The Interval

The S&P 500 closed the week at 7,718.60, nine hundredths of one per cent above where it started. That is the whole equity move, in a week in which the oil price rose 9.6 per cent, freight rates rose 13.9 per cent, the entire American yield curve sold off and two countries exchanged fire in the world’s most important shipping lane. Shares did not fall. They did not do anything at all.

Underneath that flat surface, the week did a great deal. Gold fell 2.2 per cent and silver 2.6, in the middle of a Middle East escalation, which is the opposite of what most people expect gold to do and has a mechanical explanation set out in The Week That Was. Freight had its own week, set out above, and finished at its highest of the year. Friday's American jobs report came in at four times the forecast, the strongest month since March, with unemployment unchanged. The numbers, and what is odd about their composition, are below.

The through-line is in the hero quote. August did not resolve the oil crisis; it stopped reporting it. Between 7 August and 28 August crude drifted, the coverage moved on, and this letter’s own energy dial scored the calm as green. The risk had not gone anywhere. It was in an interval, and the interval ended on a Sunday afternoon over a small island most readers had never heard of.

Six of the eight signals in the crash framework are green, the yield curve is still red, and the energy dial has turned amber for the first time since July.

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

The call was the thirty-year Treasury yield at this Friday’s close, against 5.15 per cent. At or above that line, the Treasury’s announcement that it would at least double its purchases of its own long-dated bonds had not put a durable ceiling over the long end. Below it, saying so had done more work than I credited.

The thirty-year closed on Friday at 5.24 per cent. It went up, not down, and the call is correct. It was in the money by seven basis points, which are hundredths of a percentage point, when it was made, and it finished nine of them clear. So I want to be careful about how much credit I take. This was not a close-run thing that broke my way. It was a level already crossed, which stayed crossed. What it establishes is narrow and worth having. Announcing an intervention at the long end of the American government bond market, in the current environment, buys you a single session and then the market goes back to what it was doing. The buying itself begins on 9 September, and that is a different test, one this letter has not yet made a call on.

The full working, every weight, every threshold and the arithmetic that reaches twenty, is in Appendix A6. A subscriber with only this page can reproduce the number.

2026 Thesis Check-In

Every sixth edition this letter checks its annual thesis against the evidence, and this is the sixth. The claim made in January was that 2026 would be the Year of the Repricing: that the spread between asset classes would matter more than the direction of any one of them, and that by December at least three of the five pre-registered proxies would have moved in different directions.

Here are the five, each a 4 September close measured against its locked 1 January baseline, and a reader can do the arithmetic without leaving the page. Crude oil up 44.75 per cent. The S&P 500 up 12.75. Gold up 2.04. High-yield corporate bonds up 1.29. The aggregate bond index down 5.04. Best minus worst is a spread of 49.8 percentage points, the widest of the year, against 36.6 at the last check-in six weeks ago.

And the split is four up, one down. That is the part I am not going to dress up. A wide range on a lopsided split is a weaker result than a narrower range with genuine two-way dispersion, because one asset can supply almost all of it. This quarter one asset did. Strip crude out and the remaining four sit inside eighteen percentage points of one another. Our own register scores that split weak and inconclusive, and I am not going to upgrade it here. The honest word is directional, not confirmed. The thesis is winning on its own terms and it is winning on one barrel. That is not the same claim, and December will decide which one was true.

The demand collapse that nobody measured

Through August the market worked from a settled premise: global oil demand had fallen away, the shipping crisis had been absorbed, and the price was drifting back to where it belonged. Crude fell from 87.06 dollars on 21 August to 83.44 on the 28th, and this letter scored the energy dial green on both its windows. That premise is the thing that did not survive the first week of September, and the reason it did not survive is more interesting than the strike that broke it.

Almost nobody measures global oil demand. It is inferred, and the most common way to infer it is from how hard refineries are running, because refinery throughput is observable and consumption is not.

Imagine estimating how much a household eats by watching how often the kitchen light comes on. For years the correlation is excellent, because the kitchen is where the food is. Then the family buys a chest freezer for the garage and starts eating out of it. The kitchen light goes on less, the meals do not stop, and your indicator now tells you the family has lost its appetite. You will keep believing that, with the numbers to back it up, right until the freezer is empty and they all walk back into the kitchen at once.

The garage freezer, in oil, is refined-product inventory held outside the developed economies. Satellites can count crude in floating storage and read the lids of tanks in the countries that publish tank data. Diesel and petrol sitting in inland storage across the non-developed world are not observable in the same way, and that is precisely the pool a rerouting crisis draws down first. The research house Goehring and Rozencwajg made this argument in their second-quarter commentary, and it is their mechanism rather than mine, so it is credited to them.

The verified figures point the same way and are worth more than the mechanism. The International Energy Agency’s own Oil Market Reports through 2026 have world oil demand falling by roughly 1.6 million barrels a day this year, and it attributes that fall to the closure of the shipping lane and to high fuel prices rather than to any loss of underlying appetite. Over the same period the Agency has global supply forecast down 4.3 million barrels a day, to around 102 million. Supply is falling almost three times as fast as demand. Take those two estimates at face value, which is more than this section has just asked you to do, and they still do not describe a market with a demand problem: both sides are shrinking and the sell side is shrinking faster. That is a market quietly getting tighter while the reported numbers make it look slack. And if the demand figure is wrong in the direction argued above, the gap is wider still rather than narrower, which is the useful thing about it.

Which produces the uncomfortable version of the peace trade. The consensus view is that a settlement out there sends oil down. If the inventory read is right, a settlement sends it down and then up, because reopening the route triggers restocking across every buyer who has been running thin at once, into a system with less supply than it had in January. Resolution would tighten the physical market before it loosened it. That is not a prediction of the price, which depends on things nobody can see, and it is a reason to be careful about treating good geopolitical news as automatically good news for the cost of moving goods.

What the rate model says

This letter’s internal read of the Federal Reserve’s next move is Hold, with a hawkish tilt. It moved hawkish this week and it is now close enough to the top of the hold band that one more hawkish input changes the verdict, which is worth stating now rather than after it happens. I am not printing the composite figure. A reader can reproduce the crash-gauge number from Appendix A6, and cannot reproduce this one from anything on the page, so publishing it to one decimal place would be asking for trust rather than offering working.

Two disciplines keep this cautious. The first is that the rate view only moves when two of its three blocks agree, and this week only one did: the market block, on a two-year at 4.37 and a services survey running hot at 55.4. The inflation block is still reading softer on three-month momentum, and the labour block is genuinely split, for the reason set out under Who Pays above: one hot month against a twelve-month average five times cooler. One month does not overturn that average, and the model does not let it.

The second is the Taylor gap, the distance between what a standard policy rule would prescribe and where rates actually are. It stands at plus 1.28 percentage points and is stable. A positive gap means a cut is not yet justified by the rule. That is the discipline that stops this letter calling a cut early, and it has not moved.

Crude oil closed at 83.44 dollars on Friday 28 August. It then closed at 85.76, 90.22, 91.01, 91.30 and 91.48. Five sessions, five higher closes, and the largest single day was Tuesday 1 September, the first trading day after the Larak strike, when the barrel added 5.2 per cent. The week finished up 9.64 per cent, and the four-week move is 17.01 per cent. On the year the barrel is up 44.75 per cent, which makes it the best-performing thing on this letter’s Scoreboard after freight, and the reason the Repricing spread widened by thirteen points in six weeks.

Natural gas rose 3.3 per cent to 2.975 dollars, which is a fraction of the crude move and is the useful contrast: this is a liquid-fuel event, tied to a shipping route, rather than a general energy event. Gas is priced regionally and does not travel through that lane in the same way.

The Baltic Dry Index measures what it costs to hire a ship to carry bulk cargo such as iron ore, coal and grain. It rose from 3,186 to 3,628, up 13.9 per cent, and unlike crude it did not climb in every session: it slipped to 3,157 on the Tuesday, then ran 3,331, 3,488, 3,628. On the year it is up 92.8 per cent, and it is the single largest mover on the board.

Two readings compete and they matter differently. The first is that a shipping crisis in the Gulf lengthens voyages, because vessels route the long way round, and a fleet that is fixed in size carries fewer cargoes per year when each cargo takes longer. Rates rise without a single extra tonne being shipped. The second is that Chinese buying of iron ore and coal has genuinely picked up, in which case the index is telling you something cheerful about industrial demand. The first is a tax; the second is growth. What separates them is the pattern by route: a rerouting effect concentrates in the vessel classes and lanes that pass the Gulf, while a demand effect shows up in Pacific capesize rates. Until that split is on the page with a verified source behind it, both readings stand and neither is chosen.

The American economy added 162,000 jobs in August, against a consensus forecast of 53,000. The unemployment rate held at 4.1 per cent, with 7.0 million people out of work. It was the strongest month since March, and it sits against an average of just 31,000 a month over the preceding twelve. This is the release that has been calling the whole labour picture into question all year, and for one month it stopped doing so.

Read the composition rather than the total. The Bureau of Labor Statistics named food service, drinking places and local government education as the sources of the gain. Every one of those jobs is real and useful, and none of them is the signature of an expansion. An economy adding staff to schools and bars is an economy in which the domestic service sector is doing the work while everything that trades internationally waits to see what happens. The number was hot. The mix was not.

Canadian retaliatory tariffs take effect on 8 September, dollar for dollar against roughly twenty billion dollars of American goods, after talks between the two governments collapsed in the third week of August. This is the second time in three editions this letter has flagged a goods-price channel that most rate commentary is not pricing. It lands four days after this edition publishes and eight days before the Federal Reserve meets, which is an awkward sequence for anybody hoping the September inflation data will be clean.

Gold fell 2.21 per cent to 4,429.80 dollars and silver fell 2.57 to 66.05, in a week containing an exchange of fire between the United States and Iran. Anyone who holds gold as protection against exactly this would be forgiven for feeling misled.

The mechanism is not mysterious and it is not about the war. Gold pays no income, so the cost of owning it is whatever you gave up by not holding something that does. Yields rose right across the American curve this week, which makes gold more expensive to hold. They moved for reasons that had nothing to do with Larak: a jobs number three times consensus, and an oil price that makes a rate cut harder. When the safe-haven bid and the interest-rate cost pull in opposite directions in the same week, the interest-rate cost usually wins over five days and the safe-haven bid usually wins over five months.

The disagreement worth knowing about. Two of the research houses this letter tracks are currently in direct, published conflict on gold. One camp argues the metal is early in a structural revaluation. The other, Goehring and Rozencwajg, argues in its second-quarter commentary that the selling is not finished, on the specific ground that exchange-traded fund holdings have shed only around 145 tonnes against 950 to 1,100 tonnes in previous full liquidation cycles, which would leave the great bulk of the forced selling still to come. Same asset, same week, opposite conclusions from people who have both done the work. This letter’s standing rule in that situation is that the view goes to amber with the tail named, and the tail here is that a Federal Reserve rise in September triggers the liquidation leg the second camp expects.

Elsewhere: Nvidia recovered 5.9 per cent to 230.36 dollars, taking it back to its own ninety-day high. Japan fell, with the Nikkei down 2.09 per cent, and the yen had its sharpest week in months, the dollar buying 156.22 yen against 160.04. Europe was weaker across the board, the DAX down 1.97 per cent and the Euro Stoxx 50 down 1.37. Mitsubishi UFJ, which this letter has had on its books since edition 25, rose 5.2 per cent on the yen and the Japanese rate story, and is now up 19.4 per cent since it was first written up.

One dial moved this week and it moved for one reason: the barrel. Everything measuring credit, volatility and positioning is where it was, which is why a nine and a half per cent move in oil lifts the composite by five points rather than fifty.

Three of the eight readings are carried from earlier dates rather than freshly measured this week: bond volatility, from 27 August; the percentage of the index above its two-hundred-day average, which is a midpoint of a published range rather than an exact print; and insider selling clusters, from edition 27. Every carry is asterisked and dated in Appendix A6.

In China, a digital replica of a livestream seller ran for seven hours and sold 7.7 million dollars of product, which was more than the human it was copied from had achieved (PitchBook, 1 September 2026). Not a demonstration, not a pilot. A commercial session, on a real platform, against a directly comparable human benchmark.

This matters because the substitution argument has spent three years short of exactly this. There is no shortage of studies showing that a model can pass an examination, and a great shortage of cases in which a specific human’s specific economic output was produced more cheaply by a machine and the difference was banked by somebody. Livestream selling is unusually susceptible: the product is a performance, the performance is already mediated through a screen, and the buyer never meets the seller. Most jobs are not like that. It is a genuine data point and it is not yet a trend, and the honest way to hold it is as the first clean example of a category rather than as the start of a wave.

The counterweight arrived in the same week and from the opposite direction. A widely read economics writer argued that full displacement requires software agents to work unsupervised over long horizons without drifting, and that current systems do not do this reliably. Both things can be true: a machine can replace a task that is short, repetitive and observable, and fail completely at a task that requires being trusted alone for a fortnight. The dividing line between those two categories is where the next five years of employment argument will be fought.

Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to raise more than 500 billion dollars of third-party financing for artificial-intelligence compute. Inside the structure is the detail worth your attention: Nvidia will provide residual-value support on up to 25 per cent of an opportunity, which means it is guaranteeing part of the future resale value of its own chips in order to help other people borrow against them.

A residual-value guarantee is what a car manufacturer offers when it wants to make leasing attractive. It is a legitimate and common instrument, and it tells you something specific: at the price and the term on offer, the lender would not otherwise take the risk. When the manufacturer of an asset has to underwrite what the asset will be worth in order for the financing to clear, the plain economics are not clearing on their own. It also creates a circular exposure, because the guarantor is the same company whose sales the financing exists to fund.

Nothing has been executed. No rate and no term have been published, and until they are this is an announcement rather than a transaction. It belongs on the page because of what its structure discloses, not because of its size.

On Sunday 30 August, American forces struck two Revolutionary Guard rocket launchers on Larak Island, a small island inside the Strait of Hormuz, as the Guard prepared to disperse sea mines into the shipping lane. The strike was confirmed on the record by Navy Captain Tim Hawkins, a spokesman for United States Central Command, and reported by CNN, the Washington Post, Al Jazeera and Stars and Stripes. Iran’s state news agency reported at least two people killed.

Iran retaliated with a combined missile and drone operation against the King Hussein and Al Azraq air bases in Jordan, and separately against the United Arab Emirates. Here the reporting diverges sharply and the divergence is the point. Jordan reported intercepting eight missiles. An American source said nearly all were intercepted with no significant impact. Only the Revolutionary Guard itself claims meaningful damage. A description of this as the largest missile attack in months is circulating and it is one commentator’s framing rather than a finding, and it is not carried here.

The formulation that survives checking is narrower and more useful: this was the first exchange of fire between the United States and Iran since late July, in the seventh month of the war. That is the sentence that matters to anyone pricing a barrel, because it re-establishes a tempo. A conflict that had gone six weeks without direct contact has now had contact, and the market spent the following five sessions repricing what six weeks of quiet had been taken to mean.

Two figures now circulating widely are not printed here and should not be relied on anywhere. A 95 dollar oil price said to trigger Chinese intervention, and an October point of no return. Both are forecasts by their authors with no primary source behind them, and both have the shape that makes a number spread: specific enough to sound researched, and unattributable.

Beyond the Strait: the clearance queue

The flashpoint that is not about oil, and which has a longer half-life. China has been delaying export clearances for germanium, quartz and rare-earth magnets bound for Taiwan, and the restrictions applied to Japan during 2026 are increasingly described by officials in both countries as coercion rather than administration. The European Union has responded with a three billion euro programme aimed at securing alternative supply.

The mechanism is worth understanding because it is not a tariff and it is not a ban. It is a queue. An export licence that takes eight weeks instead of two is not a refusal, cannot easily be challenged at the World Trade Organisation, and is deniable in every forum. It also does exactly what a ban would do to a manufacturer trying to hold a production schedule. Control of a chokepoint does not require closing it; it requires being the only one who decides how fast things move through it, which is the same insight that makes the Strait of Hormuz worth a carrier group.

This cuts across a position this letter holds. Mitsubishi UFJ is on the books as a bet on Japanese banks re-rating, and a sustained Chinese squeeze on Japanese industrial inputs is a headwind to Japanese corporate earnings that has nothing to do with interest rates. The position is up 19.4 per cent and the thesis is intact. The cross-current is named because a thesis that only ever collects supporting evidence is not a thesis.

On Friday, shares in BioNTech closed at 103.76 dollars, valuing the company at about 26 billion dollars. Sitting inside it, doing nothing at all, was 16.63 billion euros of cash and investments, a figure the company disclosed in its own second-quarter results on 4 August 2026.

Convert the cash at the euro-dollar rate BioNTech’s own buyback disclosure implies for the quarter, where it reports repurchasing shares at 89.50 dollars equal to 77.85 euros, or about 1.15 dollars to the euro, and the cash is worth roughly 19 billion dollars. Which leaves the market paying somewhere around 7 billion dollars for everything else the company owns: fourteen pivotal clinical trials, five late-stage cancer programmes, 7,807 employees and a partnership with Bristol Myers Squibb worth up to 11.1 billion dollars.

Hold that number against one other. In June 2025 Bristol Myers Squibb agreed to pay 1.5 billion dollars up front, plus 2 billion of further non-contingent payments through 2028 and up to 7.6 billion in milestones, for half of one BioNTech drug, splitting development, manufacturing and worldwide profits down the middle. A large pharmaceutical company with its own scientists and its own lawyers committed 3.5 billion dollars that it owes whatever happens, for a half share of a single asset. The public market is currently pricing the whole of the rest of BioNTech, that asset included, at roughly twice what Bristol Myers committed for the half.

Either the market is wrong or Bristol Myers is. That is the case study.

The archetype: the Studio Producer

Uğur Şahin and Özlem Türeci build the sound and hand the record to someone else to take on tour. It has happened twice already. Their first company, Ganymed Pharmaceuticals, was sold to Astellas in December 2016 for 422 million euros up front and up to 860 million in milestones. Their second company made the vaccine the world took in 2021, and the name on it, and the chief executive who did the interviews, belonged to Pfizer. In March 2026 they announced they will leave BioNTech by the end of the year to found a third company, and in August the board named Guido Oelkers, a commercial operator who has run the Swedish rare-disease firm Sobi since 2017, to take over by 1 February 2027 at the latest.

A studio producer’s skill is the platform, the technique and the room, not the performance. It is an unusually accurate piece of self-assessment, and most founders never make it. It also has a characteristic failure mode, which is too many tracks in progress and a reluctance to scrap the ones that are not working. Hold on to that. It is where this story ends up.

The uncertainty, part one: the decision nobody was paid to make

On Friday 24 January 2020, the Lancet published a paper describing a family in Shenzhen infected with a new coronavirus. One of them had never been near the market in Wuhan. One was a child who was infected and had no symptoms at all. Şahin read it.

His own account, given to a Danish science publication in April 2022, is the sentence this whole case study turns on: the article convinced them a pandemic was coming and that they should act, “regardless of the risk to our company.”

Here is what that risk actually was, from the filings rather than the legend. At the end of 2019 BioNTech held 519.1 million euros of cash. It had lost 179.2 million that year and burned 198.5 million from operations, which is roughly two and a half years of runway before you add anything to the workload. It had been a listed company for fifteen weeks, and the listing had gone badly. It filed to sell 13.2 million shares at 18 to 20 dollars, then cut the deal to 10 million shares at 15 to 16. It priced at 15.00, the bottom of the reduced range, and closed its first day at 14.24, down five per cent.

And there was no customer. Fosun’s money arrived on 16 March. Pfizer’s 185 million dollars arrived on 9 April. The European Investment Bank’s 100 million euros in June, the German government’s 375 million in September. Every single source of funding for the vaccine programme arrived after the decision to run it. By Monday 27 January a team of forty had been told to move.

One correction to the received version, and it matters because a case study that manufactures conflict is worthless. There is no board opposition on the record. None. The largest shareholder is on record as supportive at the time, and Türeci’s own account describes persuasion rather than resistance. Nobody in a position to stop it has ever said it was reckless. The uncertainty here was not internal politics. It was two people committing a company with thirty months of cash to a problem that did not yet have a name, and being right, which is a much less comfortable story than a boardroom argument because it does not tell you what to do next time.

The uncertainty, part two: the one that is still open

In March 2026 BioNTech said it planned to contribute mRNA rights and technology to the founders’ new company “on an arm’s length basis in exchange for a minority stake”, and that a binding agreement was expected by the end of the first half of the year.

In the second-quarter filing of 4 August, five months later and past its own deadline, the language reads: “As of August 2026, discussions regarding the new company and potential contributions by BioNTech are ongoing.” From “we plan to contribute” to “potential contributions”. No name, no defined technology, no stake, no economics.

The founders are, in substance, negotiating the transfer of BioNTech’s technology with themselves, and the deadline they set for settling it has passed. There is no allegation here and none is implied: related-party transactions of this kind are normal, are supervised, and are disclosed precisely so that people can watch them. What is worth noticing is the sequencing. A clean hand-over would have this settled, the chief medical officer’s job filled and the successor in post. This one has none of the three, and the people leaving are the people the technology is being transferred to.

What the money actually did

BioNTech converted a pandemic into one of the largest cash piles ever assembled by a company that had never sold a product of its own. Second-quarter revenue this year was 105.6 million euros, down 59.5 per cent on a year earlier, and full-year guidance was cut on 4 August to 1.6 to 1.9 billion euros, the second cut of 2026. Set that against guided research spending of 2.0 to 2.3 billion and overheads of 0.7 to 0.8 billion, and at the midpoints the company is running an operating loss of something like 1.15 billion euros for the year. First-half net loss was 1.35 billion.

None of which is alarming on that balance sheet. The question is not survival, it is judgement, and the record on judgement is mixed in an instructive way.

Three things were done well. It sold half its best asset to Bristol Myers rather than trying to fund a global lung-cancer programme alone, which is a founder-led company admitting it cannot do the commercial part by itself. It cut its own research budget by 200 million euros in August on stated prioritisation, and said the saving continues into future years. And it started buying back its own shares, 151.6 million dollars in the second quarter at an average of 89.50, which is a considered admission that not every euro has a better home inside the company.

One thing was not. On 28 August 2026, one week before this edition, BioNTech stopped the Phase 2 trial of autogene cevumeran, its individualised mRNA cancer therapy, in patients whose bowel cancer had been surgically removed. An independent safety board found a numerical imbalance in overall survival between the treated and untreated groups and judged that continuing would not change the result. No new safety signal was found. The trial had already crossed its futility boundary, the pre-agreed point at which a study is judged unlikely to succeed, in October 2025.

Ten months elapsed between the boundary and the stop.

The transferable lesson

A platform is a machine for generating options, and an option only pays if somebody exercises it. Fourteen pivotal trials on a 2.2 billion euro research budget is not a strategy, it is a portfolio of unexercised options, and the discipline that decides which ones to abandon is scarcer and less celebrated than the science that creates them. The hardest capital allocation decision is not what to fund. It is what to stop funding, and a company rich enough never to have to choose will usually choose late.

That is the lesson available to anyone running anything, and it costs nothing to apply. Look at what you are still paying for after the evidence arrived, and count the months.

Where this sits

Materiality first, because a case study should be able to say why the company deserves the chair. BioNTech passes on significance and not on growth, and the two should not be blurred. The research budget alone is larger than the entire stock market value of most mid-sized biotechnology companies. Fourteen pivotal trials at once is a scale very few firms of any size attempt. The Bristol Myers deal, on the headline terms both companies published when they announced it in June 2025, is among the largest licensing transactions the industry has done. And the class of drug the partnership is chasing is aimed squarely at replacing Merck’s Keytruda, which on Merck’s own half-year reporting sold 16.40 billion dollars in the first six months of 2026 alone. Revenue, meanwhile, is falling by design, because the pandemic product is over.

This letter ran BioNTech through its standard company framework before writing any of the above, as it does for every company it names. The framework validates three things and only three: the balance sheet, the valuation arithmetic set out at the top, and the materiality. It contradicts any reading in which BioNTech is winning. It is behind on both of its two lead bets, its individualised-therapy flagship stopped last week, and its bispecific antibody, a single molecule engineered to do two jobs at once, trails a rival programme that has an American regulatory decision listed for November. The framework’s verdict is a watch, not a buy, and nothing here is a recommendation to do anything.

Which returns us to the studio. Two people who have twice built the instrument and handed it to someone else to play are about to do it a third time, leaving behind sixteen billion euros, a stopped flagship, an unfilled medical chair and an unsigned agreement about what they are allowed to take with them. It is not a failure of ambition. It may be the most accurate thing they have ever said about themselves.

We are no longer giving this gauge a score out of one hundred. A single number made the situation look more precise than the evidence behind it really was. Instead we ask a simpler question, and answer it with what we can actually see.

What would we have to see before we could say the artificial-intelligence hardware shortage is genuinely starting to break?

Two things would change our view. Neither has happened.

1. Memory prices start falling

Status: not happening yet · last checked 5 September

Memory chips are the clearest place to see whether supply is catching up with demand, because they are made by a handful of firms, sold on contracts and priced in public. Supply is not catching up.

Contract prices are still going up. They rose by roughly 90 to 95 per cent in the first quarter of this year, 58 to 63 per cent in the second, and another 13 to 18 per cent in the quarter just printed. The suppliers are guiding to further increases in the current one.

The rate of increase is slowing sharply. The prices themselves are not falling. Those are different things, and this condition tests for the second one. It is not met.

2. The companies financing the build-out start to struggle to raise money

Status: we cannot measure this properly yet

There is a second way the shortage could break, and it has nothing to do with chips. Somebody has to pay for all of this. If the firms funding new hardware found it harder and more expensive to borrow, that would be an early warning that the boom is running out of money rather than running out of demand.

We do not yet have a measure of that we would trust. So we are not going to say the condition is met, and we are not going to say it is unmet either. It is not yet measurable, and a missing measurement is not the same thing as a reading.

One thing a reader deserves to know this week. As currently defined, this condition cannot see a manufacturer promising to underwrite the future resale value of its own hardware so that other people will lend against it. That is exactly what was announced on the Nvidia financing platforms in The Speed of Now above.

What about the big falls in some of these shares?

Separate question, and we keep it separate on purpose.

Six of the seventeen companies we track are more than twenty per cent below their own high of the past ninety days, measured on our price record to Friday. SK Hynix is down 43.6 per cent, Talen Energy 27.6, Bloom Energy 26.9, Hua Hong Semiconductor 25.5, Vertiv 21.6, and the semiconductor index itself 20.6.

That tells you investors have become more selective. It does not tell you the hardware shortage is ending. Nvidia and Constellation Energy are sitting at their own ninety-day highs in the same week. A view about where the bottleneck is and a description of what the shares have done are two different facts, and averaging them produces a number that answers neither question.

What is happening to the cost of renting the chips?

Some rental prices are falling. The open-market price for an hour on an H100, the workhorse chip of this build-out, closed Friday at 1.267 dollars for capacity that can be taken away from you at any moment, across 53 live offers. On Friday it went below 1.30 dollars for the first time since we began keeping this record. Every print since 14 August has been below our working estimate of what it costs an operator to own and run one of these machines for an hour, which is roughly 1.65 dollars.

That is worth watching, and we are not scoring it. Our cost estimate is not backed by a published build we could defend line by line, so it is shown and it earns no verdict. Three other price series that would sit beside it are recorded as not observed rather than merely missing. The newer chip generations return a usable reading roughly two weeks in eight. And one token-price series printed the same figure five weeks out of six, then moved fivefold, which is what a change in sampling looks like rather than a change in price.

A price we cannot reliably capture is worse than no price at all, because it looks exactly like one.

So what is the picture?

The evidence still points to scarcity rather than to scarcity breaking, and the shape of it is worth holding on to.

Renting the machine is getting cheaper. The components you need to build one are not. That split is the whole state of the trade at the moment, and it has held for four consecutive weeks. The binding constraint is memory, and it did not move this week.

What would change our mind

One. A memory contract price printing down month on month on the published quarterly series. Falling, not merely rising more slowly.

Two. The first capital raise by either of the two large firms financing this build-out that carries structure rather than plain equity, which would say the money is getting harder to come by.

Three, and this one is a market price rather than a break condition. Whether the H100 rental price stays below 1.30 dollars an hour for two consecutive weeks, or bounces back above it.

And what would push it back the other way: a memory supplier withdrawing or delaying capacity it has already announced, or a new bottleneck appearing somewhere these two conditions do not look, which is what happened to every previous version of this list.

Positions changed by this gauge since it began: none. That line stays on the page every week until it is not zero.

What each of the two conditions would take to satisfy, and why this instrument reports conditions rather than scoring them, is set out in Appendix A11.

Every dollar the world produces now takes less than half the energy it took in 1970, so an oil shock today passes through a far less energy-hungry economy than the one it hit in the seventies.

In 1970, on the eve of the first oil shock, the world used 2.50 kilowatt hours of energy for every dollar of output it produced. In 2022, the most recent year in the series, it used 1.21. The improvement did not stop when it stopped being news: 2.34 in 1980, 1.77 in 2000, 1.52 in 2010, 1.36 in 2015, 1.21 in 2022. Think of a car that used to do twenty-four miles to the gallon and now does fifty. Fuel at ninety-one dollars a barrel still stings every time you fill up, and it takes a smaller bite out of the same wage. That is the entire argument, and it is why a barrel at ninety-one dollars in 2026 is a smaller tax on the world economy than a barrel at thirty-five was in 1980, once inflation is taken out of the comparison. Nobody announced it. It was bought by fifty years of unglamorous engineering.

Two limits, because this section is a counterweight and not a comfort. The series measures all energy, not oil alone, and much of the gain came from cleaner electricity and from services replacing heavy industry, neither of which helps a haulier buying diesel. And the improvement has slowed: 1.36 to 1.21 across those seven years is about 1.6 per cent a year, well under the pace of the decades before it. Resilience is not immunity: efficiency took decades and the shock arrives in a fortnight.

On 1 September, most of the staff of LIV Golf worked their last day. More than three hundred people were employed at the league’s peak; sixty-day notices went out in July; a spokesman said the league was scaling back operations as it transitioned to its next chapter. The day before, the Financial Times reported that LIV could seek bankruptcy protection as soon as the week beginning 7 September, which is the week after this edition.

Four months before that, on 30 April, Saudi Arabia’s Public Investment Fund confirmed it would end funding at the close of the 2026 season. Its own words: the “substantial investment required is no longer consistent” with the fund’s strategy, a decision made “in light of PIF’s investment priorities and current macro dynamics”. Reported spending since 2022 is around five billion dollars on Front Office Sports' 1 September tally, on course for six by the end of this year, at roughly a hundred million dollars a month, with thirty million of prize money per event. All of those are press estimates rather than disclosed figures and are carried as such. The league is now in talks with BC Partners Credit for 250 to 350 million dollars of new money.

Roughly five to six billion dollars spent. Roughly three hundred million sought. That ratio is the story: about five pence of new money for every pound already gone.

What the money bought, and what it did not

This column exists to look at assets whose supply is fixed or artificially constrained while the pool of capital chasing them grows. Golf is the inverse case, and it is more instructive than the usual one, because it shows what happens when a very large buyer identifies the wrong constraint.

Nothing about golf is scarce in the way LIV assumed. Courses are abundant. Tournaments are abundant. Elite players turned out to be straightforwardly purchasable, which was the whole premise and it worked. What is genuinely fixed is the licence: world ranking points, eligibility for the four major championships, and the accumulated legitimacy that flows from both. LIV controlled who turned up. It never controlled what turning up was worth. It did not obtain world ranking points at all until 2026, four years after launch, and when they came they were capped and small.

The audience figures are the proof. LIV averaged 386,000 viewers across fourteen rounds on Fox in 2026, up twelve per cent on the previous year. Coverage of the PGA Tour on CBS and NBC averaged 2.5 million across the regular season, roughly six and a half times as many. Both figures are trade-press aggregations of ratings data rather than primary, and are used here as an order of magnitude rather than a measurement.

Meanwhile the incumbent did the opposite trade. In January 2024 Strategic Sports Group, fronted by Fenway Sports Group and including the owners of the Mets and the Cubs, agreed to invest up to three billion dollars into the newly formed PGA Tour Enterprises. The initial 1.5 billion for 11.62 per cent implies a valuation just over 12.9 billion dollars; the Tour’s own stated figure is twelve billion. Nearly two hundred players were granted equity, illiquid and vesting, worth more than 1.5 billion in aggregate. Not one new tournament was added.

Two entities, one sport, one four-year window. One manufactured supply and is heading for a claims process. The other sold a share of the sanction it already held and took an initial 1.5 billion dollars of institutional money at a valuation of twelve to thirteen billion, with a further 1.5 billion committed but conditional on a settlement that never came. The capital went to the one that added nothing.

The precedent, and the honest objection

This has happened before and the pattern is consistent. The Indian Cricket League launched in 2007 with money, players and the same format the Indian board then adopted itself; the board withheld recognition, banned participants, and the league was gone by 2009. World Series Cricket in 1977 is the exception that proves it: Kerry Packer signed the best players in the world and it worked, but it worked as leverage, and he settled in 1979 for exclusive broadcast rights and disbanded the competition. Manufacturing supply has never been a standing business in sport. It has only ever been a way of buying the sanction, and that is exactly what the never-consummated 2023 framework agreement was reaching for when the PGA Tour rejected the fund’s 1.5 billion dollar offer in April 2025.

Now the objection, which is serious and I am not going to bury it. The withdrawal is over-explained. Everything above is consistent with the structural reading. It is equally consistent with a simple fiscal one: an oil price that spent much of the period below seventy dollars, a large Saudi budget deficit, and a fund cutting capital spending across its whole portfolio. Golf was one line in that portfolio among many, and most of the others have no sanction dimension at all. Both readings predict precisely what happened, which means the evidence does not choose between them.

What would choose between them, inside a year: if the sanction reading is right, LIV stays impaired even with new capital, because 250 to 350 million dollars buys tournaments and does not buy eligibility. If the fiscal reading is right, the league is viable at a lower cost base under a sponsor with a healthier balance sheet. Watch which one happens.

Two further concessions the record demands. The wall was porous rather than absolute: LIV players played the majors throughout on exemptions, and one of them won the 2024 US Open. And the fund has not left sport. It reaffirmed sport as a priority sector in the same statement, and is reported to be in talks on a boxing venture valued at four to five billion dollars, structured around bringing the sport’s four sanctioning bodies together, with the fund taking a minority stake. That is the same money buying the sanction instead of manufacturing the supply, one league later. Whether that is a lesson learned or a coincidence will take about four years to establish.

What scarcity is worth when you own it

The prices in the sanctioned leagues, over the same period, went the other way. The Seattle Seahawks sold in July 2026 for 9.61 billion dollars, a record for a controlling sale of a National Football League club on Sportico’s tally. Sportico puts the average NFL franchise at 9.34 billion, up thirty-one per cent in a year. The Los Angeles Lakers were sold in October 2025 at a ten billion dollar valuation, approved unanimously and closed. Ten months later, in August 2026, they were agreed again, to a group led by Josh Kushner and Bob Iger, at 12.5 billion. That second sale is not yet done. It needs league approval at a meeting on 15 and 16 September, it requires the buyer to divest a stake in another club, and it is being contested by a member of the selling family who says she never agreed to it.

Non-correlation, the property this column keeps testing, is not a property of sport. It is a property of sanctioned fixed supply. Thirty-two closed franchises with a broadcast cycle and a rule change admitting new money behave one way. An open-ended cost centre with no eligibility behaves another, and it went to zero in the same twelve months in which the first went up thirty-one per cent. That is the distinction, and this year supplied both halves of it at once.

Three pillars to the Augmenter and one to the Displacer, on four separate observations, and the running total goes to 41 against 19.

Pillar one, how fast this is being adopted. The Augmenter. A widely read economics writer set out the case on 1 September that full displacement requires software agents to run unsupervised over long horizons without drifting off task, and that current systems do not do that reliably enough to be left alone. That is friction in the adoption curve rather than in the technology, which is exactly what this pillar is built to detect. It is the difference between a tool that is capable and a tool that can be trusted while nobody is watching, and only the second one replaces a salary.

Pillar two, what is happening to jobs. The Augmenter. August payrolls came in four times consensus, with unemployment unchanged, as set out earlier. This is an economy-wide measurement rather than an absence of evidence, and it points one way. The strongest contrary item of the week is genuine and is named in The Speed of Now: a digital replica outselling the human it was copied from in a Chinese livestream. It is one seller in one niche, measured against a national payroll survey, and it is not scored here for that reason. It is the class of evidence that will eventually win this pillar, and it has not won it yet.

Pillar three, what kind of shock this is. The Displacer. American private non-residential construction spending is contracting once data centres are stripped out, with data-centre construction up 46 per cent while everything else shrinks (PitchBook, 31 August 2026). That is capital concentrating in one theme while the rest of the built economy goes backwards, which is the demand-side concern this pillar exists for: investment that shows up in aggregate figures while the activities that employ most people are quietly getting smaller.

Pillar four, the compute ceiling. The Augmenter. Nvidia’s financing platforms are intended to mobilise more than 500 billion dollars of third-party capital, and Nvidia has said it will provide residual-value support on up to a quarter of an opportunity. A vendor offering to underwrite the resale price of its own product in order to get third-party lending across the line is evidence that at current costs the economics do not clear unaided. The compute ceiling is not gone. It is being papered over with credit enhancement, which is a different thing and arguably a more fragile one.

Week 34: the Augmenter 3, the Displacer 1. Running total: the Augmenter 41, the Displacer 19. That total runs from the reset in Week 20 and nothing before it is counted. Fifteen weeks at four pillars a week is sixty, and 41 plus 19 is sixty, so the ledger reconciles. Four pillars, four separate observations, none used twice.

What would flip the score. A named large employer attributing a specific headcount reduction to artificial intelligence in a filing hands the Displacer the labour pillar outright. A memory contract price printing down month on month hands the Augmenter’s compute ceiling to the Displacer and would probably take the shock pillar with it. And evidence that construction outside data centres has turned back up would take pillar three straight back.

What the Displacer and the Augmenter are, and how the four pillars are scored, is set out in Appendix A13.

No new company is added this week, and both open calls stay where they are. Neither had a qualifying catalyst: a named, company-specific event in the past seven days that has not already been used to justify a previous appearance. A share price moving the right way is not a catalyst; it is the definition of a position you already hold.

Both calls, their entry prices, their benchmarks and where they currently stand are set out in the Portfolio Watch table further down this edition, between the Scoreboard and the appendix.

The candidate that did not make it

Something worth showing, because the work that does not produce a call is usually invisible and this week it was the most useful work done.

A company came through this letter’s bottom-up screen with a profile that is exactly what the screen exists to find: a highly ranked, fast-growing medical business, revenue up more than twenty per cent, a record quarter, raised full-year guidance, and a share price a third below its high. On the surface, a good growth company having a bad year in the market. The mandatory pre-publication work is four steps and takes several hours, and it exists precisely so that surfaces do not decide anything.

The first three steps found a thin competitive moat and no technical setup, which would have made it a watch rather than a call. The fourth step killed it outright, and not for anything in the business. It failed on the quality of its own financial reporting: the audit and disclosure-control review did not clear, and the company is therefore not a candidate at any price until it does.

That is a statement about whether the numbers can currently be relied upon, not about whether the business is any good, and the two are genuinely separate questions. The company is not named here, deliberately, and the detail is not published either. There is no call attached to it. Naming a company in order to publish its problems, while declining to publish a view on it, is the least defensible use of a name that exists; and setting out those problems in detail while withholding the name is not better, because a reader cannot check what they cannot identify.

The transferable habit is the whole point of telling you. When a growth story looks cheap for no obvious reason, the reason is often filed in the least-read section of the annual report. Read the auditor’s opinion on internal controls before you read the revenue growth. It takes ninety seconds and it is the highest-yielding ninety seconds in the document.

The pipeline, said plainly

It has now been nine editions since this letter opened a new scored position, against a standing target of one every three. That is a real gap and it is recorded here rather than left for a reader to notice. The screen this week produced two candidates through two independent routes, and both were declined on the evidence, one of them for the reasons above. A dry pipeline caused by discipline and a dry pipeline caused by inattention look identical from the outside, which is why the count is printed.

Both open calls are scored at their thesis horizons, which are in 2027, and the running record on both horizons is published in full at each quarterly reckoning rather than in fragments each week. Neither call is a recommendation to buy or sell anything.

Somebody at the Baltic Exchange has had a very good fortnight, and almost nobody has written about it.

The index that measures what it costs to hire a ship has risen 92.8 per cent this year, which makes it the best-performing line on this letter’s Scoreboard by a distance of forty-eight points. It has beaten gold, it has beaten every stock market on the board, it has beaten the oil that half of it is arguably about, and it has done so while attracting roughly none of the attention that any of those received. There are no Baltic Dry exchange-traded funds advertised on the underground. Nobody has written a book about the index. It is a daily survey of what shipbrokers say it costs to move rocks across water, and this year it has quietly outperformed everything anyone did write a book about.

Elsewhere in the week’s small print: two of the assets on this letter’s Scoreboard printed a Monday close identical to the previous Friday’s, to the penny, and an automated check flagged both as probable feed errors on the ground that comparable venues had traded normally. They had. Monday was the August bank holiday, and the London and Baltic exchanges were shut. The machine had every price it needed and no calendar, which is a reasonable description of a great deal of financial analysis.

We had filed the oil crisis as done,
and the rate cut as all but begun.
Then a strike in the strait
and a jobs print too great,
so the filing has come all undone.

Three things to watch next week. The two-year Treasury at Friday’s close, against the 4.45 per cent this letter has put its name to. Canadian retaliatory tariffs taking effect on 8 September, dollar for dollar, four days after this edition and eight days before the Federal Reserve meets. And the enlarged Treasury buyback operations beginning on 9 September, where the buying itself is finally tested rather than the promise of it. If all three move the same way, meaning a two-year above 4.45, tariffs landing as announced and the buybacks failing to hold the long end, then the market is pricing a central bank that has to tighten into a supply shock it cannot fix, and that is the least comfortable configuration available. If they diverge, and the likeliest divergence is a two-year that stalls below the line while the long end steadies, then the oil move is being read as a tax on growth rather than a source of inflation, and the whole rate argument in this edition needs rewriting from the top. One of those two weeks is coming. We find out on Friday.

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

Freight and crude now hold the top two places on the board, at plus 92.8 and plus 44.8 per cent, and the two of them moved 13.9 and 9.6 points in a single week. The gap between best and worst across the twenty-six lines is 110.0 points, up from 87.3, and almost all of the widening came from one shipping lane.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 4 September 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. Five rows are named by an abbreviation: MSCI ACWI (All Country World Index) is the broadest global share index; AGG is the US aggregate bond market; LQD is investment-grade corporate bonds; HYG is high-yield, the riskier corporate borrowers; and TLT is long-dated US government bonds. Nothing else is added, dropped or substituted mid-year, which is the only way a year-to-date table means anything. The single-company calls in Portfolio Watch are tracked separately. Annualised volatility shows how much each line typically swings in a year, judged from its last eight weeks: higher means bumpier, not worse, and a dash means we do not yet have enough weeks to measure it honestly.

Every close here is drawn from the same price record the table itself reads from, so the numbers on this page and the numbers in our record cannot drift apart. All closes are Friday 4 September 2026. MSCI EM is derived from the EEM exchange-traded fund close of 68.70 multiplied by the locked index ratio of 28.367, which gives the 1,948.81 published above; the fund price and the index level are two scales of one series, and the ratio is locked and never re-derived. Baltic Dry is the Baltic Exchange Dry Index as published by Hellenic Shipping News, 4 September, 3,628. 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.