The Standard

The S&P 500 finished the week up 0.49 per cent at 7,711.76, recovering last week’s fall, and spent Friday morning higher still before a speech in Wyoming turned it around. Gold fell 2.04 per cent on the week and silver 2.42, both selling off hardest on Friday, alongside the index itself, the largest chipmaker, semiconductors as a group and bitcoin. Six things that usually disagree with each other went down together in the same afternoon.

Nvidia reported on Wednesday evening, and on the revenue and guidance figures in its own results release of 26 August it is the strongest quarter of the artificial-intelligence cycle so far. The shares rose 8.7 per cent on Thursday, then gave back 4.57 on Friday. The semiconductor sector finished the week down 1.30 per cent despite it. And on Friday morning the Bureau of Labor Statistics said its preliminary annual benchmark indicates American employment in March was overstated by 79,000 jobs, where economists surveyed by Bloomberg had expected a revision upward of 183,000.

The through-line is in the hero quote, and it runs through three separate sections below. Crude oil fell more than four per cent while the diesel refined from it set an all-time record on 17 August and has not come back down. The volatility index printed 14.43, its second-lowest reading of the year and two tenths off the low it set a fortnight ago, while the gap between it and the volatility of individual shares reached a record high in July. Even the company in this week’s Case Study reports a pre-tax margin of seventy-seven per cent, which is not a number most screens will show you, because the listed company reports the whole business and holds only part of it. In each case the number being reported is not the number that matters.

Seven of the eight signals in the crash framework are green, and the one that is red has been red for months.

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 test was narrow and it has an answer. The two-year Treasury had to be read against 4.35 per cent, and a close at or above that line would say Jackson Hole had turned the Federal Reserve hawkish and the front-end easing story was over. The two-year closed Friday at 4.34. One hundredth of a percentage point below the line.

So the tell did not fire, and I am going to score it as exactly that. It also did not matter, because everything the tell was designed to detect happened anyway, in front of an audience, on the record, at eleven o’clock on Friday morning. A line missed by a basis point is still a line missed, and the honest version is that the instrument was too finely set rather than that the reading was wrong.

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

The standard, and the things he did not say

The speech is called In Our Time and the Federal Reserve has published it in full, which means anyone can check this rather than take my word for it. The sentence that repriced the curve was not a threat. It was a condition: “Here is my standard: We must be confident that underlying inflation is moving to our objective, clearly and at sufficient speed. Otherwise, we have work to do.”

Then he set out, in the Fed’s own numbers, why he is not confident. The personal consumption expenditures index, which is the inflation measure the Fed actually targets, is running at 3.7 per cent over twelve months and 4.1 over six. A six-month reading above a twelve-month reading means the recent trend is worse than the year, not better. Fifty-four per cent of the 199 components the Fed tracks are rising at more than three per cent a year, against a pre-pandemic average of thirty-two. On the summer’s better inflation prints, which many people including this letter had read as encouraging, he was blunt: they do not tell him that underlying trends have meaningfully improved.

And then the line that did the real damage to the rate-cut case. He would be hard pressed, he said, to describe broad financial conditions as restrictive. Credit spreads near historic lows, banks easing lending standards, equities near highs. The argument is that policy is not squeezing anything, so there is nothing to release.

Now the more useful exercise, which is reading what is absent. I went through the published text and its eighteen footnotes looking for the things a chairman in his position might be expected to address. Three are missing.

Tariffs. Not once. Not the word, not the concept, not the American and Canadian talks that collapsed on 21 August and whose retaliation lands on 8 September. A chairman explaining stubborn goods inflation did not mention the policy that is raising the price of goods.

The Fed’s independence. Not addressed. The whole of his engagement with it is seven words, “especially in light of recent developments”, and he never says what the developments are.

A reaction function. Explicitly refused, and this is the part that ought to change how professionals model him. He wishes, he said, that our understanding of the economy were precise enough that some simple rule could be rigorously relied upon, but our knowledge does not extend that far. That deserves a sentence of honesty from me, because this letter runs an internal rate model anchored to exactly that kind of rule. The chairman has now said in public that it is not how he decides. Our model currently reads Hold with a hawkish tilt, which happens to be right, and its anchor just got weaker, and you should hear that from me first.

The fuel, not the crude

Here is the thing I think is being under-weighted, and it runs against the comfortable reading.

Crude oil fell 4.16 per cent over the week, to 83.44 dollars. On that basis energy inflation looks like it is fading, and every dashboard that reads oil prices, including ours, scored it benign.

Eleven days earlier something happened that those dashboards do not see. On 17 August the American diesel crack spread, which is the margin between crude oil and the ultra-low-sulphur diesel refined from it, reached an intraday record of 102.20 dollars a barrel on the US Gulf Coast benchmark against WTI, the first time that spread has ever exceeded a hundred. The previous peak was in the high eighties, set in October 2022 after the invasion of Ukraine. It has stayed around or above a hundred since, on the most recent published reading; we hold no verified Friday 28 August print for the spread, and say so rather than carry the record forward as if it were today’s number. Distillate inventories, on the Energy Information Administration series, stood at 103.4 million barrels for the week ending 21 August, the lowest level for the time of year since 1996.

Why the refined product matters differently from the crude: a crack spread is a market-implied refining margin, and it tells you the tightness is in the fuel rather than in the oil. Diesel is what moves freight, farm machinery and most shipping, so a sustained record margin raises distribution costs across a wide range of goods. Whether it reaches shop prices, and how fast, depends on inventories, refinery utilisation, imports and how quickly the margin normalises. It is a signal of product tightness, not an automatic tax.

So the picture is this. A central bank has told us it needs to see underlying inflation falling clearly and quickly. It is looking at a crude price that fell over the week. Meanwhile the refined fuel that actually carries energy costs into the shops set an all-time record eleven days ago and has not come back down. The bottleneck moved downstream, and the headline energy number does not register it.

What the rate model says

Our internal reading of the Federal Reserve’s next move is Hold, with the bias moved to neutral-to-hawkish, and the two-year at 4.34 confirms that rather than contradicting it. The composite sits at 9.2 on a scale running from minus 100, a cut being certain, to plus 100, a rise being certain, with anything between minus 15 and plus 15 read as a hold. It was 3.8 a week ago, and that is the largest single-week move it has made.

Sitting behind it is the Taylor gap, which is worth one sentence of explanation because it is the discipline that stops this letter calling a cut too early. It is 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, meaning a cut is not yet justified by the rule, and that was true before the chairman told us he does not use the rule.

The two-confirmation discipline this letter uses says the rate view moves only when two of the three blocks agree. Inflation trajectory and market pricing both moved hawkish this week. Labour is the block that did not, and it is the one to watch, for reasons The Week That Was sets out. Both things cannot stay true for long, and that tension is next week’s story.

Forced Flows

On Friday alone, and these are single-day moves rather than weekly ones: gold fell 2.29 per cent, silver 3.37, bitcoin 3.03, Nvidia 4.57 and semiconductors as a group 3.47, while the S&P 500 gave up a morning gain of about half a per cent to finish the day lower. All six moved down inside the same afternoon, and the first of them turned within twenty-five minutes of Warsh beginning to speak in Wyoming.

When gold falls with shares, the reflex is to call it a liquidation, meaning forced sellers dumping whatever they can. There is a simpler explanation available, and a competing one, and it is worth holding both.

Gold pays you nothing. Its only cost is the interest you gave up by not holding cash instead. When a central bank sounds more likely to raise rates, that forgone interest rises and gold becomes more expensive to hold. The same rise in rates makes future company profits worth less today, which pushes down shares, and it strengthens the dollar, which pushes down everything priced in dollars. One cause, several victims. The dollar index rose 0.38 per cent on the day, and gold stopped almost exactly on its two-hundred-day average of 4,527, closing at 4,529.90.

The competing explanation is positioning, and it is not weak. Gold miners had risen nearly forty per cent in August and were on course for their best month since 2020. A crowded trade unwinding on any pretext looks identical, on the day, to a repricing driven by interest rates.

The tape is consistent with a discount-rate shock and does not on its own prove one. Here is the test that separates them over the coming fortnight. If gold recovers while real yields stay high, it was positioning. If gold stays down with yields, it was the cost of money. That is checkable, and the test is more use to you than the verdict.

Who Pays

The Bureau of Labor Statistics published its preliminary annual benchmark on Friday morning, the exercise that compares the monthly employment survey against a count drawn from tax records. It indicates that the March 2026 level of American employment may have been overstated by 79,000 jobs, with private employment overstated by 178,000. Economists surveyed by Bloomberg had expected an upward revision of 183,000.

Two things this is not, because the distinction does real work. It is not a revision to the published monthly figures: those are unchanged, and the final benchmark is not incorporated until February 2027. And it is not a measurement of job losses. It is the gap between two independently produced counts, each with its own errors, and what it establishes is measurement risk rather than a weakening labour market.

Even read that carefully, the tension is real. The Federal Reserve chairman told an audience that same morning that labour markets are consistent with full employment, and he has good grounds: unemployment is 4.1 per cent and claims are near multi-decade lows. But the market spent Friday pricing rate rises off the inflation half of the Fed’s mandate on a day when the reliability of the employment half was called into question. That is a strange combination to hold for long.

Where Value Is Captured

Nvidia reported on Wednesday evening. Adjusted earnings of 2.22 dollars a share against a sell-side consensus of 2.09, revenue of 96.22 billion dollars, and guidance for roughly seventy per cent revenue growth into its 2028 financial year, all as stated in the company’s own results release of 26 August. It is the strongest print of the artificial-intelligence cycle so far. The shares rose 8.7 per cent on Thursday, the best day for both the S&P 500 and the Nasdaq since 4 August.

They then fell 4.57 per cent on Friday, giving back about half of it, and the semiconductor sector finished the week down 1.30 per cent.

I searched the Friday tape and the company newsflow and found no Nvidia-specific bad news that day, which is not the same as establishing there was none. What there was, was a higher discount rate, which hurts richly valued growth companies more than anything else. The clean way to state it: Nvidia’s results did not change, the financing backdrop did. The best quarterly print in the cycle bought the sector a single day and left it lower on the week. If you want one fact that captures how much of this market is a bet on the price of money rather than on the technology, that is the fact.

Capacity

The Baltic Dry Index, which measures the cost of shipping dry bulk cargo, rose 12.14 per cent on the week to 3,186, the largest weekly move on our board. Crude oil fell 4.16 per cent in the same week.

That looks like a contradiction and is not. The move is almost entirely in Capesize vessels, the largest class. Their average earnings rose 5,313 dollars a day to 46,601, and that is a Thursday 27 August fixture rather than a Friday one, because the daily Capesize print lands a session behind the headline index. The index itself, at 3,186, is Friday’s and is verified. Capesize means iron ore and coal. The named cause, from the Baltic Exchange’s daily fixture commentary, is Pacific iron ore demand from the major miners combined with charterers rushing to secure vessels ahead of a tropical depression in the northern South China Sea, following a typhoon that did the same thing a fortnight earlier.

So the move is concentrated in one vessel class, one commodity route and two weather events. That is tightened supply of ships rather than broad new demand for cargo, and the fixture data is specific enough to show it. A double-digit weekly rise in shipping rates reads like an industrial recovery and, on this week’s evidence, is not one.

Cost Structure

One sized data point on an asset class with nothing to do with any of the above. The Los Angeles Lakers were agreed for sale on 12 August at 12.5 billion dollars, the highest price ever agreed for an American sports franchise, to a group led by Josh Kushner and Bob Iger. Mark Walter agreed to buy the club from the Buss family at a ten billion dollar valuation in June 2025 and completed the purchase on 30 October 2025, after the league approved it. So the holding period is about ten months on a completed basis, or fourteen from one agreement to the next. Either way the same asset has repriced twenty-five per cent, and Walter takes roughly 2.5 billion dollars.

For scale: Jerry Buss bought the Lakers for 67.5 million dollars in 1979, and the Boston Celtics sold last year for just over six billion. The sale still requires approval from the league’s board of governors, which next meets in September, so it has not completed. We will return to what fixed-supply trophy assets are telling us when it does.

Three of the eight readings are carried from earlier dates rather than freshly measured this week: the percentage of the index above its two-hundred-day average, measured 7 August; insider selling clusters, from edition 27; and the high-yield spread, which is an August level rather than a precise Friday print. Every carry is asterisked in Appendix A6.

Commoditisation

Nvidia is reported to be in talks to buy Hugging Face for about 12.9 billion dollars. The Information broke it on 26 August and it has been picked up widely. Neither company has confirmed it and the talks are reported as not yet producing a signed agreement, so treat it as reported rather than done.

Hugging Face is where open-source artificial-intelligence models are published and downloaded. It is the distribution point for the free alternatives to the paid frontier systems. Nvidia, which sells the chips those systems run on, took part in its funding round three years ago and would now be buying it outright. On the revenue figure reported alongside the deal, of roughly 150 million dollars annualised, the price implies a multiple in the region of eighty times sales, which is a number to treat as reported rather than established.

Read it either way and it is interesting. If open models are going to commoditise the expensive laboratories, the chip company that sells to everyone benefits and is buying the distribution channel for the winner. If they are not, it has spent thirteen billion dollars insuring against something it does not believe. Nvidia’s own customers, the large laboratories, are designing their own silicon to reduce their dependence on it, which makes this look less like an acquisition and more like a hedge.

Enforcement

The bigger change in Friday’s speech was not about rates at all, and it has been largely overlooked. Warsh argued that forward guidance, the practice of central banks telling markets what they intend to do next, has overstayed its welcome.

If markets take their cue from the Fed, and the Fed takes its cue from market prices, then both are reading a mirror. He called it a hall of mirrors and cited the academic work behind it. His conclusion is that a quieter central bank sees the economy more clearly, because it is no longer looking at its own reflection.

Then he made the point most commentary skipped, and it is the one that matters for anyone who does not trade for a living. Market participants, he said, are unlikely to bear the biggest costs of the hall-of-mirrors problem. The most serious harm is likely to fall on those without financial assets, who are left to deal with inflation that is too high or jobs that suddenly look less secure.

The practical consequence: the question stops being what the chairman has promised, and becomes what reaction function you can infer from how he responds to data. That is a harder job and a less comfortable one, and it is now the job.

The Strait of Hormuz, which carried about a fifth of the world’s oil and liquefied natural gas before the war, remains closed. On Tuesday the Iranian and Omani foreign ministries announced an interim framework for resuming ship transits. On Wednesday a Revolutionary Guard spokesman said agreements had been reached on each country’s share of the strait’s waters and its revenues.

Those are not the same statement. The diplomats described a framework and did not mention fees; the military described a completed revenue split. Iranian officials separately said any navigation deal would not mean immediate reopening, because normalisation is conditioned on the United States lifting sanctions and a naval blockade.

Crude fell more than four per cent on the week partly on this. The market has priced a reopening that the two Iranian statements do not agree has happened.

In the autumn of 1977, a few months into his career as a market maker on the floor of the American Stock Exchange, a thirty-three-year-old Hungarian immigrant heard a trader offer three hundred call options on DuPont, slightly out of the money, two days from expiry. Nobody bid. His own pricing sheet said they were worth about twenty dollars. He bid 12.50.

A moment later another trader offered to buy five hundred at 37.50. He said sold, selling five hundred contracts, two hundred of which he did not own. Then DuPont halted. The news, when it came, was an earnings beat and a three-for-one stock split. When trading resumed, the options he had sold at 37.50 were worth 450.

Thomas Peterffy lost 75,000 dollars in a matter of minutes, close to half of everything he had.

Who he was at that moment

He had arrived in New York on 12 December 1965, aged twenty-one, with no English. His first job was drawing road maps for a highway engineering firm in Queens for sixty-five dollars a week. When the firm bought a programmable calculator, an Olivetti Programma 101, he volunteered to learn it for a reason he still gives: he figured it would be easier to learn than English.

Ten years of programming later he had saved 200,000 dollars. He spent 36,000 of it on a seat on the American Stock Exchange and deposited most of the rest, about 165,000, as trading capital. The DuPont trade took close to half of that within months. It took him until 1982 to rebuild it. He gave up smoking the same night, and he stopped speculating entirely.

The conclusion he drew is the one that matters, and it is not the obvious one. He did not decide he needed better information. He decided that human judgement standing in a pit was the thing that had failed, including his own, and that the answer was to remove it.

The archetype

Peterffy is the invasive species. Not a predator that hunts the incumbents, but an organism that arrives from outside the ecosystem, thrives on conditions the natives take for granted, and has replaced them before they have worked out that a competition is under way. Everything that follows is that pattern.

The guild, and how it defended itself

What happened next is usually told as a story about technology. It is better understood as a story about rules.

In 1983 he built handheld touchscreen devices, roughly the size of a hardback book, gold wires sandwiched between transparent plastic, and put them in the hands of his traders on the exchange floor. A clerk ran each device two blocks back to the office every few trades to be reloaded. One of them caught fire in the pit. His competitors objected, not that the devices were unfair, but that they had sharp edges in the crush of the crowd. He rounded the corners.

The Chicago Board Options Exchange passed a rule that analytical devices may not be used on the trading floor. Peterffy’s response, in his own words: how can you say such a thing?

The New York Stock Exchange in 1985 permitted screens, but only mounted on the back wall, thirty feet from the pit. Sitting at his kitchen table with a mug of coloured pencils, he re-encoded every digit as a colour, so his traders could read fair values across a crowded room. They learned the language in a day or two.

Then in 1987 he spliced into a Nasdaq terminal’s data line, ran the prices through his algorithms and sent orders back down the same cable: the first fully automated trading system on Wall Street. An official visiting a fast-growing client walked in, found a computer trading with no human present, and telephoned the next day. The rule book said orders must be entered via the keyboard. He had one week to comply.

His team worked every night for a week. They mounted a camera above the screen to read prices and built a frame of metal arms and motors over the keyboard: rubber fingers that typed, he says, like a machine gun. The official came back, watched it for a while, and left without a word. Staff wore ear plugs.

Notice what none of those objections were. Not one was an argument about market quality, or fairness to investors, or systemic risk. Sharp edges. A rule against analytical devices. Thirty feet. Enter it via the keyboard. Each is a procedural defence of an information advantage, written to look like safety. Each was satisfied to the letter and defeated in substance.

The uncomfortable second act

The version usually told next is that he heroically killed his own profitable business to build the better one. That is not what the record shows, and the truth is more interesting.

He founded Interactive Brokers in 1993 to run brokerage alongside the market-making arm. At the 2007 flotation, fourteen years later, market making still generated eighty per cent of group revenue. He floated only about a tenth of the company, using a Dutch auction rather than a bank syndicate, because he wanted advertising rather than capital. He announced the closure of the market-making business only in March 2017, and sold it that October.

Why then? Because by the mid-2000s the contest had moved from analysis to speed, and rivals were spending hundreds of millions on microwave towers. His account of the decision: he thought it would cost him billions of dollars. And, more revealingly, that he knew everything there was to know about market making and it no longer interested him.

His edge was never the automation. It was the period during which he was the only one automated. When everyone had the same tools, the advantage moved to a kind of capital expenditure he did not want to make, and he left. The displacer became the displaced inside twenty years.

Which raises the obvious question, and it is worth answering rather than leaving. If the head start was the whole moat, the successor business should have been commoditised too. It has not been. Its pre-tax margin went from seventy-five per cent to seventy-seven in the last year while customer accounts grew thirty-four. The second moat is a different animal from the first: very high fixed costs, almost no marginal cost per additional account, and regulatory permissions in more than a hundred and seventy markets that are slow and expensive to replicate. A head start expires. Scale does not.

What generalises

The transferable lesson is the third act, and it is the part almost nobody tells.

The floor traders he had spent two decades routing around did not disappear into unemployment. They needed a way to keep working from an office instead of a pit, and they trusted him. In his words: they knew he was an honest business person, and they needed a way to continue their business on a computer from their office, so they became his customers, and that is how Interactive Brokers became the broker for professional traders.

The people the machine displaced were the machine’s first market. That is worth carrying into any argument about automation and jobs, in either direction.

The scale, sized properly

Interactive Brokers reported its second quarter on 21 July 2026. Pre-tax profit margin of seventy-seven per cent, on total net revenues of 1,896 million dollars and income before taxes of 1,456 million, both figures from the company’s own results release. The arithmetic reconciles: 1,456 divided by 1,896 is 76.8 per cent.

Customer accounts stand at 5.185 million, up thirty-four per cent on the year; customer equity at 930.3 billion dollars, up forty; daily average revenue trades at 4.824 million, up thirty-six. The average commission per cleared order is 2.64 dollars. Peterffy told an interviewer in 2012 that twenty years before that, the same trade would have cost you fifty or sixty dollars. Asked what he is proudest of, he said: the money we have saved people in getting markets to be more efficient.

And here is the part that is rarely put beside the margin. In that same quarter, net interest income was 1,057 million dollars against commission revenue of 673 million. Interactive Brokers earns more from the spread on the cash its customers leave with it, and on lending against their holdings, than it does from trading commissions. That is fifty-six per cent of revenue. It is the quiet answer to how a broker charges 2.64 dollars a trade and earns seventy-seven per cent: the low commission is part-funded by the interest on the balances behind it. The consumer surplus is real and so is the mechanism, and a profile that reports the margin admiringly without the mix is telling half of it.

The firm has also been fined for compliance failures: thirty-eight million dollars across three American regulators in August 2020 over anti-money-laundering deficiencies, including a failure to file at least a hundred and fifty suspicious activity reports. Separately, after the April 2020 negative-oil-price episode, its systems could not handle negative prices; it voluntarily repaid more than 102 million dollars to affected customers before the regulator concluded its case.

What this gauge measures, and why scarcity inverts before it breaks, is explained in the appendix. This section opens on what the dials did, which this week is nothing, and on the one event that looked like it should have moved them.

What moved

The reported Hugging Face talks, described in The Speed of Now above, do not move the model-layer dial. The red band on this gauge requires open weights to reach parity with the closed frontier, and an acquisition is not a benchmark. What a deal at that price establishes is the strategic conviction of an unusually well-informed buyer, not the technical parity this dial measures. Those are different claims and only one of them is scored here.

Worth holding the distinction, because it is what keeps the gauge honest. A dial that moved on news rather than on evidence would have told you scarcity was breaking three separate times this summer, and it would have been wrong each time.

The business test: third Friday under the line

Captured automatically from the open market on Friday 28 August. The full register is Appendix A12.

The interpretation, and why the two are kept apart

That is the third consecutive Friday below the line, counting from 14 August. The run is no longer than that: the rent was above the cost line on 31 July at 1.867 and on 7 August at 1.733, which broke the earlier run, so the 17 July baseline row in A12 showing a negative margin is not part of this sequence. The test this section set for itself in August was four consecutive prints. One more and it is met, which means next Friday is the one that matters rather than this one.

Guaranteed capacity, as opposed to the interruptible spot price, sat at 2.27 dollars an hour, so the gap between the two remains wide. Buyers who need certainty are still paying a large premium; buyers who can tolerate interruption are renting below cost. That spread is the useful number, because it says the shortage is in reliable compute rather than in compute.

The reader-facing instruction is unchanged: watch chip prices, not share prices. Memory contract prices are still not printing down month on month, and until they do, the upstream scarcity that this entire trade is priced on has not broken. A rental market can go underwater on oversupply of a three-year-old part while the bottleneck one layer up is completely intact, and this week that is exactly what the two readings are saying.

Falsification and what we are watching

A memory capacity delay, or a demand surge that empties inventories again, would re-tighten scarcity and drop this score. The score would rise if the Hugging Face deal completes and an independent benchmark establishes open-weight parity, which are two separate events and only the second one moves the dial.

Three exact conditions. A memory contract price printing down month on month, down rather than merely rising more slowly. A fourth consecutive Friday with the rental below 1.64. And any crack in the extreme ultraviolet lithography monopoly, still the only development that would take this score above fifty-five.

The answer a single number gives you here is the wrong one, and the reason is worth four minutes.

The VIX, the index that measures how much turbulence traders expect in the American share market over the coming month, touched 14.2 on Friday and closed at 14.43. That is its second-lowest level of 2026, behind the 14.25 it set on 14 August. It printed that number after the Federal Reserve chairman said financial conditions are not restrictive and set a test for cutting rates that current inflation fails.

Taken alone you would call that complacency and move on. It is more interesting than that, because the index is not measuring what most people think it is measuring.

The distribution, not the average

Underneath the calm index, three things are reported to be true at once: the spread between American large-cap volatility and technology-index volatility is at a record; put buying in the semiconductor sector’s main exchange-traded fund is at the highest level in that fund’s history; and the gap between the VIX and the average volatility of individual shares reached 34.14 points in July, an all-time high. The third of those is the one this section relies on and the only one measured from the options market’s own published series; the first two are reported figures we have not independently reconstructed, and they are offered as reported rather than as verified.

That last measure is the one to understand, and the mechanism is below.

This is not a calm market. It is a market where the average is quiet and the distribution is not, where the biggest single sector is reported to be hedged more heavily than at any point on record, and where the headline measure of fear is structurally incapable of registering any of it.

Why this is a distributional fact rather than a measurement error

It would be easy to call the VIX broken. It is not broken; it is answering a narrow question accurately. The VIX is derived from the prices of options on the S&P 500, so what it measures is how much the market expects the index to move over the coming month, not how much the companies inside it will move.

Those are different quantities, and the gap between them is correlation. When five hundred shares move together, the index moves as much as they do. When they move in opposite directions they cancel out, and the index can sit still while every constituent is in violent disagreement. A low reading in those conditions is not telling you that nothing is happening. It is telling you that a great deal is happening and it is not pointing the same way.

Which means the reading is most misleading precisely when dispersion is highest, and dispersion is highest when the market is in the middle of deciding something. The index will look calmest at the moment the argument underneath it is fiercest.

The takeaway is small and portable. When someone tells you volatility is low, ask whose volatility. If you own the index, they are right. If you own eight shares, they may be describing a market you are not in.

On 11 August 1987, Paul Volcker left the Federal Reserve and Alan Greenspan took his chair. Greenspan was a fortnight into the job when the American share market reached its peak, on 25 August 1987, the last week of August, the Dow at 2,722.42, the market at a record, measured volatility conspicuously low.

Ten days later, on 4 September, the new chairman raised the discount rate from 5.5 to 6 per cent. It was his first policy act. Contemporary accounts, including the Federal Reserve’s own history, describe it as a move made in part to announce his presence.

On 19 October the Dow fell 508 points, 22.6 per cent, in a single session.

What is the same, and it is not the crash

The parallel is not the outcome and I am not going to pretend it is. Nobody can tell you the odds of an October 1987 from an August 1987, and anyone who says otherwise is selling something.

The parallel is the setup, and it is closer than the calendar coincidence alone would suggest, though not exact: Greenspan was a fortnight into the job at the peak and Warsh is around day one hundred. A brand-new chairman inside his first hundred days. A market at or near record highs. Measured volatility at conspicuously low levels. A long bond under pressure. And a chairman whose first substantive act is to establish anti-inflation credibility rather than to respond to a visible emergency.

Warsh is on approximately day one hundred. Volatility sits within two tenths of the year’s low, for the reasons the Noise Barometer sets out above. The long bond is under visible strain, having touched a nineteen-year high this month.

What is different, and it matters

Greenspan raised rates. Warsh explicitly did not, and explicitly refused to say whether he will. In 1987 the dollar was falling and the policy response was partly about defending it; on Friday the dollar rose on the speech. And the 1987 market was priced on earnings that had grown far slower than the index, whereas today’s concentration sits in companies with real and growing profits, whatever you think of the multiple.

What happened next, which is the analytical payload

The useful part of 1987 is not the crash. It is the seven weeks before it. The market did not fall when Greenspan raised rates on 4 September. It drifted, then broke. The repricing came late and all at once, in a market that had spent months looking quiet.

The lesson is about sequencing rather than outcome. A new chairman changing the policy regime does not produce an immediate verdict. It produces a period in which the market has not yet worked out what the new reaction function is, and prices as though the old one still applies. That interval is where 1987 did its damage, and it is where we are now.

Take the finding from the section above as the input: compressed index volatility is not evidence of agreement underneath. In August 1987 it was not either.

What these two forces are, and how the four pillars work, is explained in the appendix.

Four pillars, four different observations, and this week they split.

Pillar 1, adoption speed. The Displacer. Warsh told Jackson Hole that reported annualised token sales for the two leading laboratories alone now exceed a hundred billion dollars, an increase of more than five hundred per cent in a year. He was relaying rather than measuring, and the attribution matters, but a Federal Reserve chairman putting that number in a prepared speech is itself evidence that adoption is outrunning the gradual-adoption forecasts. Point to the Displacer.

Pillar 2, labour dynamics. The Augmenter. Friday’s preliminary benchmark indicates American employment in March was overstated by 79,000, with private employment overstated by 178,000, concentrated in retail trade, education and health services, manufacturing and business services. Those are not the categories where artificial intelligence substitutes for cognitive work, and not one of the revisions was attributed to automation. This is the weakest kind of evidence, an absence, and it sits in genuine tension with the Challenger figures in A7, where artificial intelligence was the leading stated reason for announced job cuts for a fifth consecutive month. Two readings of the same labour market, and the pillar goes to the Augmenter because the payroll benchmark is a measurement of what happened while the Challenger series is a count of what employers said. So it comes with a positive fact beside it: software engineering job postings, the category most exposed to substitution, have continued to grow. Point to the Augmenter, narrowly.

Pillar 3, type of economic shock. The Augmenter. Warsh reported that more than half of this year’s growth in capital spending on equipment and intangibles can likely be ascribed to the artificial-intelligence build-out, with the four-quarter change running near nine per cent, the highest since 2021. Capital spending of that size is a supply expansion. It is money spent on capacity, hiring construction and electrical trades, and raising future output. That is the Augmenter’s mechanism, not the Displacer’s demand destruction. Point to the Augmenter.

Pillar 4, compute constraints. The Displacer. The open-market rent on a leading artificial-intelligence chip closed the week at 1.33 dollars an hour against an all-in cost of 1.64, a third consecutive Friday below breakeven. Compute becoming cheaper than the person at a cognitive task is the Displacer’s core condition. The rent itself has not fallen, and the register in A12 prints it at 1.333 for a second straight Friday, and back at the same level it held in mid-July; what has crossed is the cost line, not the price. Renting compute below cost is still that condition met, even when the reason is oversupply rather than efficiency. Point to the Displacer.

Week 33 result: a two-all draw. Running total, the Augmenter 38, the Displacer 18.

Fourteen weeks at four pillars is fifty-six, and thirty-eight plus eighteen is fifty-six. The ledger reconciles. Note what that check does and does not establish: it verifies the bookkeeping, not the judgements. Each award above is a reading of which side this week’s evidence favours, and the four rest on four different observations, which is the rule this section holds itself to.

What would flip the score. A named, large employer attributing a headcount reduction specifically to artificial intelligence would hand the Displacer pillar two immediately and decisively. Conversely, compute rents recovering above the all-in cost for two consecutive weeks would return pillar four to the Augmenter by removing the substitution-threshold argument.

Arista Networks closes, and the number is the point

On 29 May this letter put Arista Networks on this list at 147.00 dollars, with a thesis about artificial-intelligence networking demand, a pre-registered condition that would prove it wrong, a benchmark it had to beat, and a fixed date at which it would be closed and scored whatever the answer. That date was Friday.

Arista closed at 195.38, up 32.91 per cent from entry. The Nasdaq 100, its pre-registered benchmark, fell 2.73 per cent over the same period. Excess return: 35.64 percentage points.

The failure condition, set on 19 June before any of this was known, required Arista to have cut its networking revenue guidance or seen gross margin fall below about sixty per cent, and for the share price to be below the 147 entry. The price finished thirty-three per cent above entry, so the second limb could not be satisfied on any reading.

Scored: plus two, vindicated. It is the first plus two in this ledger.

The denominator, because a track record you cannot reconcile is not a track record

That verdict is the seventh thesis-horizon close since this scoring began. Six of the seven have been correct and one was a double miss, MP Materials, which failed on both horizons and was reported here at the time with the same prominence as any winner. The ledger holds twenty-one calls in total: these seven closed at their horizons, two still open, and twelve macro calls made in April that were closed administratively because they were entered without a benchmark. Across all twenty-one, cumulative excess return against benchmarks stands at 50.0 percentage points, and the twelve unbenchmarked April calls contribute nothing to it in either direction. Every one of those calls is in the running record, and the losses are counted in the same column as the wins.

What was actually right, and what was luck

The thesis was that the market was pricing Arista as a networking company in an artificial-intelligence market, rather than as a bottleneck in it. That has held: the networking layer kept its pricing while the sector around it was repriced twice.

What I did not forecast, and should not claim: that the benchmark would fall. Of that thirty-five-point excess, 2.73 points, under a tenth, is the Nasdaq 100 going backwards over three months, which had nothing to do with Arista. It was a good call in a market that helpfully agreed to make it look better.

No new call is added this week. The two open positions are tracked in Portfolio Watch, in the appendix below.

In December 1972 a widowed mother of ten named Jean McConville was taken from her flat in Belfast in front of her children. She was not seen again for thirty years.

That is the first chapter, and Patrick Radden Keefe spends the next four hundred pages on what happened, who did it, and what it did to everyone who touched it. Say Nothing reads like a thriller and every word of it is reported. There is no invention here at all: it is built from interviews, archives and a set of oral histories recorded at Boston College by former paramilitaries who believed the tapes would stay sealed until they died, and which were subpoenaed while they were still alive.

I am recommending it for the reason this section exists, which is that it is extraordinarily hard to put down. And for a second reason that has more to do with this letter’s usual subject matter than it appears.

Keefe is doing something unusual with evidence. He is assembling a conclusion about a specific event from testimony given by people with reasons to lie, decades after the fact, whose memories have been reshaped by the need to live with what they did. He shows you the assembly. You watch him weigh a source, discount it, corroborate it against another, and arrive somewhere he can defend. It is the most rigorous piece of reasoning under uncertainty I have read in narrative form, and the discipline it demonstrates is exactly the discipline that separates a real analytical judgement from a confident-sounding one.

A word on the subject. This is the Troubles: real people, real killings, and a political memory that is still live for a great many readers. Keefe is scrupulous and the book is admired across the divide, but it is not neutral background in the way a mountaineering disaster is. I am recommending the book and its method, not adjudicating anything, and I would not want the recommendation read as anything else.

The chairman gave nothing away,
Just a standard the numbers must weigh.
    The crude price came down,
    The diesel stayed crowned,
And the lorries will pass on the pay.

Three things to watch next week. Friday’s payrolls, the first labour reading since the preliminary benchmark called the series into question. The thirty-year Treasury at Friday’s close, against the 5.15 per cent this letter has put its name to. And the enlarged Treasury buyback operations beginning on 9 September, which will be tested by the market within hours of starting.

Suppose all three point the same way: a soft payroll, a thirty-year above 5.15, and buybacks that do not hold the long end. Then a government is losing control of its own borrowing costs at exactly the moment its central bank has ruled out helping. The price of money becomes the only story that matters for the rest of the year. If they diverge, and the likeliest divergence is a firm payroll against a long end that steadies, then September passes quietly and the diesel spread turns out to be the thing we should have been watching all along.

I do not know which. That is why they are worth watching, and why we will tell you in seven days which one it was.

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

Freight leads the year at plus 69.3 per cent, a long way clear of the Turkish lira at plus 36.2 and crude at plus 32.0, and the gap between best and worst across the twenty-six lines is 87.3 points. One commodity is doing most of that widening, which is a narrower fact than a broad repricing.

26 assets ranked by year-to-date return · baselines locked 1 January 2026 · close of Friday 28 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. 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 28 August 2026. MSCI EM is derived from the EEM ETF (exchange-traded fund) close of 67.1402 multiplied by the locked index ratio (28.367); Baltic Dry is the Baltic Exchange Dry Index as published by Hellenic Shipping News (28 Aug, 3,186). Every year-to-date figure is recomputed from the locked 1 January baselines rather than carried forward, so an error cannot compound week to week.

Every company that has appeared in On the Radar remains tracked here until its thesis horizon. From this month every new call carries a horizon of at least twelve months, because a shorter window judges the weather rather than the climate; the calls entered before that change 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. Arista Networks, whose thesis horizon closes on 28 August, is written up in full in On the Radar above rather than repeated here. Benchmark returns are recomputed each week from verified closes rather than carried. Mitsubishi UFJ is measured against the iShares MSCI Japan fund from 2 July. YPF is measured against Alphabet, the holding the investor behind that thesis sold to fund it, from 18 June. Each date is the first verified close in our price series on or after entry. Both returns in this table were recomputed this week from entry and the verified Friday close, because the stored figures had drifted and we would rather recompute than carry. Every return is measured against a benchmark chosen and fixed at entry, because a call that beats nothing you could have held instead is not a win.

Running record, and the denominator named so you can reconcile it: two calls are open, first mentioned in weeks 23 and 25, and one closed on Friday. Mitsubishi UFJ is up 13.53 per cent against a benchmark up 2.93, an excess of 10.6 points. YPF is down 0.22 against a benchmark down 5.83, an excess of 5.6 points. Both are ahead of the benchmark fixed for them at entry; only one is up in absolute terms, which is the distinction this table exists to draw. Seven calls have now closed at their thesis horizons since week 13, and the closed book is the one to read hardest. The most recent is Arista Networks, closed on 28 August at plus two, vindicated, the first plus two in this ledger: up 32.91 per cent against a Nasdaq 100 down 2.73, an excess of 35.64 points, with its pre-registered failure condition unable to fire. Of the seven, six were correct and one was a double miss. The twenty-one are these seven, the two still open, and twelve April macro calls closed administratively for having been entered without a benchmark. Across all twenty-one, cumulative excess return against benchmarks stands at 50.0 percentage points, of which the two positions here and the Arista call just closed account for 51.8 and the closed losers for the difference. That figure moves down as easily as up, which is the point of printing it.

Forward-looking only. Every dated item that has already happened has been removed rather than left standing: a catalyst table inviting you to look forward to a speech this edition has just taken apart is furniture, not information.