01
The Magazine · 3 min read
Executive Summary
Good inflation news, and long money got dearer anyway. The reason was published the same morning.
+
Executive Summary
Good inflation news, and long money got dearer anyway. The reason was published the same morning.
The one thing. On Wednesday the American inflation figures came in almost exactly as everybody had hoped. Consumer prices rose one tenth of one per cent in July. The core measure, which strips out food and fuel, rose two tenths, and stands two and a half per cent above a year ago. Producer prices, published the next day, did not rise at all. Wage growth was the softest of the year. This is the news the bond market has spent eighteen months asking for. And the interest rate on thirty-year American government debt went up, from 5.19 per cent to 5.25 per cent.
There were several stories this week and that is the one that matters, because it is the one that reaches you. Long-dated government yields set the floor under fixed-rate mortgages, corporate loans and the financing of every data centre being built. When they rise on good inflation news, something other than inflation is setting the price. That something was published the same morning, in the Treasury’s monthly account of what it took in and what it spent.
What you can safely ignore this week. The records. The S&P 500 set one on Thursday at 7,798.99, a single point short of 7,800, before easing to 7,785.76 on Friday, and the Russell 2000 and the Dow both touched highs of their own. The coverage will tell you this was a week of exceptional strength. It was a week of falling short-term rates, which is a different thing and a more fragile one. The one piece below genuinely worth your time if you read nothing else is the Stack Inversion. There is a shortage running through the computer industry that has already reached the shop floor, and no tariff has anything to do with it.
The call I am putting my name to. The thirty-year Treasury must hold above 5.00 per cent through Friday 21 August. The Weekly Tell sets out why, and what it would mean if it breaks. We score it next Saturday, in public.
Everything below is the working. The Analytical Takeaway carries the rate argument; the appendix carries the arithmetic, so you can check the number yourself.
The week resolved into a single awkward pairing. The inflation problem receded on every measure that was published, and the price of long-dated government borrowing rose anyway. Short rates fell, long rates did not follow, and the gap between the two is now the widest it has been this year.
Underneath that sat a Treasury statement almost nobody read: a $432bn deficit in one month, net interest costing more than national defence, and a customs duty line that came in negative because tariff refunds ordered by the courts now exceed tariff collections.
Elsewhere: the Los Angeles Lakers changed hands at a valuation twenty-five per cent above the record the same franchise set fourteen months ago; a memory-chip shortage reached American consumer prices; and the compute margin on the workhorse artificial-intelligence chip went below its cost line for the first Friday since we started measuring it.
02
The Magazine · 9 min read
Analytical Takeaway
Inflation cooled and long money got dearer. The Treasury published the reason the same morning.
+
Analytical Takeaway
Inflation cooled and long money got dearer. The Treasury published the reason the same morning.
The gauge rises five points to twenty, because the oil price moved enough in one week to turn the energy dial amber, while the yield curve stays the standing red it has been all summer.
No Credible Crash Signal
Six of the eight signals are green, one is red and one turned amber. The score identifies preconditions, not outcomes: it says whether the next ninety days deserve more caution than the last ninety, and conditions can dissipate without a crash.
Every yield in this edition is a Friday close on the US Treasury constant-maturity series, unless it says otherwise.
Scoring last week’s tell
Last week the tell was the two-year Treasury at Friday’s close, against a line at 4.20 per cent. It closed at 4.17. The tell fired on the dovish side, and it has now fired dovish twice running. Three weeks ago the same instrument sat at 4.37 and we resolved this theme the other way, hawkish, and said in print that the two-year was the instrument and nothing more was needed. It has since gone 4.37, 4.28, 4.19, 4.15, 4.17, and it now sits three basis points under the line we drew for it. That earlier resolution is superseded, and it is superseded by the test we published, not by a change of mind.
The yield curve stays red, as it has all summer. The ten-year now yields 0.51 percentage points more than the two-year, the widest gap of the year. A yield curve is simply the price of lending for a long time against the price of lending for a short time, and normally the long loan pays more, because more can go wrong. For most of the last two years it did not, which was the warning. It has now un-inverted decisively, and un-inversion is not the all-clear people take it for. Curves usually steepen fastest on the way into trouble, not out of it.
The energy dial turned amber. Crude rose 5.4 per cent over the week, to $82.40, which sits inside the band we set in advance. Over four weeks it is flat, down a tenth of a per cent. The signal reads the worse of the two windows, so amber it is, and both windows are printed in the appendix, so you can see the choice and disagree with it.
Howell’s liquidity cycle: maturing. Levels broadly flat, momentum in his own words this week supportive but no longer improving. It is the right qualifier on a credit reading this calm: spreads are tight because money is still plentiful, not because the tide is still coming in.
The full working, every weight, every threshold and the arithmetic, is in Appendix A6. A subscriber with only this page can recompute the twenty.
The argument
Start from what the long end did, because it is the only thing this week that behaved strangely.
Over three weeks the front end has fallen about twenty basis points while the long end has barely moved. That is a bull steepener, and it is what a market looks like when it is pricing rate cuts. Over this week the shape flipped: the two-year fell two basis points (hundredths of a percentage point), the ten-year rose three, the thirty-year rose six. That is a bear steepener, and it is what a market looks like when it is worried about something the central bank does not control.
Both are true, at different distances, and the distinction is the whole argument. The Federal Reserve’s hike case is draining away. The price of long money is not following it down.
The reason landed the same morning as the inflation figures, and got a fraction of the attention. The Treasury published its account for July: $334bn in, $766bn out, a deficit of $432bn in a single month. Two lines in it matter more than the total.
The first is that net interest cost $104.16bn and national defence cost $86.05bn. The government now spends more servicing what it has already borrowed than it spends on its armed forces.
The second is stranger, and I had to read it twice. Customs duties came in at minus $8.55 billion. Not a small number. A negative one. The government paid out $33.38bn in tariff refunds during July and collected less than that in tariffs. It was the third consecutive month of net outflows, and June was worse, at minus $25.56bn.
The mechanism is not mysterious once you look. On 20 February the Supreme Court held, by six votes to three, that the emergency powers act does not authorise the tariffs imposed under it, and sent the question of refunds back down. On 4 March the Court of International Trade directed Customs to process those refunds through its ordinary administrative channels, so importers did not have to sue individually to be repaid. More than $160bn had been collected by the date of the judgment. Roughly $100bn has gone back so far, with statutory interest accruing and the mechanics still being litigated. A revenue line that was supposed to help close the deficit is currently widening it.
One correction before anyone does the arithmetic: about $99bn of August benefit payments landed in July because the first fell on a non-business day. Treasury’s own calendar-adjusted deficit for the month is $333bn, up eighteen per cent on last year, not forty-eight. The number is smaller than the headline. The direction is not.
So the shape of the week is this: the inflation problem is receding and the borrowing problem is not, and the long end has started pricing the second. For anyone refinancing a mortgage, a business loan or a data centre, the good news on prices did not arrive at your door, and on this reading it is not going to for a while.
Where I could be wrong
Two ways, and the second is the one that would cost most.
The first is that this is supply, not fear. August is a heavy issuance month and thin summer books exaggerate small imbalances. If the thirty-year retraces below 5.00 in the next fortnight as issuance clears, then I have dressed a plumbing story as a fiscal one. The tell that distinguishes them is below.
The second is the competing explanation I like least because it is the most plausible. A steepening curve with falling front-end rates is also exactly what an ordinary, healthy easing cycle looks like, and I may be reading a term premium (the extra yield lenders demand purely for lending far into the future) into what is simply a market pricing cuts and reaching for duration risk elsewhere. The observable that separates them: if the thirty-year falls while the two-year falls, it was always a cutting story and the fiscal reading is decoration. If the two continue to move apart, it is not. I am watching the gap, not the level.
The thirty-year Treasury has not closed below 5.00 per cent since 20 July. The tell this week: if it closes below 5.00 on any session before Friday 21 August, the long-end regime I have described has broken, the fiscal premium was smaller than I claimed, and the argument above is wrong. If it holds above 5.00, the case that the long end has stopped pricing inflation and started pricing debt survives another week. We score it next Saturday, either way.
July inflation was benign on every published measure. The two-year fell two basis points, the ten-year rose three, the thirty-year rose six. The Treasury ran a $432bn monthly deficit, with net interest above defence and customs duties negative.
The long end is no longer pricing inflation. It is pricing the quantity of borrowing, and the tariff refunds have removed a revenue line that was supposed to narrow the gap.
A heavy August auction calendar clearing well would suggest this was supply and thin summer books, not conviction. A thirty-year that falls alongside the two-year would mean it was always a cutting story.
A correction on the record: last week’s edition reported the thirty-year as having fallen two basis points to 5.19. The 5.19 was right and the comparison was not; it was struck against a Thursday reading of 5.21 where the verified Friday close a week earlier was 5.27. The thirty-year fell eight basis points that week, not two. The underlying cause, a rate series that did not carry the long end daily, has been fixed.
03
The Magazine · 2 min read
The Week That Was
Inflation cooled, a tariff line went negative, and a basketball team repriced twenty-five per cent without playing a game differently.
+
The Week That Was
Inflation cooled, a tariff line went negative, and a basketball team repriced twenty-five per cent without playing a game differently.
Who PaysCustoms duties came in at minus $8.55bn in July, the third negative month running. Whoever the tariffs were meant to be paid by, the answer this quarter is the Treasury. The mechanism is in the Analytical Takeaway.
Forced FlowsReddit rose 10.4 per cent after S&P Dow Jones Indices said it joins the S&P 500 before the open on 18 August. Index inclusion obliges every fund tracking that index to buy, regardless of price or opinion. It is the purest example of a move with no view in it.
Where Value Is CapturedThe Los Angeles Lakers changed hands on Wednesday at a reported $12.5bn, agreed with a group led by Josh Kushner and Bob Iger, and subject to approval by three quarters of the NBA’s thirty governors. Two things are usually got wrong about it. It is a valuation, not a cheque: the Buss family’s roughly fifteen per cent is not being bought. And the previous record holder was the same franchise, fourteen months ago, at $10bn. An asset with fixed supply repriced twenty-five per cent in a year without playing a game differently.
CapacityYankee Global Enterprises agreed $2.6bn of credit and equity with Apollo Sports Capital, Apollo’s largest American sports investment. The shape matters more than the size here: baseball bars private equity funds from owning more than fifteen per cent of a club, so the money arrives mostly as credit. When the equity door is only ajar, capital comes through the debt window.
CommoditisationNu Holdings rose twelve per cent on a record quarter, its first with net income above a billion dollars. Globant fell 12.6 per cent on a weak one. The two are the same business model at different ages: the first still taking share, the second discovering that services priced per hour do not compound when the hours can be automated.
04
The Magazine · 6 min read
Bubble & Risk Scan
The crash gauge has six of its eight signals green, the yield curve red and the energy dial newly amber. The cards below are the fuller picture, and three of them sit amber.
+
Bubble & Risk Scan
The crash gauge has six of its eight signals green, the yield curve red and the energy dial newly amber. The cards below are the fuller picture, and three of them sit amber.
The extra interest a riskier company pays to borrow compared with the government; it widens when lenders turn nervous. It did not move at all this week, holding at the same level as last Friday. Lenders are charging riskier borrowers no more than they were a week ago, in a week when the government’s own cost of long-dated borrowing rose. Corporate credit and sovereign credit are being priced by different crowds, and only one of them is worried.
The standing red, and the only one left. Think of the yield curve as a savings account that normally pays you more the longer you lock your money away; when it stops doing that something is wrong, and when it starts doing it again after a long spell of not, a downturn has historically been close. The ten-year pays 51 basis points more than the two-year, five more than a week ago and the widest of the year, and the way it got there is the interesting part. The two-year, which tracks what the central bank does next, fell two basis points on benign inflation. The ten-year rose three. The thirty-year, which tracks what the economy does over a generation, rose six, and sits at 5.25 per cent. So the bond market got the inflation news it wanted and charged more for long money anyway. That is not a slowdown being priced. It is the price of lending to a government being repriced, and it is why this gap and not the level is the thing to watch.
The share market’s fear gauge closed at its lowest of the year. The bond-market equivalent, the MOVE index, is carried at 66.6 from late July for want of a fresh print, and we flag that rather than let you find it. A fear gauge at the year’s low in the week the labour market turned negative is worth sitting with: it is either well-founded calm or a market that has decided bad news is good news and has stopped pricing the possibility that bad news is simply bad.
The share of the index trading above its long-term trend line fell from 66 per cent to 61.6 in a week the index itself made a record. Fewer shares carrying a higher index is the classic early sign of a narrowing market. It scores green because the threshold is 60, but it is 1.6 points from amber and it moved the wrong way, which is exactly the kind of detail a composite score can flatten.
ISM is the Institute for Supply Management, whose monthly survey of American factory purchasing managers is the oldest reliable read on industrial demand; above 50 means demand is growing. New orders rose to 56.7, and the headline factory index reached its highest since May 2022. Factory employment expanded for the first time in thirty-three months. Hold that against the payroll figure: the goods economy is accelerating while the aggregate jobs number is shrinking, which locates the weakness in services and government rather than in industry.
The ten largest companies are 38 per cent of the index, and the cyclically-adjusted valuation, which compares today’s price with ten years of averaged earnings to strip out the cycle, reads 41.8, the second-highest in roughly a hundred and fifty years. A record index made on narrowing breadth at that valuation is not a crash signal, but it is the reason the crash gauge at 20 should be read as “no trigger is currently pulled” rather than as “this is cheap”.
July car sales ran at an annual rate of sixteen point three million, down from June and below forecast, though still well above the fifteen million level at which we would flag it as a priority. Cars are the most useful single consumer indicator because buying one requires both confidence and credit. A softening here alongside a negative payroll print is a coherent picture, and it is the mechanism by which a labour-market slowdown reaches the real economy.
The dial reads the worse of the one-week and four-week oil moves, and both are green for the first time since June. Last week this dial sat thirty-seven cents of crude away from turning amber, and we said so. This week the barrel moved, the one-week window reads plus 5.4 per cent, and the dial is amber. That is the gauge working as intended: read the dials and not only the number, and the working in Appendix A6 lets you move the line yourself.
Three of the eight readings, the MOVE index, the percentage above the 200-day average 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.
What this means in practice
A composite score of 20 out of 100 says the same thing in plain language: there is no credible crash signal in the machinery. One dial is red, the yield curve, red all summer; one turned amber, the price of crude. What the gauge does not measure is the cost of government borrowing, which is the thing that actually moved this week.
A composite of twenty out of a hundred says the plumbing is sound. Credit spreads are tight, volatility is low, and three fifths of the American market sits above its long-run average price. Set that beside a government bond costing 5.25 per cent to issue and you have no contradiction at all. You have an economy where nothing is breaking and everything is becoming dearer to finance, which is a description most households would recognise before most strategists would.
The valuation dial is where the discomfort sits. The cyclically adjusted price-to-earnings ratio on the S&P 500, which averages profits over ten years to strip out the cycle, stands at 42.3. That is roughly the ninety-eighth percentile since 1881, and second only to the peak of 1999 and 2000. The top ten companies are thirty-eight per cent of the index. Neither is a timing signal and neither has ever been one. Expensive markets stay expensive for years. They simply do not have much room for disappointment, and a long bond at 5.25 per cent is a higher bar for every equity to clear.
05
The Magazine · 2 min read
The Speed of Now
A tariff refund that reaches the middle of the supply chain and stops there.
+
The Speed of Now
A tariff refund that reaches the middle of the supply chain and stops there.
CommoditisationThe clearest signal in enterprise software this week was not a product launch but a price. Services billed by the hour do not compound when the hours can be automated, which is the mechanism underneath Globant’s 12.6 per cent fall and underneath a good deal of the rerating happening across the sector.
Four minutes. Paste this into Claude:
US customs duties were negative in the July 2026 Monthly Treasury Statement, at about minus $8.5 billion, because tariff refunds exceeded collections after the Supreme Court struck down the IEEPA tariffs. Walk me through who actually receives those refunds, what an importer has to do to claim one, and who ultimately bore the cost during the period the tariffs were being collected. Then tell me what you are uncertain about.
What I found. The refund does not go to the person who paid the higher price. It goes to the importer of record, which is the company that filed the customs entry, and only for entries it can document. The household that paid more for the goods in 2025 has no claim at all and no mechanism to make one. The money comes back to the middle of the chain, not the end of it. The same pattern appears wherever a policy is reversed: the reversal almost never runs backwards along the path the cost took forwards.
The uncertainty the model flagged, correctly, is the one I would flag too: how much of the tariff was passed through to consumers in the first place is contested, and the answer decides whether this refund is a windfall or a genuine repayment.
06
The Magazine · 2 min read
Geopolitical Watch
One newsletter says a grain corridor closed three weeks ago. If it is right, almost nobody outside shipping has noticed.
+
Geopolitical Watch
One newsletter says a grain corridor closed three weeks ago. If it is right, almost nobody outside shipping has noticed.
Since 22 July, according to the newsletter Doomberg, Russia has effectively closed the Odessa port complex to shipping, and Ukrainian agricultural exports may fall by half in the coming months. That is one source and it is worth saying so plainly: rail, road and river cannot substitute for a deep-water port, so if the account holds the consequence is severe. The trigger it describes is a Ukrainian drone campaign against Russian-linked shipping in the Sea of Azov in mid-July, including grain cargoes.
Three weeks on, the Baltic Dry index has fallen 7.3 per cent on the week, from 3,089 to 2,863. It is the largest decline anywhere in the Scoreboard.
The temptation is to connect the two. I am not going to, because a freight index covers the world and one port does not move it alone, and the position I can defend is that the timing is suggestive and the causation is not established. What is established is the freight move. What is reported, by one paywalled source and not yet corroborated, is the port closure. Those deserve different levels of confidence and this section is not going to blur them.
07
The Magazine · 8 min read
Case Study, Horizon3.ai
Once a year, a large company pays a team of specialists to break into its own network. That cadence is what this company is selling against.
+
Case Study, Horizon3.ai
Once a year, a large company pays a team of specialists to break into its own network. That cadence is what this company is selling against.
Once a year, a large company pays a team of specialists to break into its own network.
They are called penetration testers, and the exercise is a simulated burglary. They spend a few weeks trying every door, and they hand over a report describing exactly how they got in and what they could have taken. It is expensive, it is genuinely useful, and it is out of date the moment somebody installs a new piece of software. Which happens the following Tuesday.
So the picture most organisations actually live with is this. They are measured on the day of the test, and unmeasured for the other three hundred and sixty. The report describes a building that no longer exists.
The archetype: the Ambush Predator. Snehal Antani spent his career on the inside of that problem and then walked round to the other side of it. He was chief technology officer of Joint Special Operations Command, the American military’s tier-one raiding organisation, before that chief technology officer at Splunk and chief information officer at GE Capital. He co-founded Horizon3.ai with Anthony Pillitiere, whom the company describes as a United States Air Force special operations veteran and its head of community engagement. The predator does not chase. It waits inside the terrain it already knows, and it moves once.
What they built is called NodeZero, and its trick is not finding holes. Automated scanners have found holes for twenty years and buried their owners in lists of them. NodeZero chains them. It runs continuously inside live production systems and works out how an attacker would string three unremarkable weaknesses together to reach something that matters: an out-of-date printer to a forgotten service account to the payroll database. A list of flaws tells you what is wrong. A chain tells you what is about to happen.
Then the money arrived, and the shape of its arrival is the story. The company raised a seed round of $2.5m in January 2020. Then $5m, then $35m, then $40m, then $100m. On 3 August it announced a Series E of $250m at a reported valuation of $2bn, co-led by NightDragon and NEA, with SAIC and Qualcomm alongside. Six rounds, and the last one is larger than the first five put together. That takes the total raised to roughly $430m, which reconciles against the company’s own statement of $178.5m at its Series D fourteen months earlier.
Those numbers deserve different levels of trust and this piece will not blur them. The round is on the record. The $2bn rests on database reports rather than a filing, and a headline valuation can carry structured terms that flatter it.
Here is the sizing, and it is a comparison rather than an adjective. Pentera, the closest direct competitor, passed $100m of annual recurring revenue in April. Horizon3 is approaching the same mark now, having started from a $2.5m cheque six and a half years ago. The caveat travels with it every time: that figure is the company’s own, unaudited, and recurring revenue is not the same thing as revenue. The growth rate the company quotes, better than a doubling year on year, dates from June 2025 and is fourteen months old.
And there is one bear point stronger than the rest, because it comes from the company itself. Horizon3 leans hard on the federal market: FedRAMP High authorisation, expansion into Secret and Top Secret workloads, and a contract under the National Security Agency’s continuous autonomous penetration testing programme. That is a genuine moat, since almost nobody else can operate there. It is also a revenue base tied to government procurement cycles, which arrive in lumps and stop for reasons that have nothing to do with how good the product is.
The unresolved issue is commercial, not technical. The technology works. The question is whether a security platform a company already owns simply absorbs it as a feature. Pentera is now shipping the same capability natively. XBOW and others are building on the identical architecture. If continuous self-attack becomes a checkbox inside the platform a company has already bought, then Horizon3 has spent six and a half years and $430m proving a feature for somebody else’s product. The evidence currently leans toward feature rather than standalone product, and this profile says so instead of offering a coin toss.
Two more things belong here because they are missing. No independent third-party validation of NodeZero’s effectiveness could be found, which is not the same as finding that none exists, and for a product whose entire claim is that it thinks like an attacker, that gap is worth naming. And no personal-stake beat is on the record. There is no documented near-death round, no bet-the-house pivot, no co-founder who walked. Antani has spoken publicly about the strain of a startup on family life and nothing more specific than that. Two case studies running, this publication has found no such moment, and the correct response is to say so rather than to write one.
The same pattern is arriving in credit, compliance, insurance underwriting and safety inspection. What Horizon3 actually sells is the conversion of a periodic assurance into a continuous one. In every one of those industries there is an incumbent whose business model is the once-a-year visit. So the question to carry out of this piece has nothing to do with penetration testing: what else are you currently measuring once a year and assuming holds in between?
An autonomous penetration tool is a weapon that has been taught to fire itself. The governance question is not whether it works, but who may point it, at what, and how anyone would know if it were pointed elsewhere.
We could not find a published company position on any of the following, and the absence is an observation and not an accusation. Who may buy it, and what vetting a purchaser passes. How it is treated under export control, given that offensive intrusion software sits close to existing arms-control regimes. What happens to the skill floor: the more autonomous the tool, the less capable its operator needs to be, which is the whole commercial proposition and also the whole risk. How a private vendor operating inside classified government environments is held accountable. And what happens when a tool that attacks live production systems causes the outage it was bought to prevent.
For most companies, the absence of a published ethics position is an oversight. For this category it is the question.
08
The Gauge · 9 min read
The Stack Inversion
The gauge holds at 32.5. But the compute margin went below the line on Friday, for the first time since we set the test.
+
The Stack Inversion
The gauge holds at 32.5. But the compute margin went below the line on Friday, for the first time since we set the test.
The gauge holds at 32.5 out of 100. First cracks. Unchanged for a second week, and unchanged is itself the finding.
Three things are being measured here and it is worth separating them before the numbers. The gauge asks whether the upstream bottlenecks are dissolving, and its answer is no: scarcity is intact. The business test asks whether renting out the workhorse chip still makes money, and this week its answer changed for the first time. The interpretation is that one is not yet evidence of the other.
The unchanged dial is the one to watch. Memory pricing stays green, meaning prices are still rising, and this week produced the strongest evidence yet for why. J.P. Morgan’s global research team published a note on 6 August projecting DRAM prices, the working memory inside every computer, up more than four hundred per cent between the start of 2024 and the end of 2026. Carry the caveat with the number: that is a projection, not a print, and it is cumulative across roughly three years.
The transmission is already in the official statistics, which is what makes it real. American consumer prices for computer software and accessories, and producer prices for storage devices, have both risen twenty-three per cent since the end of 2024. The laptop you buy this Christmas will cost more than the one you bought last year, and no tariff explains it. A shortage of memory does. This is why the gauge reads chip prices and not share prices: a shortage pays its suppliers best in the quarter before it ends.
Three dials are held deliberately, and the reasons run in print. The lithography monopoly stays green: the prototype report that prompted the question is dated 27 July and moving now would be a re-reading rather than news. Memory capacity holds amber: the Chinese capacity story is six weeks old and the company at its centre has refused to cut prices, which is scarcity intact. Positioning holds amber rather than red, because the alarming leverage ratio in circulation does not verify, regulators have already imposed curbs, and volumes have collapsed.
The business test: a second month on the line
Last week this section asked whether renting out an H100 makes money, answered nine cents an hour, and printed the test that would settle it: a sustained month below the line would be the inversion. Four Fridays have passed, so it can now be answered for the first time. Here is the result.
On Friday the open-market interruptible rent, the price at the margin where independent operators compete for work, was $1.47 an hour, taken as the daily median across fifty-four live offers. Against a cost of $1.64 to own and run the chip, that is a margin of minus seventeen cents.
One Friday is not four. The weekly series runs on Friday closes, and on the last four it has read $1.33, $1.87, $1.73 and now $1.47. Inside this week alone the same daily measurement went $1.73, $1.33, $1.40, $1.33, $1.71, $1.73, no poll, and then Friday’s $1.47. A Saturday capture came back at $1.60 and sits outside the weekly series. The swing inside one week is wider than the margin being measured.
The test has started; it has not passed. Taking the two months of Fridays together, the marginal H100 has earned approximately its cost of capital and nothing beyond it, crossing the line in both directions and settling on neither side. A shortage does not end with a crash. It ends with the marginal unit quietly ceasing to be worth buying, and this is what that looks like from the inside, long before it reaches a contract price or an earnings statement.
The cost stack, recomputed this week and not carried over.
| Line | $ per hour | Assumption |
|---|---|---|
| Rental, interruptible | 1.47 | Daily median across fifty-four live offers, Friday. The guaranteed rate was $2.27 |
| Chip depreciation | 1.18 | $35,000 all-in per chip including its share of server, networking and installation, straight line over four years, at eighty-five per cent utilisation, which is 7,446 billable hours a year |
| Power | 0.10 | Seven hundred watts for the chip, lifted to nine hundred and eighty for cooling and facility overhead, at ten cents a kilowatt hour |
| Hosting, bandwidth, staff | 0.15 | Facility rent, network transit and operations, per billable hour |
| Financing | 0.21 | Nine per cent on an average outstanding balance of half the purchase price across the four-year life |
| All-in cost | 1.64 | The four cost rows added together |
| Hardware margin | minus 0.17 | The rental less the all-in cost. A loss of seventeen cents for every hour the chip is rented |
Those four costs produce an all-in hourly cost of $1.64, leaving a seventeen-cent loss at Friday’s rent. The full derivation is in A12. Two assumptions carry most of the weight: the long end rose this week, which lifts the financing row by a fraction of a cent and changes nothing, and depreciation remains the most contested line, and a five-year write-off instead of four would save about twenty-four cents an hour and restore the margin outright.
The interpretation, and why the two are kept apart
The rental test sits deliberately outside the composite. It measures the economics of one ageing chip; the gauge measures upstream bottlenecks across the whole stack. They are kept separate because a falling H100 rent could mean either obsolescence or a genuine easing of scarcity, and mixing them would let the first masquerade as the second.
Which is the competing explanation, and it is a strong one. The H100 is a 2022 part and Blackwell is shipping. The observable that separates them: if the newest silicon rents at a widening premium while the H100 sinks, it is displacement. If the premium compresses too, it is the tide going out. We do not yet have a clean current-generation rental series, so this is a test we have specified and not yet run. Naming it now is the point; it is the first thing this section will add.
This is a margin and scarcity gauge, not a value gauge: it can say the toll bridge is losing its monopoly, not whether what crosses it is worth anything. The rental test measures one ageing chip, not the whole stack. Neither is a crash forecast, and a reading can sit at “first cracks” for a long time.
1. Memory contract prices printing down month on month. Not decelerating, down. The single load-bearing signal for the whole trade, and nothing is printing down yet.
2. The premium on current-generation silicon. The one reading that separates H100 obsolescence from a broad easing of scarcity, and the series we do not yet have.
3. Any crack in the extreme-ultraviolet monopoly. Structural, and the only development that would take this score above fifty-five.
The score falls if a memory capacity delay, a demand surge that empties inventories again, or the Chinese tool programme slipping a year on quality re-tightens scarcity.
The full eight-signal working is in Appendix A11. The price register is in A12, where the memory contract price row remains absent and says so.
09
The Magazine · 2 min read
The Long View
Britain has been here twice, and neither time ended in default. What it actually took is the interesting part.
+
The Long View
Britain has been here twice, and neither time ended in default. What it actually took is the interesting part.
After Waterloo, British government debt approached 180 per cent of national income, and servicing it consumed close to two fifths of every pound the government collected.
Set that against the American figure in this week’s Treasury statement, where net interest took a little under a third of July’s receipts, and the striking thing is not the distance between the two. It is how short the distance is. Britain carried the heavier load, and the century that followed was the one that built the railways, the factories and the cities. By 1914 the ratio was twenty-nine per cent.
Britain has twice carried a debt burden heavier than America’s today, and twice walked it back down without defaulting.
Data: Our World in Data, UK government debt as a percentage of GDP, 1727 to 2016.
The caveat is mandatory, and it makes the section better. Britain did not grow out of it in the comfortable sense. Over the thirty years after 1946, nominal debt still rose from £27bn to £64bn. The ratio fell because nominal national income rose more than twelve hundred per cent, of which only 2.3 per cent a year was real growth and 6.5 per cent was inflation. The effective interest rate on the debt sat below inflation in twenty-four of those thirty years, sustained by capital controls and restrictions on bank lending, alongside primary surpluses averaging 1.6 per cent of national income. None of those three tools is available to the United States today.
So the counterweight is not “relax”. It is this: the arithmetic has been beaten before, and here is exactly what it took.
10
The Magazine · 4 min read
Frontier
Silicon cannot make light. That single physical fact sits underneath every fibre-optic link in an AI data centre, and underneath three circulating claims about a shortage.
+
Frontier
Silicon cannot make light. That single physical fact sits underneath every fibre-optic link in an AI data centre, and underneath three circulating claims about a shortage.
Silicon cannot make light. Push electricity through it and you get heat, which is why every fibre-optic connection in an artificial intelligence data centre ends in a small piece of a different material called indium phosphide.
A modern data centre is not one computer. It is tens of thousands of chips that must talk to each other constantly, and copper cannot carry that much conversation. So they talk in light, down glass, and something has to switch that light on and off millions of times a second.
One clarification before the numbers, because every confused claim about this shortage comes from mixing up three different things. The bottleneck is not the metal itself, and it is not optical components in general. It is one narrow chain, and it has three steps.
First, refined indium, the raw metal. Second, the wafer, a polished disc of indium phosphide, which is the blank material a chip is built on. Third, the laser, the finished component that actually turns an electrical signal into light. Each step is worth several times the one before it, because that is what manufacturing does, so the three are wildly different in size and none of the numbers below converts into the others. When a claim moves between steps without saying so, it stops meaning anything. The shortage is real at the third step, the finished laser.
| The step | What it is | Size | How that figure is built |
|---|---|---|---|
| 1. Refined indium | The raw metal, recovered as a by-product of zinc smelting. Not narrow: most of it goes to touchscreens, not to this chain at all | ~$410m in 2025 | 1,100 tonnes of world refinery output at the $370 a kilogram annual average, both from the US Geological Survey’s 2026 summary. That multiplication is the whole of it, and you can do it yourself. China refines 760 of the 1,100 tonnes, which is 69 per cent |
| 2. The wafer | A polished disc of indium phosphide, the blank a chip is built on | ~$198m in 2025 | Independent research on the substrate market alone. It counts discs sold, and explicitly excludes everything built on top of them |
| 3. The laser | The finished component that turns an electrical signal into light | $1.9bn in 2025 $22.75bn in 2030 | Not a market. Six named suppliers’ own production capacity, restated in dollars, added up. The 2030 column is what those suppliers agreed to build |
Read down that table and the confusion explains itself. All three are 2025 figures, deliberately, because a table like this is worthless the moment one column borrows a different year. Step two is smaller than step one because the wafer market counts only the discs, while the metal figure counts every use of indium, most of it touchscreens rather than anything in this chain. Step three is nine and a half times step two because a finished laser is worth many times the blank it is grown on. None of these three numbers is a bigger or smaller version of the others, and none of them can be checked against another. Each one can only be checked against its own source.
Three claims are circulating, and two are wrong in instructive ways.
The first inverts its own story. It is repeated as five factories running more than thirty per cent below capacity, which would mean idle plants and departing customers. What Lumentum’s chief executive Michael Hurlston actually said, on the 5 May earnings call, is that output runs more than thirty per cent below demand. The plants are flat out and fully allocated. A quiet factory tells you the orders have gone; a factory at full tilt that still cannot fill its book tells you the opposite. Three caveats travel with the figure and rarely do: it is specific to one laser line rather than the whole business, only four of the five fabs are in production, and the fifth, at Greensboro, does not open until 2028.
The second is not a market and not growth. It circulates as an optics market rising from $1.9bn in 2025 to $22.75bn by 2030. Those figures come from Rosenblatt Securities, and they are six named suppliers’ own production capacity restated in dollars. Divide the second by the first and you get 11.97, which is precisely the twelvefold capacity build those suppliers agreed to after Nvidia asked them for twenty. Both ends of that line are the same measurement, six years apart, and nothing about a market is being forecast at all. The check a reader can run is the headline one: $22.75bn over $1.9bn is 11.97. The supplier columns underneath sum to $1.84bn and $21.95bn, which divide to 11.93, the same twelvefold from the other direction. Those itemised sums are in the broker note and not in public coverage, so treat them as corroboration and not as something you can open. One warning if you go looking: divide the itemised 2030 total by the rounded 2025 headline and you get 11.55, and the other crossing gives 12.36. Neither means anything. Both ends of a ratio have to come from the same column.
It also matters which of the three steps you are reading, and this is where the number gets quoted wrongly. Rosenblatt’s $1.9bn is capacity to build finished lasers. Independent research puts the wafer market, the bare discs one step earlier, at roughly $198m in the same year. Those two are not in conflict and neither is wrong. A wafer costs a fraction of the laser built on it, so a gap of that size between the two steps is ordinary manufacturing, and anyone treating the smaller figure as a correction of the larger one has made the same mistake in the opposite direction. The number is only checkable once you say which step it belongs to. Calling the capacity series sixty-four per cent annual growth deletes the only interesting fact in it, which is that Nvidia asked those six suppliers for twenty times more and got twelve.
The third claim is the real one. Indium is a by-product of zinc smelting, so you cannot simply decide to mine more of it, and China refined about sixty-nine per cent of world supply last year. That constraint sits upstream of every fab and no capital expenditure fixes it quickly.
The strongest objection is who is making the case. The most vivid claim comes from a chief executive whose company’s valuation improves the longer the shortage lasts, and the sizing comes from a single broker note. Against them, the independent forecaster declines to call it. LightCounting expects component prices to be broadly stable through 2026 and 2027, and says the industry will probably swing from shortage to glut at some point in the future, without naming one. The bottleneck is real and it is narrower, better sourced at the component layer, and less dated than either the headline version or anyone claiming to know when it ends.
11
The Debate · 2 min read
The Displacer vs the Augmenter
Two all this week. The unemployment rate improved for the wrong reason.
+
The Displacer vs the Augmenter
Two all this week. The unemployment rate improved for the wrong reason.
This week: two all. Running total, the Augmenter 35, the Displacer 13.
Adoption speed, to the Augmenter. Average hourly earnings rose one tenth of one per cent in July, the softest month of the year, with no productivity break anywhere in the same release. Technology adoption fast enough to be displacing labour at scale shows up as a discontinuity in wages or output. There is not one.
Labour dynamics, to the Displacer. The unemployment rate improved from 4.2 to 4.1 per cent, and it improved for the wrong reason. Household employment fell by 87,000. The labour force fell by 264,000. Participation slipped from 61.5 to 61.4 per cent. Nobody got hired; a shrinking denominator did the work.
Type of shock, to the Displacer. July’s Treasury statement showed net interest above defence spending and customs revenue negative. Fiscal room to cushion a demand shock is narrowing while the shock has not arrived. That is the demand-destruction argument in its least dramatic and most credible form.
Compute constraints, to the Augmenter. Memory prices are projected up more than four hundred per cent across three years and the main Chinese challenger has refused to cut. A rising cost of compute is a ceiling on substitution, and it is rising.
The score changes if a named layoff wave attributed to artificial intelligence hands the Displacer labour outright; if a memory price rolls over month on month, handing it compute and making it a clean sweep; or if a downstream margin turns positive without a price rise, handing the Augmenter the shock pillar.
What the two forces are, and how the four pillars are scored, is set out in the appendix.
12
Income · 6 min read
ERDR Standing Dashboard
Even-week format: twelve strategies, dashboard only. The Deep Dive is held, and the reason is printed.
+
ERDR Standing Dashboard
Even-week format: twelve strategies, dashboard only. The Deep Dive is held, and the reason is printed.
Even-week format this week, twelve strategies, dashboard only.
The Deep Dive is held, and this is a decision and not an omission. This is an odd week and a Deep Dive is due. The register behind these twelve rows has not been rebuilt, and a confident note written on stale inputs would fail this publication’s own freshness rule while looking exactly like a good one. The dashboard therefore runs with its dated freshness note unchanged, and the note stays until every row is individually re-sourced. The rebuild is scheduled for tomorrow.
| Strategy | Indicative yield | Spread vs IG | Week on week | Action |
|---|---|---|---|---|
| 1. Active income fund plus Lombard borrowing | 9.0 to 13.6% | +357 to 817 bp | Quoted levered, and the sum is worth showing because it is the whole point of the row. The fund itself yields 6.8 to 7.6 per cent. Lombard borrowing costs 4.60 per cent, which is the 3.80 per cent overnight funding rate plus a typical 80 basis point margin. At two times assets to equity that is twice the fund yield less one turn of borrowing, so 9.0 to 10.6 per cent; at three times it is 11.2 to 13.6. It stays a Watch rather than an Add for the same arithmetic read backwards: at three times, every 100 basis points the funding rate rises costs 200 basis points of return | Watch |
| 2. Bundled corporate loans (high-quality CLO tranches) | 6.2% | +77 bp | Floating coupon; the first row that loses income if the cut arrives | Hold |
| 3. Listed infrastructure debt and equity | 5.6% | +17 bp | Long end rose to 5.25 per cent; discounts narrowing | Add |
| 4. Private debt funds (business development companies) | 10.7% | +527 bp | Floating and fully exposed to a cut; credit quality still the watch item | Hold |
| 5. Agency mortgage REITs (real estate investment trusts) | 13.1% | +767 bp | Quoted at the levered equity level, as every row here is; the underlying mortgages are government-guaranteed and yield only about 150 basis points over investment grade, so the 13 per cent is the borrowing, not the asset. A steeper curve widens the gap between what the mortgages pay and what the funding costs, which is the mechanism that helps this row, and the curve did steepen five basis points this week. How much of that reaches an equity holder depends on the leverage and the hedge book, and those differ enough between managers that the improvement should be read as a direction and not as a number | Watch |
| 6. Senior secured leveraged loans | 8.4% | +297 bp | Floating; default rates still benign but a weakening labour market is the risk | Hold |
| 7. Preferred shares and hybrid capital | 7.1% | +167 bp | Fixed coupon, long duration; the cleanest beneficiary if the cut is real | Add |
| 8. Real asset royalties | 6.4% | +97 bp | A quiet week on both legs: gold up nine tenths of a per cent and crude up 5.4, so the energy leg helped rather than hurt for the first time in a month | Watch |
| 9. Emerging market hard-currency sovereign carry | 7.6% | +217 bp | A softer dollar path is the tailwind here; the yen intervention complicates the funding leg | Hold |
| 10. High-yield municipal bonds | 6.3% | +87 bp | Fixed coupon and tax-advantaged; local government payrolls fell in July, which is a credit signal to note rather than act on | Watch |
| 11. Private credit direct lending | 10.2% | +477 bp | Floating, and the row where the compute profit-and-loss above matters most: lending against chips that now rent for less than they cost to own and run, and are obsolete in four years, is a collateral question and not a yield question | Watch |
| 12. Trade and supply chain finance | 7.4% | +197 bp | Freight rates down 7.3 per cent on the week and a grain corridor shut since 22 July; the volume argument that supports this row is the one to watch, not the coupon | Hold |
Each strategy is explained in full when it is the week’s Deep Dive.
Freshness note, carried: these twelve rows were last individually re-sourced on 8 August 2026. They are carried at that date rather than refreshed, and this note comes out only when every row has been re-sourced from the register.
13
Accountability · 2 min read
On the Radar
No new call, for the second week running. The three that did not make it are worth more than the names would have been.
+
On the Radar
No new call, for the second week running. The three that did not make it are worth more than the names would have been.
No new call this week, for the second week running.
Three companies were on the list. None of them made it, and the reasons are worth more than the names would have been.
The first had already done what we would have said it was going to do. It has run hard over the past year and now trades above where the analysts covering it think it should. Its chief executive and its largest outside holder sold into that entire rally, and not one of them bought a share. Being right about a company a year late is not an idea, it is a history lesson.
The second had the better story and the worse timing. The argument for it, that a closed shipping strait lengthens voyages and makes tanker owners busier even as cargo falls, is a good argument. It has also been the front page of the shipping trade press for weeks, and there is a listed fund that exists purely to express it, sitting at a record. When your non-consensus view is available in an exchange-traded wrapper, it is not non-consensus.
The third we simply could not verify. Three sources gave three different price histories that contradicted each other and their own arithmetic. We do not publish numbers we cannot stand behind, so it stays in the drawer until we can.
A correction to the record on MP Materials. Two weeks ago this letter closed that call at minus one, failed-soft. That was the wrong value and it has been corrected to minus two. The rules define minus one as finishing behind your benchmark, and MP finished 3.63 percentage points ahead of the rare-earth index while losing a quarter of its value. The pre-registered condition for the thesis being wrong had fired on both limbs, and a fired condition is a minus two. Beating an index that fell 27.58 per cent does not vindicate a thesis whose entire claim was that the company would stop being priced like that index. The verdict is worse than we first published it, and the ledger now says so.
14
And Finally · 3 min read
And Finally
Five lines, three things to watch, and one of ours closes on Sunday.
+
And Finally
Five lines, three things to watch, and one of ours closes on Sunday.
The data all came in benign,
Inflation obediently in line,
But the thirty-year rose,
As the Treasury shows,
You can’t refinance a decline.
Three things to watch. The thirty-year Treasury, and whether 5.00 per cent holds as a floor or gives way. Memory contract prices, the single dial that decides whether the artificial intelligence trade is still a scarcity story. And the August auction calendar, which will tell us whether this week’s long-end move was fiscal conviction or simply too much paper in a thin month.
If all three move the same way, with the thirty-year holding above five, memory still rising and the auctions clearing badly, then the long end is repricing government credit and not inflation, and that is the regime for the rest of the year. If they diverge, and the thirty-year slips while the auctions go well, this was plumbing and I have over-read it.
One of ours closes on Sunday. USA Rare Earth went into the notebook on 16 May at twenty-one dollars, benchmarked against the rare-earth index, with a scoring date fixed that day and never touched since. It closed Friday at twenty. We score it on Monday and print the verdict next week, whichever way it falls.
Until next week. Stay curious and stay hedged.
Anthony Rosenthal
15
Evidence · reference layer, scan or search
Scoreboard & Appendix
26 assets ranked year-to-date, five active calls tracked against the benchmark each was given at entry, and every number behind the edition.
+
Scoreboard & Appendix
26 assets ranked year-to-date, five active calls tracked against the benchmark each was given at entry, and every number behind the edition.
Freight is still the year’s best line at plus fifty-two per cent, though it gave back seven of those points this week, while crude, which led for four months, sits fourth. The gap between best and worst is just under ninety points, 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 14 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.
| Rank | Asset | 1 Jan baseline | Week 31 close | YTD | 8wk vol |
|---|---|---|---|---|---|
| 1 | Baltic Dry Index | 1,882.00 | 2,863.00 | +52.13% | 55% |
| 2 | USD/TRY | 35.40 | 47.88 | +35.25% | 1% |
| 3 | Nikkei 225 | 51,830.00 | 68,713.80 | +32.58% | 27% |
| 4 | WTI Crude | $63.20 | $82.40 | +30.38% | 62% |
| 5 | Russell 2000 | 2,481.91 | 3,068.42 | +23.63% | 11% |
| 6 | Nasdaq 100 | 25,200.50 | 30,046.14 | +19.23% | 23% |
| 7 | MSCI EM | 1,595.20 | 1,889.53 | +18.45% | 19% |
| 8 | Copper | $5.682 | $6.599 | +16.15% | 14% |
| 9 | MSCI ACWI | 140.58 | 162.29 | +15.44% | – |
| 10 | Euro Stoxx 50 | 5,740.15 | 6,539.59 | +13.93% | 12% |
| 11 | S&P 500 | 6,845.50 | 7,785.76 | +13.74% | 13% |
| 12 | Swiss SMI | 13,248.10 | 14,390.67 | +8.62% | 10% |
| 13 | FTSE 100 | 9,948.30 | 10,750.10 | +8.06% | 9% |
| 14 | DAX | 24,540.20 | 26,440.31 | +7.74% | 14% |
| 15 | HYG | 78.15 | 79.71 | +2.00% | 2% |
| 16 | Gold | $4,341.10 | $4,380.40 | +0.91% | 25% |
| 17 | Nifty 50 | 24,420.00 | 24,366.00 | -0.22% | 10% |
| 18 | LQD | 109.02 | 106.12 | -2.66% | 5% |
| 19 | AGG | 102.15 | 97.48 | -4.57% | 3% |
| 20 | Hang Seng | 26,340.00 | 25,116.85 | -4.64% | 21% |
| 21 | USD/ZAR | 17.55 | 16.18 | -7.79% | 10% |
| 22 | Silver | 70.61 | 64.99 | -7.96% | 51% |
| 23 | TLT | 94.27 | 82.04 | -12.97% | 7% |
| 24 | Natural Gas | $3.514 | $2.733 | -22.23% | 26% |
| 25 | Bitcoin | $87,850.00 | $62,975.59 | -28.31% | 22% |
| 26 | Ethereum | $2,967.00 | $1,880.65 | -36.61% | 38% |
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 14 August 2026. MSCI EM is derived from the EEM ETF (exchange-traded fund) close of 66.61 multiplied by the locked index ratio (28.367); Baltic Dry is the Baltic Exchange BDI as published by Hellenic Shipping News (14 Aug, 2,863). 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.
Portfolio Watch, active calls
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.
| Company | Entry | Week | Close (14 Aug) | Return | Benchmark | Excess | Original thesis | Score date |
|---|---|---|---|---|---|---|---|---|
| YPF Sociedad Anónima (NYSE: YPF) | $50.31 | Wk 23 | $50.05 | −0.5% | −6.0% | +5.5pp | Vaca Muerta shale, Argentina LNG and a sovereign re-rating, mispriced as a spot-oil emerging-market cyclical | Jun 2027 |
| Eight weeks in and back to within half a per cent of entry, having recovered most of last week’s fall as the crude price came back. The excess figure deserves a warning rather than a victory lap. It widened by more than two points this week almost entirely because Alphabet, the single stock this call is benchmarked against, fell. A single-name benchmark swings harder than an index, and this is the only one of the four open calls that moved in our favour, which is exactly why it gets the scrutiny. Nothing structural has moved. | ||||||||
| Mitsubishi UFJ Financial (NYSE: MUFG) | $20.17 | Wk 25 | $23.06 | +14.3% | +5.4% | +8.9pp | Japanese rate normalisation and repatriation flows re-rate the megabanks; the under-priced legs are the flow and the jumbo-hike option | Jul 2027 |
| Six weeks in and now almost nine points ahead of Japan itself, having added nearly three points of excess in a single week. The intervention of a fortnight ago is the thesis arriving through an unexpected door: a government defending its currency alongside the Americans is a government that intends to keep normalising rates, which is the whole argument. That the yen has since given most of that back, closing the week at 157.75, does not weaken the point. It sharpens it, because a defence that does not hold is a government that will have to reach for rates instead. | ||||||||
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 five and thirty-seven 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. Two 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.
Economic indicators
| Indicator | Latest | Prior | Direction |
|---|---|---|---|
| Consumer price inflation, headline (July, rel. 12 Aug) | +0.1% | – | Month on month. The week’s hinge, and the print the bond market had spent eighteen months asking for |
| Core consumer price inflation (July, rel. 12 Aug) | +0.2% m/m +2.5% y/y | – | Core strips out food and fuel. Benign on both windows |
| Producer price inflation, final demand (July, rel. 13 Aug) | 0.0% m/m +4.7% y/y | – | Flat on the month; no pipeline pressure sitting behind the consumer print |
| Average hourly earnings (July, YoY) | +3.2% | +3.3% | The softest wage growth of 2026; +0.1% on the month |
| Ten-year breakeven inflation rate (14 Aug) | 2.27% | 2.40% | What the bond market expects inflation to average over ten years. Prior reading is 1 June; expectations fell and the long end rose anyway, which is the edition’s argument in two numbers |
| Federal deficit, July (Monthly Treasury Statement, rel. 12 Aug) | $432.3bn | – | Net interest $104.16bn now exceeds national defence at $86.05bn. Roughly $99bn of August benefits were pulled into July because 1 August fell on a non-business day; Treasury’s calendar-adjusted figure is $333bn, up 18% year on year, not 48% |
| Customs duties, July | −$8.55bn | −$25.56bn | Third consecutive month of net outflows, after $33.38bn of refunds. Prior is June. The Supreme Court struck down the IEEPA tariffs in February and roughly $100bn has been returned |
| Nonfarm payrolls (July, rel. 7 Aug) | −23k | +20k | Last week’s hinge, carried here because the rate argument still rests on it: the first fall in months, and May and June revised down 103,000 combined |
| Fed funds rate (current) | 3.50 to 3.75% | 3.50 to 3.75% | Held 9-3 on 29 July, three members preferred a rise; our read is Hold with the bias moving to a cut |
This week’s releases are the July consumer price index (12 August), producer prices (13 August) and the July Monthly Treasury Statement (12 August), all verified against the primary publications rather than press summaries. A dash in the prior column means we hold no verified prior on the same basis and would rather print nothing than print a comparison we cannot stand behind. No core producer price figure is published here: two independent retrievals disagree on whether the ex-food-and-energy measure was released, and an unresolved series does not go on the page.
Fixed income & yield curve
The curve did not move as one. The two-year fell two basis points on the inflation news, the ten-year rose three and the thirty-year rose six. That is a bear steepener, and it is the opposite shape to the three weeks that preceded it.
| Tenor | Yield | Week on week |
|---|---|---|
| 2-year Treasury | 4.17% | Down 2 basis points from 4.19; three basis points under the 4.20 dovish line we drew, and the tell fired dovish a second week running (14 Aug) |
| 5-year Treasury | 4.36% | Up 1 basis point from 4.35 (14 Aug) |
| 10-year Treasury | 4.68% | Up 3 basis points from 4.65 (14 Aug) |
| 30-year Treasury | 5.25% | Up 6 basis points from 5.19, on the week inflation came in benign on every measure; it has not closed below 5.00 since 20 July (14 Aug) |
| Yield curve (10Y − 2Y) | +51bp | Dis-inverted and the widest of the year, up 5 basis points on the week; the standing red (both legs 14 Aug) |
| HY OAS (high-yield option-adjusted spread, the extra yield over government bonds) | 271bp | Unchanged on the week; far below the 350bp danger zone (Thursday 13 Aug, see note) |
| IG OAS (investment-grade, the same measure for the safest corporate borrowers) | 78bp* | Carried; no fresh print located this week (last fresh 6 Aug) |
Treasury yields are the constant-maturity rates for Friday 14 August, taken from our own rates series rather than from press reports. That distinction earned its keep this week: figures circulating on 14 August showed the two-year, ten-year and thirty-year all falling, and those were Thursday’s readings. On Friday the ten-year and the thirty-year both rose. The high-yield spread is the Thursday 13 August observation, the most recent published, because the index publishes with a one-day lag; it is not a Friday reading and is not presented as one. *The investment-grade spread is carried from 6 August and marked as carried.
Commodities
Crude took back last week’s fall and freight gave back its rise, which is last week’s picture run in reverse. The metals barely moved at all, and the one large number on the board is the Baltic Dry index falling seven per cent while a grain corridor stayed shut.
| Commodity | Close (14 Aug) | WoW | YTD |
|---|---|---|---|
| WTI crude oil | $82.40/bbl | +5.4% | +30.4% |
| Gold | $4,380.40/oz | +0.9% | +0.9% |
| Silver | $64.99/oz | +2.6% | -8.0% |
| Copper | $6.600/lb | +0.4% | +16.2% |
| Natural gas (Henry Hub) | $2.733/MMBtu | +2.7% | -22.2% |
| Baltic Dry Index | 2,863 | -7.3% | +52.1% |
Commodity closes are Friday 14 August. Baltic Dry is the Baltic Exchange BDI as published by Hellenic Shipping News (14 Aug, 2,863); it is off-feed, so it is named and dated rather than pulled. Gold has crossed back above its 1 January baseline of $4,341.10, which is why a rise of nine tenths of a per cent on the week and a year-to-date of nine tenths of a per cent are the same number and not an error.
Upcoming catalysts
| Date | Event | Relevance |
|---|---|---|
| 16 Aug | USA Rare Earth thesis-horizon verdict, the week 18 call at its three-month horizon | The rare-earth pair’s second test, scored in full next Saturday whichever way it falls |
| to 21 Aug | This edition’s own tell: the thirty-year Treasury against 5.00 per cent | If it closes below 5.00 on any session before Friday 21 August, the long-end argument above is wrong and we say so on 22 August |
| to 21 Aug | The two-year Treasury against 4.10 per cent | A close below 4.10 would say the rate market has fully priced the cut the labour data implies |
| late Aug | The remainder of the August Treasury auction calendar | Whether this week’s long-end move was fiscal conviction or simply too much paper in a thin summer month. The single best test of the argument this edition makes |
| 28 Aug | Arista Networks thesis-horizon verdict, the week 20 call at its three-month horizon | The book’s best open call, graded in full and in public |
| 5 Sep | August jobs report | Whether July’s minus 23,000 was a level shift or a one-month scare; the single most consequential number of the next month |
| mid-Sep | Next FOMC decision | Where the cut this edition argues for either lands or does not |
| Nov | China’s wider export restrictions come out of suspension | Bears directly on both rare-earth positions; a live catalyst rather than background, and dated here so it is not forgotten |
FX
| Pair | Rate | YTD | Driver |
|---|---|---|---|
| USD/TRY | 47.88 | +35.25% | Lira weak on domestic inflation; the slowest grind on the Scoreboard and the most reliable |
| USD/ZAR | 16.18 | -7.79% | Rand firmer on a softer dollar path and the gold move; year-to-date measured against the locked baseline |
| USD/JPY | 157.75 | – | The week’s second story: a joint Japan-US intervention on 31 July, confirmed publicly on 3 August, the first yen-buying operation of its kind since 1998 |
| DXY (dollar index) | ~98* | ~flat | Softer as the front end fell; carried, no separate verified print at production |
*The dollar index is carried and marked; every other row is a verified 14 August close.
Volatility, risk indicators & the crash-gauge working
| Indicator | Level | Signal |
|---|---|---|
| VIX | 14.25 | The year’s low, well below the 20 caution line (14 Aug) |
| MOVE index (bond volatility) | 66.6* | Well below the 110 caution line (carried from edition 28, last fresh 24 Jul) |
| High-yield credit spread (OAS) | 271bp | Unchanged on the week; far below the 350bp danger zone (Thursday 13 Aug, see note) |
| S&P % above 200dma | 61.6%* | Above the 60% line (carried, last fresh 7 Aug) |
| Yield curve (10Y − 2Y) | +0.51 | Dis-inverted and the widest of the year; the standing red (14 Aug) |
| Insider clusters (net selling) | 0 sectors* | No net-selling cluster (carried, last fresh 17 Jul) |
| Energy shock (WTI, two speeds) | 1wk +5.4% | Amber on the one-week window; four-week minus 0.1%, green. The rubric reads the worse of the two (14 Aug) |
| Crash probability score | 20.0/100 | No credible crash signal; up 5 on the week |
The rubric, and this week’s working
Each of the eight signals scores 0 when green, 5 when amber and 10 when red. Multiply each score by its weight, add the eight together, then multiply by ten so the scale runs from 0 to 100. (If every signal were red that is 10 × 1.00 × 10 = 100, which is why the scale tops out there.) This week only the yield curve is red and only the energy dial is amber, so the sum is 15% × 10 = 1.5 for the curve plus 10% × 5 = 0.5 for energy, which is 2.0, and 2.0 × 10 = 20.0. The six remaining signals are green and contribute nothing. There is no change to the rubric this week, so there is no restatement: last week’s number and this week’s are measured the same way. *Carried readings. MOVE, the percentage above the 200-day average and the insider-cluster count are carried from an earlier date, marked with an asterisk and dated in the table above. The high-yield spread is a Thursday 13 August observation, because the index publishes with a one-day lag; it is not a Friday reading and is not presented as one.
| Signal | Weight | Green (0) | Amber (5) | Red (10) | This week | Score |
|---|---|---|---|---|---|---|
| High-yield credit spread (OAS) | 15% | <350bp | 350 to 450bp | >450bp | 271bp* | 0 |
| MOVE index (bond volatility) | 15% | <110 | 110 to 130 | >130 | 66.6* (carried) | 0 |
| ISM new orders | 12.5% | >50 | 48 to 50 | <48 | 56.7 (July, fresh) | 0 |
| Yield curve (10Y − 2Y) | 15% | <−0.25 | −0.25 to 0 | >0 | +0.51 | 10 |
| VIX | 12.5% | <20 | 20 to 28 | >28 | 14.25 | 0 |
| S&P % above 200dma | 10% | >60% | 40 to 60% | <40% | 61.6%* (carried) | 0 |
| Insider clusters | 10% | 0 sectors | 1 sector | 2 or more | 0 sectors* (carried) | 0 |
| Energy shock (WTI, two speeds) | 10% | 1wk <+5% and 4wk <+10% | 1wk +5 to 10% or 4wk +10 to 20% | 1wk >+10% or 4wk >+20% | 1wk +5.4%, 4wk −0.1% | 5 |
* Three inputs are carried, not fresh: the MOVE index (from edition 28, last fresh 24 July, no new print located this week), the percentage of the S&P above its 200-day average (last fresh 7 August) and insider clusters (from edition 27, last fresh 17 July). We would rather tell you than let you find it. Every other reading is this week’s.
The arithmetic: one dial is red and one is amber. The yield curve scores 10 at a weight of 0.15, so 0.15 × 10 = 1.5. The energy dial scores 5 at a weight of 0.10, so 0.10 × 5 = 0.5. Every other dial scores zero. The eight contributions sum to 2.0, and 2.0 × 10 = 20.0 out of 100. The bands: 0 to 30, no credible crash signal, normal volatility expected · 30 to 55, elevated caution · 55 to 75, pre-crash conditions assembling · above 75, high crash probability. The score identifies preconditions, not outcomes: conditions can assemble and then dissipate without a crash. It tells you whether the next 90 days deserve more caution than the last 90. The restatement: none this week. The rubric is unchanged from last week, so last week’s 15.0 and this week’s 20.0 are measured the same way, and the week-on-week change is a rise of five points. How close it was: the energy dial turned amber on the one-week window, at plus 5.4 per cent against a 5 per cent threshold. It cleared the line by four tenths of a percentage point. The four-week window is minus 0.1 per cent, comfortably green, and the rubric reads the worse of the two, so amber it is, contributing half of its ten per cent weight, which is the 5.0 that took the composite from 15.0 to 20.0. Red would need a one-week move above 10 per cent, or a four-week move above 20 per cent.
Where each number comes from. High-yield spread: FRED, ICE BofA index, Thursday 13 August, the most recent published, because the index carries a one-day lag. MOVE: ICE, the .MOVE close, carried from edition 28 and last fresh 23 to 24 July; a low-confidence web read this week is consistent with that level and did not move it. ISM new orders: the ISM July manufacturing report, released 3 August, cross-checked against FRED series NEWORDER. Yield curve: US Treasury constant-maturity rates, the 10-year minus the 2-year, both legs 14 August. VIX: the ^VIX close, 14 August. Percentage of the S&P above its 200-day average: Barchart, carried, last fresh 7 August. Insider clusters: OpenInsider, trailing four weeks, carried, last fresh 17 July. Energy shock: our own verified WTI closes, the 14 August close of $82.40 against $78.18 on 7 August (one week, +5.40 per cent) and against $82.49 on 17 July (four weeks, −0.11 per cent). Every one of them is free and public, and you can pull every one yourself.
The Stack Inversion working
The gauge scores how close the artificial-intelligence hardware shortage is to ending. Each of the eight signals scores 0 when scarcity is intact, 5 when a crack is opening and 10 when supply has arrived. Each contributes its weight multiplied by its score divided by ten, so an amber contributes half its weight and a red contributes all of it; the eight are added. Higher means scarcity is inverting, which is the risk to everything priced on the shortage. There is no rubric change this week, so no restatement.
Three terms in the table, in plain English. DUV (deep ultraviolet lithography, the machines that print circuit patterns onto silicon) is the older, widely available generation. EUV (extreme ultraviolet lithography, the only tool able to print the very finest chips) is the one bottleneck nobody has broken. HBM (high bandwidth memory, the stacked memory that sits beside an AI chip and feeds it data) is where the shortage bites hardest.
| Signal | Weight | Scarcity intact (0) | First crack (5) | Supply arrived (10) | This week | Score |
|---|---|---|---|---|---|---|
| Memory capacity expansion | 20% | none | announced or funded | online and shipping | Unchanged: the Chinese memory raise and the announced Korean state bid are funded, not shipping. No new capacity event this week | 5 |
| Domestic lithography (immersion DUV) | 15% | none | first tools, small numbers | at scale, quality-competitive | Unchanged: first domestic tools reported, small numbers, quality unproven | 5 |
| EUV chokepoint | 15% | monopoly intact | credible challenger | alternative shipping | Monopoly intact; no challenger reported | 0 |
| Memory pricing (DRAM and HBM) | 20% | rising or firm | rolling over | falling | Firm. No new contract print this week, and the corroborating evidence points the same way: Arista disclosed on 4 August that it has secured memory into 2027, which is what a buyer does in a shortage | 0 |
| Leading-edge foundry access | 10% | stalled | incremental | at volume | Incremental | 5 |
| Model-layer commoditisation | 10% | frontier closed | open weights gaining | open and cheap at parity | HELD at amber, and the decision is worth stating. Every model at the open frontier is now Chinese and free to download, and second-quarter margin data puts the model and application layer at roughly minus 59 per cent against plus 41 upstream. Both are evidence of open weights GAINING, which is the amber definition. The red band requires parity, and no benchmark comparison establishing it was published this week, so the dial does not move | 5 |
| Power as the bottleneck | 5% | not binding | value migrating to power | power is the priced scarcity | Value migrating to power and the physical build | 5 |
| Scarcity-premium positioning | 5% | modestly priced | crowded | euphoric or levered | Crowded; the complex rose with the whole market on the rate trade while index breadth narrowed | 5 |
The arithmetic: memory capacity 20 × 5 ÷ 10 = 10.0; lithography 15 × 5 ÷ 10 = 7.5; the EUV chokepoint 15 × 0 ÷ 10 = 0; memory pricing 20 × 0 ÷ 10 = 0; foundry 10 × 5 ÷ 10 = 5.0; the model layer 10 × 5 ÷ 10 = 5.0; power 5 × 5 ÷ 10 = 2.5; positioning 5 × 5 ÷ 10 = 2.5. The eight contributions are 10.0, 7.5, 0, 0, 5.0, 5.0, 2.5 and 2.5, and they sum to a composite of 32.5 out of 100. Last week’s 32.5 is measured on the same rubric, so the gauge is unchanged. The model-layer dial was the week’s live question and was held at amber rather than moved to red: the red band requires open weights at parity with the closed frontier, and this week’s evidence, that the open frontier is entirely Chinese and that the layer is loss-making, shows the gap closing rather than closed. The bands: 0-30, scarcity intact · 30-55, first cracks, supply announced or funded but not delivered · 55-75, inversion underway, prices and not just share prices turning · above 75, scarcity broken. The load-bearing dial is memory pricing and it is still green: the reader instruction is to watch chip prices, not share prices.
Where each number comes from. Memory pricing: TrendForce and DRAMeXchange DRAM and HBM contract and spot, most recent quarterly contract guidance, carried; corroborated this week by Arista’s own disclosure of 4 August. Capacity and lithography: named primary reporting on fab raises, listings and tool shipments, carried from edition 29 with no new event. EUV: ASML disclosures. Foundry: SMIC and Hua Hong node progress. Model layer: second-quarter layer-margin data published 7 August and the open-weight model rankings of the same week. Power and positioning: this edition’s own breadth reading and the week’s sector moves. Every carried reading is dated at source and marked as carried above.
The Stack Inversion price register
A gauge made of judgements needs a spine made of prices. This register locks the baseline at first publication and reports the change every week thereafter, so that a reader can watch the actual cost of artificial-intelligence compute rather than the share prices of the companies selling it. Every figure is machine-captured from a named source on a fixed schedule and is never hand-keyed.
| Price | Baseline (17 Jul) | Prior week (7 Aug) | This week (14 Aug) | Week | Since baseline |
|---|---|---|---|---|---|
| H100 open-market rent, interruptible, per hour | $1.333 | $1.733 | $1.467 | −15.3% | +10.1% |
| H100 open-market rent, guaranteed, per hour | $2.161 | $2.059 | $2.268 | +10.2% | +5.0% |
| All-in cost to own and run one H100, per hour | $1.64 | $1.64 | $1.64 | – | – |
| Compute margin, rent less cost, per hour | −$0.31 | +$0.09 | −$0.17 | −$0.26 | +$0.14 |
| Live offers behind the median | 64 | 55 | 54 | – | – |
| Frontier-class model, per million input tokens | $2.50 | $2.50 | $2.50 | flat | flat |
| Second frontier vendor, per million input tokens | $3.00 | $3.00 | $3.00 | flat | flat |
| Budget tier, per million input tokens | $0.10 | $0.14 | $0.14 | flat | +40.0% |
| Open-source hosted floor, per million input tokens | $0.05 | $0.02 | $0.02 | flat | −60.0% |
The working, so the margin line can be checked. Rent is the daily median of live open-marketplace H100 listings, captured automatically each morning; the count of offers behind each median is printed above so a thin day is visible. The cost line is computed, not quoted, and every assumption is printed in full in the Stack Inversion section. Thirty-five thousand dollars all-in per chip, written off straight-line over four years at eighty-five per cent utilisation, gives $1.18 an hour. Power at nine hundred and eighty watts and ten cents a kilowatt hour gives $0.10. Hosting, bandwidth and operations are $0.15, and financing at nine per cent on an average outstanding balance of half the purchase price gives $0.21. Those four add to $1.64. The margin line is simply this week’s interruptible rent less that figure. The capture stores $1.467, shown as $1.47 elsewhere in this edition; against $1.64 that is a loss of $0.173, which we print as minus seventeen cents. The third decimal is kept here so nobody has to wonder why $1.47 less $1.64 is described as exactly seventeen cents. The cost stack was recomputed this week rather than carried forward, and it came back at the same $1.64: the long end rose, which lifts the financing row, but at these weights the effect is a fraction of a cent. Any change to the cost line will be stated here rather than absorbed silently.
No restatement this week. Every figure in the four price rows above comes from the same automated Friday poll of the same marketplace that set the baseline on 17 July, on the same basis, to the same precision. The baseline has not been re-keyed and no prior published figure has moved. When one does, it will be named here alongside the number it replaced.
The token rows join the register this week, with two disclosures attached. First, they are aggregator-derived rather than read from vendor pages, which puts them at medium confidence, below the H100 rows. Second, the open-source hosted floor is held at $0.02 for continuity of the series even though a cheaper paid tier was observed this week at $0.010. Whether that tier belongs in a floor built to track the open-weight alternative is a definitional question we have not ruled on, and until we do, moving the line would be a change of definition dressed as a change of price. The observation is recorded here so that the ruling, when it comes, can be checked against it.
What is still not in this register, stated plainly. One row belongs here and is absent: the contract price of memory, which is the single most load-bearing dial in the gauge this register is meant to spine. It is tracked and it has no fresh automated capture this week. A register whose figures are typed in by hand is worse than no register at all, because it looks identical to one that is measured. That row joins when the capture does, and this note comes out when it does.
AI & technology data points
| Company / event | Data point | Relevance |
|---|---|---|
| AI layer margins, Q2 2026 (Apollo, 7 Aug) | Upstream (chips, memory, networking, power) around +41%; models and applications around −59% | The single most useful frame in the edition; held the model-layer dial at amber, because the red band requires parity |
| Hyperscaler capex as a share of GDP (Apollo, 6 Aug) | ~3% of US GDP across 2027 to 2029, against 0.3% in 2019; telecoms peaked at 1.2%, housing at 6.6% | Unremarkable in level, unprecedented in speed; the +2.7pp swing is the largest of the three |
| Arista Networks (NYSE: ANET), Q2 results 4 Aug | Revenue $3.04bn, +37.7% YoY, the first $3bn quarter; FY guide raised a third time to ~$12.6bn; multi-year commitments $9.7bn, near-triple YoY; memory secured into 2027 | Primary source: company release filed with the SEC. The memory disclosure is the item that matters beyond the beat |
| Open-weight frontier | Every model at the open frontier is now Chinese | The toll booth at the model layer is the one most exposed to a free substitute |
| H100 open-market rent (14 Aug) | $1.47 an hour interruptible, the daily median across fifty-four live offers, against a computed all-in cost of $1.64; a margin of minus seventeen cents | The compute profit-and-loss turned negative for the first time. Working shown in the Stack Inversion section and A12 |
The H100 rent row is this week’s, captured Friday 14 August. The four rows above it are carried from edition 30 and are dated in their own cells: the two Apollo readings to 6 and 7 August, the Arista results to 4 August, and the open-weight observation to the week of 7 August. No new artificial-intelligence data point of this class reached us between 8 and 14 August, and it is better to say that than to reprint four items as though they were new.
Geopolitical radar
| Flashpoint | Status | WMP assessment |
|---|---|---|
| Odessa port closure | This week’s lead. The Odessa port complex has been effectively closed to shipping since 22 July, leaving Ukraine functionally landlocked. Officials warn agricultural exports could fall by half, because rail, road and river cannot substitute for a deep-water port. The trigger was a Ukrainian drone campaign in the Sea of Azov in mid-July that struck dozens of commercial vessels, including Russian grain carriers | A grain corridor closed three weeks ago and almost nobody outside shipping has noticed. The Baltic Dry index fell 7.3 per cent this week, the largest decline anywhere on the Scoreboard, and the temptation is to connect the two. We are not going to: a freight index covers the world and one port does not move it alone. The timing is suggestive and the causation is not established, and those are different claims |
| US–Iran / Hormuz | Carried from edition 30, no new development this week. A threatened bombardment announced and fully withdrawn over the weekend of 1–2 August. An Iranian parliamentary committee is drafting conditions restricting the Strait by vessel flag. Crude has since retraced the fall, closing 14 August at $82.40, up 5.4% on the week | A ceasefire is not peace; a signature is not compliance; and now, a cancellation is not a policy. Four reversals in five months. Treat as a risk that keeps being priced and unpriced, not as a resolved one |
| Japan–US joint yen intervention | Carried from edition 30. Coordinated MoF and US Treasury yen purchase on 31 July, confirmed publicly 3 August. The yen closed 14 August at 157.75, weaker than the roughly 155.20 it reached immediately after the operation. First joint yen-buying operation since 1998 | A monetary flashpoint rather than a military one. It changes what a yen move signals for anyone funded in yen: no longer a Japanese story global markets can read as noise. That the level has retraced within a fortnight is the observation, not a footnote to it |
| Critical minerals policy | White House mining roundtable, 7 August: over $2bn in commitments; the largest rare-earth-specific item, $150m, went to rare-earth-FREE magnet technology | Directionally supportive of the sector and quietly funding its substitute. What the state is buying is independence from China, not a future for any particular mine |
| China export-control list | MP Materials and USA Rare Earth added June 2026; implementation of the wider restrictions suspended until November 2026 | Bears directly on the book’s two rare-earth calls. The November date is a live catalyst, not background |
| Corporate decoupling | Carried from edition 30. SpaceX and Tesla reported writing China exclusion into supplier contracts and legal entities | Supply-chain separation moving from tariff lines, where it is visible and arguable, into contract terms, where it is neither |
Odessa is the only new entry this week and is the reason this radar runs at all; the remaining four rows are carried from edition 30, dated 7 August, and each says so in its own cell. A radar that reprints the same five flashpoints every week without marking which of them actually moved is a habit, not an assessment.
Consumer health dashboard
Six monthly indicators of the American consumer, updated as new releases drop. No new print reached this dashboard in the week to 14 August. All six readings are carried, and every row is dated with the month it belongs to. Auto sales, added last week at an annual rate of 16.3 million, remain the most recent addition. No high-priority flags. The watch item is unchanged and it is the interaction between a negative payroll month and a savings rate at three per cent: households with little buffer and less hiring is the mechanism by which a labour slowdown reaches the shops. A dashboard of monthly series will have quiet weeks, and the discipline is to say which week was quiet rather than to redecorate it.
| Indicator | Current | Prior | Direction | Release |
|---|---|---|---|---|
| Auto sales SAAR (seasonally adjusted annual rate) | 16.3M | 16.5M | Softened, above the 15.0M flag line (carried from 7 Aug) | Jul 2026 |
| Retail sales MoM | +0.2% | +1.0% | Positive; +0.7% ex-petrol. The July release is not yet in this dashboard, so the row holds its June reading (carried from 7 Aug) | Jun 2026 |
| NY Fed 1-year inflation expectations | 3.7% | 3.5% | A three-year high. The July survey is not yet in this dashboard (carried from 7 Aug) | Jun 2026 |
| Conference Board consumer confidence | 91.2 | 91.2 | Above the 85 flag line (carried from 7 Aug) | Jul 2026 |
| NY Fed % worse off than a year ago | 48.0% | 48.0% | Near half of households (carried from 7 Aug) | Jun 2026 |
| Personal savings rate | 3.0% | 3.0% | Above the 2.5% flag line (carried from 7 Aug) | Jun 2026 |
Sources: Cox Automotive / JD Power; US Census Bureau; NY Fed Survey of Consumer Expectations; Conference Board; BEA. All six readings are carried from edition 30, dated 7 August. None of the six is a new week 31 observation. Each is labelled with the month it measures. Two July releases, retail sales and the NY Fed survey, are due and have not yet reached our capture; their rows hold their June readings and say so rather than being quietly advanced a month.
How to check this edition
The Scoreboard is not typed out by hand. When this page loads, the table above fetches the week’s closes directly from our price record and draws itself from what it finds. The numbers you are reading and the numbers in our record are therefore the same numbers, by construction. One honest caveat, because we would rather tell you than be caught: a typed copy of the table also sits in this page as a fallback. If the live fetch fails (an ad-blocker, a corporate network, a printout) you are reading that copy instead. It is generated from the same record and it matched at publication. Closes come from a direct market-data feed taken after the close, never from a search, and every year-to-date figure is recomputed from the baselines we locked on 1 January rather than carried forward, so an error cannot compound week to week.
What is carried this week, in full. In the crash gauge, three of the eight inputs: the MOVE index (from edition 28, last fresh 24 July), the percentage of the S&P above its 200-day average (last fresh 7 August) and insider clusters (from edition 27, last fresh 17 July), each asterisked in A6. In the currency table, the dollar index. In the consumer dashboard, five of the six indicators, each labelled with its release month. In the Stack Inversion working, the memory-pricing, capacity, lithography and foundry readings are carried from last week with no new event, and each is marked as carried in A11. The high-yield spread in A2 is the Thursday 13 August observation, the most recent published, because the index carries a one-day lag; the investment-grade spread beside it is carried from 6 August. Both are dated as such. Frontier’s sources, which this register omitted at publication: Lumentum’s 5 May 2026 earnings call for the thirty per cent figure; Rosenblatt Securities for the supplier capacity series; independent wafer-market research for the $198m substrate figure; the United States Geological Survey’s 2026 Mineral Commodity Summaries for indium tonnage, price and the Chinese share; and LightCounting’s May 2026 newsletter for the counterweight. Everything carried is marked where it appears and logged in our exceptions register for the Monday re-check. Two derived figures, stated plainly: MSCI EM is the EEM ETF (exchange-traded fund) close multiplied by the locked index ratio (28.367); Baltic Dry is the Baltic Exchange reading as published by Hellenic Shipping News.
The Repricing line in the masthead tracks five asset classes: the S&P 500, the Bloomberg Aggregate bond index, gold, WTI crude and the high-yield credit market. Dispersion is the year-to-date gap between the best and worst of the five; the split is how many are up and how many are down. This week the range is just under 35 points (WTI +30.4 at the top, the aggregate bond index −4.6 at the bottom) with four up and one down. That is a stronger reading than last week’s 28 points, and the widening is almost entirely the oil going back up, which is the same lever that narrowed it a week ago. The split changed too: gold crossed back above its 1 January baseline, so what was three up and two down is now four up and one down. Two consecutive weeks in which one commodity moves the reading by seven points in each direction is worth stating plainly, because a dispersion measure that oil can swing on its own is telling you less about five asset classes than the number implies.
Charts and outside sources. Four charts here (the masthead sparklines, the Scoreboard bars, the yield curve and the commodity moves) are drawn in our own house style from the figures above. On the Radar additionally embeds three live TradingView price charts, one per company. We reproduce no third-party chart as an image. Outside research and reporting cited this week: the Bureau of Labor Statistics employment situation release of 7 August; Arista’s and MP Materials’ own results releases filed with the Securities and Exchange Commission; the ISM July manufacturing report; Apollo’s published economic charts; the Japanese Ministry of Finance statement of 3 August; the White House fact sheet of 7 August; the International Rice Research Institute, the FAO and USDA, and a 2023 paper in Nature Reviews Earth & Environment; reporting by Dialogue Earth and Climate Home News on rice carbon credits; TrendForce on memory contract prices; Hellenic Shipping News; and the Gooding Christie’s Pebble Beach catalogue. Where a source is client-only research you cannot open, we say so rather than cite it as though it were public.
The calls. Every directional call is logged at the moment it is made, at the price it was made, and scored twice: once at four weeks to test the timing, and once at a declared horizon to test the analysis. Losses are published with the same prominence as wins. On the Radar entries are what I am watching and why. They are not recommendations to buy or sell.