chinese laboratories. This is the quiet story of the year and it is worth being precise about why it matters commercially rather than politically. A closed model is a toll booth: you pay per use, forever. An open model of similar quality removes the toll booth. If the gap between the best paid model and the best free one keeps closing, the pricing power sitting inside the application layer, already at minus fifty-nine per cent margins, gets harder to recover, not easier.

Capital spending is now measurable as a share of the entire American economy, and the speed is the story. The same Apollo research puts hyperscaler capital spending at around three per cent of American gross domestic product across 2027 to 2029, against zero point three per cent in 2019. Compare that with history and the level is unremarkable: the telecoms build-out peaked at one point two per cent and the housing boom at six point six. It is the change that is unprecedented. Going from zero point three to three per cent is a rise of two point seven points, though that is not quite like for like: one point two and six point six are the peaks those booms reached, not the distance each travelled. What is comparable is the multiple. This is roughly a tenfold increase in the share of the economy going into one kind of asset. The bubble question is usually asked as “is this too big?” The better question, and the one this data supports, is “is this too fast?” An economy can absorb a large investment. It struggles to absorb a large investment arriving all at once, because the things it needs, transformers, turbines, electricians, skilled construction labour, cannot be conjured at the same speed as the capital.

The oil price fell nearly eight per cent this week because a bombardment did not happen. Over the weekend of the first and second of August a threatened Iranian strike was announced and then fully walked back, and the market took the reversal as evidence that the Strait of Hormuz would stay open. That is the week’s dominant geopolitical fact and it is why the energy dial on the crash gauge turned green.

It is also, on the record of the last four months, the least reliable kind of good news. In June a memorandum was signed in Islamabad that reopened the Strait and granted waivers on Iranian crude; in July those waivers were revoked after renewed attacks on shipping. This publication has run the phrase “a ceasefire is not peace” since April and added “a signature is not compliance” in July. This week supplies the third: a cancellation is not a policy. An attack that is announced and then withdrawn inside forty-eight hours tells you about the decision-making, not about the intention, and the decision-making has now oscillated four times in five months. An Iranian parliamentary committee is meanwhile drafting conditions that would restrict the Strait to certain flagged vessels. The right way to hold this is not as a resolved risk but as a risk that keeps being priced and unpriced, and the practical consequence is that anything in a portfolio that depends on the oil price staying below eighty dollars is exposed to a headline, not to a trend.

Beyond the barrel: two currencies and a rulebook

The breadth guard on this section exists precisely so that it does not become a monthly bulletin about one waterway, and this week supplies a genuine second flashpoint, which is monetary rather than military.

Japan and the United States intervened together in the currency market, and that is a rarer event than an oil scare. The operation itself was the previous Friday, but its confirmation landed on Monday the third of August in a statement from Japan’s finance minister, which said in plain terms that the Ministry of Finance had purchased yen “in coordination with the U.S. Department of the Treasury”. It is the first joint yen-buying operation since 1998, and the first coordinated currency action of any kind between them since 2011. Japan had already intervened alone in April and in May with only brief effect; the difference this time is the second signature. The American Treasury Secretary went further, saying the United States supports correcting what he called the substantial undervaluation of the yen.

The mechanism deserves explaining, because a currency intervention sounds technical and is not. For years, money has been borrowed in Japan, where it is cheap, and invested wherever returns are higher, a trade that works as long as the yen does not rise. When the yen strengthens sharply, the loan is suddenly worth more than the asset bought with it, and the trade unwinds, and it unwinds by selling the asset. That is why a yen move can arrive in a European share index or a South African bond that has nothing to do with Japan. A currency intervention is like a landlord suddenly raising the rent on every tenant who borrowed to furnish the flat: the shock is felt in the furniture market. The reason to watch it now is not the level but what happened next, which is the honest first verdict on the second signature: by Friday the yen had given back the whole of the intervention gain and a little more. Two governments said they would not hesitate to do it again, and the market sold into it inside four sessions. Either the market is calling a bluff that will be answered, in which case the next operation is larger, or coordinated intervention is worth less than it used to be against the sheer size of the carry trade. Both readings are live, and the level, a hundred and fifty-seven and seventy-five at Friday’s close, is the scoreboard on which one is right.

And the rulebook is tightening in the other direction. Reporting this week records both SpaceX and Tesla writing China exclusion into supplier contracts and legal entity structures, which is the corporate sector doing quietly and contractually what governments have been doing loudly and legislatively. It is a small item and a large signal: the cost of supply-chain separation is moving from a tariff line, where it is visible and arguable, into contract terms, where it is neither.

In late 2022 a group of investors, Temasek, Breakthrough Energy Ventures, GenZero and a venture builder called 100x100, went out into the rice paddies of Vietnam and Indonesia and asked farmers what would actually change their behaviour. The answer came back in three parts, and none of them was the climate: better yield, a higher price for the crop, cheaper inputs. The investors wrote a plan around those three answers, incorporated a company in Singapore, and only then went looking for someone to run it. They found Dhruv Sawhney, and he arrived about six months after the company he now leads had already been designed.

That sequence is unusual enough to be the reason this profile exists. Almost every company case study is the story of a founder with an obsession who persuaded capital to follow. Rize is the inverse: capital with a thesis that went looking for a founder. It is a structure with a name in the industry, the venture build, and it is becoming more common precisely because the problems that attract climate capital are ones nobody happens to be personally obsessed by.

The problem, stated sharply

Rice is grown in standing water, and standing water is where methane comes from. Flooded paddies are one of the largest single sources of methane from human activity, roughly a tenth of the global total depending on whose inventory you use. Methane matters disproportionately because it traps far more heat than carbon dioxide over a short horizon and then breaks down, which makes reducing it one of the few climate levers that works within a career rather than within a century.

And there is a technique that works, which is the whole reason this is interesting. Let the field dry out for a few days between floodings instead of keeping it permanently submerged, and the methane falls by thirty to seventy per cent with no loss of yield, on the International Rice Research Institute’s numbers; a 2023 synthesis in Nature Reviews Earth & Environment puts the average at fifty-three per cent. This is not speculative technology. It is a change in when you open a sluice gate, and it has been understood for twenty years. So the problem was never the science. The problem is that seventeen thousand smallholders, each farming a couple of hectares, will not change a two-thousand-year-old practice because a spreadsheet in Singapore says the planet needs it.

That is the gap Rize was designed to sit in, and it is a distribution problem dressed as a climate problem.

The uncertainty, and it is structural rather than personal

The honest strategic uncertainty here is not a night before a pivot. It is whether a company assembled by its investors can develop the thing that makes founder-led companies work, which is somebody who will not let it fail. Sawhney has the operating history: he built a food supply-chain business called WOTU in the mid-2010s that was acquired by Zomato in 2018 and relaunched as Hyperpure, and he co-founded the agricultural platform nurture.farm in 2020. He has said on the record what took him there. At Hyperpure, he told an interviewer in December 2022, “I realised the growers who cultivate, nurture, and produce the food we eat are short-changed at every supply chain step.” That is a real conviction with a real origin, and it happens to fit the mandate he was hired into. Whether hired conviction behaves like founded conviction under strain is the question the next three years answer, and it is unanswerable today. What can be said is that no published account records a personal stake at Rize of the kind this section usually looks for, and we are not going to invent one.

Now the size of it, which is where honesty is expensive

Rize raised thirty-one million dollars in July 2026, twenty million of equity led by BNP Paribas Asset Management’s alternatives arm with the Rockefeller Foundation, Temasek and Breakthrough Energy Ventures, and eleven million of debt from UOB, BIDV and Temasek Foundation, taking the total raised to forty-seven million. It says it works with seventeen thousand farmers across more than fifty thousand hectares and has shipped about fifteen hundred tonnes of low-emission rice to Europe, Canada, Australia and Singapore.

Set those against the world. Global rice paddies cover roughly a hundred and seventy million hectares. Fifty thousand hectares is three hundredths of one per cent of that. World rice production runs to about five hundred and forty million tonnes a year; fifteen hundred tonnes is three ten-thousandths of one per cent. A closer and more useful comparison: Vietnam’s own government programme for low-emission rice had passed three hundred and fifty-four thousand hectares by the end of 2025, seven times Rize’s footprint across both its countries, and plans another eight hundred thousand by 2030.

This publication has a rule that a company earns the marquee either by being materially significant in its industry or by scaling fast enough that it will be. Rize fails the first test by three orders of magnitude and its case rests entirely on the second. It says it has grown tenfold in two years. We can partly check that: in May 2024 it told Bloomberg it serviced about two thousand five hundred hectares, and fifty thousand is twenty times that, not ten, which means the tenfold figure is measured on something else, farmers or revenue, and the company has not said which. In the same 2024 interview its stated abatement goal was a hundred million tonnes of carbon dioxide equivalent; the July 2026 release says five hundred million. A goal raised fivefold in twenty-six months, with no explanation offered, is the sort of thing that should be noticed out loud.

The transferable lesson, and it is the reason to read this rather than a press release: the mechanism can matter enormously while the company matters hardly at all, and the two questions must be kept separate. Alternate wetting and drying, applied across the world’s paddies, would be one of the largest single climate interventions available to humanity. Rize is currently a rounding error inside it. Both statements are true, and the investment question is only ever the second one.

Arista told the market this week that it has locked up its memory supply into 2027, which is what a company does when it believes a shortage is real and lasting; this gauge exists to tell you, in a number, how close that belief is to expiring.

What this gauge is, and why it has its own section

Start with a bridge. If you own the only bridge across a river, you can charge whatever the traffic will bear, and your bridge is worth a fortune. The bridge is not valuable because it is well built. It is valuable because there is no second bridge. The day a second bridge opens, your toll collapses, and nothing about your bridge has changed.

The entire artificial-intelligence trade is a bet on bridges. Every layer of what the industry calls the compute stack, the memory chips, the logic chips, the machines that print them, the power to run them, the cooling, the networking, the models themselves, earns a premium only for as long as it is a bottleneck. The premium starts falling the day supply arrives. That is the whole idea, and this gauge is a count of how many second bridges are under construction, and how close any of them is to opening.

It is deliberately paired with the Displacer and the Augmenter, which sits two sections below. That framework scores the demand side of artificial intelligence: does it replace workers or make them more productive? This one scores the supply side: is the shortage that everything is priced on real, and for how much longer? Two sides of one trade. Three things it is not, and the first is the important one. It is a margin and scarcity gauge, not a value gauge. It can tell you that the toll bridge is losing its monopoly; it cannot tell you whether what crosses the bridge is worth anything, and a high reading is not an instruction to sell. It is not a view on whether artificial intelligence works. And it is not permanent: it is a gauge of the build-out phase, and when the build-out ends it will be retired rather than quietly repurposed.

The reading runs from nought to a hundred, and it runs the opposite way from instinct. Higher is worse for anyone who owns the scarcity. The four bands are: 0-30, scarcity intact and the tolls are safe · 30-55, first cracks, meaning new supply announced and funded but not arrived · 55-75, inversion underway, showing in prices rather than only in share prices · above 75, scarcity broken, and the whole complex reprices from a monopoly multiple to a commodity one.

The eight dials, in plain English

• Memory capacity, the heaviest weight. Is new memory-chip factory capacity being announced, funded, or actually shipping? Announcements are cheap. Shipping is what ends a shortage.
• Home-built lithography. Lithography machines print the circuits onto silicon. China is building its own; the question is whether they work at volume and at acceptable quality, or whether there are five of them in a shed.
• The EUV (extreme ultraviolet lithography) chokepoint. One Dutch company makes the machines needed for the most advanced chips. Nobody else is close. This is the deepest moat in the stack, and the only dial that can take this gauge above fifty-five.
• Memory pricing, and this is the one that decides everything. Not capacity announcements, not share prices, the actual contract price of memory, both the ordinary working memory called DRAM and the high-bandwidth memory, HBM, stacked beside an AI chip. More on why below.
• Leading-edge foundry access. Can Chinese foundries make advanced chips in volume, or only in samples?
• Model commoditisation. Are the free, downloadable models catching the paid ones? If they do, the toll booth at the model layer disappears.
• Power as the new bottleneck. When chips stop being the constraint, electricity becomes it, and the premium migrates from the silicon to the substation.
• Positioning. How crowded and how levered is the trade that owns all of this? It does not change the physics; it changes how violent the adjustment is.

Why memory pricing carries the most weight, and the sentence to remember: watch chip prices, not share prices. Record earnings in a shortage are the most dangerous kind of earnings, because the cure for high prices is high prices. A firm earning extraordinary margins on a scarce component is, by earning them, funding every competitor’s decision to build more of it. The share price tells you what people believe about the shortage. The contract price tells you whether it still exists. This week the shares in the complex rose with everything else, and the memory price did not move.

The authority stamp is worth exactly one sentence. When the Federal Reserve held rates at the end of July, the chair named the prices of logic and memory chips as an inflation signal he was watching. The load-bearing dial on this gauge is now also being watched by the person who sets the interest rate.

The compute profit and loss, and the working shown

A gauge that only scores dials is an opinion. So here is the arithmetic that turns it into a sum a reader can check. The question is simple: does renting out the workhorse artificial-intelligence chip, an H100, currently make money or lose it?

The revenue line. On Friday, renting an H100 by the hour on the open marketplace, where independent operators compete for work, cost one dollar seventy-three per hour at the interruptible rate, taken as the daily median across fifty-five live offers. The guaranteed, uninterruptible rate on the same marketplace was two dollars and six cents. We use the interruptible number because it is the price at the margin, the one that moves first.

The cost line, every assumption stated so you can disagree with any of them. The table below is the whole sum: the rental at the top is the revenue, the four middle rows are the cost, and the last line is what is left.

One dollar eighteen plus ten cents plus fifteen cents plus twenty-one cents is one dollar sixty-four per hour. Against Friday’s rent of one dollar seventy-three, that leaves nine cents an hour, about five per cent of revenue. For context, that is a worse margin than a supermarket earns on groceries, on an asset that cost thirty-five thousand dollars and will be obsolete in four years.

Change any assumption and the answer moves, which is the point of printing them. Cheaper power at six cents a kilowatt hour, which a large operator with its own generation can achieve, saves four cents. A five-year write-off instead of four saves about twenty-four cents and turns a five per cent margin into a twenty per cent one, which is exactly why the depreciation schedule is the most contested number in the entire artificial-intelligence build-out. Financing at fifteen per cent instead of nine, which is closer to where a small levered operator borrows, costs fourteen cents and wipes the margin out entirely.

What the four-week series shows. On the last four Fridays the open-market rent has read one dollar thirty-three, one dollar thirty-three, one dollar eighty-seven and one dollar seventy-three. It has spent the month oscillating across the line rather than settling on either side of it. That is the honest reading and it is more interesting than a trend: the marginal H100 is currently earning roughly its cost of capital and no more, which is what the end of a shortage looks like at the very beginning, before it reaches the contract prices or the earnings statements. A sustained month below the line would be the inversion arriving in cash.

This gauge tells you which bottleneck is dissolving, not that a crash is coming, and it can sit at “first cracks” for a long time. What would re-tighten the scarcity and drop the score: a delay to the announced capacity, a fresh wave of demand that empties inventories, or China’s machine programme slipping a year on quality. The full eight-signal working is in Appendix A11, and the price register in Appendix A12, so you can recompute both numbers yourself.

On Friday, Gooding Christie’s published the catalogue for its Pebble Beach sale, and lot thirty-nine carries an estimate written the way auction houses write them when they are certain: in excess of twenty-five million dollars. Not a range. A floor.

The car is a 1964 Shelby Cobra Daytona Coupe, chassis CSX2300. Six were built. This is the one Carroll Shelby kept for himself, from 1975 until 1998. It sells on the fifteenth of August, and if it makes its number it will be the most expensive American car ever sold at auction.

The record it is aiming at is twenty-two million dollars, set by a 1935 Duesenberg SSJ at Gooding’s Pebble Beach sale in 2018. The record before that was thirteen and three-quarter million, for the first Cobra ever built, CSX2000, sold by RM Sotheby’s at Monterey in 2016. Two houses, two weekends in August, ten years apart: thirteen point seven five, then twenty-two, and now an ask above twenty-five. Note what that third number is. It is not a price. It is what a house that has already spoken to the bidders is willing to put in print eight days beforehand, which is a different kind of information and, on the record of the last decade, a fairly reliable one. On the two prices that actually happened, two years apart, the top of this market compounded at about twenty-six per cent a year. If the ask is met next Saturday, the ten-year number is a little over six per cent. Those are two very different investments, in an asset class that produces no cash flow, costs money to store and cannot be valued by any conventional method, and which one it turns out to be will be known on the fifteenth.

Why this section exists, and what these three numbers are actually measuring

The argument this section makes, every time it runs, is the same one: a fixed or artificially constrained supply, meeting a growing pool of capital that has run out of places to find genuine scarcity. Six Daytona Coupes exist. Six will always exist. No amount of demand can call a seventh into being, which is the opposite of almost every other asset a portfolio can hold, where high prices summon supply.

But the interesting part is not the scarcity, which has been constant since 1964. It is what changed on the demand side. The reason these prices moved in the last decade is not that more people came to want old cars. It is that the market around them became legible: provenance databases, standardised condition reporting, published auction results, and a small number of houses whose word on authenticity is accepted. That is the pattern this section watches for in every category it covers. Grading arrives, opacity ends, and the moment a market becomes verifiable, institutional money can enter it, because an institution cannot buy something it cannot mark. The price step follows the standard, not the enthusiasm.

The honest caveat, because six per cent a year is less impressive than a twenty-five-million-dollar headline sounds. Over the same decade a plain index of global shares did at least as well, with dividends on top, daily liquidity, no insurance premium and no storage. A trophy asset is not a superior investment; it is a differently-correlated one that happens to be enjoyable to own. Anyone buying one as a portfolio decision rather than as a preference is telling themselves a story. What the number genuinely tells you is something else, and it is macro rather than automotive: when the price of unique, unproductive, unreproducible objects rises steadily for a decade, it is usually a statement about the quantity of capital looking for a home rather than about the objects.

The same mechanism, in a stadium

One other item from the week belongs here for the contrast. On Thursday, Arsenal renewed its naming-rights, shirt and training-kit partnership with Emirates through to 2033, pulling forward a deal that had been due to run to 2028. The existing arrangement is worth about forty-five million pounds a year. The new terms have not been disclosed, so no figure should be attached to it, and this publication will not attach one.

What is disclosable is the behaviour, and it is the more informative half. A sponsor that renews five years early, against no competitive pressure it has told anyone about, is buying certainty at a price rather than testing the market for the best one. That is what a party does when it expects the asset to become more expensive, not less. Sixty thousand seats. One name on the building. The supply is as fixed as the six Cobras, and the buyer moved early.

Here is a number almost nobody is talking about, and it has been the best call of 2026 by a distance. The iShares Russell 1000 Value fund has returned twenty-two point two per cent this year, measured on price from the first trading session of 2026. Its growth twin has returned five point eight. The gap is sixteen point four percentage points, and for most of the period since the financial crisis that sentence would have read the other way round.

Some definitions, because the labels do real work. “Value” means shares that are cheap relative to what the business earns or owns: banks, insurers, energy companies, industrials, utilities. “Growth” means shares priced for what the business will earn later: software, semiconductors, the large technology platforms. For most of the period since the financial crisis, growth won so consistently that value investing became a slightly embarrassing thing to admit to at dinner. The professional consensus arrived at a settled explanation, which was that cheap companies are cheap for a reason and technology had permanently changed the economics of a business.

The consensus may still be right about the long run. It has been wrong about 2026 by sixteen points.

The mechanism, which is duller than the story

The reason is not that markets rediscovered virtue. It is arithmetic, and it runs through the interest rate. A growth company’s value sits mostly in profits it has not made yet, which have to be discounted back to today; a value company’s sits mostly in profits it is making now. When the discount rate rises, the distant profits lose more of their present value than the near ones, mechanically, without anybody changing their opinion about either business. Rates have stayed high all year. That is most of your sixteen points, and it has nothing to do with whether the technology works.

Which is exactly why this week matters and why the section runs now. The two-year yield fell through our dovish line on Friday, and in the same week the gap narrowed by two point six points, from nineteen point zero to sixteen point four, its sharpest weekly compression of 2026. Growth did the work: the value index gained a little and the growth index gained a lot, which is what a falling discount rate does to a company whose profits are mostly in the future. If the market has genuinely repriced toward a rate cut, then the single largest engine behind value’s outperformance has just gone into reverse. The contrarian position at the start of the year was to own the boring things. The contrarian position now is to notice that the reason for owning them has just weakened, in one week, and that nobody has noticed because the number is still a large positive.

The counterweight, and it is real. The dispersion between asset classes this year, from freight at plus sixty-four to Ethereum at minus thirty-six, is a hundred points wide, and inside equities the value-growth spread is a second, independent axis of the same repricing. One weak week for value inside a sixteen-point lead is not a turn; four would be, and we will count them here. And there is a version of this where value keeps winning for a different reason: if the labour market really is cracking, the companies that earn money today are worth more than the ones promising to earn it in 2030, cut or no cut.

Each week the WMP scores the central economic argument about artificial intelligence as a contest between two forces. The Displacer is the case that AI substitutes for human labour, concentrating the gains in capital and eventually destroying the spending power the economy runs on. The Augmenter is the case that AI raises human productivity, expands output, and spreads the gains broadly. Both sides agree AI is powerful; they disagree about whether it is a demand shock or a supply shock. It is the demand-side twin of the Stack Inversion gauge above, which scores the supply.

A draw this week, and it is worth saying why, because the first version of this score was three-one to the Augmenter. Two of the four pillars had been awarded off the same observation, the profit-pool split reported in The Speed of Now, pointing in opposite directions on each. One week’s fact cannot be evidence twice. Re-scored so that every pillar rests on something different, the week is level, which is the honest reading of a week in which employment fell and nobody could name a machine that caused it.

• Pillar 1, adoption speed, the Displacer. The evidence is the speed of the build rather than its size. Capital spending on this technology is heading for about three per cent of American output against zero point three per cent in 2019, and that swing of two point seven points is larger than either the telecoms build-out or the housing boom managed. An S-curve has friction in it by definition, and a two-point-seven-point swing in six years is not what friction looks like. The Augmenter’s standing counter, that regulation and organisational resistance slow real deployment, produced nothing new this week.
• Pillar 2, labour, the Augmenter, and this is the week it was properly tested. Payrolls went negative. If the Displacer’s thesis were arriving, this is what its arrival would look like. But the declines the Bureau of Labor Statistics reports are in local government education and in retail trade, with health care still growing, and no named layoff wave anywhere cited artificial intelligence as the cause. The rule this framework holds itself to is that the Displacer scores on displacement, never on inference; a weak labour market is not evidence of a machine taking a job unless somebody names the machine. And there is a positive fact rather than only an absence: in the same week, factory employment expanded for the first time in thirty-three months, in the part of the economy that has been automating longest.
• Pillar 3, type of shock, the Displacer. Its clearest point in a month, and it rests on the layer margins reported above. Income is accruing to the owners of the scarce inputs, and the layer that is supposed to eventually pay for all of it is not yet paying for any of it. That is the demand-destruction risk in its earliest form: capital income rising, the return still hypothetical. Note that this is a different fact from the one that took pillar one, and it has to be, because a single week's observation cannot be evidence twice.
• Pillar 4, compute cost, the Augmenter, and the argument is about volume rather than price. An H100 rents for a dollar seventy-three an hour, which undercuts almost any wage on earth, and a thin margin for the owner makes it cheaper still, not dearer, so the price argument runs the Displacer’s way and should be conceded. The Augmenter’s point is that substitution at the scale the Displacer’s thesis needs requires the fleet to multiply, and the fleet is rationed by memory and by power rather than by willingness to pay. Arista is locking up memory into 2027. That is the ceiling: not that compute costs too much, but that there is not yet enough of it, and the constraint is physical rather than financial. Which is why the memory price is the number that would flip this pillar, and why it is the heaviest dial on the gauge two sections above.

What would flip the score. A named layoff announcement citing artificial intelligence as the cause would hand the Displacer the labour pillar immediately, and after this week’s payroll print that claim would land far harder than it would have a month ago. A fall in memory contract prices that dropped the cost of compute decisively below the cost of a worker would hand it the compute pillar, and that would be a clean sweep. Going the other way, a downstream margin that turned positive without a price rise would hand the Augmenter pillar three, and any hard evidence of enterprises slowing their deployment plans, rather than merely losing money on them, would take back pillar one.

The ERDR (Equity Return for Debt Risk) framework tracks twelve income strategies that aim to earn an equity-like return for taking on debt-like risk. The twelve-strategy Standing Dashboard runs every week with each strategy’s current yield, its spread over investment-grade credit, and any change in thesis; a Deep Dive on a single strategy joins it in alternate weeks.

This week the macro cross-current reversed. For two months the story here has been that rising long-term rates were lifting the yield available across the income complex. On Friday the whole curve fell, the two-year by nine basis points and the ten-year by ten on the week, and if that repricing holds, the arithmetic of this table changes in two opposite directions at once. Existing holdings gain in price. New money is offered less. The strategies with floating coupons, the loan and private-credit rows, lose income immediately if the central bank cuts; the fixed-coupon rows, preferred shares and municipal bonds, keep their coupon and gain in price. That divergence is the single thing worth watching in this table over the next quarter, and it is why the two Add ratings this week are both on the fixed side.

A rate cut is not one event for an income portfolio; it is a gain for anything with a fixed coupon and a pay cut for anything that floats.

On the freshness of these twelve numbers, plainly. They move with the rate curve, which is this week’s, and with credit spreads, which are Thursday’s. The strategy assessments behind them, the risk ratings and the vehicle-level yields, come from our own register, and most of those entries were last reviewed in the spring. So read this as a set of indicative levels that are correctly ordered and correctly signed, rather than as twelve freshly-struck quotes. Where a row is load-bearing this week, the Lombard row, its arithmetic is printed beside it so you can check it rather than take it. The register is being rebuilt in full before the next Deep Dive, and this note comes out when the twelve rows are individually re-sourced.

Each strategy is explained in full when it is the week’s Deep Dive. Yields are indicative and quoted at the level an investor would actually hold the strategy, which for the two levered rows, the Lombard row and the mortgage REITs, is the levered level. Every spread in this table is struck against one number and the same number, so the column can be compared down the page and checked with a subtraction: investment-grade corporate credit at 5.43 per cent, which is this week’s ten-year Treasury at 4.65 plus the investment-grade index spread of 78 basis points. Subtract 5.43 from any yield above and you get its spread. Two rows need a word beside them. The municipal row is tax-free, so its 85 basis points understates it for a US taxpayer, for whom the tax-equivalent yield is nearer 9 per cent. And the two levered rows, the Lombard row and the mortgage REITs, are quoted at the levered level, so their spreads are the return on borrowed money and carry the risk that goes with it.

Three of the five open calls had a genuine, dated, company-specific event in the past seven days, which is the test a company has to pass to appear here rather than in the tracking appendix. The other two sit in Portfolio Watch with the Scoreboard. Nothing new is added this week: the candidate that cleared the Thursday screen has not been through the analysis this publication requires before a name is written down, and a call made to fill a slot is not a call.

These are analytical frameworks for watching companies, not recommendations. Every entry carries a pre-registered condition that would prove it wrong and a fixed date on which it is scored in public, whatever the answer.

Arista Networks · NYSE: ANET · entered week 20 at $147.00

The catalyst, dated: second-quarter results on 4 August, filed with the Securities and Exchange Commission. Revenue of three point zero four billion dollars, up thirty-seven point seven per cent on the year and the company’s first three-billion-dollar quarter; adjusted earnings of a dollar and two cents against seventy-three cents a year ago; full-year guidance raised for the third time this year to about twelve point six billion. Multi-year purchase commitments reached nine point seven billion dollars, nearly triple a year ago.

What the market is still missing. The thesis when we entered was that networking is the quieter, higher-margin chokepoint of the artificial-intelligence build-out: everyone models the chips, almost nobody models the switches that connect them. That has now largely been paid for, with the shares up twenty-eight per cent from entry against a technology index that has gone nowhere over the same period. What has not been paid for is the sentence buried in the earnings call: the company has secured its memory supply for 2026 with visibility “well into 2027”. Read against the Stack Inversion section above, that is a networking company telling you its binding constraint is the same memory shortage the whole trade is priced on. If that shortage ends, Arista’s costs fall. It is one of the few names in the complex for which the end of the scarcity is good news rather than bad, and it is not priced that way.

What would change the thesis: the pre-registered condition, set at entry and unaltered, is that the thesis is wrong if by 28 August Arista has cut its artificial-intelligence or Ethernet revenue guidance or gross margin has compressed below about sixty per cent, and the shares are below the $147.00 entry. Neither has happened; guidance went the other way. Scored publicly on 28 August, at its thesis horizon, whatever the answer.

MP Materials · NYSE: MP · entered week 17 at $67.21

The catalyst, dated: second-quarter results on 6 August. Revenue of a hundred and eight and a half million dollars, up eighty-nine per cent on the year. Adjusted earnings before interest, tax, depreciation and amortisation, a rough measure of what the core business earns before financing and accounting choices, turned positive at twenty-eight and a half million against a loss of twelve and a half a year ago. That is a swing of forty-one million dollars in twelve months. Adjusted earnings per share came in at minus one cent. Alongside it, a long-term supply agreement for separated gadolinium with a new American aerospace customer.

The honest arithmetic, and the part that is being missed. Look inside the segments, because the headline hides a division. The materials business tripled: neodymium-praseodymium oxide and metal revenue rose two hundred and seventy-seven per cent. The magnetics business, which is the entire strategic argument for this company existing, fell seventeen per cent on the year to sixteen and a half million dollars. Selling more oxide at a higher price is a commodity miner having a good quarter. Selling more magnets is a strategic asset re-rating. This quarter delivered the first and not the second, and that distinction is the whole thesis.

And the sharp edge in Friday’s policy news. The White House mining roundtable announced over two billion dollars of critical-minerals commitments and the sector rose seventeen per cent on it. But the single largest rare-earth-specific cheque in the package, a hundred and fifty million dollars, went to a Minnesota company building permanent magnets that contain no rare earths at all. The same policy impulse that is subsidising this company’s supply chain is simultaneously funding the technology designed to make it unnecessary. That is not a reason to abandon the thesis, because substitution technologies take a decade. It is a reason to be precise about what the government is actually buying, which is independence from China rather than a future for any particular mine.

The verdict, scored here rather than deferred

The thesis horizon is 9 August, which falls on a Sunday, so the deciding close is Friday the seventh, which this edition already reports. The condition written down in week 17 and never altered: the thesis is wrong if, at that date, the shares are below the $67.21 entry with no progress on contract reclassification or a strategic-utility re-rating, with the market still pricing it as a commodity miner. Both limbs are satisfied on this page. The shares closed at $51.11, twenty-four per cent below entry. And the magnetics revenue that was the entire strategic-utility argument fell seventeen per cent while the oxide business tripled, which is the market being right to price it as a miner.

So the call failed on the condition we pre-committed to. There is a second number and it does not rescue it: measured against the rare-earth basket we chose as its benchmark at entry, the position is three point six percentage points ahead, because the sector fell further than the company did. That is worth exactly what it says. The thesis was that this company would stop being priced as a miner. It was not right about that, and it was not wrong about rare earths. We score the thing we wrote down. The formal close is logged at its date and the full ledger entry, with both horizons, runs next week.

The one thing that will not happen is a quiet extension: the date was fixed at entry and a published goalpost does not move.

USA Rare Earth · NASDAQ: USAR · entered week 18 at $21.00

The catalyst, dated: the two-step merger with Texas Mineral Resources completed on 7 August, filed with the Securities and Exchange Commission, and a separate announcement on 6 August that second-quarter results will be released after the close on 10 August. A shareholder vote on a further proposed merger is set for 28 August.

What the market is missing. This one is simpler than it looks, and the interesting number is the one that is not the share price. The stock rose twenty-nine per cent this week; the specialist rare-earth exchange-traded fund rose sixteen point eight. So roughly three-fifths of the move was the sector and two-fifths was the company. That ratio is the honest way to read a week like this and it is the reason this publication scores every call against a benchmark rather than against zero. Against that benchmark the position is twelve percentage points ahead since entry, while being down eight per cent in absolute terms, which is a genuinely uncomfortable sentence and the correct one.

What would change the thesis: the pre-registered condition is that the thesis is wrong if, at 16 August, a competition ruling blocks the Serra Verde acquisition or the Defense Department purchase floor or Energy Department grant lapses, and the shares are below the $21.00 entry. Scored on 16 August. Monday’s results are the last material input before that date.

Why no new call this week

One listed candidate cleared the screen and has not yet cleared the analysis, and from this month every new call commits to a horizon of at least twelve months, because a shorter window judges the weather rather than the climate. A twelve-month commitment made in a hurry is worse than no commitment. It waits.

Spare a thought this week for the economic forecaster, who was asked what would happen to American employment in July, said “up eighty thousand”, and watched it come in at minus twenty-three. And then spare a rather larger thought for the fact that the market rose sharply on the news. There is a version of finance, taught in books and believed by outsiders, in which the economy does well and the shares go up. We have not lived in it for some time. This week a hundred and three thousand jobs were quietly removed from the record for May and June, and the immediate consequence was that gold, shares, bonds and bitcoin all rallied together, because everyone concluded that a central bank would now be kinder. It is a strange machine. It is at least an honest one, in that it tells you exactly what it is pricing, provided you are willing to hear that what it is pricing is the price of money and not the state of the country.

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

Freight has quietly become the year’s best asset at plus sixty-four per cent while crude, which led for four months, has slipped to fourth; the gap between best and worst is a hundred points wide, and no single shock is doing the widening any more.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 7 August 2026. The basket is a fixed set, chosen on 1 January and unchangeable during the year: twelve equity indices, four bond and credit funds, six commodities, two currencies and two cryptocurrencies. Four rows are named by their ticker: MSCI ACWI (All Country World Index) is the broadest global share index; AGG is the US aggregate bond market; LQD is investment-grade corporate bonds; HYG is high-yield, the riskier corporate borrowers; and TLT is long-dated US government bonds. Nothing else is added, dropped or substituted mid-year, which is the only way a year-to-date table means anything. The single-company calls in Portfolio Watch are tracked separately. Annualised volatility shows how much each line typically swings in a year, judged from its last eight weeks: higher means bumpier, not worse, and a dash means we do not yet have enough weeks to measure it honestly.

Every close here is drawn from the same price record the table itself reads from, so the numbers on this page and the numbers in our record cannot drift apart. All closes are Friday 7 August 2026. MSCI EM is derived from the EEM ETF close (65.64) multiplied by the locked index ratio (28.367); Baltic Dry is the Baltic Exchange BDI as published by Hellenic Shipping News (7 Aug, 3,089). Every year-to-date figure is recomputed from the locked 1 January baselines rather than carried forward, so an error cannot compound week to week.

Every company that has appeared in On the Radar remains tracked here until its thesis horizon. From this month every new call carries a horizon of at least twelve months, because a shorter window judges the weather rather than the climate; the calls entered in weeks 17, 18 and 20 keep the shorter horizons they were given at entry, because a published date does not move. A company moves from On the Radar to this table when there is no new catalyst that week: the analytical call is intact, but there is nothing fresh to add. The three companies with a dated event this week are written up in full in On the Radar above. Every return is measured against a benchmark chosen and fixed at entry, because a call that beats nothing you could have held instead is not a win.

Running record, and the denominator named so you can reconcile it: four calls are open, tracked since first mention in weeks 18, 20, 23 and 25, and all four are currently ahead of the benchmark chosen for them at entry, by between one and thirty percentage points. A fifth, MP Materials from week 17, closes in this edition and is scored above, which is why it has left this list. Three of the four are nonetheless down in absolute terms, which is what beating a falling sector looks like and is stated here rather than left for a reader to work out. Five calls have now closed at their thesis horizons since week 13, and the closed book is the one that should be read hardest: two beat their benchmark, three did not, and MP Materials, scored above, failed the condition it was given at entry. The full ledger, both horizons and every loss, is published in the Quarterly Reckoning.