Start with what the week was for.

The method, and why it transfers

Everything this week came out of documents anyone can download, for nothing, without a subscription or a terminal.

The cash flow statement. The annual report. The results release. The proxy. An earnings call transcript. That is the entire toolkit, and the tests run on it are arithmetic rather than judgment: divide this by that, put the answer next to the trading range, read the paragraph the chief executive wrote freely, count what changed.

None of this requires an opinion about a business. It requires the patience to open eight filings instead of one.

That is the transferable part, and it is why the specific company matters less than the sequence.

What they were good at

Start where the evidence is strongest, because a review that only finds faults is not a review.

This management team ran the business well. Revenue roughly quadrupled over the departing chief executive’s tenure. In fiscal 2022 the company earned $854.8 million on stockholders’ equity of $3.15 billion, a return on equity near 27%, which is a level most large companies never reach. Gross margin sits above 60%. The brand became strong enough that walking past a store tells you nothing is obviously wrong.

What that tells you: the operating half of the job was done properly. Product, stores, brand, people. Whatever went wrong later did not go wrong because nobody could run a shop.

That matters for what follows, because the rest of this is not a story about incompetence.

What they were expensive at

The other half of a chief executive’s job is deciding where the money goes, and this is where the record separates.

They bought one company and wrote most of it off. MIRROR cost $500 million in 2020. By the fourth quarter of fiscal 2022 it had been written down by $442.7 million after tax. The company had disclosed the goodwill risk itself, in the annual report filed in March 2021, along with the observation that its management had limited experience in that area.

They bought their own stock as it got more expensive. Across fiscal 2021 to 2025, 15.2 million shares for about $4.62 billion, an average near $304. The stock traded below $100 last week. The largest single year, $1.6 billion in fiscal 2024, came at the second-highest price of the whole series.

And the capital base test confirms it from the owner’s side. Between fiscal 2024 and 2025 the share count fell 3.9% while earnings per share fell 9.4%. The buybacks were pushing that number up and it went down anyway.

One fact cuts the other way and belongs here. Through the most recent half, with sales at existing stores falling 9%, the company opened nine net new stores. It also cut planned net new openings for the year from 40 to 35, and pop-ups from 65 to about 40, publicly and with a reason. A team under pressure that reduces its own expansion plan is doing the harder thing.

What that tells you: two different skills, and this team had one of them. A 27% return on equity says the business converts capital into profit efficiently. Paying $369 a share for that business in 2021 was a separate decision, made by the same people.

That distinction is the single most useful thing a run of annual reports will give you, and it is invisible in any one year.

The announcement, and what it did not say

One document from that acquisition is worth reading closely, because it teaches something that outlasts this company.

The 8-K filed on 29 June 2020 says the purchase will advance the company’s strategic vision by strengthening its omni guest experiences through digital sweat, and will bolster its digital sweatlife offerings.

Two phrases there need translating. Omni guest experiences is retail for selling to the same person through more than one channel: shop, website, app, and now a screen on their wall. The sweatlife is the company’s own coined term for its customer’s life around exercise.

Strip both out and the strategy underneath is coherent, and I want to be fair about that. We sell clothes to people who exercise. Today we see them when they need leggings. If we sell them the workout, we are in their house every day and a subscription pays us monthly rather than occasionally. That is a real idea, and Peloton had shown the model could work.

So the purchase was not frivolous in concept. The questions sit elsewhere: $500 million for a business launched two years earlier whose own forecast for that year was $100 million, bought while gyms were shut and home fitness was among the most crowded trades in retail.

And here is what the announcement does not contain. A number. No expected return on the $500 million. No payback period. No revenue or margin target. The only financial expectation attached to the deal that day came in a television interview rather than in the filing.

So a test, and it takes two minutes on any acquisition announcement. Translate the jargon into plain language, then ask what the rationale promises in figures.

The translation matters as much as the figures, because jargon is not merely ugly. It is untestable. “We expect to convert 5% of our guests to a $39 a month subscription” is a sentence you can check in two years. “Deepen our roots in the sweatlife” is a sentence nobody can ever be wrong about.

A rationale you cannot translate into a checkable claim is a rationale nobody can be held to. That is the warning, and it is separate from whether the strategy is any good.

What they said about it

A record tells you what happened. The account tells you how a team handles what happened, and those are separate pieces of evidence.

The forecast. At the acquisition, more than $100 million of revenue that year and break-even the next. Raised to more than $150 million. Raised again to $250-275 million. Cut to $125-130 million. Then written off. It rose twice before it fell, which means confidence increased while the evidence was turning.

The description. At the purchase, the business was fuelling the company’s growth plan. Explaining the write-off, the finance chief called it a very small portion of management’s five-year plan. Same business, same team, the stated importance falling with the value.

The statement. In the release carrying a $442.7 million write-off, the chief executive’s own paragraph described strong results, continued high performance and enduring brand strength. The impairment was disclosed elsewhere in that release and he named it on the call when asked. It is not in the paragraph he chose the words for.

What that tells you: nothing improper, and a consistent preference. Given a choice about emphasis, this team took the favourable one, every time, in the places where the choice was theirs to make.

The three I could not settle

Three tests did not produce a clean answer, and in each case the reason is worth more than the result would have been.

Is the decline the industry’s or the company’s. Inconclusive, because four of the six companies competing for this customer are private and file nothing. The one comparable public business, Athleta, fell exactly the same 12%. What would settle it: filings that do not exist. If you hear someone state confidently that Lululemon is losing share to Vuori or Alo, ask where the number came from.

Did they ever act on a limit, at a cost. Not proven. Form 8-K Item 5.02 records senior arrivals and departures within four business days, and it gives who and when but never why, because the SEC considered requiring reasons and dropped it over defamation risk. What would settle it: an appointment where a leader hands a function to someone who has fixed that exact problem elsewhere, read alongside the letter that explains it.

Owner earnings. Retired. The method is to treat depreciation as maintenance spending and capital spending above it as growth. On this company it does not hold: between $167 million and $183 million a quarter went into a distribution centre project, new stores, relocations, renovations and technology, and a relocation is maintenance and growth at once. What would settle it: a business with a stable asset base, where the proxy works. Not this one, not this year.

Not proven is a verdict, not a gap. It means the evidence does not exist, rather than that it came back against them, and treating those as the same thing is how a framework starts convicting people for being new.

So how do I read this management team

Competent operators who were poor at the part of the job that compounds.

That is not a soft verdict and it is not a harsh one. It is the specific finding that emerges from reading the filings in sequence rather than one at a time, and it is worth being precise about why it matters.

A business earning 27% on equity throws off cash. What happens to that cash is the whole question, because over ten years it becomes most of what you own. Buffett’s arithmetic puts it at more than 60% of all capital in the business after a decade at modest retention rates. This team put $4.62 billion of it into their own shares at roughly three times today’s price, and $500 million into an acquisition that was written down by $442.7 million after tax.

The operating skill built the cash. The allocation decisions are what happened to it.

And the account they gave has a consistent shape. Forecasts revised upward before they were cut. A business described as central on the way in and peripheral on the way out. A write-off disclosed where required and absent from the paragraph written freely. None of it improper. All of it in the same direction.

That direction is the finding. One favourable choice is a coincidence. Three separate ones, across three different documents, over three years, is a disposition. And a disposition is the thing you can expect to continue, because it is the part that does not depend on the business.

What Buffett’s own criteria would say

Last Monday I started with the list Berkshire printed in the back of its annual report for years, because it is the shortest honest statement of what a buyer of whole businesses wants. It is worth finishing the week by running this company against it.

Demonstrated consistent earning power. Passed, historically. Revenue quadrupled, margins were strong for a decade, the record is real. But the same criterion says future projections are of no interest and, in the same breath, no turnaround situations. That is the sentence that bites. A business with comparable sales falling 12% in its largest market, a category shifting underneath it, and a new chief executive brought in to fix it is a turnaround, whatever else it is.

Good returns on equity while employing little or no debt. Passed, and comfortably. A return on equity near 27%, and the balance sheet carries no borrowings under its credit facility.

Management in place. This is the one the whole week has been about, and the answer is the most interesting of the three. The criterion exists because Buffett does not want to supply management. He is buying a team as much as a business. A company that changed chief executive last Tuesday has management in place in the literal sense and not in the sense the criterion means, which is a team whose record you can read.

So the verdict against that list is clean and it is not close. The economics pass. The turnaround exclusion fails. And the management criterion returns the same answer this week has returned throughout: not yet knowable.

That is not a reason to walk away. It is a reason to know which question you are waiting on. Most of what has been written about this company since the announcement is about whether the new leader is the right choice. The criterion says something narrower and more useful: you cannot know yet, and here is what would tell you.

What a change of management does to a review

This is the part I would want if I were reading rather than writing, because the situation is more common than it looks. Somebody leaves, somebody arrives, and every review written in that window is confused about which record it is describing.

Some tests survive a change and some reset.

The business does not change when the chief executive does. Return on capital, gross margin, the strength of the category, the balance sheet, the competitive set. All of that is still true on Wednesday. Everything I found about this business survives the transition intact.

Everything about the people resets to zero. The allocation record, the candour record, the forecasting record, the trait. Not to a bad score. To no score.

The departing record becomes the bar, not the verdict. This is the distinction that took me most of the week to say properly. Calvin McDonald’s capital decisions tell you nothing about Heidi O’Neill. What they tell you is what a good outcome and a poor one look like at this specific company, which is the yardstick you hold up to whatever she does next.

And a transition hands you something you do not usually get. A new chief executive has an incentive that a continuing one does not: problems named early in a tenure belong to a predecessor, and problems named later belong to her. Write-downs, restructurings and honest disclosures cluster early in a tenure for exactly that reason.

Which cuts both ways, and you should know both.

The favourable reading is that you are about to learn more about this business in two quarters than the previous two years disclosed, because it is temporarily in her interest to tell you.

The unfavourable reading is that the same incentive makes early candour cheap. Naming a problem you did not create costs nothing. The test that matters is the second one, when the problem is hers.

That is why the watchlist below runs on documents rather than on the next few months of announcements.

Why this matters if you invest for yourself

Because it is the part of the work that nobody does for you.

Analyst coverage will tell you what the business might earn next year. Nothing in it will tell you what this team did with $4.62 billion, whether they raised a forecast before cutting it, or whether they named a loss in the paragraph they wrote themselves. Those are all knowable, all free, and all sitting in documents that require nothing but the patience to open them.

A self-directed investor’s real advantage is not better information. It is being willing to read the boring documents that contain it.

There is a second advantage, and it is the one this week was really about. You can do this before the price moves. Everything here was available months ago. None of it required knowing what the stock would do. When a price does fall far enough to interest you, the management work is either already done or it is being done in a hurry, and hurried work is how people talk themselves into things.

What happens next

Heidi O’Neill started on Tuesday. She has run this company for days, and nothing in this week’s evidence is hers. I am not going to grade her on a record she did not make.

What I will do is watch four documents.

The next results release, and specifically the paragraph attributed to her. Does it name what is not working, in the quarter it happens.

The next repurchase disclosure. About $713 million of authorisation remains and the company has said it expects to keep buying at 2025 levels. The test is not the amount. It is the average price, which is the cost divided by the change in share count, and it appears in the same paragraph every time.

Her first proxy statement, which will set out the terms of her arrangement: what vests on a change of control, what severance looks like, and how much of her pay depends on results that have already happened.

The 8-K stream, for who she hires and what she gives up to get them. The record will not say why, but it will say who and when, and if she brings in someone who has fixed this exact problem elsewhere, that is a decision with a cost attached.

Those four are dated, filed and free. Between them they would answer most of what this week could not.

And one thing this week did not touch at all. I have not valued this business. Everything above is one test of several, run deliberately in isolation so that a falling price could not make the evidence look better than it is. Whether the price is attractive is separate work, and if I do it, it will be its own week with its own standard of proof.

Which leaves the honest summary of six days, in a sentence.

A good business, run well and allocated badly, handed to someone with no record at it, at a price that has fallen far enough to make all of that worth knowing.

Three of those four things I can tell you from filings. The fourth arrives over the next year, in documents that are already scheduled.

Not investment advice. The subscriber decides.

Yesterday I said a record tells you how a team behaved but not whether the behaviour is still happening. Before I get to that, the record needs to be complete.

Why one year was not enough

On Wednesday I showed you a single number: 2.2 million shares bought during fiscal 2021 at an average of $369.16, for $812.6 million.

On its own that tells you very little, and I want to be careful about why.

Fiscal 2021 was an excellent year for the business. Revenue grew sharply, the brand was strong, and the shares had risen because the results deserved it. Nothing about the year was bad. What was expensive was the price paid for the stock, and any company can pay a high price once.

What a single year cannot tell you is whether that was a slip or a method, and the difference is everything. So I went back through the annual reports and pulled the same two figures out of each one: how many shares they bought, and what they paid.

Both numbers sit in the Financing Activities paragraph of the annual report, which is where a company explains what it did with cash. Divide the second by the first and you have the average price paid. It is the arithmetic from Wednesday, run ten times.

Eleven years, six filings

Here is what came out, with the implied average alongside.

In fiscal 2015 the company bought 5.0 million shares for $274.2 million, about $55 each. In 2016, 0.5 million shares for $29.3 million, about $59. In 2017, 1.9 million for $100.3 million, about $53.

Then the price of the shares began to climb, and so did the buying.

In 2018, 4.9 million shares for $598.3 million, about $122. In 2019, 1.1 million for $173.4 million, about $158. In 2020, 0.4 million for $63.7 million, about $159.

In 2021, 2.2 million shares for $812.6 million. The company disclosed that average directly: $369.16.

In 2022, 1.4 million shares for $444.0 million, about $317. In 2023, 1.5 million for $558.7 million, about $372. In 2024, 5.1 million shares for $1.6 billion, about $314. In 2025, 5.0 million shares for $1.2 billion, about $240.

The company spent more on its own stock in fiscal 2024 than in any other year in the series, at a price close to six times what it paid in 2017.

And fiscal 2025 is the year worth stopping on. That was Calvin McDonald’s last full year. Americas comparable sales fell 3%. Operating margin fell 380 basis points to 19.9%. Earnings per share fell from $14.64 to $13.26.

The company spent $1.2 billion buying its own stock during it.

Three things about those numbers before anyone builds a case on them. From fiscal 2022 onward the reported cost includes commissions and excise taxes, so the implied price is slightly higher than the true purchase price. The share counts are rounded to a tenth of a million, which makes the smallest years imprecise, and I have marked those approximate. And fiscal 2025 comes from the most recent annual report and the full-year results release, both of which state the same figures.

The five years that matter

Take fiscal 2021 through 2025 together. Five full years, each from the same paragraph of an annual report.

15.2 million shares. About $4.62 billion. An average of roughly $304 each.

The stock traded below $100 last week.

Buying accelerated as the price rose. That is the sentence, and it is arithmetic rather than opinion.

What I am not saying

I am not saying the shares were worth less than $304 when they were bought. Nobody can establish that from outside a company, and I would be pretending to knowledge I do not have.

In 2021 and 2024 this was a business growing revenue at double digits, throwing off cash, with a brand that looked untouchable. Even in 2025 revenue still grew 5%. Buying your own stock in that situation is a defensible decision and thousands of management teams make it.

Nor am I saying returning cash to owners is wrong. It is one of the five things a company can do with a dollar, and often the best one.

What I am saying

Last week I set out the test that decides whether a repurchase helped the owners who stayed. Two conditions, both required. The company has funds beyond what the business needs. And the stock is selling at a material discount to what it is conservatively worth.

The first condition looks comfortably met in the years I pulled: this was a business generating cash with a strong balance sheet throughout. The second is the one that gets skipped, and it is skipped in a specific way: a rising share price feels like confirmation that the business is working, which makes buying feel prudent exactly when it is most expensive.

A buyback is the only capital decision where the price is public before you make it. Which means it is the only one where getting the price wrong is a choice rather than an accident.

That is what a run of filings tells you and a single year does not. In 2023 the company paid its highest average price of the series, about $372. In 2024 it spent $1.6 billion, more than in any other year. In 2025, with the Americas already shrinking and margins falling, it spent $1.2 billion more.

What the interim team did with it

Calvin McDonald left at the end of January. Meghan Frank and André Maestrini have run the company since, and the quarter reported on 3 September is theirs.

In that quarter, the company repurchased 2.7 million shares for $330.0 million. That is about $122 a share, computed the same way.

It is the lowest average price paid since fiscal 2018, and it is still above where the stock traded a fortnight later.

In the same quarter they opened nine net new stores, ending with 825 against 784 a year earlier, and spent $149.7 million on capital projects. Then they cut the full-year outlook for the second quarter running.

On the call, the company said repurchases remain its preferred method of returning cash and that it expects 2026 buyback levels in line with 2025.

They committed to keep buying in the same breath as they told investors the business would be smaller than they thought.

The thing they did decline

One test from last week is what a management team says no to, and this one said no to something.

Net new store openings for the year were cut from 40 to 35. Pop-up stores were cut from 65 to about 40. Both were announced publicly with a reason attached.

That is a real mark in their favour and it belongs here. A team under pressure that reduces its own expansion plan is doing the harder thing, and I would rather report it than leave it out because it complicates the argument.

The bar

Which brings me to the person who started on Tuesday.

Heidi O’Neill has run this company for days. Nothing in this week’s evidence is hers, and grading her on it would be the error I warned about on Monday. What the evidence does is set the height of the bar, and the bar has three parts.

Capital discipline, measured by price rather than by amount. The company has roughly $713 million of repurchase authorisation remaining and has said it intends to use it. Whether that is a good decision depends entirely on what it pays. If the next annual report shows a large sum spent at a price that later looks high again, nothing has changed. That is checkable, it is dated, and it appears in the Financing Activities paragraph where the rest of this series came from.

The core category, which the company itself named. In the last quarter the company reported a greater-than-expected slowdown in core categories and named leggings specifically. This is not a cost problem or a store problem. Comparable sales in the Americas fell 12% because fewer people came in, fewer of those bought, and the ones who bought spent less. That is a product and brand problem, and it is the one thing on this list that cannot be fixed by allocating capital better.

The account, when the next thing goes wrong. Something will. What this week established is that the record of how this company describes its own failures is mixed: the write-off was disclosed as required, acknowledged when asked, and absent from the paragraph the chief executive wrote freely. The first letter and first results release under new leadership will show whether that changes.

What getting back to good actually requires

Not growth. Growth is what a recovery looks like from outside, not what produces it.

The business still generates cash: $589.3 million from operations in the first half, against $209.7 million a year earlier, with $1.4 billion in the bank. The gross margin is intact. This is not a company fighting for survival, which is precisely why the next few decisions are worth watching. Nobody is forced into anything.

A company with $1.4 billion in cash and a falling share price is a company whose management is about to reveal what it believes.

Tomorrow

The last piece.

What can and cannot be known about the incoming chief executive, where to look for the first evidence, and the documents that will settle it, with dates.

Not investment advice. The subscriber decides.

Yesterday I set out two decisions this management team made with shareholders’ money, and said today would be the account they gave of one of them. This is that, and the account turns out to have three parts.

Start with why an old forecast is worth anything at all.

Why old forecasts are evidence when new ones are not

Berkshire’s acquisition criteria have said in print since the early 1980s that future projections are of no interest. A plan is a claim about people who have not yet done the thing.

That rule is about not trusting a forecast as evidence about a business. It says nothing about forecasts a company has already made, and those are a different document entirely. They are dated. They are specific. The outcome is known.

A projection is worthless as evidence about a business and valuable as evidence about the people who made it.

Where to find them

Old forecasts live in four places, all free. The transcript of any earnings call, usually posted on a company’s investor relations page and carried by several sites going back years. The press release attached to a results filing, where guidance is usually given in a table near the end. Investor day presentations, which contain the most specific and most forgotten numbers a company ever publishes. And the announcement of any acquisition, where a management team explains what it expects the thing to do.

Pick a decision that is at least three years old and large enough to have been discussed. Write down what they said it would do, with the date. Then find what it did.

The trail, once

Lululemon announced it was buying MIRROR on 29 June 2020, for a purchase price of $500 million.

That same day the chief executive told CNBC the business expected more than $100 million of revenue that year and would break even or be slightly profitable the next. At the results announced in September 2020 the company raised that to more than $150 million. The annual outlook was later set at $250 to $275 million. At the results announced in December 2021 it was cut to $125 to $130 million. In March 2023 the business was written down by $442.7 million after tax.

That is the sequence. Now the part that matters, which is not the miss.

Three things to look for in a revision

First, the direction as evidence arrived. A team that holds a forecast while the world moves is being stubborn. A team that raises one is telling you its confidence went up. Here the number went up twice before it came down, which means that as data accumulated through 2020 and 2021, the read got more optimistic rather than less.

That is not dishonesty. It is a question about instruments. A management team either had information that supported the raise, or it was reading its own enthusiasm. You cannot tell which from outside, and you do not have to, because the next revision tells you.

Second, the size of the correction when it came. A guidance range moving from $250 to $275 million down to $125 to $130 million is roughly a halving in a single step. Small corrections mean a team is adjusting continuously. One large correction means the gap had been open for a while before it was closed.

Third, whether the description of the decision changed. This is the one worth learning, because it is the least visible and the most telling.

The importance downgrade

When Lululemon announced the purchase, it described the acquisition as building on its vision, strengthening its omni guest experiences, and fueling the Power of Three growth plan. That is a business being placed near the center of the strategy.

When the impairment was explained, the chief financial officer described the business as a very small portion of management’s five-year plan.

Same business. Same management. Between those two statements, its stated importance to the company fell.

Watch for that. A decision that was central when it was made and peripheral when it failed has been reclassified, and the reclassification is a choice about how the record reads.

I am not saying either description was untrue. A business can be strategically central when it is bought for $500 million and marginal after $442.7 million of it has been written off. Both statements can be accurate. What the pair tells you is how this team frames a decision on the way in and on the way out, and framing is the part they control completely.

The second half of the test

You now know what a team predicted and what happened. The remaining question is what they said when the answer arrived.

Yesterday I said I would read the letter from the year the write-off landed. That needs a correction, and the correction is worth more than the original promise.

Lululemon does not publish a shareholder letter. Berkshire’s annual letter is famous partly because writing one at that length is a choice rather than a requirement, and a test built on reading one assumes a document a company may not produce. When it does not, the nearest equivalent is the paragraph attributed to the chief executive in the results release, and it does the same job for the same reason: it is the passage a management team writes freely, in a document otherwise governed by accounting rules.

When a test assumes a document a company does not produce, find the place where the same choice gets made. Do not skip the test and do not pretend the document exists.

Find the results release covering the quarter the loss landed. Skip the headline. Skip the tables. Go to the paragraph attributed to the chief executive, which is the one part of the document a management team chooses the words for, and which is the passage most likely to be quoted everywhere else.

Three outcomes, easy to tell apart. They name the loss and say what they are doing about it. They gesture at it in language general enough to describe any year. Or it is not there.

Lululemon’s release of 28 March 2023 reported fourth-quarter revenue up 30% to $2.8 billion, GAAP earnings per share of $0.94 and adjusted earnings per share of $4.40. In his statement, McDonald described strong results across the business, a continued high level of performance, the enduring strength of the brand, and optimism about sustained growth.

The impairment does not appear in that paragraph.

Being fair about this

Three things need saying and leaving them out would make the point stronger and the article worse.

The write-off was disclosed. It is in the release, in the annual report, and in a note of its own. Nobody hid anything, and under the accounting rules nobody could.

McDonald named it on the call. Asked about the quarter, he said the company was taking an impairment charge related to assets and goodwill associated with Mirror. That is an acknowledgment and it counts.

And a chief executive’s statement in a results release is a marketing document. Every company writes them the same way, and none of them lead with the bad news.

So the claim is narrow. The company disclosed what it was required to disclose, and the chief executive acknowledged it when asked. The paragraph he chose the words for did not contain it.

What all this tells you about competence

Not that the acquisition was a mistake. The write-off already told you that and it took no skill to notice.

What the exercise gives you is three readings you cannot get from the financial statements.

How they read their own information. Confidence rose for eighteen months into a category that was turning. Either the data supported that or it did not, and the size of the eventual correction is the clue.

How fast they close a gap. One halving rather than a series of trims.

How they describe a decision once it has gone wrong. Central on the way in, a very small portion on the way out, and absent from the paragraph they wrote freely.

None of those is a character judgment and none requires guessing at anyone’s motives. All three are dated, filed, and available to anyone willing to spend an afternoon.

A management team’s forecasting record is a record like any other. It is just the only one they publish about themselves in advance.

What this cannot tell you

Every document in today’s piece is at least two years old, and the person who wrote most of them left the company in January.

A record tells you how a team behaved. It does not tell you whether the behavior is still happening, and the people running this business today are not the people who bought MIRROR.

So the test has to be run again on the present, which is a harder thing to do, because the results are not in yet and the write-off has not happened.

Tomorrow: the last ninety days. The quarter reported two weeks ago, what the interim team did with the company’s money while the stock was falling, and which number carried the headline this time.

Not investment advice. The subscriber decides.

Yesterday ended on the only evidence left, so here it is. But two words in that paragraph are doing a lot of work, and they are worth ninety seconds before anything else.

Goodwill, and what it means to write it off

When one company buys another, it pays a price. Then the accountants add up what it bought: the buildings, the equipment, the inventory, the cash, the patents, anything you can point at and value on its own.

The price is almost always higher than that total.

The difference has a name. Goodwill. It is what the buyer paid above the value of the identifiable things, and it stands for everything that is not a thing: the brand, the customer list, the team, and above all what the buyer expects the business to become.

Goodwill goes onto the balance sheet as an asset, and it sits there at the price paid. It is not worn down year by year the way a delivery van is. It stays at full value until somebody decides it should not.

Goodwill is the price of a company’s expectations, recorded as an asset.

That is where the second word comes in. Accounting rules require a company to test that asset regularly and ask whether the business it bought is still worth what was paid. If the answer is no, it has to write the value down, and the write-down is called an impairment.

An impairment is not a cash payment. The money left years earlier, at the moment of purchase. What the charge does is force the company to say so on the record: we paid this much, it is not worth that, and here is the difference.

It is the accounting system making a company admit in public that it overpaid.

So when a filing says a significant portion of a purchase price was allocated to goodwill, it is telling you that most of what was bought was expectation rather than machinery. And when it says impairment charges may be required if the acquisition does not yield expected returns, it is telling you what happens if the expectation does not arrive.

Both of those sentences appear in Lululemon’s own annual report, about this acquisition, before any of it happened.

Why the cash flow statement and not the income statement

A company’s earnings tell you what a year produced. They do not tell you what anyone decided.

Every year a business keeps some of its profit rather than paying it out, and somebody chooses where that money goes. Reinvest it. Buy something. Pay down debt. Pay a dividend. Buy back stock. Those five choices, made repeatedly over years, become the company you own.

Calvin McDonald held the job from 2018 until earlier this year. That is eight years of those choices, and two of them are large enough to see from the outside without any interpretation at all.

The acquisition

In the second quarter of fiscal 2020, Lululemon acquired Curiouser Products Inc., trading as MIRROR, an in-home fitness company selling a wall-mounted screen with live and on-demand classes. The timing is worth holding. This was the middle of 2020, when in-home fitness was a category with an obvious tailwind and gyms across much of the world were shut.

By the fourth quarter of fiscal 2022, the company was running an impairment test on it. Impairment testing was completed as of 29 January 2023, and the result was $442.7 million of post-tax impairment and other charges, including $407.9 million against goodwill and other assets.

A business bought at the top of a category was written down by hundreds of millions of dollars two and a half years later. That is not unusual. What follows is.

They wrote the warning themselves

Go back to that annual report, filed in March 2021 and covering the fiscal year in which the purchase was made, and read the risk factors. Companies are required to disclose what could go wrong, and much of that section is boilerplate.

This part was not boilerplate. The filing stated that a significant portion of the purchase price had been allocated to goodwill, and that if the acquisition did not yield expected returns the company might be required to record impairment charges, which would adversely affect its results. It went on to note that its management team had limited experience in the area.

I want to be careful about what that establishes, because it is easy to overreach.

A risk factor is not a prediction. It is a list of things that could happen, drafted by lawyers, and every company files pages of them. The presence of a warning does not mean anyone expected the outcome.

But it does establish two things that matter for a management assessment.

The risk was identified, in writing, by the people who took it. Nobody can say afterwards that this came out of nowhere.

And the company said, in its own filing, that its management had limited experience in this area. That is an unusually direct statement, and it is the company’s own characterization rather than mine.

The question a management review asks is not whether a decision went wrong. It is what was known when it was made. Here, the company told us what it knew.

What it cost beyond the money

The $442.7 million is the number that gets quoted. The effect on the year is the part that shows what a write-off does to a business.

Operating income in the fourth quarter of fiscal 2022 was $314.4 million, or 11.3% of net revenue, against $590.6 million and 27.7% a year earlier. Adjusted to exclude the impairment, operating income for that quarter was $785.3 million, or 28.3%.

The company’s own annual report puts it more plainly than I could. Lululemon generated approximately 24% of its full-year operating profit in the fourth quarter of fiscal 2022. Excluding the impairment, it would have been approximately 44%.

That is a single decision, made in 2020, removing roughly twenty points of the fourth quarter’s contribution to the year’s profit in 2022.

The second decision

The other place a company’s money goes is into its own shares, and this one is measurable to the cent.

From the company’s own disclosure: during fiscal 2021, Lululemon repurchased 2.2 million of its own shares at an average price of $369.16 per share, for a total cost of $812.6 million. As at 30 January 2022 it had $187.5 million of authorization remaining, which it completed in the following quarter. In March 2022 the board approved a new program for up to $1.0 billion.

Last week the shares traded below $100.

I am not going to tell you the shares were worth less than $369 in 2021. Nobody can establish that from outside, and a management team buying stock in a business growing quickly is doing something defensible on its face.

What I can tell you is the arithmetic. $812.6 million of shareholders’ money bought 2.2 million shares at an average of $369.16. At under $100, those shares are worth roughly 27% of the average price paid.

That is not a forecast that failed. It is a purchase, and the price is a matter of record.

The test I could not finish

Tuesday I said I would run the first test from the management series on this record: whether earnings grew faster than the capital base. A business that adds to its equity every year produces rising earnings almost automatically, so growing profit on a faster-growing balance sheet is not the achievement it appears to be.

I could not assemble that from primary sources in the time I had. The revenue and earnings history is available in aggregators, and under the rule I set this week, an aggregator is a lead and not a citation. The figures I would need sit in eight separate annual reports and I have not read all eight.

So I am not going to give you a return-on-equity number I have not verified. That test stays open, and I will come back to it when the filings are in front of me rather than a summary of them.

What this establishes

Two decisions, both large, both chosen rather than forced, both made by the same person.

An acquisition made in mid-2020, written down by $442.7 million after the company had itself disclosed the goodwill risk and its own limited experience in the area.

And $812.6 million spent on its own shares at an average price now more than three times the market.

Neither was a response to a crisis. Both were made from a position of strength, in years when the business was growing quickly and revenue was rising.

A record of growth and a record of allocation are different records, and this company has been graded almost entirely on the first one.

Tomorrow

What he said would happen, against what happened.

There is a specific public forecast attached to that acquisition, made at the time of purchase, and it can be checked against the outcome. Then the letter from the year the write-off landed, to see how the company described it in the document it wrote freely.

Not investment advice. The subscriber decides.

Start with why any of this matters before I get to the numbers.

The rule has to work in both directions

Yesterday I wrote that a long record can be luck riding a good industry, and that outcomes can be borrowed from a rising tide.

If I will not credit a management team for growth the whole sector enjoyed, I cannot blame one for a decline the whole sector is suffering. Same test, opposite direction. Without that check there is no argument I could make this week that could be proved wrong. Every poor number would count against management and every decent one would count for them, which is not analysis, it is decoration.

Two of six

The companies competing for this customer are Vuori, Alo Yoga, Athleta, Rhone, Gymshark and Nike.

Vuori, Alo Yoga, Rhone and Gymshark are all private. They file nothing. Everything published about their growth comes from the companies themselves, from funding announcements, or from trade estimates, and none of it is audited or filed. I have no way to check any of it, so none of it is in this piece.

That leaves Athleta and Nike.

Four of the six brands most often named as taking this customer cannot be examined at all. That is not a footnote. It shapes everything below.

The closest comparison reported the same number

Athleta is owned by Gap, which reports it as a brand within its results. Women’s activewear, sold mainly through American malls, to broadly the same customer.

For its quarter ended 1 August 2026, Gap reported Athleta net sales of $264 million, down 12%, with comparable sales down 12%.

Lululemon’s Americas comparable sales, for its quarter ended 2 August 2026, fell 12%.

Same category. Same market. Quarters ending one day apart. The same number.

I did not set out to find that and it complicates what I expected to write.

The consumer did not stop buying clothes

Inside Gap, in the same quarter, Gap brand comparable sales rose 10% and Banana Republic rose 3%. Old Navy fell 4%. The company’s overall comparable sales fell 1%.

One retailer, one quarter, the same malls and broadly the same shoppers. The denim brand grew double digits and the activewear brand fell 12%.

I want to be careful about what that proves, because I nearly claimed more than it does. Denim and activewear are not the same category. The research firm Circana, whose data the industry runs on, separates them: activewear on one side, non-active apparel including jeans on the other. So Gap brand growing tells you nothing directly about whether activewear is soft.

What it does establish is narrower and still useful. People were in the malls and they were spending. The consumer did not stop buying clothes. Whatever happened to Lululemon in the Americas, it did not happen because Americans stopped shopping for apparel.

The category itself is shifting

For the category, there is separate evidence, and it is a step further from the source than I would like.

Retail Dive, citing Circana’s Consumer Tracking Service, reported earlier this year that United States activewear apparel sales were up 2% against the prior year while non-active apparel was down 2%. Jeans, in the second group, grew 4%.

So activewear overall was still outrunning the rest of the wardrobe. Not collapsing.

But in the same reporting, a Circana apparel analyst, Kristen Classi-Zummo, described a clear shift away from leggings, sports bras and other performance products, even as activewear continued to outpace the apparel market.

That figure is from earlier in the year, not from the August quarter, and it comes through a publication rather than from the firm directly. I am giving it the weight that deserves, which is directional rather than precise.

Nike sits between the two. For its fiscal year ended 31 May 2026, Nike Brand apparel revenue rose 2% in constant currency, and North America apparel rose 4%. North America comparable store sales fell 2%, not 12%. But Nike apparel is a far wider basket than tights, and I am not presenting Nike as healthy either: the same filing shows Nike Direct down 6%, brand digital down 12%, Converse down 31% and net income down 3%.

Activewear did not stop working. The performance products at the centre of it appear to be where the softness sits, and that is most of what this company sells.

The explanation covers a sixth of the business

Yesterday I set out the geography without comment. Here is the comment, and I want to be careful with it because the numbers can be made to say more than they do.

China did slow. Comparable sales there fell 8%, and revenue fell 2% in constant dollars. It rose 4% only in reported dollars, which is a currency effect rather than a sales one. So her answer was addressing something real.

The question is how much of the problem it addresses.

China Mainland is roughly a sixth of this company’s revenue. The Americas is 67% of it, down from 70% a year ago, and Americas comparable sales fell 12%, four points worse than China.

And the filing gives that larger decline a different reason. It attributes the Americas fall to traffic, conversion and average order value, all moving the wrong way at once. Nothing in that sentence is about brand sentiment.

Brand noise in China is an answer about a sixth of the business. Two thirds of it fell harder, and the filing explains that part differently.

I should be fair about one thing. She was asked about China, and an executive answering the question put to her is not dodging. Analysts choose what to ask. But an explanation that covers the smaller part of a decline is still only an explanation of the smaller part, and a reader who heard the call and not the filing would come away with the wrong impression of where this company’s problem sits.

Athleta is a North American business. It fell 12% in the same quarter. Whatever is happening to Lululemon in the Americas is happening to the one comparable public company in the same market, and no event in Beijing accounts for either.

About that word

I have used her word twice now without examining it.

Noise describes weather. Something that happens around you, that you did not cause and cannot control, that passes. It also has the useful property of containing no claim, which means there is nothing in it to check.

There is a specific thing it may refer to. According to Forbes, on 30 May the company staged a yoga event on the Great Wall of China with the actor Zhu Yilong and roughly two thousand attendees, promoted as honouring Chinese culture. The performance featured a drum that critics identified as a Japanese taiko rather than a Chinese dagu. The criticism drew more than fifty million views on Weibo. Forbes reported that the company took over two weeks to acknowledge it and apologise, and that its statement said limitations in its professional knowledge had left it unable to identify potential controversies.

That event sits inside the quarter that ended on 2 August. I cannot tell you it caused anything, and one article I read quoted people expecting little short-term sales impact.

What I can say is that it was not noise. An event the company planned, staffed, promoted and then took two weeks to address is not something that happened to it. It is something it did.

The two-week gap is the part I will come back to on Thursday, because how long a company takes to say a thing went wrong is a decision, and decisions are what this week is about.

What this establishes

Less than a clean answer, and more than I expected.

It is not the consumer. Denim grew 10% in the same malls, in the same quarter, to the same shoppers.

It is not all of activewear either, which Circana had up 2% while the rest of the wardrobe fell.

It is not only this company. The one comparable public business selling women’s activewear in North America reported exactly the same 12% decline in the same quarter.

Something is wrong in the performance end of North American women’s activewear, and both of the public companies in that description have it.

That is as far as filings can take me, and the reason is the four private companies. If Vuori and Alo and Rhone and Gymshark are taking this customer, their filings would show it. There are no filings. The evidence that would separate a category problem from a share problem does not exist in public.

Which makes the rest of the week matter more, not less

If the sector had collapsed, there would be no management question worth asking.

If the sector had been fine, the decline would point straight at the people and the rest of the week would be a formality.

Neither happened. Part of this looks like something happening to a category. Part of it is happening in a market nobody at the company was asked about, with an explanation in the filing that has nothing to do with the answer given on the call.

So the only evidence that can separate those is what this company did with its own money, what it said about its own results, and what it took while saying it. That evidence is filed, it goes back a decade, and nobody has to guess at it.

Tomorrow

Eight years of Calvin McDonald’s capital record.

Ten years of cash flow statements opened at once. What they bought, what they wrote off, and which way the share count moved. There is one acquisition in there that the whole day turns on.

I will also run the first test I described last Monday, which is whether earnings grew faster than the capital base, because a business that adds to its equity every year produces rising earnings almost on its own, and that flatters a record that may not deserve it.

Not investment advice. The subscriber decides.

Let me be careful about what my own eyes are worth here.

Last week I described a method. This week I run it.

I ended last week on a line I want to pick up this morning. Every position is a bet on people, whether you admit it or not, so the size of the bet should match the evidence you have on them.

That is easy to say. Getting the evidence is the work, and it is the part nobody watches anybody do.

Over five days last week I set out where that evidence lives. Where the retained cash went. What price they paid for their own stock. Whether they named the bad year in the year it happened. What they took while writing it. And what to do when the record is too short to contain a bad year at all.

So this week I go and get it, on a real company, in public, and you watch what it produces.

The company is Lululemon, and the timing is what makes it worth six days. Most reviews of a management team look at people who are in the job. Here the chief executive of the last eight years left in January, two interim leaders have been running it since, and the incoming one starts tomorrow morning. There is no team in place to grade.

That sounds like a reason to wait. I think it is the reason to look now, because a company between leaders is the only time you can see the whole record laid out and nobody has yet started rewriting it.

A method nobody has seen run is a diagram. This is the same method with a company attached and the results printed whether they flatter it or not.

One promise before any of it. Where a test I published last week cannot be run here, or runs and tells me nothing, I will say so. A week that only demonstrates the tests that happen to work is an advertisement, and you would be right not to trust the next one.

What I saw is a reason to look, not a finding

One person, one mall, a handful of visits. That is an anecdote. It is not a sample, it does not represent the country, and I have no idea whether the shop I walked past is typical of the 825 company-operated stores the company reported in its quarterly filing on 3 September.

Philip Fisher had a name for this kind of observation, which I wrote about last Monday. Scuttlebutt. You learn about a business from the people around it rather than only from what the company publishes, and the point of it is not to replace the filings. It is to notice when the outside picture and the official one disagree, because the disagreement is the finding.

Here they appear to disagree, and I bought Dillard’s and Ralph Lauren off the same instinct years ago. Being a customer gave me a reason to look at both. It did not give me the answer. The filings gave me the answer.

So the question this week is not whether I saw a busy store. It is what a busy store and a 12% decline can both be true at the same time.

There is at least one answer that requires nothing surprising. Comparable sales measure what the same stores sold against the same period a year earlier. A shop can be full and still sell less than it did last year, if people are buying fewer items, or cheaper ones, or if last year was busier still. Traffic and spending are different measurements and only one of them is in the release.

That is a possible explanation, not a demonstrated one, and I am not going to pretend I have established it from a car park.

What the company reported on Thursday

These are the company’s own figures, from the quarterly report it filed with the Securities and Exchange Commission on 3 September and the call held the same day.

Net revenue fell 4%, or 5% in constant dollars, to $2.4 billion from $2.5 billion. Global comparable sales fell 9%, or 10% in constant dollars. In the Americas, comparable sales fell 12%. International comparable sales fell 3%, with revenue there up 4%.

Earnings per share came in at $2.92 against $3.10 a year earlier. That figure beat what analysts expected and it is lower than last year, which are both true at the same time and worth holding onto.

Then the numbers that moved the stock. The company cut its full-year outlook for the second consecutive quarter, to revenue of $10.35 to $10.5 billion, a decline of 5% to 7%, from a prior range of $11 to $11.15 billion. Full-year earnings per share guidance came down to $9.48 to $9.73 from $10.95 to $11.15. Third-quarter revenue is guided to fall 10% to 11%.

Operating margin was 18.8%, against 20.7% a year earlier. That figure includes $134.5 million of tariff refunds. Take them out and operating income is roughly $319 million on $2.42 billion of revenue, about 13.2%, which is more than seven points below last year.

The stock fell around 18% and traded under $100 for the first time since 2018.

Three more from the same disclosures, which matter more than they first appear.

Markdowns rose 70 basis points in the quarter, and the company expects them up again in the third quarter. A markdown is a price reduction, and a brand charging a premium for a logo does not usually need more of them.

The company ended the quarter with 825 company-operated stores, up from 811 at the start of the fiscal year and 784 a year earlier. It opened stores through a half in which sales at existing stores fell 9%.

And this, which is the line I keep returning to. The filing says the decline was driven mainly by reduced traffic, lower conversion, and lower average order value in the Americas.

Read that again slowly. Fewer people came in. Of the ones who came, fewer bought something. Of the ones who bought, they spent less than they used to.

Three separate things went wrong at the same time, at three separate points in the same transaction. That is not one problem. It is the whole funnel.

Nor is it confined to a channel or a country. Digital revenue fell 6%. United States revenue fell 8% and Canada fell 11%. Women’s revenue fell 4% and accessories fell 13%. China Mainland revenue rose 4% in reported dollars.

I am not going to tell you whether the price is attractive. I have not done that work and it is not what this week is about.

Why I am spending the week on management

Because a falling price raises a question that price cannot answer.

A brand this strong does not usually stop working overnight, and a business doing $10 billion of revenue with a 60% gross margin is not a broken machine. Something is wrong, and the range of possible somethings is wide: the product, the competition, the consumer, the pricing, the execution, or the people deciding all of it.

Management is one test among several. It is not the whole analysis and I would be misleading you to present it that way. But it is the test I spent last week building, and it happens to be the one that matters most when a company has to change direction, because changing direction is a decision and decisions are made by people.

There is one more reason, and it comes from the company itself. Asked on Thursday’s call about weakness in China, interim co-chief executive Meghan Frank attributed it to brand noise affecting sentiment and a softer Tmall shopping event, rather than to macro issues.

That is management saying this is not the industry and not the consumer. It is a claim about their own brand and their own execution, which is a claim I can test.

And this company has an unusual amount of management to look at.

Three records inside eight months

This is the thing to hold onto, and it is a habit worth applying to anything written about this company. When you read a claim about Lululemon’s management, check which of the three records it is describing, because they are separate and they answer different questions.

Calvin McDonald ran the company from 2018 until 31 January 2026. Revenue went from $2.6 billion to $10.6 billion under him. He left, and no filing I have seen explains why, which is a sentence I will come back to on Wednesday.

Meghan Frank and André Maestrini have run it as interim co-chief executives since. Thursday’s quarter is theirs. So is every decision made about capital in the last seven months.

Heidi O’Neill becomes chief executive tomorrow, 8 September. She spent more than twenty-five years at Nike, most recently as president of consumer, product and brand. She has run this company for zero days.

Three people, three records, and they answer different questions. Grading O’Neill on the last eight years is a category error. Grading the company on her Nike years is the opposite one. And the interim period is a real record in its own right. It contains the most recent decisions anyone has made with this company’s money, and it is the one I have seen least written about.

Last Monday I wrote that a projection is a claim about people who have not yet done the thing, and a record is a claim about people who already did. Two of these three have records. One does not, yet.

Tomorrow

Whether this is a Lululemon problem or an industry problem.

Last Monday I wrote that a long record can be luck riding a good industry, and that outcomes can be borrowed from a rising tide. That rule has to work in both directions. A bad record can be a bad industry, and if I skip that check, everything I write for the rest of the week is unfalsifiable.

So before I grade anybody, I put this company against its sector, using competitors’ own filings. The rest of the week is in the card above, and I will use notes throughout for figures that would otherwise pack the articles.

What this week is not

It is not a valuation. I have not done that work and I will not blend it into this one, because a low price has a way of making a management team look better than the evidence supports, and I would rather run the two separately and see whether they agree.

It is not a recommendation, and it will not become one by Friday.

And it is not a verdict on a brand. I saw a busy store. That observation survives everything I am about to write, and so does the 12% decline.

A great brand and a capable management team are different things, and a company can have one without the other in either direction.

Not investment advice. The subscriber decides.

Four days of tests, and each one needed a decade of decisions to run on. Plenty of companies have not existed that long, and the ones I find most interesting are often the youngest ones. Treating that as a closed question would be a lie about what I do, because I keep looking at them.

Start with what does not disqualify someone, because this is where the thinking usually goes wrong.

The wrong test is the industry

The soap salesman was Tom Murphy, a product manager at Lever Brothers, and the job came through a friend of his father’s. He took it with no broadcasting experience of any kind.

Seven years later he needed someone to run operations and hired a man from the new products division at General Foods, who had been testing Jell-O on his wife to work out what would sell, and who had never set foot inside a television station.

Murphy and Dan Burke went on to build Capital Cities and to buy ABC, a company several times their size, for $3.5 billion.

Here is the part that matters for our purposes. By William Thorndike’s accounting in The Outsiders, one dollar invested when Murphy became chief executive in 1966 was worth $204 by the time he sold to Disney in 1996. That is a 19.9% annual return against 10.1% for the S&P 500.

And against 13.2% for an index of leading media companies.

That second comparison is the one to hold. On Monday I warned that a long record can be luck riding a good industry, and that outcomes can be borrowed from a rising tide. Broadcasting was a rising tide. Murphy beat the tide by more than six points a year for three decades, which is the version of the evidence that survives my own objection.

A man with no experience in the business outperformed the people who had spent their lives in it. Whatever he lacked, it was not the thing that mattered.

So new to the sector cannot be the disqualifier. If it were, it would have caught the chapter that opens Thorndike’s book.

The right test is whether they have ever decided anything

The disqualification is not the wrong industry. It is the absence of any record of deciding.

You would not hire a head coach who has never coached. Not because he came from another sport, but because nobody has ever watched him choose a play with the game on the line.

Being excellent at the thing is not evidence of being good at deciding where the money goes. A founder whose entire record is technical achievement has demonstrated the first and nothing about the second.

So three questions, none of them about their sector.

Have they ever had to say no to growth that was available to them. Have they run anything through a contraction rather than an expansion. Have they built something and lived with the consequences of their own decisions rather than inheriting someone else’s.

What does not count is the resume. A resume is a list of rooms someone was in.

Murphy and Burke were explicit about this and applied it as employers rather than only as a philosophy. Thorndike records that both men preferred intelligence, ability and drive over direct industry experience, having had the benefit of that judgment themselves. Bill James was thirty-five with no radio experience when he was handed WJR. Phil Meek came from Ford at thirty-two with no publishing background to run a newspaper. Bob Iger was thirty-seven and had spent his career in broadcast sports when he was given ABC Entertainment.

They were not gambling on inexperience. They were declining to treat industry tenure as the qualification.

The exception, and why it is real

There is a version where I back an unproven team anyway, and it rests on something Buffett wrote in 1980.

He observed that with few exceptions, when a manager with a reputation for brilliance takes on a business with a reputation for poor fundamental economics, it is the reputation of the business that stays intact.

That line gets quoted as a warning, and it is one. Read the other direction it is a permission slip. If the economics of a business dominate the quality of the people running it, then an extraordinary business asks less of its managers than a mediocre one does. The moat does work the manager would otherwise have to do.

That is the honest basis for buying a wonderful business with an unproven team. It is the same asymmetry Buffett spent decades exploiting, pointed at a different variable.

It has two limits, and the limits are where I think the current enthusiasm is thin.

Limit one is the price

Margin of safety is the gap between what you pay and what a business is conservatively worth. In practice that gap is your budget for being wrong.

Management error is one of the things the budget pays for. Buy at a real discount and the team can make an expensive mistake while you still do acceptably. Buy at a price that already assumes everything goes right and there is no budget left, so the first bad allocation decision comes straight out of your return.

Margin of safety is what pays for management mistakes. At a price that requires perfection you are not buying a business, you are underwriting people you have never met.

So price does not only set the return. It sets how much management record you need before acting at all.

Limit two is how much cash the business must spend

This is the limit that gets skipped.

The 1980 logic works cleanly for a business that produces cash and needs little reinvestment to hold its position. There, allocation happens at the edges and the moat does the compounding.

A business that must deploy an enormous amount of capital every year is a different animal. Allocation is not a side activity a strong moat absorbs. It is the main event, and the company you own in ten years is close to the sum of what management chose to build.

You cannot route around a manager whose principal job is spending.

Which is why the two dials belong together. Cheap enough, and undemanding enough of capital, and the business can carry a team you cannot check. Priced for perfection and consuming capital at scale, and the team is the investment, whatever the product happens to be.

I am not reprinting the technology capital spending figures I gave you on Monday. They came forward from an earlier series and I have not re-sourced them this week, which under my own standard means they do not go in again. Look them up in current filings for whichever company you are weighing, and place it on the second dial yourself. The structure is the point, and the structure is checkable.

Every management team has a hidden assumption

I built the Firewall to name the single hidden assumption a stock price depends on. Not the ten things that could go wrong. The one thing that, if false, breaks the case.

It took me until this week to see that management teams have one too.

To be running a large public company there has to be a story. Not a marketing story, a story the team tells itself about why it is the right group for this problem at this size. Underneath it sits one assumption carrying the load.

For a founder whose company grew into something enormous, it is often that the hard part was the invention and the capital deployment is a detail. For an operator brought in from a mature industry, it is the discipline that travels. For a team in a capital-hungry business, it is frequently that money will keep being available on the terms they have grown used to.

Name it, and you can treat it the way we treat a price assumption. Set the confirm line and the break line before the news arrives, so you are reading evidence rather than reacting to it.

The question is not whether management is good. It is what would have to be true about these people for this to work, and what you would see if it stopped being true.

The one that started this

I graded Nebius, found a business growing around 454% a year, and could not buy it. I said the price was the reason.

Working through this week, I think the price was half my reason and it was the half I could articulate at the time. The other half is that at a price demanding perfection I needed a management record long enough to believe perfection was achievable, and I did not have one to read.

That is a statement about what I can verify, not a criticism of anyone running that company. I could be wrong about it. The point of writing the standard down now is that next time the reasoning is on paper before the decision instead of after.

The scorecard

Five days, five questions, and they run on any company in an afternoon.

Where the retained cash went, from ten years of the cash flow statement. What price they paid for their own stock, from the repurchase dollars against the change in share count. Whether they named the bad year, from the letter for the year the numbers already told you was bad. What they took while writing it, from the proxy. And when there is no record at all, what the business and the price have to be before you accept that.

Underneath all five is the thing I said on Monday and came back to on Thursday. I am not looking for brilliance. I am looking for people who know what they are not, and who did something about it that cost them.

Run it this weekend on a business you already own. Not one you are considering. One you own.

Two tracks

The short list has always had two columns on it. Quality, and price. This week was about the third one, and it is the column that decides whether the first two survive contact with the people in charge.

When the evidence runs out, and it will, the rule I am left with is short enough to say in one line. Every position is a bet on people whether you admit it or not, so size the bet to the evidence you have on them rather than to how much you like the idea.

You know how I end these. Hunting new ones, and waiting to buy more of the ones I already hold when fear puts them on sale. Same two tracks as always. The difference is that a week ago I could only grade one half of what I was buying.

Not investment advice. The subscriber decides.

There are two kinds of document in a company’s file, and the difference between them is the reason this piece exists.

The financial statements are a record. They are governed by accounting standards, audited by someone paid to disagree, and largely not a matter of choice. When a company overpays for an acquisition, the goodwill write-down appears whether anyone wants it to or not.

The letter to shareholders and the proxy statement are self-report. Nobody is required to be interesting, specific, or forthcoming in either. They are the company describing itself, and description is where a management team has the most freedom and therefore reveals the most.

Three days of this week have been spent on what a team did. Today is about what they say they did, and the gap between the two is not a rounding error. It is the finding.

Sixteen

The cleanest version of this comes from Buffett auditing his own vocabulary.

In his 2024 letter to Berkshire shareholders, he reported that across the 2019 to 2023 period he had used the words mistake or error sixteen times in his letters, and added that many other large companies had not used either word once over the same five years.

He named a single exception. Amazon’s 2021 letter, which he credited with some hard observations about itself. The general run of corporate communication to owners he described as pleasant talk and photographs.

Then the detail that gives the count its weight. He wrote that he had served as a director of large public companies where mistake and wrong were effectively forbidden words at board meetings and on analyst calls, and that the implied claim of managerial perfection made him uneasy.

His own qualification stays attached, because he included it. There are circumstances where legal exposure makes limited discussion the sensible course. This is a litigious country and that is a real constraint.

Buffett credited Munger with the harder half. Munger’s position was that the sin is not the mistake, it is delaying the correction of it, which he called thumb-sucking. Problems cannot be wished away and they require action, however uncomfortable that is.

Which tells you what the count is measuring. Naming an error in writing is the cheap part. It matters because a team unwilling to say the word in a letter, where saying it costs nothing but discomfort, is not likely to be the team that moves fast when the same problem shows up in the business, where acting costs real money.

The count is useless without a denominator

Here is where I have to be careful, because a raw number tells you nothing.

Sixteen is meaningful for Berkshire because Buffett published it about himself over a defined period. You cannot take that figure to another company and grade against it. A team with three admissions in five years is not automatically worse than one with eight, and a company with zero might have had an uneventful five years.

The denominator you need is not a benchmark from outside. It is the work you already did.

You have ten years of cash flow statements from Tuesday. You know what they bought, what they later wrote off, and which way the count moved. You have the average price paid for their own stock from Wednesday. Somewhere in that decade there is at least one year where the numbers went badly, and you know which year it was, from a document the company did not get to write freely.

So the test is not how many times they used the word. It is whether they used it about the thing you already found.

Pull the letter for that specific year. Three outcomes, and they are easy to tell apart.

They named it, described what went wrong, and said what they were doing about it. They mentioned it in language so general it could have described any year at any company. Or the year that produced a write-down was characterised as a period of investment, discipline, or repositioning, and the write-down does not appear.

A team that admits a mistake you had not found is candid. A team that admits the one you were always going to find is doing arithmetic of a different kind.

That last distinction is worth holding. Timing is part of the test. An error named in the year it happened is a different signal from the same error acknowledged three years later, once the market has priced it and the admission costs nothing.

What the research supports, and what it does not

There is data pointing the same direction, and I want to state it at exactly the weight it can carry.

Feng Li, in the Journal of Accounting and Economics in 2008, found a statistically significant relationship between annual report readability and company performance. Firms with lower earnings produced reports that were harder to read, and firms whose reports were more readable had more persistent earnings. Researchers in the field call the pattern obfuscation.

The measurement has been challenged. Loughran and McDonald argued in 2014 that the Fog Index used in much of that work scores ordinary financial vocabulary as complexity, which makes it a poor instrument for business text.

So the direction has held up under other measures and the instrument is disputed. That supports a modest claim rather than a sweeping one. How a company writes about a bad year is not noise, and it is not proof.

Two smaller tells, both free

The headline metric. Every company chooses which number to lead with, and Buffett has been blunt about his own choice, dismissing EBITDA as a flawed favorite of Wall Street and reporting operating earnings instead. The tell is not which measure a company picks. It is whether the pick moves. A team that has led with the same number for a decade through good years and bad is reporting. A team whose headline figure migrates to whatever looked best that year is presenting.

The outside account. Monday I introduced Fisher’s scuttlebutt. Here is where it earns its place, because the letter is the company’s version of the year and customers, suppliers, competitors, trade press and job postings are versions written by people with no reason to improve it. Where the two agree you have learned little. Where they diverge, the direction of the divergence tells you which one was managing an impression.

What they took while they were writing it

The proxy statement is the second self-report, and it is the one with numbers in it.

In the United States it is filed as a form called the DEF 14A, ahead of the annual shareholder meeting. It sets out executive compensation in detail, the terms attached, and what shareholders are voting on. It is written by the company about the company, which puts it on the same side of the line as the letter.

Four things in it are worth ten minutes. Severance multiples and change-of-control provisions, meaning what an executive receives if the business is sold or they leave. Whether equity vests on a single trigger, releasing on the sale alone, rather than a double trigger requiring both the sale and a qualifying departure. Share pledging, where executives borrow personally against company stock, which quietly converts an owner into someone who needs the price to stay up. And how much of the package depends on results that already happened rather than on remaining employed.

Shareholders have been voting on some of this and the votes have moved. According to the Harvard Law School Forum on Corporate Governance’s review of the 2026 proxy season, average support for advisory votes on change-in-control pay fell from roughly 86% in 2024 to about 74% this year, with a record ten of those votes failing.

Then read the two self-reports against each other, because they cover the same twelve months and they can disagree. A letter naming a decision that went wrong, filed alongside a proxy showing the full package paid as though nothing did, is two accounts of one year. So is the reverse.

Why this is the day the week has been pointing at

Monday I said the trait I look for is narrower than intelligence. Whether a person knows what they are not good at, and acted on that knowledge when bluffing would have been easier. I said it was the thread running through every test in this series, and then I spent three days on arithmetic without coming back to it.

Here is why it had to wait.

Self-awareness does not appear in a cash flow statement. It cannot. A write-down records that a company paid too much; it says nothing about whether anyone involved understood why. The share count records that a repurchase happened; it does not record whether the team knew what the shares were worth when they bought them.

The trait only becomes visible at the moment a management team is free to describe itself and chooses to describe itself accurately. That moment is the letter. It is the one document where the cost of honesty is entirely borne by the person writing it, and where the reward for the alternative is immediate.

Every other test this week measures a decision. This one measures whether they can see the decision clearly, which is the thing that predicts the next one.

What this still cannot tell you

Everything so far has assumed a record exists.

A decade of financing lines to total. Enough share counts to see a direction. A shelf of annual letters long enough to compare one against another. A proxy with years behind it rather than a single filing.

Some of the businesses I most want to own have none of that. The company is young, the team is new, and there is no decade to read. That is not a hypothetical and it is where the framework this week has built either holds or breaks.

Tomorrow, the exception.

Not investment advice. The subscriber decides.

Start with the standard, because it is stricter than it first appears.

Two conditions, and one of them is not enough

In his 2011 letter to Berkshire shareholders, Buffett set out when he favours repurchases. Two conditions, and he wanted both.

First, the company has ample funds for the operating and liquidity needs of the business. Second, the stock is selling at a material discount to intrinsic business value, conservatively calculated.

Intrinsic value, in plain language, is what the business is worth based on the cash it can produce over its life, as opposed to what the market happens to be charging for it today. Buffett’s word “conservatively” is doing real work in that sentence. A management team that wants a repurchase can always find a valuation that justifies one.

Notice what the second condition does. It makes the buyback a price decision rather than a policy. He put this plainly in his 2023 letter, writing that all repurchases should be price-dependent, and that what is sensible at a discount to business value becomes foolish at a premium.

A buyback is not a way of returning cash. It is management buying one particular stock, and the only stock they are allowed to buy is their own.

He also named who gains when a company overpays. The shareholders who sold, and the banker who recommended it. The owners who stayed are the ones who paid.

The number the announcement never gives you

Here is where I go past yesterday.

Tuesday I told you to track the share count rather than the buyback dollars. That is the beginning of the test, not the test. The count tells you whether the buyback did anything. It does not tell you whether it was done well.

For that you need the price they paid, and you can compute it yourself from figures the company already publishes.

Take the total dollars spent on repurchases over a period. Divide by the number of shares the count fell by over that same period. That gives you the average price management paid per share retired. Now set that number against the stock’s trading range across those years.

If the average sits in the lower part of the range, this is a team that bought when the market was unenthusiastic. If it sits in the upper part, they bought alongside everyone else, at the prices everyone else was paying.

One caution on the arithmetic. If the company issued shares over the same period, and most do through compensation, the count fell by less than the number repurchased. Your computed average price will be too high. Use the repurchase share figure from the cash flow statement or the equity note when it is disclosed, and treat the divided number as an estimate rather than a measurement.

That caution is also the finding. A company can spend an enormous sum on repurchases and end the decade with the same share count it started with. The dollars went out. Your ownership stake did not move. What was purchased was not your position in the business. It was the offsetting of shares issued to employees.

That is a real and common structure, and it is not hidden. It is visible to anyone who puts the dollars and the count side by side.

What the evidence says about how this goes

I want to be careful here, because this is a point where confident assertions outrun the data.

Researchers at the University of Kentucky examined 5,498 firms that repurchased stock in at least one quarter between 1984 and 2010, in a paper titled “Wiser to Wait: Do Firms Optimally Execute Share Repurchases?” They found strong evidence that share prices were higher, and valuation ratios less attractive, during the quarters when firms were repurchasing than during the quarters when they were not. Their summary was that firms on average buy when they should not.

McKinsey’s own work, published through the Harvard Law School Forum on Corporate Governance, reached a similar conclusion on timing. A majority of the companies they observed bought back shares when prices were high rather than low.

Now the part that complicates it, which I am including because leaving it out would be dishonest. That same McKinsey work found no compelling evidence that share buybacks damaged long-term value creation for investors overall. So the timing finding is robust and the consequence is contested. Poor timing on repurchases is not the same thing as a company being destroyed by them.

There is also a mechanical reason for the pattern that has nothing to do with anyone being foolish. Companies tend to have the most spare cash when business is good, and business tends to be good when share prices are high. The behaviour follows the cash. According to data from S&P Capital IQ reported by CFO magazine, during 2009, with the S&P 500 below 700, only 53 buybacks of $300 million or more were announced.

Buffett has made the sharper version of this charge himself. He has written that American chief executives have an embarrassing record of committing more company money to repurchases when prices have risen than when they have fallen. That is his assessment, and I am attributing it to him rather than adopting it as a measured fact.

What it looks like when someone does it properly

Henry Singleton ran Teledyne, and I mentioned him on Monday only as one of Thorndike’s eight. The record itself deserves its own space.

Through the 1960s Singleton used Teledyne’s expensively priced stock as currency, acquiring roughly 130 companies while the shares traded at high multiples of earnings. Then the market turned, the multiple collapsed, and he reversed direction.

Between 1972 and 1984, across eight separate tender offers, Teledyne repurchased approximately 90% of its outstanding shares, spending on the order of $2.5 billion. Shares outstanding fell from 88,827,372 in 1971 to 22,564,756 by 1980. Earnings per share rose roughly fortyfold between 1971 and 1984.

Read those two phases together, because separately they mean less. He issued stock when it was expensive and bought it back when it was cheap. That is the same discipline applied twice, in opposite directions, which is harder than it sounds and rarer than it should be.

One detail I am including because it cuts against the clean version of the story. The final tender, in 1984, was priced at $200 a share, roughly $30 above the market at the time. Singleton was not mechanically a low-price buyer at every moment across twelve years. The record is excellent. It is not spotless, and a story that has no rough edges usually has had them removed.

He was not right about the price every time. He was thinking about the price every time. That is the part you can check.

The 2026 version

Monday I gave you the buyback figure for the first four months of this year, reported by Bloomberg citing Birinyi Associates as the largest start to any year on record. I gave it without comment because the frame was not built yet. It is now.

A record dollar total tells you about volume. It tells you nothing about price, and price is the entire question. The same headline is consistent with a market full of disciplined allocators buying value, and with a market full of teams buying alongside each other near highs. The number does not distinguish between them.

What distinguishes them is available per company, for free, and it is arithmetic rather than opinion. Dollars spent. Change in share count. Average price paid. Trading range over the same period.

The afternoon, part two

Take the same company you used yesterday.

Find the repurchase line in the financing section of the cash flow statement for each of the last ten years. Find the share count on the cover of each annual report. Compute the average price paid. Put it next to what the stock did in those years.

You will end up with a single number that tells you whether this team treats their own shares as an investment or as a policy.

What this still cannot tell you

The arithmetic gives you the decision. It does not give you their account of the decision.

A team that overpaid and then said so in the next annual letter is a different proposition from a team that overpaid and described the year as a success. The numbers are identical. The management is not.

Tomorrow I leave the financial statements and go to what they wrote, and to the document that records what they took while they were writing it.

Not investment advice. The subscriber decides.

Yesterday I gave you the figure from Buffett’s 1987 letter without showing the work: ten years in, at an ordinary retention rate, one chief executive has deployed more than 60% of the capital in the business.

I should have shown the work. It is arithmetic and it takes one line.

If a company retains 10% of net worth a year, the capital base grows by a factor of 1.1 each year. After ten years it is roughly 2.6 times its starting size, which means the original base is down to about 39% of the total and everything else arrived on this person’s watch. The share is one minus 1.1 to the power of negative ten, or 61.45%. At year five it is already 37.9%.

The number is not a rhetorical flourish. It is a division problem, and you can run it on any retention rate you like.

That is the reason I now open the cash flow statement before the income statement.

The part of that letter I skipped

There is a line in the same letter I let stand yesterday without the qualification it needs.

Buffett’s warning about chief executives who go looking for help with allocation, from staff, from consultants, from bankers, is not an argument against seeking expertise. Read that way it becomes an excuse for exactly the behavior this week is trying to catch, which is a manager bluffing through something they do not understand.

The warning is narrower and better than that. The allocation judgment itself cannot be handed to someone whose incentives differ from the owners’. An investment bank is paid when a transaction happens, so it is structurally in favor of a transaction. That is a fact about the arrangement rather than an accusation about anyone inside it.

Buffett also gave the problem its sharpest image, comparing a manager promoted into capital allocation to a gifted musician whose final promotion is not a performance at Carnegie Hall but the chairmanship of the Federal Reserve.

Hire past your weakness, keep the decision. Those two things together are the test, and either one alone fails it.

Five doors

Every dollar a company keeps goes through one of five doors.

It gets reinvested in the existing business. It buys another company. It pays down debt. It goes out as a dividend. Or it buys back the company’s own stock.

That is the entire menu. Management picks, every quarter, forever.

Grading the picks does not require a model. It requires reading ten years of the cash flow statement in one sitting, which takes an afternoon.

It is worth knowing how uncommon that is, and I would rather give you a number than an impression. Tim Loughran and Bill McDonald, writing in the Journal of Behavioral Finance, used the SEC’s own EDGAR server logs covering 2003 to 2012 to count how often investors requested company filings. The average publicly traded firm had its annual report requested 28.4 times in the period immediately after filing.

Not 28,400. Twenty-eight.

Two limits on that figure, because it should not be stretched further than it goes. It counts direct requests to EDGAR, so it misses anyone reading the same document through a broker, a terminal, or a data provider. And the sample ends in 2012. What it does establish is that first-source filings were being pulled far less than the volume of commentary about those companies would suggest.

Here is what I look for.

Total the acquisitions over the decade and compare that total to what the acquired businesses are contributing now. If a company spent $4 billion buying things and the segment those things live in has not grown, that is an answer. Look for goodwill write-downs, which are the accounting admission that a company paid more for something than it turned out to be worth. A decade with several is a decade of overpaying.

Compare capital spending to depreciation. When a company consistently spends far more on plant and equipment than it is writing off, it is either growing hard or running to stand still, and the two look identical in the cash flow statement until you ask which.

Track the share count. Not the buyback dollars, the count. Dollars spent tell you what they did. The count tells you whether it worked.

The number that separates growth from running in place

This is where Buffett gave us a tool, in an appendix to his 1986 letter, and it is the one I use most.

He called it owner earnings. Take reported earnings, add back depreciation, amortization, and other charges that did not involve cash going out the door, then subtract the capital spending the business needs simply to hold its competitive position and its unit volume. That last piece is maintenance capital expenditure, meaning the money spent to keep what you already have rather than to build something new.

What is left is roughly what an owner could take out of the business in a year without weakening it.

The distinction matters because the common measure, free cash flow, subtracts all capital spending and therefore treats a dollar of growth investment the same as a dollar spent replacing a worn-out roof. Those are not the same dollar. One is optional and should be judged on its expected return. The other is a cost of staying in business.

Buffett was direct about why he bothered. Owner earnings, not the reported figures, are the relevant number for valuation, both for investors buying stocks and for managers buying whole businesses.

The catch is that companies do not report the split. Estimating how much of capital spending is maintenance and how much is growth is a judgment call, and two careful people will land in different places. That is a feature rather than a bug. The exercise forces you to ask where the returns are coming from, using a split the company did not choose for you.

A business that must spend heavily every year just to stay where it is has less to give you than its earnings suggest. That gap is not reported. You have to compute it.

The tests travel. The documents change.

Everything I have described so far assumes a particular kind of company. One that generates cash, keeps some of it, and spends it on things you can point at. That describes a lot of businesses. It does not describe all of them.

I want to be direct about this, because a method that only works on Berkshire-shaped companies is not a method. It is a preference.

The five questions generalize completely. Where did the capital go, what return did it earn, how much of the spending was just standing still, what price did they pay for their own stock, and is there a quick way out. Those hold for any company anywhere. What changes is which document answers them.

For a company that has never retained a dollar, the whole thing inverts. There is no retained cash to trace, so the allocation record is what they raised and what they surrendered to raise it. The share count going up is the document, not the share count coming down. Dilution history is the cash flow statement of a business that has not started generating cash yet.

For a business whose main investment is people rather than plant, the spending runs through the income statement as research and sales rather than appearing as capital expenditure. The maintenance-versus-growth question is still exactly right. You will not find the answer in the cash flow statement, because there is nothing to find there.

And the return-on-equity test needs a correction that Buffett’s own framing can obscure. Penalizing a company for growing its equity base is correct for a mature business and wrong for one earning high returns on every incremental dollar, where retaining everything is the right decision. The general form of the question is the return on new capital, not the level of return on all capital.

One more, for anyone holding something listed outside the United States. There is no proxy statement to read. Barclays, to take one I own, discloses the equivalent in a UK remuneration report. Different document, same question.

The instrument is not the test. If you cannot find the filing, that does not mean the question stopped applying.

Why decent, intelligent people spend money badly

None of this explains why capital gets misallocated so consistently. Bad managers would explain a few cases. It would not explain the pattern.

Buffett’s answer came in his 1989 letter, in the section reviewing his first twenty-five years of mistakes, and he called it his most surprising discovery. He named it the institutional imperative.

His starting assumption had been that decent, intelligent, experienced managers would make rational business decisions more or less automatically. He learned that this is not so, and that rationality tends to wilt when the imperative comes into play.

He described four mechanics. An institution resists any change in its current direction, the way an object in motion resists a change in course. Projects and acquisitions materialize to absorb whatever funds happen to be available. Any wish of the leader, however poorly founded, gets supported by detailed rate-of-return studies prepared by the people who work for that leader. And the behavior of peer companies, whether in expanding, acquiring, or setting pay, gets copied with little thought.

Then the line that makes it useful rather than cynical. These are institutional dynamics, not venality or stupidity.

That distinction is why I can write about this without picking a side about anyone. Nobody in this description is a villain. The mechanism runs on ordinary incentives inside ordinary organizations, and it runs hardest when there is a lot of money around and everyone in the industry is moving the same direction at once.

Buffett’s response was structural rather than moral. He wrote that after some expensive mistakes he tried to organize and manage Berkshire in ways that minimized the imperative’s influence, and to concentrate investments in companies that seemed alert to the problem.

That last clause is a management test in five words. Does this team seem alert to the problem.

What that means to look at right now

I will describe the mechanics and leave the conclusions to you.

Yesterday I put two numbers in front of you without comment, the capital spending guidance across the largest technology companies and the record start to a year for buyback announcements. Here is the comment.

Hold those numbers against the second and fourth mechanics. Projects materialize to absorb available funds. Peer behavior gets copied.

I am not saying the spending is wrong. I have no idea, and neither does anyone writing confidently about it. I am saying that this is the exact set of conditions under which the imperative operates most strongly, and that the question worth asking about any individual company in it is narrow and answerable: does this particular management team behave like a group that is alert to the problem, or like a group that is matching its peers.

The difference shows up in what they say no to. A team that has declined something available to it, publicly, and explained why, has told you which one it is.

The afternoon

Pick a company you own. Not one you are considering, one you own.

Open ten years of cash flow statements. Total the acquisitions. Find the write-downs. Compare capital spending to depreciation. Track the share count, not the dollars.

You will finish with a view of that management team built from what they did rather than what they said, and you will have it before the price moves rather than during.

The short list I keep is supposed to be businesses I could act on quickly. For years that list was built on quality and price. This is the column I never filled in.

There is one thing this afternoon cannot tell you, and I want to name it rather than let you find out later.

Everything above reveals what management did. Totals, write-downs, the direction of the share count. None of it reveals whether they got a good deal. A company that spent $4 billion on acquisitions and a company that retired 8% of its shares both show up as activity, and activity is not the same as judgment. The price they paid is a separate question, and it is the one that separates a disciplined allocator from a busy one.

Tomorrow I take the fifth door on its own, because it is the only one of the five where the price is visible to you in advance.

Not investment advice. The subscriber decides.

Andrew Kilpatrick spent three decades recording what Buffett and Charlie Munger said at Berkshire annual meetings. In his account, Buffett reduced the evaluation of a management team to two things: how well do they run the business, and how well do they treat the owners.

That is the whole frame. Every other test I will write about this week is a detail hanging off one of those two hooks.

But there is a rule that sits above both questions, and I want to put it first because it decides what evidence you are allowed to use.

Buffett printed the rule in the back of the annual report

Since at least 1983, Berkshire’s annual report has carried a short list of acquisition criteria on the page after the shareholder letter. Six items. Two of them are about people, and both say the same thing.

The second criterion asks for demonstrated consistent earning power, and then adds a clause most readers skim past: future projections are of no interest, and neither are turnaround situations. The fourth criterion is four words long. Management in place. The parenthesis after it explains why: he cannot supply it.

Read those two together and you have a policy, not a preference. Buffett is not saying he prefers a proven team. He is saying that a plan is not admissible evidence, and that he will not underwrite someone’s potential with his own money.

A projection is a claim about people who have not yet done the thing. A ten-year record is a claim about people who already did it, repeatedly, through conditions nobody chose.

I want to be careful not to overstate what a record proves. A long record can be luck riding a good industry. That is why the record has to be read as decisions, not outcomes. What did they do with the cash in the fat years. What did they do in the year the business got hit. Did they buy their own stock when it was cheap or when it was expensive. Outcomes can be borrowed from a rising tide. Decisions cannot.

William Thorndike made this measurable in his book The Outsiders. He studied eight chief executives, including Henry Singleton at Teledyne and Tom Murphy at Capital Cities, and found their companies returned an average of 20.1% a year to shareholders during their tenures against 12% for the S&P 500 over the same windows. Compounded, that gap meant beating the index by roughly twentyfold.

None of the eight is remembered for charisma. What they shared was a long, visible, boring record of deciding where the money went.

There is one more thing worth naming before the week starts, because it is what I am looking for underneath all of this.

A record of decisions is evidence of something narrower than intelligence. It is evidence of whether a person knows what they are not good at, and whether they acted on that knowledge when it would have been easier to bluff. That trait is the one I trust, and it is the thread running through every test in this series.

Where the cash went tells you whether they knew the limits of what they could build. What they paid for their own stock tells you whether they knew the limits of their own optimism. Whether they name a mistake before the market finds it tells you whether they can see themselves clearly at all.

I am not looking for brilliance. I am looking for people who know what they are not, and who have done something about it that cost them.

That last clause is the whole guardrail. A chief executive who sounds humble on a conference call is performing humility, and performance is free. I want it in the form of a decision that gave something up.

The first question is not about charisma

Most coverage of a chief executive is coverage of a personality. The stage presence. The vision. The interview where they seem impressive.

None of that is the first question.

The first question is a number, and Buffett named the number as far back as his 1977 letter to Berkshire shareholders. He pointed out there that companies love to announce record earnings per share, and that a record is close to meaningless on its own. A business that adds to its equity base every year will produce rising earnings almost automatically, the same way a dormant savings account produces rising interest through compounding.

His preferred measure was return on equity capital. Return on equity, or ROE, is simply the profit a company earns each year divided by the money the owners have left in the business. It answers a plain question: for every dollar of mine you are sitting on, how many cents did you make this year.

A record earnings number tells you the company got bigger. Return on equity tells you whether it got better.

That distinction is the entire first question. A management team that grows earnings by 5% while the equity base grows 10% has not run the business well. It has run a larger version of the business worse, and the headline hid it.

The second question is the one almost nobody asks

How well do they treat the owners.

This is not about dividends, and it is not about whether the chief executive seems likable on an earnings call. It is about four concrete behaviors, and I will spend the rest of the week on them.

Do they allocate capital well, meaning do they put the retained profit somewhere that earns a decent return. Do they buy back stock at a price that helps me, or at a price that helps the optics. Do they tell me the truth early when something breaks. And do they take a share of the company for themselves through compensation that I would not have agreed to if anyone had asked me.

Those four are testable from public documents. Not one of them requires access, a conference, or a call with investor relations. They are in the annual letter, the cash flow statement, and the proxy statement.

Why I care about this now, specifically

I did not pick this topic because it is timeless. I picked it because 2026 is, as far as I can tell, the largest capital allocation moment any of us will watch in our investing lives.

Two numbers, both from the reporting I have already cited in this series and in the financial press. The Big Five technology companies are guiding to somewhere in the range of $775 to $800 billion of capital spending this year, most of it flowing to a narrow set of recipients. And according to Bloomberg, citing data from Birinyi Associates, American companies announced $665 billion of share repurchases in the first four months of the year, the largest start to any year on record.

Set aside every opinion about whether that spending is wise. The mechanical fact is enough: an extraordinary amount of shareholder money is being deployed right now by people who will not be asked to justify it for years.

Capital allocation is not a boring corner of governance. In 2026 it is the loudest thing happening, and almost nobody is scoring it.

The job nobody interviews for

Buffett made the sharpest version of this point in his 1987 letter, and it has stayed with me since I first read it.

His observation was that most people reach the top of a company by excelling at something else. Marketing. Production. Engineering. Administration. Sometimes internal politics. Then, on the day they become chief executive, they inherit a job they have likely never done and that is not easy to master, which is deciding where the company’s money goes.

He put a number on how much that job matters. A chief executive whose company retains earnings equal to 10% of net worth each year will, after ten years on the job, have been responsible for deploying more than 60% of all the capital at work in that business.

Read that again slowly. After a decade, the majority of the company you own was assembled by that person’s allocation decisions, not by the founder, not by the brand, not by the moat. By their choices about where the cash went.

Buffett added the part that stings. Chief executives who recognize they lack the skill often turn to their staff, to consultants, or to investment bankers. In his and Munger’s experience, that help tended to make the problem worse rather than better.

What my own record taught me

I own Dillard’s at an average cost of $33.43. It trades around $560 today. I own Ralph Lauren at $73.17, now around $378. Those are the two I point to most, and I want to be careful about why.

They are not evidence that I predicted anything. They are evidence that I was ready when fear put businesses I understood on sale, and that I did the work before the moment arrived rather than during it. The reward was for being ready, not for guessing.

But here is the piece I did not appreciate at the time. I bought a retailer and an apparel brand in a market that had decided both were structurally finished. The reason both worked was not that the market was wrong about the sector. It was that in both cases the people running the business made a long series of unglamorous capital decisions that a different management team would have made differently, and worse.

I shop at Dillard’s. I have bought Ralph Lauren there at deep discounts for years. Once I bought a Herschel suitcase with a $325 tag for $113, and the discount made me happier than the luggage did. Being a customer gave me a reason to look. It gave me conviction when the screen looked ugly. It did not give me the answer. The filings gave me the answer, and the part of the filings I underweighted was the management scorecard.

That is the gap I am closing this week.

Why I am not just copying Buffett

Sam Walton said he spent more time in his competitors’ stores than in his own. He was not there to admire them. He was there to take what worked, bring it back, and improve on it.

That is the method behind this whole week. Find whoever does a thing better than everyone else, understand exactly how, then build your own version that fits your situation. I am not writing a summary of Buffett’s views. I am building a guideline for vetting management that any subscriber can run, and he is the first source because he is the most documented, not because he is the only one.

He also has a specific gap, and it happens to sit exactly where I am hunting.

His own published acquisition criteria include a preference for simple businesses, with a parenthetical admitting that if there is a lot of technology involved he will not understand it. That is an honest limit and it has served him. But I am a value investor who leans toward the picks-and-shovels layer of a technology buildout, and I have some technology background of my own. A framework that routes everything technical to the too-hard pile cannot help me where I most need help.

So the second source is Philip Fisher, and Buffett has said openly that a meaningful part of his own approach came from him.

Fisher was the growth investor, working in technology companies in the 1950s and 1960s, and his fifteen-point checklist is heavily weighted toward management quality: depth of the team beyond one person, whether they plan on a long horizon or a short one, how they treat their people, and whether they are candid when things go wrong.

His research method is the one Sam Walton was practicing without the name. Fisher called it scuttlebutt. You learn about a business by talking to the people around it, meaning customers, suppliers, competitors, and former employees, rather than only reading what the company publishes about itself.

Management has every incentive to describe itself favorably. Their competitors have none. When the outside picture contradicts the official one, the contradiction is the finding.

That is not a soft technique. For most of us it is the only primary research available, and it is more accessible now than it was in 1958. Customer reviews, industry forums, job postings, employee reviews, and trade press are all scuttlebutt, and they are free.

Buffett gives me the discipline about records and capital. Fisher gives me a way to work in businesses that Buffett would decline to analyze at all. The guideline this week is built from both, and I will keep adding sources to it as I find people who do this better than I do.

What this week covers

Four places the record shows up, and one piece about what to do when there is no record at all.

Tuesday: capital allocation, the one job that cannot be delegated, and how to read ten years of a cash flow statement to grade it. Plus the force Buffett called the institutional imperative, which explains why decent, intelligent managers spend money badly without any villain in the story.

Wednesday: share buybacks, and the two conditions Buffett has said must both be true before a repurchase helps you. This is the record across a full cycle, and it is where the $665 billion number gets uncomfortable.

Thursday: candor, which is the record in a bad year. How a management team writes about its own mistakes, and what the proxy statement says about them that the annual letter never will. The letter is the words. The proxy is the receipts.

Friday: the exception. Sometimes I want to own a business whose management has no long record to read, because the business itself is extraordinary. That is a real situation and pretending otherwise is dishonest. So Friday asks what the business has to be worth before I will accept an unproven team, where I draw that line, and why the line has to be drawn in advance. The full scorecard closes the week.

The point of doing this before the price moves

I do not forecast crashes. I prepare for them. That means keeping cash ready and keeping a short list of businesses I understand well enough to act on quickly, so that when fear does the pricing I am not starting my research from zero.

A short list built only on business quality and valuation is half a list. The other half is knowing, in advance, which management teams I would trust with a decade of retained earnings and which I would not. That work is slow, it is unglamorous, and it is impossible to do properly in the middle of a panic.

History gives you the setup, not the date. The scorecard is something you can finish before either one arrives.

Two tracks, as always. I am hunting new great companies I do not own yet. And I want to buy more of the great ones I already own when they go on sale, and lower my average on the ones I intend to hold forever. Both tracks run through the same question, which is whether the people inside will treat my money the way I would.

Tomorrow, the one job.

The Long View · The reckoning at the end of the series · Evidence, not prediction

The Long View makes no forecast here and names no target. This is a reading of public data and an honest conclusion. Verify everything and decide for yourself.

Where this started

A few weeks ago I graded Nebius, an AI company growing more than four hundred percent a year, one of the most impressive businesses I had ever scored. And I could not buy a share, because the price demanded everything go right forever. A wonderful company at a terrifying price.

That left me one question I could not put down: what would ever make a company this good come back to a price a value investor could pay? I meant it about one stock. It became this series, because the answer kept pulling me further out, to the bond market, the national debt, the consumer, and finally the machinery of the whole economy. This is where that road ends.

But before I show you where it leads, you should know what kind of investor is doing the looking, because it is the reason I went looking at all. I am a value investor, and if I am honest, I am a value everything. I bought a large Herschel suitcase at Dillard’s last weekend, 325 dollars on the tag, and I paid 113 for it, sixty-five percent off. The discount made me happier than the luggage. I am not cheap. I just cannot stand overpaying for something I know the real worth of, whether it is a suitcase or a business.

That instinct is not a hobby. It is the entire method, and it is what makes a piece like this matter, because a value investor spends most of a career waiting, and only acts when fear puts a great business on sale. I know it works, because I have done it.

When the world was terrified in 2020, I was buying.

My average cost on Dillard’s is about 33 dollars, and it trades near 560 today. Ralph Lauren, an average near 73, now around 378. Cardinal Health, around 56, past 230. Coca-Cola near 48 and AbbVie in the 120s, both since more than doubled.

I have shopped at Dillard’s for years and bought more Ralph Lauren there at deep discounts than I care to admit, so when both went on sale in 2020, I was not reading a screener. I was buying businesses I already knew as a customer.

But let me be clear about what that pattern is not. I did not predict COVID. I did not forecast the crash. I simply had cash set aside and a short list of businesses I understood, so that when the moment came, I was ready to act while others were frozen. Not predicting the storm, but being prepared for it. That is the entire edge, and it is what this series is built around.

Underneath all of it ran the most boring habit in investing: dollar-cost averaging into my index funds, the same amount every month, through the fear and long after. My average cost in the Vanguard technology index sits near 47, and it trades around 120 today. I timed none of it. I just never stopped. A down market is the best time to lean into that too, more shares at lower prices, which matters most if you are building toward a goal or nearing retirement.

So that is the lens. And here is why I am telling you now, at the start. In 2020 I was ready by temperament and system alone, no dashboard of warning lights, just cash and a discipline. This time I have that same preparation, plus something I did not have then: the instruments are already flashing. The data does not make me a value investor, my system does that.

But a set of readings this abnormal is a gift on top of it, a signal the environment may be turning my way sooner than usual. That is what sent me down this road, and it is where the rest of this piece goes.

The point where the data started to concern me

I have no interest in alarm, so let me do the opposite. I will name every number and its source, so you can check me, and let the weight speak for itself.

Start with the consumer, because in this country the consumer is not one part of the economy, the consumer is the economy. Household spending drives roughly two-thirds of it. When the people who do that spending start to strain, that is not a side story. It is the main one.

And the strain is not my opinion. The Federal Reserve Bank of New York’s Household Debt and Credit report puts total household debt at a record 18.8 trillion dollars. In its early-2026 data, the New York Fed recorded the highest auto-loan delinquency it has ever measured, credit-card delinquency near the levels last seen at the peak of the 2008 crisis, and student-loan delinquency at its worst since before the pandemic pause.

By the New York Fed’s own figures, serious credit-card delinquency sits at 7.1 percent, up from 5.3 percent before the pandemic, and auto-loan delinquency at 3.0 percent, up from 2.4 percent.

Now the honest complication, and it may be the most important number here. On the surface, the aggregate looks fine. Bloomberg reported that overall thirty-day delinquency improved to 4.7 percent in mid-2026, and a fair reader would ask how I square that with everything above.

The average is hiding the distribution, and the Federal Reserve says so itself. A New York Fed economist noted that mortgage deterioration is concentrated in lower-income areas and areas with declining home prices. A Fed presentation this spring put it plainly: overall performance remains strong, but delinquencies are rising fastest in lower-income areas and regions with weakening labor markets.

A consumer-finance analyst at LendingTree, quoted by CNN on the same report, described it best, that a lot of people are doing just fine and spending because they feel secure, but an awful lot of people are really struggling. That split is the whole point.

And I would go one step further, from what I see with my own eyes. Even some of the people who look fine are running on borrowed money, good credit scores and open cards holding up spending that paychecks no longer cover. Borrowed money spends exactly like cash, right up until the moment it does not. So when I hear that the comfortable half is still spending, I do not find it as reassuring as it sounds. Some of that comfort is financed, and financed comfort is not the same as strength.

Set that beside the government’s balance sheet. Interest on the national debt now costs more than the entire military budget, and the deficit is running near two trillion dollars a year with unemployment near four percent, borrowing at a pace usually reserved for recessions during what are supposed to be the good times.

Any one of these readings you could explain away. Together, from sources as sober as the Federal Reserve and the Congressional Budget Office, they describe a system running hot with a large part of its population falling behind. That is not what a healthy expansion looks like. It is a system with far less margin for error than the headline numbers admit.

The comfortable majority is masking a struggling minority, and the struggling minority is the high-spending working class an economy like ours runs on.

The number that explained the rest

For a while I had the symptoms without the mechanism. Then I found the number that ties them together, from the least dramatic source imaginable, a Federal Reserve data series.

It is called the velocity of money: how fast a dollar moves through the economy, how many times it changes hands. A dollar that reaches someone who spends it becomes someone else’s income, which becomes their spending, and around it goes. That circulation is what an economy really is. And it has been slowing for a quarter century. By the Federal Reserve’s own data, the velocity of the M2 money supply has fallen by roughly a third since its late-1990s peak, to near the lowest on record.

Money that pools instead of circulating is a slowing economy, written as a single number.

The Federal Reserve’s own research explains why, in language far drier than the implication. A 2025 Fed study found that the wealth gains of recent years flowed disproportionately to higher-income households, whose propensity to consume is lower, and concluded that the rise in wealth did not translate into the same level of consumption it would have if it had been more evenly distributed.

Put that next to Bureau of Labor Statistics and Congressional Research Service data showing that since 1979 productivity has risen far faster than wages, and you have the mechanism behind every symptom above. The money is not reaching the people who spend it. The circulation that drives a consumer economy is quietly stalling.

The loop that looks like growth

Here is where it connects to AI, and where I had to be careful, because this part is easy to overstate.

The AI build is real, and much of its growth is real, paid for by real customers. Hold onto that. But watch how the money moves inside it.

By industry estimates compiled from company guidance, the largest technology companies are on track to spend near eight hundred billion dollars on AI infrastructure this year, more than the entire output of Switzerland, and most of it flows to a very small set of recipients.

And those companies are entangled. A chipmaker takes a stake in an AI lab, the lab commits to a cloud provider, the provider buys the chipmaker’s chips. Analysts measure these arrangements in the hundreds of billions, and the problem is plain: the same dollar can show up as a chipmaker’s revenue, a startup’s funding, and a cloud’s backlog at once, making end demand look larger and more independent than it is.

Set the two side by side. Inside the AI loop, capital moves at tremendous speed and scale, but among a handful of companies, much of it debt-financed, some flowing in a circle. Outside that loop, the circulation that reaches ordinary consumers has been slowing for twenty-five years.

The growth everyone is counting on to carry the debt and justify the valuations is running hot inside a sealed circuit that never reaches the consumer the story depends on. That is not a prediction. It is how the money is moving right now, and it is not normal.

The cushion that softened past blows is thinner now

There is one more piece most market commentary leaves out. When the consumer has buckled before, the government has stepped in to soften the blow, and that rescue does something people forget: it flows downstream into company revenue and asset prices, including the technology names at the center of this boom.

COVID is the vivid example. The Tax Policy Center puts the total federal response near 5.6 trillion dollars, and the Congressional Budget Office estimated it lifted real GDP by nearly five percent in 2020. That money did not stop at the kitchen table. It became consumer spending, then corporate revenue, then support under stock prices across the market, tech and AI included. When Washington cushions the consumer, it quietly holds up the whole market at the same time.

But the cushion is thinner this time. That same COVID response drove federal debt from about 79 percent of GDP in 2019 to 97 percent by 2022, on the Tax Policy Center’s figures, and we are above that now with the interest bill already exceeding the military. A government can only run the same play so many times before the play becomes the problem.

This is not a prediction that help would not come. It is a plain observation that there is less room to provide it without deepening the very debt at the root of the story, which means the margin for error is thinner precisely where we have leaned on it hardest before.

What the data is telling me, and why I will not ignore it

I spent a career in analytics, so let me be clear about what I am and am not doing. I am not predicting. Predicting means naming the event and the date, and no one can truly do that. But reading data and reaching a conclusion you act on is a different thing entirely, and it is the whole point of the discipline.

Think about 2008. The data told the story well before the break. The housing leverage, the subprime deterioration, the tangle of derivatives on top, it was all visible, and the people who read it carefully knew the structure would fail if the conditions held. Not one of them called the exact day. That was never the point. The data was reliable, the picture was past any single worrying number, and the honest response was to prepare rather than look away. The ones who got ready were not lucky when it came. They were positioned.

That is where I am now, and I will say it plainly because Monday I promised you an honest answer rather than a shrug. When this many independent instruments read abnormal at once, the deficit at full employment, the delinquencies past their 2008 marks, the savings rate on the floor, the housing split, velocity stalling, and a partly circular loop dressed up as organic growth, the data has crossed from a collection of curiosities into a coherent story.

And the story is that the system has unusually little margin for error. I cannot tell you when that matters. I can tell you that the math is the math, and that the readings are real.

What history does, and does not, tell us

This series has leaned on history the whole way, the debt we grew out of after the war, the inflation of the nineteen-forties, the dot-com wiring of 1999. So let me end the historical thread where it naturally points, and let me be disciplined about what it can and cannot say.

History does not tell you when. It tells you what the setup looked like the last few times, so you can recognize it. And the setup before the three great crashes of the modern era, 1929, 2000, and 2008, is unusually consistent. Two things were present each time: leverage growing faster than the real economy, and value concentrated in a small handful of names.

In 1929 it was a cluster of glamour stocks. In 2000 the top technology names. In 2008 concentrated exposures in the banks. Different decade, same shape.

Now the readings today, by the same yardsticks. Household debt sits at a record, above the level that preceded 2008. The Buffett indicator, the ratio of the market’s value to the size of the economy, which Buffett himself called the best single measure of valuation, ran near 150 percent before the dot-com crash and above 100 percent before 2008. Today it is above 200 percent.

And the concentration of the market in a few AI names is as extreme as any of those earlier episodes. By the measures history hands us, the readings are not merely elevated. They are past where they stood before the events we all remember.

I am not telling you that means a crash, and I want to be precise about why. The honest truth about a bubble is that you cannot confirm it was one until it pops. History gives you the pattern, not the date, and that pattern can persist far longer than anyone expects, or ease off without breaking at all.

So take the information for exactly what it is. Here is what the data looked like before, here is what it looks like now, and here is the plain fact that the two look a great deal alike. Do with that what you will. The only response I cannot defend is pretending the resemblance is not there.

Which is really the whole argument for preparing rather than predicting. If you get ready and the strain resolves quietly, you have lost very little, some patience, a bit of yield on cash. If you get ready and it does not resolve quietly, you are one of the few people positioned instead of trapped. Prepared, you are safe either way. Head in the sand, you are only safe in one.

Prepared, you are safe either way. Head in the sand, you are only safe in one.

Why this makes me more interested, not less

If I stopped there, you would close this piece uneasy. That is not where the data leaves me. Working through all of it, I went from concerned to energized about it.

Here is why. Everything I have just described is the exact condition under which great companies finally go on sale. The best businesses do not become affordable because they get worse. They become affordable when the environment cracks, when a stretched system wobbles and prices reset all at once, the strong dragged down with the weak.

The very fragility that unsettles everyone is the thing a patient investor waits years for. The old wisdom, and Warren Buffett has built a career on the idea, is to be cautious when everyone else is greedy and bold when everyone else is afraid. Fear is not the thing to run from. For someone who has done the work, it is the thing to be ready for.

Return to the question that started all of this. What would have to happen for a company like Nebius to come back to a price a value investor could pay? For most of this series I could only gesture at the answer. Now I can say it plainly. What would make these great companies cheap is precisely the moment the market stops assuming all of their growth is real, durable, and independent, and reckons with the part that is circular and borrowed.

That reckoning does not require any company to fail. It only requires the story to be seen clearly. And when it comes, the price of even the best of them could fall to a level that finally offers a margin of safety.

Cheap and crisis tend to arrive together. That is not a reason to fear the crisis. It is a reason to be ready for the cheap.

So, is any of this sustainable

Let me answer the question in the honest two parts it deserves, because they are not the same question.

Is the technology real and durable? Yes. AI does real work, earns real revenue, and is being woven into the economy for good. I would not bet a dollar against it as a technology.

Is the current form sustainable, the valuations priced for perfection, the growth financed with debt and flattered by circular arrangements, resting on a consumer whose circulation is stalling? Not like this, not forever. But unsustainable is not the same as collapse, and anyone who claims to know the timing is guessing or selling something. Unsustainable describes fragility, not a schedule. The right response is not to predict the day it breaks. It is to refuse to pay a price that requires it never to.

Where that leaves us

So here is where the series lands. I am not predicting a crash. I am reading a set of instruments that, together, are more abnormal than any I have seen in years, and positioning for what such readings have historically preceded, without pretending to know when.

That means holding cash, not out of fear but as a loaded position, paid a real return to wait. It means doing the research now, in the calm, so I know exactly what each business I want is worth before any storm arrives. And it means the discipline to do nothing until a price offers a real margin of safety, then the resolve to act when everyone around me is too frightened to.

That is the whole series in one sentence. The environment is showing more strain than it has in a generation, and for the investor who has done the work and kept the powder dry, that strain is not the thing to fear. It is the opportunity of the cycle, arriving in the disguise it always wears.

So here is what comes next, and it is where this series becomes the beginning of something rather than an ending. I am always hunting for new great companies, the kind that clear the firewall and earn a place in a portfolio for years. That hunt never stops, and the supplier layer we spent this week on is full of candidates.

But here is the part people forget: some of the old names I already own are still great, and a downturn does not make them worse. It makes them better, because it lets me buy more of a proven winner at a lower price.

I would love nothing more than to lower my average on the companies I already believe in. I bought Dillard’s, Ralph Lauren, and Cardinal Health when the world was scared, and I am not done with any of them. A real dislocation would be a gift, a chance to add to the businesses I know cold, at prices only fear can produce. That is why I am not afraid of what the data shows. I am ready for it.

So the next phase of The Long View starts now. I am building the list, the new names worth owning and the old ones worth owning more of, and I am putting a price on each, the number that would finally make it a buy. When the environment gives us that number, I do not want to be forming an opinion in the panic. I want to be acting on one I made in the calm.

I started this series unable to buy a company I admired. I am ending it with a growing list of companies I admire, a price I would pay for each, and the patience and the cash to wait. That is not a gloomy place to end. For a value investor, it may be the best place to stand.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker. Thank you for reading this series. The comments are open, and this is where it gets good.

Not investment advice. The subscriber decides.

Yesterday I said the value in this boom is one layer down, in the companies that supply the build. Today is not a list of them. It is something more useful and harder to find: what to look for in that layer, the handful of traits that separate a durable business you could own at the right price from a crowded trade wearing a picks-and-shovels costume. I will use real companies as examples, but the names are not the point. The lens is the point, because a lens you can use on anything outlasts a list that goes stale the day after I publish it.

Where yesterday left us

Yesterday made the case that the surest way to own this boom is to own what the build consumes, the power, cooling, and electrical layer that gets paid no matter which AI model wins. Today I could hand you a list of those companies and call it a service. I am not going to, because a list is the least valuable thing I could give you.

A list tells you what I think today. It goes stale the moment a price moves, and worse, it invites you to buy without understanding, which is how good investors turn into bag holders. What lasts is the lens: the specific traits that make a supplier worth owning at the right price, and the ones that mark a trap. Learn those and you can evaluate any company in this layer, including the ones that do not exist yet, long after this piece is forgotten.

So this is a teaching piece, in keeping with everything the series has been about. Not which stock, but how to think. Here is what I look for, and what I look out for, with real companies as illustrations of each.

Trait one: it gets paid whether or not the winner is picked

The first question is the one that has driven this whole series. Does this business get paid based on how much gets built, or based on which specific AI company wins? You want the former. You want a toll on the whole road, not a bet on one car.

A company like Vertiv is a clean illustration of the trait, it supplies the power and cooling nearly every data center needs, so its demand tracks the build itself rather than any single model. Eaton shows the same trait from the electrical side, the switchgear and distribution gear that has to go into the building regardless of whose chips arrive. The lesson is not “own these two.” It is that before anything else, you check whether a company is levered to the volume of the build or to the fortunes of one customer. Volume of the build is the trait you want. One big customer is the fragility you avoid.

Trait two: diversified enough to survive a slowdown

The build will not proceed in a straight line. So the second trait is whether the business can survive the pace slackening without falling apart.

Here the contrast teaches more than any single name. A pure-play tied entirely to data-center spending gives you the most exposure to the boom and the most damage if it cools. A diversified company, where data center is a real but not exclusive slice, gives up some of the upside for a great deal more durability. Eaton is a useful illustration again, its electrical business rides the build, but it also earns from aerospace, grid modernization, and broad industrial electrification, so a data-center slowdown dents it rather than breaks it. The lesson is to know which kind you are holding. There is nothing wrong with a pure-play, as long as you understand you are trading durability for exposure, and you are paying a price that respects the risk.

Trait three: the price leaves room to be wrong

This is the trait the crowd forgets, and it is the heart of the whole approach. A wonderful business is not a wonderful investment at any price. The third thing to look for, and the hardest to accept when a stock is running, is whether the price still leaves a margin of safety.

The examples here are a warning. Vertiv has climbed more than 250 percent in a year and trades at a forward earnings multiple in the forties. Comfort Systems has run well past two hundred percent. These are strong businesses that have become expensive, and buying a strong business at a euphoric price is the exact mistake this series began with, the Nebius error, one layer down. The lesson is not that these are bad companies. It is that the picks-and-shovels idea is no longer a secret, it is on magazine covers, and once a theme reaches the cover the easy money in the obvious names is usually gone. The trait you are hunting is the gap between a price and the durable demand beneath it. When that gap closes, you wait, no matter how good the story sounds.

Trait four: it sits where the bottleneck is deepest

Not all layers are equally scarce. The fourth trait is whether a company sits at the tightest bottleneck in the build, because scarcity is what protects both pricing and demand.

Right now the deepest bottleneck is power itself. You cannot wish a gigawatt into existence, and the grid takes years to expand. So the companies that generate power and make the heavy equipment to move it, the turbine and grid names, the utilities and independent power producers, sit at the point of greatest scarcity. They tend to be slower, duller, and more regulated than the cooling darlings, which is precisely why the crowd has often overlooked them and why they can still trade at saner prices. The lesson is to look for the tightest constraint in any supply chain, because that is where durable pricing power and the least crowded valuations tend to coincide.

Trait five: it has delivered this before, somewhere else

Here is a trait the crowd rarely looks for, and it is where some of the best-value research often hides. Look for the proven operator repositioning into this build, rather than only the established name everyone already calls an AI stock.

The AI buildout does not require inventing everything from scratch. It needs power, cooling, electrical work, precision manufacturing, complex construction, and systems integration, all things certain companies have done well for decades in other fields. A business that has already delivered at scale in an adjacent technology, and is now turning that proven capability toward the AI build, can capture a real slice of this spending long before the market relabels it an AI company and reprices it accordingly. That window, between delivering and being recognized for it, is where value lives.

But hold the discipline, because a track record elsewhere is necessary, not sufficient. These are mission-critical builds with long qualification cycles and no tolerance for failure. Operators cannot drop untested equipment into a live data hall. So the trait to hunt is both halves together: a company that has truly delivered this kind of work before, and can get qualified into the AI deployment path. Proven capability plus a credible route in. That combination is rarer, and more valuable, than either one alone.

The trait that sits underneath all of them: the environment

Every one of those traits is read against the environment we have spent the series mapping, and this is the part most stock write-ups leave out. This whole layer is levered to how much gets built, and how much gets built depends on the cost of money, the weight of the debt, a stretched consumer, and external pressures on energy and shipping that can move input costs without warning.

That changes what a fair price is. A richly priced supplier is doubly exposed, once to its own stretched multiple and once to any slowdown in the build the environment might force. So the final trait is really a discipline: demand a price with enough room that you survive being wrong about the environment, not just about the company. And keep cash ready, because the same environment that could pressure these names is the one that would eventually hand you the good ones at a price worth paying.

Why this is really about being ready

Here is the part that ties this piece to everything we have been building toward. Owning a lens for this layer is not academic. It is preparation for a specific moment that this whole series has argued is more likely than the market assumes.

Walk the logic back. The famous AI names are too expensive for a value investor. What could bring them down is not a company stumble but the environment cracking, the debt, the strained consumer, the cost of money, the pressures on energy and shipping. When that happens, and cheap and crisis tend to arrive together, the correlated selloff will not spare this supplier layer. Everything will fall at once. That is exactly the moment the lens pays off.

Because in that moment, most people freeze or sell. The person who has already run these five questions on a handful of names does the opposite. They already know which supplier is a fragile pretender and which is a durable business being handed to them at a discount it does not deserve. They are not scrambling to form an opinion in the middle of a panic. They are pouncing on a decision they made in the calm, with cash they were paid to hold while they waited. The homework is what turns a frightening selloff into the best buying window of the cycle.

That is the whole strategy of this series in one line. You cannot control when the environment hiccups. You can control whether you are ready when it does. And everything this series has laid out, the debt, the strained consumer, the geopolitics, the stretched valuations, is the environment starting to signal that the time to build that list is now, not after the move. The research takes weeks. The window may not.

What this is really teaching

Put the traits together and you have a lens, not a list. Does it get paid on the build or on one winner. Can it survive a slowdown. Does the price leave room to be wrong. Does it sit where the bottleneck is deepest. Has it delivered this kind of work before and can it get qualified in. And does it all hold up against an environment that could break more than one way. Run any company in this space through those five questions and you will know far more than a ticker list could ever tell you, and you will know it next year too, when the list would have been useless.

So keep the lens close and the list short. Run these five questions on the handful of names in this layer you find most durable, write down the price that would finally make each one a buy, and then wait, in cash, paid to be patient, ready to move the day the environment gives you your price.

Tomorrow I close the series where it began. I go back to the question that started all of this, the one about a great company at a price I could not pay, and I give you my honest answer: what I now believe could bring that repricing, how close I think we are, and what a value investor does about it. It is the reckoning this whole series has been building toward.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker.

Not investment advice. The subscriber decides.

The Long View · Where the real opportunities in the build are hiding · Not a recommendation, and not transaction advice

The Long View names no positions here and does not broker deals. This is an investing idea examined in full, and a reminder to do your own work before acting.

The value investor’s honest position

Let me start with the uncomfortable truth about where people like me stand in this boom. We look for good businesses at defensible prices with a margin of safety. By that standard, the marquee AI names are not buys. Either we did not get in early enough, back when the price still made sense, or the valuations have run so far past the fundamentals that no disciplined investor can touch them now. Both of those can be true at once, and for most of the famous names, they are.

That is exactly where this series began. I graded Nebius, a great business growing more than four hundred percent a year, and I could not bring myself to buy a single share, because the price demanded that everything go right forever. A wonderful company is not a wonderful investment. The gap between those two things is the whole game, and in the AI trade that gap has become a canyon.

And the canyon does not sit in still air. These prices assume a calm, cooperative world, and the numbers describe something else. The federal debt is roughly the size of the economy, the interest bill runs over a trillion dollars a year, and the consumer underneath it all is stretched thin, with savings near record lows and delinquencies rising. Then there is the part that never makes the earnings call. Normally about 20 million barrels of oil a day move through the Strait of Hormuz, roughly a fifth of what the world consumes, per the U.S. Energy Information Administration. In the first quarter of 2026 that flow fell nearly 30 percent from a year earlier, to 14.6 million barrels a day, and Brent crude climbed from the low seventies in late February toward the mid nineties within weeks. I take no side on any of that. I only note the number, because it is the kind of shift that reprices energy, shipping, and inflation expectations while the market is looking the other way.

I raise the environment for a specific reason, and it is the reason this piece exists. A stretched, distorted environment is exactly where prices get manipulated and where traps get set. When money is cheap and stories are loud, weak businesses get dressed up as strong ones, circular arrangements get counted as real demand, and a richly priced name can look inevitable right up until the environment shifts under it. The value investor’s job in a moment like this is not to predict the shift. It is to hunt for the trap, the place where the price is quietly resting on something that only holds if nothing goes wrong. Which is why the safest ground is often the boring, physical, essential layer that gets paid no matter which story turns out to be true.

Most investors respond to that canyon by staring harder at the same names, waiting for a dip, asking which model will win, which chip, which hyperscaler. I think that is the wrong question, and I think asking it is exactly how you get led away from the real opportunities.

The question that pays

Here is the better question, the one this whole series has been sharpening. Not which AI company wins. Instead: who gets paid no matter which one wins?

You do not have to predict the winner of a gold rush to make money. You can own the outfit selling picks, shovels, denim, and rail freight to every prospector on the mountain, and get paid whether any single miner strikes it rich or goes home broke. The AI boom has the same shape. Underneath the handful of enormously expensive names sits an entire physical economy that gets paid to build the thing, and a lot of it is unglamorous enough that the crowd has not bid it into the stratosphere.

Follow that question to its logical end and it leads somewhere that sounds absurd and is completely serious.

The purest version: it might not be a stock at all

The scarcest input in the entire buildout is not chips or capital. It is skilled labor, the electricians, pipefitters, and mechanical crews the whole thing physically cannot proceed without. I made that case yesterday and will not re-argue it here.

So the purest way to own that scarcity is to own the business that employs it. A local electrical or mechanical contractor, the kind whose owner is heading into retirement, changes hands for roughly three to eight times earnings. The AI names trade at twenty to sixty times sales. Same tailwind underneath both. One of them is priced for perfection. The other is priced like a used truck, and it owns the one thing the largest construction boom in history cannot get enough of.

That is the most extreme expression of the value idea, and it is real. It is also not for everyone, and I want to be honest about why in one breath rather than a manual: it is an operating business, not a ticker. It needs someone to run it, and the value can walk out the door with the seller if you are careless. It rewards an operator, not a passive holder. If that is you, it may be the best risk-adjusted asset in this whole story. If it is not you, do not force it.

And if that is not for you, look at what the build cannot happen without

Here is the part that brings it back to the stock market, because most readers do not want to run a plumbing company, and that is fine.

Those skilled workers cannot build a single data center with stock options. They build it with tangible things: transformers, switchgear, and busway to move the power. Chillers, cooling loops, and air handlers to haul away the heat. Generators, turbines, and grid connections to feed it. Steel, concrete, fiber, and the trucks and logistics to stage it all. Project management and engineering to sequence it. Every one of those is made or provided by a company, and many of those companies are public, ownable, and tied directly to the volume of the build rather than to which chatbot wins.

That is the same principle as the plumbing company, just expressed as shares instead of a business. Own what the build consumes. Own the layer that gets paid per data center, not per victory in the model wars. The demand reaching those suppliers is set by how much gets built, and right now the plans call for more building than the world has the power, the parts, or the people to deliver.

Here is why this matters for the question driving the whole series. We have spent weeks on what would have to happen for the famous AI names to ever fall to a price a value investor could pay, and that answer is still unfolding, it depends on the environment finally catching up with the valuations. But there is a second, quieter answer hiding in plain sight. The value you are hunting may not require the frontier names to crash at all. It may already exist, in real dollars and cents, one layer over, in the companies that supply the build. The hype, and the punishing multiples that come with it, concentrated on a handful of frontier names. A lot of the suppliers underneath them were never bid up to anything like the same extreme, because they are dull, industrial, and easy to overlook. That is the whole opportunity. The market’s attention is a spotlight, and value tends to survive in the parts of the room the spotlight never swings to.

I am not saying every supplier is cheap, some have been discovered and bid up too, which is tomorrow’s warning. I am saying the measurable value, the kind you can count on a balance sheet rather than hope for after a crash, is far likelier to be found in this layer than in the names on the front page. That is a real answer to the series’ question, and it does not require you to wait for anything to break.

And here is the part that makes this the strongest ground in either direction. Suppose the thing this whole series has been circling does happen, and the environment finally cracks the frontier valuations. That shock would not spare the suppliers. They would sell off too, dragged down in sympathy, because in a panic correlation goes to one and almost everything falls together. For the frontier name that started at sixty times sales, a forty percent drop may still leave it expensive. But for a sound supplier with real earnings, real customers, and an already-sane multiple, that same drop does something very different. It turns a reasonable price into a gift. A good business, essential to the build, suddenly on sale because it had the bad luck to be standing next to the names everyone was fleeing.

That is where the value is truly hiding in these scenarios, and it is why the work matters now, before anything happens. The winners in a repricing are not bought during the panic, they are identified before it, so that when the correlated selloff comes, you already know which supplier is a fragile pretender and which is a durable business being handed to you at a discount it does not deserve. The frontier names falling is not the risk to this strategy. For the right supporting companies, it is the entire opportunity. Cheap and crisis tend to arrive together, and the disciplined investor’s job is to have already done the homework on who is worth buying when they do.

And this is the assumption underneath all of it, the one that makes buying into a selloff sound rather than reckless. AI is not going away. Whatever happens to the valuations, the technology is real, it does useful work, and it is being woven into the economy for good. What a repricing would change is the pace, not the existence, of the build. The frenzy might cool. The spending might slow from a sprint to a walk. But a walk still needs power, cooling, switchgear, and skilled hands, year after year. That is the difference between this and a true bubble in something with no underlying use. You would not be buying a supplier to a fad that vanished. You would be buying a supplier to an essential industry that merely stopped sprinting, marked down because the market cannot tell the two situations apart in a panic. Telling them apart is the whole job.

Are we being led away from the real opportunities

This is the thought I keep coming back to. The opportunities in this boom are not hidden. They are simply boring, and the noise machine that surrounds AI is built to keep your eyes on the exciting, expensive names instead of the dull, essential ones. Every hour spent debating which frontier model wins is an hour not spent asking who supplies the power and the cooling to run all of them.

I think a lot of ordinary investors are being distracted, not maliciously, just by the gravity of the hype, from the parts of this build where a value investor can still find a defensible price. The distraction is expensive. It keeps people either paying any price for the famous names or sitting the whole thing out because those names scare them. There is a third door, and it is standing open in the least glamorous part of the room.

The market priced me out of the AI trade. It did not price me out of the AI build, and it certainly did not price me out of the companies that supply everything the build depends on. That is the long view: the best investment on the board is rarely the one everyone is looking at. Sometimes it is the one nobody thinks to get excited about.

There is one more piece to this posture, and it is the least glamorous of all. While you do the homework and wait for the price, cash is not a failure to invest. It is a position. With short-term rates sitting around three and a half to four percent, you are paid something real to hold it, and in an environment of sticky inflation, where the market now leans toward rates holding or rising rather than falling, that yield is less likely to melt away under you than it would be in a fast-cutting cycle. I will be honest that inflation eats part of that return, so cash is not free, it is patience with a modest coupon. But patience with a coupon, aimed at businesses you have already identified, is a very different thing from fear sitting on the sidelines. One is waiting for a pitch. The other is refusing to swing at all. The value investor waits, gets paid a little to do it, and keeps the list ready.

There is a live tension worth naming here, and it takes us straight back to the five roads out of the debt. With inflation still sticky, the direction of rates is truly contested right now. The Fed has held steady for several meetings, but its own projections have turned more hawkish, with several officials penciling in a possible hike, even as a softer jobs number has others arguing for a cut. I am not going to predict which way it breaks, and anyone who tells you they know is guessing. What matters for a saver is that the two roads have opposite consequences. If the authorities raise rates or hold them high to tamp down inflation, that rewards the saver, the coupon on cash holds or grows. If instead they let inflation run while keeping rates below it, which is one of the quiet roads a government can take to shrink a debt this size, that is the road that punishes the saver, because inflation eats the coupon and then some. So even cash, the most boring position on the board, is a bet on which road gets taken. I hold it with my eyes open, not because it is safe, but because it is patient and, for now, paid.

And here is the part that closes the loop. Look at what each direction hands me. If rates stay high or climb, I keep getting paid to wait. If rates fall, they will not fall in a vacuum, the authorities cut because something in the economy cracked, the labor market or credit or the market itself, and that is exactly the dislocation that finally reprices the expensive names and puts sound companies on sale. In other words, the same event that lowers the yield on my cash is the event that gives my cash something worth buying. High rates pay me to wait. Falling rates tell me it is time to deploy. The cash wins in both directions, as long as I have done the homework and know what I am waiting to buy. The only way to lose this setup is to not hold the powder, or to spend it early on the very names this whole series has warned are too expensive.

Where this goes tomorrow

I have kept this deliberately at the level of the map, the layers, the principle, the right question, because tomorrow I get specific. I will name the parts of the supply chain worth watching, from the electrical and cooling layers to the power and grid names that feed them, and I will be honest about the trap waiting even here: some of the obvious picks-and-shovels names have already been discovered and bid up, and buying them at those prices repeats the very mistake this series keeps warning against. The right question gets you to the right room. Tomorrow we look at what is in it.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker.

Not investment advice. The subscriber decides.

The Long View · The AI economy, the debt, and you

The Long View names no positions here. This is analysis, not financial or career advice, and not a prediction of any particular outcome.

Yesterday I left you with a promise. The whole hopeful case for the debt rests on one piece of history: after the Second World War we carried debt bigger than the entire economy and simply grew our way out of it. We did. So the question that decides everything is whether we can do it again, because the entire plan, the one painless road out of five, is to grow out of this debt the way we grew out of that one.

That is the part of this series’ driving question I want to take up today. If part one was what could make these companies cheap, and part two is what that change would cost, this piece sits right on the hinge between them: can the one painless escape even work anymore? I went back and looked at what really powered the postwar miracle, and then at the economy we would run the same play with now. They are not the same machine. Not even close. And once you see the difference, the hopeful story gets a lot more honest, and it points somewhere useful for what you personally do next.

Let me start by dismantling the comfortable story, because it deserves dismantling.

The precedent everyone leans on

After the Second World War, the United States carried federal debt of more than one hundred percent of its entire economy, a level it would not see again until now. And over the following decades, it did not default, it did not collapse, it grew out of it. The debt shrank against a booming economy until it was a manageable fraction of output. This really happened, and it is the single most cited reason to believe today’s debt is survivable.

But look at what powered that growth, because the mechanism is everything. In 1945, manufacturing was about thirty percent of the economy and nearly forty percent of all jobs. It was a nation that made physical things. And the way you grow an economy that makes things is by employing more people to make more of them. Wages rose across the whole workforce, manufacturing workers gained about a quarter more real income during the war years alone, and those wages became the demand that bought the next round of production. Growth and employment rose together, in a single self-reinforcing circle. More jobs meant more income meant more demand meant more jobs. The tide truly lifted all boats, because the engine of the growth was mass employment itself.

It was also helped by things that will not repeat: a young workforce with the enormous baby-boom generation just ahead, and a world whose factories had been reduced to rubble while America’s stood intact, making the country the workshop of the planet.

The economy that has to do it this time

Now look at the economy that has to pull off the turnaround today. Manufacturing is about ten percent of output and eight percent of jobs. Services are roughly eighty percent of the economy, and consumer spending is around two-thirds of it. We do not primarily make things anymore. We serve each other, and we buy from each other, and the whole machine runs on people having enough income to keep spending.

That difference is not a detail. It reverses the entire logic of the turnaround.

In 1945, you grew the economy by adding workers, and their wages became the demand. The growth shared itself automatically, because it was built from labor. AI grows a service economy the opposite way: by delivering the service with fewer workers. That is not a flaw in the technology, it is the entire point of it, the reason companies spend hundreds of billions on it. But it means the growth engine and the demand engine, which were the same people in 1945, now work against each other. You grow output by removing the workers whose incomes were the demand. The virtuous circle of the postwar boom, more jobs, more income, more demand, more jobs, runs in reverse.

So what does this do to the turnaround

It does not make growing out of the debt impossible. But it removes the thing that made 1945 work: the automatic, broad, self-sharing quality of labor-intensive growth. In 1945 the broad outcome, where the gains reached everyone, was the default, because that was simply how the growth was built. Today the concentrated outcome is the default, because AI growth flows first to the companies that own it and the shareholders behind them, not to the workers it replaces. The gains no longer share themselves. They have to be shared on purpose, through policy, wages, or design, and on-purpose is always harder than automatic.

There are two more headwinds the 1945 economy did not face. A service economy has fewer rungs left to climb to, because services were the rung everyone climbed to when manufacturing left the country. If AI now pressures services, there is no next great labor-absorbing sector visible behind it. And demographics have flipped from tailwind to headwind: instead of a young workforce and a baby boom ahead, we have the baby boom retiring, fewer workers supporting more beneficiaries, which strains the debt no matter what AI does.

So the honest verdict is this. The comforting analogy is broken. The postwar growth was broad, labor-intensive, goods-based, and demographically blessed. The potential AI growth is narrow by default, labor-replacing, service-based, and demographically burdened. It can still help the debt. But left to its natural shape, it produces exactly the least helpful result: concentrated gains sitting on top of a thinning consumer, which is not how you grow your way out of anything durably.

Where do the displaced workers go

Which brings us to the question that really matters to a human being reading this: if the machine grows by shedding workers, and the service rung is the one under pressure, where does a displaced worker go?

The honest answer is not “nowhere,” and it is important not to overstate the doom. The most cited labor forecast, from the World Economic Forum, projects roughly ninety-two million jobs displaced by the end of the decade, but around one hundred seventy million created, a net gain of about seventy-eight million. Work is not ending. It is moving. But notice the word: moving. The new jobs are not the same jobs, in the same places, for the same people. The labor market is splitting into two tiers, high-skill, high-trust, interpersonal work rising, and routine, rules-based, middle-skill work shrinking. The middle is what hollows out, and the middle is where a great many paycheck-to-paycheck households currently live.

So the answer to “where do they go” is: somewhere different, if they can make the move. And that “if” is the whole ballgame.

What really resists the machine

Here is where the research gets truly useful, because the durable ground is not a mystery. The work that resists automation clusters in a few recognizable families, and they share a logic: AI is extraordinary at routine, digital, predictable tasks, and weak where the job requires a body, a licensed human judgment, or a relationship of trust.

The skilled trades sit near the top, and this surprises people who assume “the future” means sitting at a screen. An electrician, a plumber, an HVAC technician, an elevator mechanic, works with their hands in unpredictable physical environments that defeat automation, and the work cannot be shipped overseas. Demand is rising, not falling, partly because the AI buildout itself, the data centers, the power, the electrification, needs enormous amounts of skilled physical labor. Many of these pay well into the six figures and require a certificate or apprenticeship, not a four-year degree and its debt.

Human care and licensed judgment is the second family: nursing, therapy and mental health, caregiving, the healthcare professions broadly. These combine physical presence, empathy, split-second judgment, and legal accountability, and they are backed by the largest demographic tailwind there is, an aging population that needs more care every year. Healthcare is projected to add tens of millions of roles this decade.

The third family is high-trust judgment and accountability, the roles where a human must own the decision because ethics, liability, persuasion, or leadership are involved. A trial lawyer reading a jury, a leader navigating conflict and values, a clinician making a complex diagnosis. And the fourth is genuine creative direction, not routine content generation, which AI now does, but the taste and vision to decide what should be made and whether it is any good.

There is a fifth path that is not a category but a posture: work alongside the machine rather than against it. The people who build, manage, and audit AI systems command a large wage premium precisely because they direct the tool instead of competing with it. Across all of these, the same handful of human capacities keep appearing as the durable core: critical thinking, emotional intelligence, complex problem-solving, persuasive communication, and adaptability. Those are the things to build, in whatever field you are in.

The gap the data is screaming about

If you want to know where the wealth really gets made in a shift like this, you look for the place where demand has run violently ahead of supply and cannot catch up, because that gap is where fortunes and careers are built. In all the data I gathered for this series, one gap stood out above every other, and it is not subtle. It is the loudest signal in the numbers.

The single biggest bottleneck in the entire AI buildout is not chips, and it is not capital. It is the shortage of electricians and skilled trades to build and power the data centers. This is not my characterization, it is the industry’s. Microsoft’s president has called the electrician shortage the number one problem slowing data-center expansion. Nvidia’s chief executive says the need runs into the hundreds of thousands. The electrical workers’ union calls it a life-or-death situation for Big Tech. The head of the world’s largest asset manager raised it directly with the government.

The numbers explain the panic. McKinsey estimates a gap of one hundred thirty thousand electricians by 2030. Reporting this month put the broader need at roughly half a million electricians, three hundred thousand welders, and five hundred fifty thousand plumbers to keep pace with the build. The country needs over three hundred thousand more electricians while about twenty thousand retire every year, and more than forty percent of the existing trades workforce is set to retire by 2031. Electrician employment is growing faster than any other construction category the government tracks, and demand is still outrunning it.

And the consequence is not theoretical. Roughly half of the data-center capacity planned in the United States for this year, about seven gigawatts out of twelve, has already been canceled or delayed, and a single delayed mid-size facility can cost its owner more than fourteen million dollars a month in lost revenue. Hundreds of billions of dollars of capital are pointed at capacity that there are not enough skilled hands to build. Capital is not the constraint. Labor is.

Sit with the irony, because it is the whole series in a single image. The companies remaking white-collar careers with AI now depend entirely on blue-collar workers to keep their infrastructure growing. The same boom that threatens the routine desk job is desperate, right now, for the electrician, the welder, the pipefitter, the high-voltage crew, the people who work with their hands in the physical world the machine cannot enter. The gap the data is screaming about is exactly the durable ground we mapped a moment ago, made concrete and urgent and extraordinarily well paid.

For a working person weighing where to point their own effort, that is about as clear a signal as economic data ever gives you: six-figure pay, no student debt, more security than most software jobs, and a structural shortage that lasts for a decade or more. And it extends past the electrician into the whole physical layer, the transformer and switchgear makers, the substation crews, the commissioning and controls specialists, the cooling and HVAC engineers, the grid-modernization work that will outlast the data-center build itself. The people quietly making a fortune off the AI boom may turn out to be the ones who never bought a single AI stock, and instead learned to build the thing the stocks depend on.

And here is the part that matters if you are reading this thinking you are too far from a tool belt to benefit. You do not have to be the electrician to profit from the electrician shortage. Every scarce tradesperson on a site needs a schedule to follow, a budget to hit, a design to build to, a quality check to pass, a permit to clear, and equipment to install, and that entire supporting layer is short too, because the build itself is short. Those are analytics and coordination roles: project schedulers who run the critical path, cost estimators, owner’s representatives, commissioning agents who verify the systems work, quality inspectors, the digital-modeling coordinators who build the plan before the concrete is poured. For someone with technical or analytical skills and even an old familiarity with how a job site works, these are among the least crowded and best paid seats in the whole boom, the brains around the build rather than the hands on it. I will lay out that map in more detail in the notes, because it deserves its own space.

The question to ask yourself

There is one diagnostic from the research worth keeping, because it cuts through the noise. Ask yourself: is the work I am doing today essentially the same as the work I was doing two years ago? If the answer is yes, then AI’s improvement curve is catching up to you faster than your own skills are moving, and that gap is the risk. Exposure is rarely about your job title. It is about where in the role you sit. The junior analyst doing repeatable tasks is exposed; the senior one who owns judgment and relationships is not. The trajectory of your skills matters more than the name of your job.

The most undervalued asset you own

Let me bring this home, because this newsletter is about one idea above all others, and it applies here more powerfully than anywhere.

Everything I write is about value: buying an asset for less than it is worth, with a margin of safety, so that being partly wrong does not ruin you. We apply that discipline obsessively to stocks. And almost nobody applies it to the asset that really funds their life, their own capability. The most undervalued, most neglected holding most people own is their own skillset, the thing they stop investing in the day they leave school, while they pour attention into stock tickers they cannot control.

You cannot control the deficit, the bond market, the Treasury, or the shape of the AI buildout. You have almost no influence over whether the macro chain tightens or the growth arrives broad or narrow. But you have complete control over one thing: which side of the polarizing labor market you stand on. And in an economy that grows by replacing routine labor, the highest-margin-of-safety investment a working person can make is not a stock at all. It is making themselves into the human the machine cannot replace, because that protects the income that every other investment, every plan, every bit of security, ultimately rests on.

I have to be honest about the hard part, because the earlier piece in this series demands it. Reskilling is easiest to preach and hardest to do for exactly the people most exposed, the paycheck-to-paycheck, debt-loaded households with no time and no cushion to retrain. That is real, and it is why the systemic answer, that the gains have to be shared, still stands and is not solved by individual effort alone. Not everyone can become a nurse or an electrician, and age, health, geography, and money all constrain the move. This is guidance, not a cure, and I will not pretend otherwise.

But at the level of the individual, the direction could not be clearer. Treat your own capability as your core position. Add to it deliberately, on the durable side of the line, the physical, the human, the judgment-based, the trust-based, the things a model in a data center cannot do. Watch your own skill curve at least as closely as you watch any stock. Because the same force that this whole series has been tracing, the one that could reprice the market, strain the government, and hollow the consumer, is also the one that will decide, household by household, who thrives in what comes next. The market is worth watching. But the best investment on the board, for most people, is the one in the mirror.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker.

We grew out of the last great debt by putting a nation to work. We are being asked to grow out of this one by a machine that works instead of us. Whether that ends well for the country depends on choices far above any of us. Whether it ends well for you depends, more than anyone likes to admit, on what you decide to become before the choice is made for you.

And it points somewhere I did not expect, which is where I am going tomorrow. If the scarce, valuable thing in this whole build is skilled hands, then the businesses that employ those hands are sitting on the rarest asset in the economy. A lot of them are quietly for sale. Tomorrow I follow that all the way to its logical, almost absurd conclusion: the best AI investment I can find might not be a stock at all.

Not investment advice. The subscriber decides.

A quick word on why this series exists, because it names the question everything here is chasing. This week I graded Nebius, a soaring AI name growing 454 percent a year, and I could not buy it, because the price terrified me. That left me with one question I could not put down, and it is the question this whole series is built to answer:

What would have to happen in the world for a company this good to ever come back to a price a value investor could pay, and what would that change cost the rest of us?

That question has two halves, and the series follows both. The first half, what could ever make these companies cheap, led into the macro machinery: the bond market, the debt, the roads a country can take out of it. The second half, what that change would cost, leads to the people underneath it all, which is where this piece goes. I do not answer the whole question today. Each piece takes one part of it, and I come back to answer it in full at the end of the series. Today is about the cost.

Where the chain left us

Yesterday I traced a chain: the AI buildout runs on borrowed money, the cost of that money is set in the bond market, the bond market is under pressure from a government borrowing more than the world wants to lend it, and that pressure runs all the way down to the price of every AI stock. Follow that chain to its source and you do not end up at a company or a chart. You end up at one thing underneath all of it: the national debt. That is the ground the whole AI trade is standing on, and it is the problem this piece and the ones after it are really about.

So let me state the problem plainly, because everything follows from it. The federal debt held by the public is around thirty-two trillion dollars, roughly the size of the entire economy. The deficit runs near one and nine-tenths trillion dollars a year, and we are running it at full employment, when the budget is supposed to be at its healthiest. The interest bill alone is about a trillion dollars a year now, more than the entire military budget, and it swallows roughly eighteen cents of every dollar the government collects.

Here is the consequence, and it is the part that makes this urgent rather than academic. High debt means a high interest bill. The interest bill widens the deficit. The wider deficit forces more borrowing, which adds to the debt, which raises the interest bill again. When interest compounds faster than the economy grows, the problem stops needing any help and begins feeding itself. We are not past that point. We are close enough to it that the only question worth asking is the practical one: how does a country really get out from under a debt like this?

The answer is not open-ended. History gives a country a bounded set of exits, and I want to walk all of them, because you cannot understand where the AI economy is heading until you see which road we are on and what it costs the person living through it.

There are only five roads out, and four of them hurt

This is not a matter of opinion or cleverness. History offers a bounded set of exits from a sovereign debt this large, and every country that has faced one has taken some mix of them. There are five. It is worth walking each one plainly, because only when you see all five do you understand why the whole country is quietly betting on a single one.

The first road is to grow out of it. If the economy expands faster than the debt and its interest, the burden shrinks on its own, even if the dollar amount keeps rising. This is how the United States escaped its post-war debt, and it is the only road on this list that does not hurt. Nobody has to be taxed harder or paid less. You simply grow, and the debt fades against a bigger economy. Hold that thought, because it is the one everything else in this series turns on.

The second road is austerity: cut spending and raise taxes enough to run a surplus and pay the debt down. It has been done, most recently in the late nineteen-nineties. But the repercussion is brutal and political. Closing a deficit this size means cutting deeply into Social Security, Medicare, and defense, or raising taxes across the board, or both. Right now the country is moving the other way, the deficit is rising, not falling, so as a practical matter this road is closed. Nobody is walking it.

The third road is to inflate it away. Let inflation run hot and repay the debt in cheaper dollars, so the debt shrinks against rising prices even though the number does not change. The government did exactly this in the nineteen-forties. The repercussion is that savers and ordinary people pay the bill through the erosion of what their money is worth, and richly-valued assets, the expensive stocks most of all, tend to get repriced hard when inflation and higher rates arrive. It works, but it works by quietly taking from everyone who holds dollars.

The fourth road is financial repression, a gentler cousin of the third: hold interest rates below the rate of growth and inflation for years, quietly, so the debt slowly erodes without a dramatic crisis. Japan is the living example, carrying debt far larger than ours, held together by captive domestic buyers and near-zero rates for decades. It is survivable. The repercussion is a long, low-growth grind and a slow bleed of anyone trying to save, and a fragility that lasts as long as the arrangement does.

The fifth road is default, simply not paying. For a country that prints its own currency this one is almost never chosen outright, because it can always create the dollars it owes. The realistic version of default is not a missed payment. It is the soft default of the third road, inflating the value of the debt away, which is why it rarely appears on the menu as itself.

So look at the five together. Austerity is shut. Inflation and repression both work by taking from savers and repricing assets, and default is just inflation wearing a darker coat. Four of the five roads out are either closed or painful, and they land hardest on the people holding dollars and the stocks priced for perfection. Only the first road, growing out of it, hurts no one. That is why the government, the market, and most of us are quietly praying for it.

And the engine everyone is counting on to deliver that growth is AI.

You already watched them choose a road this week

Here is where it stops being theory. In yesterday’s piece I noted that the Treasury stepped in this week to buy back its own long-term debt, and that yields fell when it did. Look at that move through the five roads and you can see the government quietly making a choice in real time.

Buying back long-term debt to hold down the interest rate on it is not growth, and it is not austerity. It is a mild form of the fourth road, financial repression, leaning on the bond market to keep the cost of the debt lower than it would otherwise be. It is being presented as routine housekeeping, and taken alone it nearly is. But read alongside a rising deficit and foreign buyers stepping back, it looks like the first small reach for the lever a country pulls when the other roads are closed and it is buying time for growth to show up.

That is the honest way to read this week. The government is not on the painful roads yet. It is managing the edges, holding the long end of the bond market down a little, keeping the trap from tightening while it waits for the one painless road to deliver. Which brings the whole thing back to a single dependency. All of this, the buybacks, the patience, the hope, rests on growth arriving, and arriving broadly enough to matter. So it is worth asking, with clear eyes, what that growth really does to the people it is supposed to save.

I believe that road is real, and it may be the best hope the country has. But there is a cost buried inside it that never shows up on the government’s ledger, and it lands squarely on the person at that kitchen table. Once you see it, the hopeful story never sounds quite the same.

What productivity growth really means

Start with what “productivity growth” is, in plain terms, because the phrase hides the thing that matters. Productivity growth means producing more with less. More output per worker. And the way AI delivers it, the reason companies are spending hundreds of billions on it, is by letting a business do the same work, or more, with fewer people. That is not a side effect. That is the product. When a company says AI made it more productive, it very often means AI let it do what used to take more employees.

At the level of the whole economy, that can be wonderful. Every major wave of automation, farm machinery, the factory line, the computer, destroyed jobs and, over time, created even more of them, at higher wages. That is the optimistic case, and it has history behind it. But it has a condition attached that people skip past: the workers displaced have to be reabsorbed, into new work, fast enough, before the damage compounds. And that is where today’s numbers should stop you.

The consumer is already at the edge

Here is the condition of the American household right now, before AI displacement has happened at any scale. Roughly fifty-nine percent of Americans live paycheck to paycheck. Total household debt sits at a record, near nineteen trillion dollars. The personal savings rate has fallen to about four percent, down from over six just two years ago. The average credit card interest rate is around twenty-one and a half percent, the highest on record, and serious auto-loan delinquencies have climbed above where they stood at the worst of the 2008 crisis. Consumer sentiment is sitting in recession territory even though unemployment is only about four percent.

Read those numbers together and one phrase from the data captures it: this is survival borrowing, not discretionary spending. A large share of households are not putting purchases on credit because they want a little extra. They are borrowing to cover the gap between what they earn and what they need, in a full-employment economy. That is the cushion, or the absence of one, that would have to absorb a wave of displacement. There is almost nothing there.

Why this is a circle, not a line

Now put the two halves together, and you get the contradiction. The workers most exposed to AI, routine white-collar work, administrative roles, customer service, scheduling, basic analysis, are disproportionately the same households described above: stretched, indebted, without savings. These are not people with a year of runway to retrain. Many are one missed paycheck from delinquency.

And here is why displacing them does not just hurt them, it hurts the very thing the growth was supposed to fix. Consumer spending is roughly two-thirds of the entire economy. A displaced worker stops being a taxpayer and starts being a cost: they pay less income and payroll tax, they draw on the safety net, they cut their spending to the bone, and they may default on their slice of that nineteen trillion dollars in household debt, which then lands on the banks. Every one of those is a negative on the government’s ledger. So the growth that was supposed to shrink the debt, if it arrives by hollowing out the workforce rather than lifting it, quietly widens the deficit through the back door and weakens the consumer economy that generates the growth in the first place.

That is the trap in one sentence. AI growth helps the government’s balance sheet by raising output, profits, and corporate tax, and it hurts the government’s balance sheet by displacing the consumers whose spending and taxes are most of the economy. The same force pulls both ways, and which way it nets out depends entirely on whether the gains are shared or concentrated.

What could get us out of displacement

The debt has a bounded set of roads out, and so does this. It is worth naming them plainly, because “AI will just create new jobs” is a hope, not a plan, and the other roads deserve to be on the table.

The first road is the historical one: new work is created faster than old work is destroyed, and displaced workers move into it. This has happened every time before. The honest question is whether AI is different, because unlike a tractor or a spreadsheet, it targets cognitive work directly, which is exactly where displaced workers have always fled to. If AI climbs the same ladder people are trying to climb, the reabsorption is harder than history suggests.

The second road is redistribution: if AI concentrates its gains in the companies that own it rather than the workers it replaces, the state taxes those gains and recycles them, through wage subsidies, transfers, or something like a universal basic income. Set aside the politics for a moment and notice the irony, this road only works if the AI profits are real and large enough to tax, which sends you right back to needing the boom to be genuine.

The third road is absorption: sharing the reduced work across more people through shorter hours, so productivity shows up as time rather than only as pulled-forward profit. The fourth is simply time, AI adoption proving slower and messier than the demonstrations suggest, so displacement unfolds over decades that natural retirement and gradual reskilling can absorb.

And the fifth road is the one to watch for, the unmanaged one, where displacement outruns creation and redistribution and reskilling all at once, leaving a large, indebted, un-reabsorbed population. On that road consumer demand falls, growth weakens, and the AI-growth escape from the debt stalls, because it undercut the consumer base that powered it. That is the road where both problems, the public debt and the private household, get worse together.

The two problems are one problem

Here is the insight that ties it together, and it is the whole point. The debt problem and the displacement problem are the same problem seen from two sides. The AI-growth road out of the debt only truly works if the growth is broad, if the productivity gains reach ordinary workers as income and not only shareholders as profit. When the gains are broad, workers keep earning, keep spending, keep paying tax, and the debt shrinks while demand holds. When the gains are narrow, workers are displaced and indebted, demand falls, the safety net swells, and the rescue eats itself.

So the answer to “what gets us out of the citizens’ displacement” turns out to be the same as the answer to “what makes the growth road out of the debt truly work.” In both cases it is this: the gains have to be shared. Concentrated AI wealth can make the government’s numbers look better on paper while making the citizen’s worse, and because the citizen is the economy, that is not a solution at all. It is the crisis being moved from the public balance sheet onto the private one, where it is harder to see and just as real.

What this means for a patient investor

I write about price and value, so let me bring it back to the ground where I live. The point of all this is not to predict the future of work. It is to correct a lazy assumption sitting underneath a lot of AI investing: that the growth story is a clean, one-directional good.

It is not. The same force that could rescue federal solvency could erode the mass consumer market that most companies, and the entire tax base, ultimately depend on. A great many businesses are valued today on the assumption of a healthy, spending consumer stretching out for years. If AI growth arrives in a way that thins that consumer out faster than it lifts them up, some of those valuations rest on a foundation that is quietly cracking, and it will show up in the data long before it shows up in the narrative.

So the discipline is simple to state. Watch the consumer as closely as you watch the AI capex. The savings rate, the delinquency rates, real wages, labor-force participation, these are not separate from the AI story. They are the other half of it. The AI buildout numbers tell you how much is being spent to create the productivity. The consumer numbers tell you whether the economy underneath can survive the way that productivity gets delivered. A patient investor who watches only the first half is reading one page of a two-page story.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker.

The growth might save the government. The question this piece leaves you with is the one the cheerful version never asks: at what cost, to whom, and whether the people who pay it are the same people whose spending the whole thing depends on. Watch the consumer. That is where you will see the answer first.

Which leaves one more question hanging, and it is the one I take up tomorrow. Everyone who tells the hopeful version leans on the same proof: we carried debt this big after the Second World War and simply grew out of it. We did. So tomorrow I go back and check whether that escape is still available to us, because the economy that pulled it off was a very different machine from the one we have now. What I found changes the odds on the whole hopeful story, and it points to a gap most people never see coming.

Not investment advice. The subscriber decides.

Government debt, hyperscaler debt, and the cost of money that ties them together. Here is the chain, link by link, and why the chain, not any company stumbling, is the thing most likely to hand a patient investor a great business at a fair price.

The Long View · The AI economy, seen from the bond market

The Long View holds none of the companies named here. This is analysis, not a recommendation, and explicitly not a prediction of any particular outcome.

We have spent a lot of words lately on price, on the idea that a great company and a great investment are not the same thing, and that the bridge between them is always what you pay. This piece is about the thing that could move those prices, not company by company, but all at once, and it starts somewhere most technology investors never look: the market for government bonds.

Stay with me, because the connection is more direct than it seems, and once you see it you cannot unsee it.

Link one: the buildout runs on borrowed money

Start with what the AI boom really costs. The largest technology companies, Microsoft, Meta, Amazon, and Alphabet, are spending on a scale with no precedent. Industry capital expenditure on AI infrastructure is projected to exceed $700 billion in a single year. And that spending has crossed a line that matters: it now consumes essentially all of the cash these companies generate, and then some.

Consider Meta. In its most recent quarter it produced about $31.9 billion in operating cash flow, an enviable figure for almost any business on earth, and it spent $31.1 billion of it on capital expenditure, leaving a mere $784 million behind. A year earlier that leftover figure had been over $8 billion. To keep building at this pace, Meta issued roughly $25 billion in new long-term debt in that single quarter, lifting its total debt to nearly $84 billion. It is not alone. Across the four giants, capital spending is rising far faster than the cash coming in, free cash flow is collapsing toward zero and in some cases through it, and most of them are tapping the debt markets to make up the difference. Alphabet posted its first-ever negative free cash flow quarter. The AI buildout, in other words, is increasingly financed with borrowed money, on the promise that the revenue will arrive later.

That is link one, and it is the foundation everything else rests on: the companies funding the AI economy are now dependent on continued, affordable access to debt.

Link two: borrowed money has a price, and that price is set by Treasuries

Here is where government enters the story. The cost of all that corporate borrowing is anchored to the yield on U.S. Treasury bonds, the interest rate the American government pays to borrow. That yield is the risk-free rate, the number against which every other loan, bond, and investment in the economy is priced. When it rises, every corporate bond issued to build a data center costs more. When it falls, borrowing gets cheaper.

And Treasury yields have been climbing toward levels not seen in two decades. The reason is partly the simplest force in any market: supply and demand. The U.S. government is running enormous deficits and issuing a torrent of new debt to fund them, increasing the supply of bonds. At the same time, some of the largest long-standing buyers of that debt are stepping back.

Link three: the foreign buyers are turning into sellers

For decades, foreign governments were reliable, price-insensitive buyers of U.S. debt. That is changing, and it is not a threat or a forecast, it is in the data. Foreign holdings of U.S. Treasuries fell by roughly $72 billion in a single recent month, the third decline in four months, led by selling from Japan and China, the two largest foreign holders. Earlier in the year, a single month saw Japan and China sell a combined $89 billion, with Saudi Arabia, Taiwan, India, Canada, and the United Arab Emirates also among the sellers.

The reasons are ordinary, which is what makes them durable. Japan’s own government bonds now yield the most since the 1990s, so Japanese institutions have less reason to send their money abroad and more reason to bring it home. A war in the Middle East has pushed oil prices up, adding to inflation and pushing Japan’s central bank to keep tightening, which lifts domestic yields further and accelerates the repatriation. When a country’s own bonds finally pay something, the appeal of American debt fades. So the foreign bid that helped hold U.S. yields down for a generation is weakening, and weaker demand for bonds, against rising supply, means higher yields.

Link four: higher yields squeeze the AI complex two ways

Now the chain closes, and it closes on the AI trade from two directions at once.

The first is direct. The hyperscalers funding the buildout with debt face rising costs to issue that debt as yields climb. A company already running its free cash flow down to nearly nothing, borrowing tens of billions to keep building, is acutely sensitive to the price of that borrowing. Higher yields make the return math on a debt-funded data center harder, and at some point they force a choice: keep spending at any cost, or protect the balance sheet. The lever a cash-conscious giant pulls first is discretionary capital spending, which is precisely the money flowing to the smaller AI-infrastructure companies underneath them.

The second is indirect, and it hits valuations. When the risk-free rate rises, the entire logic of paying a high price for far-off future profits weakens, because now you can earn a real return with no risk at all, just by holding a government bond. The stocks most exposed to this are exactly the ones priced for perfection: the fast-growing, unprofitable, no-dividend names trading at extreme multiples of their sales. Their value lives almost entirely in profits many years away, and a higher discount rate shrinks the present value of those distant profits the most. When money is expensive, the market’s patience for a story that pays off in a decade grows short.

Where this lands: a great company at a price you can finally pay

Put the chain together and you have the mechanism a patient investor should understand cold. Government deficits and foreign selling push Treasury yields up. Higher yields raise the cost of the debt funding the AI buildout and squeeze the cash flow of the giants financing it. Those giants, protecting themselves, trim the discretionary spending that flows to their smaller suppliers. And the same higher yields independently compress the rich multiples of the most expensive AI stocks. None of that requires a single company to fail, to miss, or to stumble. It is transmitted entirely from the outside, through the cost of money.

This is the answer to a question we get often: how does a magnificent business, one growing several hundred percent a year with real margins, ever fall to a price a value investor could pay, if it never stops executing. The answer is that it may never fall because of anything it does. It falls because the environment it lives in reprices. The discipline that keeps us out of these names at today’s prices is the same discipline that would let us in tomorrow, if this chain tightens and drags the whole complex down with it, leaving the best businesses on sale for reasons that have nothing to do with the businesses themselves.

The honest other side

Now the truth protocol, because a chain of plausible links is not a prophecy, and anyone who sells you certainty here is selling you something.

Foreign selling of Treasuries has been forecast to trigger a crisis for well over a decade, and it has not, because every time foreign demand softens, other buyers appear: domestic institutions, pension funds, even the stablecoin issuers who now hold well over a hundred billion dollars in Treasury bills. American government debt remains the reserve asset of the world, with no real substitute at scale. The selling so far has been gradual, a repricing at the margin, not a stampede. And there is a scenario that runs the other way entirely: a sharp economic scare would send money rushing into Treasuries as a haven, pushing yields down, easing the pressure, and extending the very buildout this chain threatens. A recession would hurt the AI trade through demand instead, a different mechanism with a different shape. The hyperscalers, too, are among the strongest companies in history, with enormous cash reserves and the ability to slow spending on their own terms rather than being forced to.

So this is a pressure, not a prediction. A map of how the pieces connect, not a claim about when or whether they will move. The value of understanding it is not that it tells you what happens next. It is that it tells you what to watch, so that if it does begin to move, you recognize it early, while everyone still staring only at the technology wonders why the best AI stocks are suddenly on sale.

And then, this week, the government pushed back

Here is how alive this is. As I was finishing this piece, the tension in the chain broke into the open. Long-dated Treasury yields had just spiked to a nearly twenty-year high, the thirty-year touching levels not seen since before the financial crisis, on exactly the forces described above: heavy issuance, a buyers’ strike in long-dated bonds since early summer, and worry about deficits and inflation. Then the Treasury Department stepped in. It announced it would more than double its buybacks of longer-dated government debt, targeting the very part of the market that had been under the most strain, and yields fell sharply on the news while stocks rose.

This matters for the argument in both directions, and honesty requires saying both. On one hand, it is the clearest possible confirmation that the pressure is real. Governments do not double their debt buybacks and single out the long end of the market unless the strain is serious. The chain I just described is not a thought experiment. The authorities acted this week to relieve exactly the force it runs on.

On the other hand, it is a reminder that this is not a one-way street. There is a policy backstop. When the long end of the bond market gets truly stressed, the Treasury and, if it came to it, the Federal Reserve have tools to lean against it, and they will use them. That can cap the rise in yields, and with it, cap the pressure on the AI trade. A patient investor waiting for the cost of money to bring great companies down to earth has to respect that the people who manage the debt do not want a disorderly rise any more than the market does.

But notice what a buyback does, and does not do. It manages the symptom, the liquidity and the price in one stressed corner of the market. It does not shrink the deficit, bring the foreign buyers back, or make the AI buildout any less dependent on borrowed money. The structural pressure is still there. This week bought some relief at the long end. It did not remove the thing generating the strain, and the fact that intervention was needed at all tells you how real the strain has become. The chain still runs. It is just now a chain the government is actively managing, which is a different thing from a chain that has been broken.

The Treasury has reached for this before

It is worth knowing that this tool is not new, because the two times the Treasury has leaned on something like it, the outcomes could not have been more different, and both are worth carrying in your head.

The first is recent and reassuring. From 2000 to 2002, the Treasury ran a buyback program, repurchasing about $67 billion of its own debt across several dozen operations. But it did so from a position of strength: the government was running a budget surplus and simply had extra cash to retire debt. The effect was benign, a modest scarcity that helped flatten the yield curve, and the program wound down on its own as the surplus faded and the 2001 recession arrived. Buybacks from surplus are a housekeeping tool, not a distress signal.

The second is older and more sobering. During the Second World War, from 1942 to 1951, the government faced the opposite problem: it had to finance an enormous war debt and could not afford high interest rates. So the Federal Reserve agreed to cap yields outright, pegging short-term rates near zero and holding long bonds at 2.5 percent, buying whatever quantity of Treasuries was needed to keep the lid on. It worked, for a while. But it turned monetary policy into a servant of the debt, and when inflation surged after the war, running near 18 percent for a stretch, the Fed was still handcuffed to those caps and could not fight it. It took a formal accord in 1951 to free the central bank to raise rates again. Holding yields down to manage debt has a price, and that price can be inflation the authorities are slow to confront.

Here is why both matter for this week. Today’s buyback is being presented as liquidity support, the benign 2000 framing. But the context is the opposite of 2000: not a surplus, but enormous deficits and a buyers’ strike in long bonds. That is structurally closer to the 1940s problem, financing heavy debt cheaply, than to the surplus housekeeping of 2000, and some analysts are already calling it a backdoor form of yield-curve management. The thing to watch is which way it goes. If it stays a modest liquidity tool, it is a footnote. If it escalates, quarter by quarter, into the government leaning ever harder on the long end to hold down the cost of its own borrowing, the 1940s are the precedent, and the eventual bill for that tends to be inflation, which is its own repricing engine for every richly-valued stock in the market. Neither history is a prediction. But the Treasury just reached for a lever it rarely touches, and a patient investor should know what happened the last two times it did.

What we watch

The hyperscalers’ free cash flow, quarter by quarter, because it is already near zero at some of them, and the moment it turns durably negative, the pressure to cut spending builds.

Their capital-spending guidance, because the first sign of discipline is a target that stops rising, and that would ripple straight down to their suppliers.

Treasury yields and the monthly foreign-holdings data, because they are the pressure gauge for the entire chain, the place the tightening would show up first.

And credit spreads, the extra yield lenders demand to fund riskier borrowers, because when they widen, the debt-funded buildout gets more expensive and more fragile at the same time.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker.

The AI future is bright. That was never the question. The question is what you pay for it, and the thing most likely to change what you have to pay is not the technology at all. It is a chain that starts in Washington’s deficits and the world’s bond markets, and ends at the price of every AI stock you have been watching and waiting to afford.

Not investment advice. The subscriber decides.

Revenue up 454 percent, a real profit margin at the segment level, and a financing trick that quietly answers our biggest worry. The business is proving itself. The price is the problem, and this week has been a lesson in why those are different questions.

The Long View · Tuesday, 18 August 2026 · Nebius (NBIS), Q2 2026 grade

All week we have circled one idea, through IBM and now through a company at the opposite end of the spectrum: the difference between a good business and a good investment. IBM taught it from the value side, a company can be worth owning because the price is low enough to forgive being wrong. Nebius teaches it from the other side. Here is a business doing almost everything right, growing at a rate that stops conversations, and the open question is not whether the company is good. It plainly is. The question is whether any business, however good, is worth what this one costs.

What we said to watch, back in May

In May we ran Nebius through the Stock Story Firewall and classified it Needs Proof. Nebius builds AI cloud infrastructure, the GPU data centers that AI companies rent to train and run their models. A GPU is the specialized chip that does the heavy computation behind modern AI, and renting access to thousands of them is the picks-and-shovels layer of the AI boom.

The hidden assumption we identified was specific: that Nebius’s revenue is scaling fast enough, and holding onto customers well enough, to justify a premium valuation before its cash is consumed by the enormous cost of the buildout. The framework we prescribed was Unit Economics Durability, which asks not whether demand exists, it clearly does, but whether Nebius can serve that demand at a profit that survives as the buildout scales. We set five tests. The company just reported, and it is time to grade them.

Test 1: Is the growth real and accelerating

Published bar: quarter-over-quarter revenue growth staying strong, with the trajectory holding rather than fading.

The number: second-quarter revenue was $582 million, up 454 percent from a year earlier, with the AI Cloud segment alone up 514 percent. Against the prior quarter, revenue grew roughly 48 percent. Proven, without an asterisk. This is not a company coasting on one good quarter. The growth is enormous and it is still accelerating, which is exactly what the bar asked for.

Test 2: Is the recurring revenue base building

Published bar: annual recurring revenue growing, signaling customers committing rather than dabbling.

Annual recurring revenue, or ARR, is the annualized value of a company’s subscription and contracted revenue, a way of turning current run-rate into a forward number. The figure: ARR reached $3.0 billion at the end of the quarter, up about 56 percent from the prior quarter’s $1.92 billion. Proven. Customers are not just showing up for one-time workloads. They are committing, and the committed base is compounding fast.

Test 3: Do the unit economics truly work

Published bar: margins expanding or holding as capacity scales, proving Nebius earns its revenue rather than buying it with underpriced compute.

The number: the AI Cloud segment posted an adjusted EBITDA margin of roughly 50 percent, up from about 45 percent the prior quarter. Adjusted EBITDA is earnings before interest, taxes, depreciation, and amortization, a rough proxy for the cash a business throws off before financing and accounting charges, and adjusted means certain one-time items are stripped out. Proven, and this is the one that matters most. A 50 percent segment margin, rising, says Nebius is not discounting its way to growth. The core service is soundly profitable at this scale. That is the single strongest piece of evidence that the unit economics are real.

Test 4: Will the cash last through the buildout

Published bar: cash reserves holding up against the capital burn, without an obvious near-term need to raise dilutive equity.

This was our biggest worry, and it is where the quarter delivered a genuine surprise. The raw numbers look alarming at first: Nebius spent $5.66 billion on capital expenditure, the money laid out to build data centers and buy chips, in a single quarter, and raised its full-year capex target to a staggering $20 to $25 billion. Against roughly $8 billion of cash, that pace of spending should light a fire.

Except it is being funded in a way we had not fully credited. Roughly 70 percent of the quarter’s deals included customer prepayments, money customers pay upfront, covering 50 to 60 percent of the associated capital cost. Nebius expects more than $9 billion in such prepayments this year. That is why the company ended the quarter with its cash pile intact and $2.3 billion in positive operating cash flow, despite the $5.66 billion of capex. The customers are pre-funding a large share of the buildout. We are moving this from inconclusive toward proven, with one reservation: the model works as long as customers keep prepaying, which ties the cash runway to continued demand. But the fear that Nebius would burn itself dry and be forced to raise equity at a bad moment looks materially weaker than it did in May. This is the most important thing we learned this quarter.

Test 5: Is the backlog converting to revenue

Published bar: the contracted backlog growing and converting, rather than sitting as a headline number that never arrives.

Remaining performance obligations, or RPO, is the contracted revenue a company has signed but not yet delivered, the backlog. The number: RPO stands in the neighborhood of $40 billion, and Nebius signed four landmark deals in the quarter averaging more than a billion dollars each. That is enormous for a company this size. But it stays inconclusive, because the specific rate at which that backlog converts into recognized revenue was not clearly disclosed, and a backlog is a promise, not a sale, until it arrives. The size is staggering and encouraging. The conversion is the thing we still cannot fully verify.

The tally, and the state that holds

Three of the five tests came in clearly proven: the growth, the recurring base, and the unit economics. The cash-runway question, our biggest worry in May, moved substantially toward proven on the strength of the customer-prepayment model. Only the backlog conversion stays truly open. On the evidence alone, this was close to a model quarter, and Nebius remains Needs Proof only because the highest-stakes question, whether all of this endures at scale, cannot be answered by two great quarters. The business is doing the proving. It is proving.

And here is the mirror to IBM

Now the part that ties the week together, and it is the whole point. We spent Monday and Tuesday on IBM, a company we own because in 2020 its price was so low it demanded almost nothing go right. Nebius is the exact opposite situation. The business is arguably far more impressive than IBM’s, growing more than four hundred percent a year at a real margin. And yet it may be the more dangerous stock, because of price.

Nebius trades at roughly sixty times its sales. Not earnings, sales, and its historical median is closer to seven. At that price, essentially everything has to go right: the growth must continue, the margins must hold, the prepayments must keep flowing, the backlog must convert, and the demand must never soften. There is no margin of safety in a sixty-times-sales price. It is the mirror image of the IBM lesson. In 2020, IBM’s low price meant I could be wrong about a great deal and still win. At sixty times sales, a Nebius buyer has to be right about nearly everything, because the price has already priced in success, and then some. It is worth noting, quietly, that the people who know the company best have been net sellers of the stock in recent months. That is not damning on its own. It is one more data point in a picture where the price assumes perfection.

This is not a knock on Nebius the business. The business is extraordinary. It is an observation about Nebius the stock, and the difference between those two things is the most valuable idea we cover. A great company and a great investment are not the same, and the bridge between them is always the price.

What we watch next

The backlog conversion rate, quarter by quarter, because a $40 billion backlog only matters when it becomes revenue, and that is the one test still open.

The prepayment model, and who is behind it, because it is now central to the whole case, and this is where my real concern sits. The prepayments funding Nebius come from a very small number of very large customers, and those customers are themselves spending far more than they earn and borrowing heavily to do it. So the cash that makes Nebius look self-funding is riding on the continued willingness, and ability, of a couple of debt-stretched giants to keep pre-paying. As long as they do, the cash-runway fear stays contained. If even one of them pulls back, the enormous capex turns from a strength into a sharp risk, fast. That is the single thread I would watch above all others.

And the price, because at sixty times sales, the stock does not need bad news to fall. It only needs good news to arrive slightly slower than perfection.

The question this leaves open, and the piece coming tomorrow

So the price is the problem. But that raises the question every patient investor really cares about, the one this grade cannot answer on its own: if Nebius is this good and this expensive, and it may never stumble, then what could ever bring it down to a price worth paying?

Sit with that, because it is not a small question. A value investor could admire this company for years and never get a chance at it, if the only thing that would make it cheap is the company failing, and it never does. But there is another way a great company reaches a fair price, and it has nothing to do with the company at all.

Remember what I just told you about the concern. Nebius is funded by prepayments from two giants who are borrowing to pay it. Those giants borrow at rates set by the bond market. The bond market is moving, right now, for reasons that reach all the way from government deficits and foreign selling of U.S. debt to the price of every AI stock on your screen. There is a chain running from Washington’s borrowing to Nebius’s share price, and if it tightens, it could hand a patient investor this exact company at a price that finally works, without Nebius ever missing a beat.

That chain is the whole story, and it is too big to bury at the bottom of a grade. So it is tomorrow’s piece, in full: every link, what could set it in motion, and the specific conditions I would need to see before a company like this one becomes a buy. If the question of when a great AI company gets cheap has ever nagged at you, that is the one to read.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker, on the companies we own and the ones we do not.

Nebius proved almost everything we asked it to prove. The business is real, the growth is real, the margins are real, and the financing is cleverer than we gave it credit for. And it might still be a poor investment from here, because a magnificent company bought at sixty times sales is exactly the kind of thing this week was about. The company is not the question. The price is.

Not investment advice. The subscriber decides.

Disclosure: the author has held IBM since March 2020 at an average cost of $113.57 and reinvests the dividend. Independent dividend analysis contributed by Dave Ahern of Dividend School. This is analysis and a bar set before results, not a recommendation.

Yesterday’s piece was the easy one to write, because it had a happy ending. This one is harder, because it refuses to let that happy ending decide anything. The most dangerous thing a long-term investor can do is let a stock’s past generosity buy it a permanent place in the portfolio. A holding earns its spot every year on its current merits, not on what it did for you five years ago. So today I set the bar IBM has to clear, in public, and I grade it against those exact lines when it reports, no more gently for the fact that I own it and it has treated me well.

Why the bar matters more now

IBM is not the company I bought. The 2020 business was a deeply cheap turnaround with a fortress dividend and a new cloud strategy. Today it is a hybrid-cloud and enterprise-AI company, and it has had a punishing year that is worth walking through, because the year itself teaches the lesson.

IBM entered 2026 near its highs, above $300 a share, after a strong 2025 that saw the stock climb roughly 35 percent. Then it took two separate blows. In February, it fell about 13 percent in a single session after a competitor announced an AI coding tool aimed squarely at the legacy-system modernization work that is part of IBM’s core consulting business, a real threat to a real revenue line. Then in July, it fell about 25 percent in a day, its sharpest single-session drop on record, after warning on second-quarter results, weak mainframe sales and a cut to full-year revenue-growth guidance. From above $300 to around $236 now, the stock is down roughly 19 percent on my cost basis for the year, and it sits well off where it started. Shareholder law firms opened the kind of investigations that routinely follow a sharp drop; that is a matter of public record, and I note it plainly and draw no conclusion, because there is nothing yet to conclude.

Here is why that year-to-date decline is not a digression but the whole point. IBM entered 2026 expensive, priced near the top of its range with little margin of safety built into the price. So when the two shocks came, there was no cushion to absorb them, and the stock fell hard and fast. That is the exact mirror of my 2020 purchase, where a rock-bottom price meant bad news had far less room to hurt. The year you are watching, in real time, is a live demonstration of what it looks like to own a good company without a margin of safety when trouble arrives. Which is precisely why the price today, and the bar the business now has to clear, deserve a hard look.

I did not sell into that drop, and to explain why, I have to tell you what kind of investor I am. Warren Buffett put it best, in his 1996 letter to shareholders: if you aren’t willing to own a stock for ten years, don’t even think about owning it for ten minutes. That is not a slogan to me, it is the whole posture. When I buy a business, I buy it as something to hold through good years and bad, crashes and boring stretches, on the assumption that I will own it for a decade or longer. An investor who thinks that way does not sell a good company because it had one ugly quarter, any more than a farmer sells the farm because it did not rain this season. A single disclosure about deferred deals is precisely the kind of near-term noise that a ten-year owner has to be able to sit through.

But, and this is the part that keeps the Buffett quote from becoming an excuse, sitting through noise is not the same as ignoring the signal. Owning for ten years does not mean owning blindly for ten years. It means holding through short-term turbulence while still checking, with clear eyes, whether the ten-year story is intact. The disciplined response to a 25 percent drop on a stock you intend to own for a decade is neither to panic and sell nor to shrug and look away. It is to define, precisely and in advance, what the next report has to show for the long-term thesis to still hold. So let me do that.

What margin of safety looks like, then and now

Before I set the bar, look at what has changed about the price, because this is the heart of it. I started buying IBM on March 23, 2020, the worst day of the COVID crash, at $93.49 a share, and I kept buying, which brought my average cost to $113.57. At that first purchase I was paying about ten times earnings, with a dividend yield above 6 percent that was well covered. Against the company’s own ten-year average of around twenty-six times earnings, that was an enormous margin of safety. The price was so low that I did not need much to go right, and when one of my reasons failed, the cheapness absorbed the blow. In 2020, the price protected me.

Today is a different picture, and honesty requires saying so plainly. IBM trades at roughly twenty times trailing earnings, with a dividend yield near 2.9 percent. That is still below its historical average, and by at least one common measure the stock looks roughly fairly valued, neither expensive nor a bargain. This is the crucial lesson, and it is the same company teaching it: the margin of safety was never a property of IBM. It lived in the price. At ten times earnings with a 6 percent yield, IBM offered a fat cushion. At twenty times with a 2.9 percent yield, that cushion is thin. The business is arguably better today, more focused, more profitable, but the protection I enjoyed in 2020 has largely been paid away by a higher price.

There is one more piece of this worth teaching, because it surprises people and it is the quiet reward for buying low. You will notice IBM’s yield went from above 6 percent in 2020 to about 2.9 percent today, and it is worth understanding why, because it is not what it looks like. The yield did not fall because IBM got stingy. The dividend rose over those years. The yield fell because the price more than doubled, and yield is simply the dividend divided by the price. When the bottom of that fraction grows faster than the top, the yield shrinks, even as the payout climbs. A 2.9 percent yield today and a 6 percent yield in 2020 can be the very same dividend, seen at two very different prices.

But here is the part that rewards the patient buyer. My yield is not 2.9 percent. That is the yield a new buyer gets at today’s price. My yield is measured against what I paid, which is called yield on cost, and against my $113.57 average it is closer to 5.9 percent, and against my first shares at $93.49 it is around 7 percent. Same dividend, but because I bought at a low price, I locked in a high yield that rising prices can never take back. This is one of the most underappreciated rewards of buying with a margin of safety: a low entry price does not just protect you on the downside, it permanently raises the income you earn on every dollar you invested, for as long as you hold. New money buys IBM’s dividend at 2.9 percent. My money, bought in the panic, still earns almost twice that.

Here is why that matters for what comes next. When the price gives you a wide margin of safety, the price does the protecting, and the business can stumble without ruining you. When the price is merely fair, the protection has to come from somewhere else: the business executing. At today’s valuation, I am no longer being handed a cushion. I am paying a fair price and trusting the company to deliver, which means the burden shifts from the price to the performance. That is precisely why the bar matters more now than it did when I bought. In 2020 I could be patient and let a cheap price carry me. Today, IBM has to earn it. So here is what earning it looks like.

The bar for October

IBM reports its next quarter on October 21. Here are the five lines, set before the numbers exist. Each has a level that confirms the thesis and a level that breaks it.

Software recurring revenue growth. IBM’s software segment carries annual recurring revenue, or ARR, the annualized value of its subscription contracts, currently around $24.6 billion growing about 8 percent a year. Growth at or above 10 percent confirms the compounding lock-in the whole story depends on. Growth below 6 percent signals the stickiness is eroding.

Red Hat growth. Red Hat, the hybrid-cloud engine, has been growing around 11 percent, with OpenShift’s recurring revenue near $2.2 billion. At or above 12 percent confirms enterprises are still adopting the platform. Below 7 percent signals it is being commoditized or pushed aside by bigger cloud rivals.

Full-year guidance. IBM already cut its constant-currency revenue-growth guidance, the growth rate stripped of currency-exchange swings, from above 5 percent to a range of 4 to 5 percent this year. Holding at or above 5 percent confirms. A second cut, or new caveats about enterprise spending, would be a serious credibility problem.

Free cash flow. Free cash flow, the cash a business generates after funding its operations and investments, is what pays IBM’s dividend and services its debt. IBM targets roughly $12 billion or more for the year. At or above $12 billion confirms the cash engine is intact. Tracking below $10 billion annualized signals deterioration.

Insider buying. After a 25 percent drop, the cleanest signal of conviction would be a named insider buying shares in the open market with their own money. Continued absence is not damning, but a purchase would say something no press release can.

Five lines. All pending until October 21. That is the bar, set in public, today.

The dividend I have to keep honest about

The dividend is why I have been content to hold through a bad year, so it is exactly the thing I have to watch most carefully, because comfort is where discipline goes to die. IBM yields around 2.9 percent today, has raised the payout for about thirty consecutive years, and I reinvest it every quarter, which means this year’s decline has quietly been buying me cheaper shares the whole way down. That is the good side. But a long streak is a point of pride companies will stretch to protect, and the honest question is whether it is funded by cash the business earns or partly by borrowing, especially with long-term debt around $56 billion after recent acquisitions.

This is where an outside view matters, because I am too invested to trust my own. Dave Ahern of Dividend School gave me his read, and on the forward question it is reassuring with one flag worth teaching in full, because it applies to any dividend stock you will ever own.

First, the reassuring part. Dave checks three things, and IBM passes all three. The free-cash-flow payout, the share of actual cash the dividend consumes, sits well below his 70 percent threshold for concern, so the cash covering the dividend has real room. Net debt is about 2.9 times EBITDA, a company’s earnings before interest, taxes, depreciation, and amortization, which is a standard way to size debt against a rough proxy for cash earnings, and 2.9 times is a level he considers fine. And interest coverage, how many times over the company’s earnings can pay the interest on its debt, is around 6 times, which he considers good. Together those say the dividend is not being squeezed by the debt today.

Now the flag, and this is the part I want every reader to carry away, because it is the single most useful early-warning sign for a dividend. Dave’s caution is that IBM’s dividend has been growing slightly faster than its free cash flow. Here is why that matters. A dividend is paid out of cash. The room between the cash a company generates and the dividend it pays is the safety cushion. If the dividend grows faster than the cash behind it, year after year, that cushion narrows, slowly, quietly, in a way that looks fine every single year right up until it does not. The payout ratio creeps from comfortable toward stretched, and one bad year can then turn a proud streak into a hard choice between cutting the dividend and borrowing to pay it. It is not a problem when the coverage starts wide, as IBM’s does. It becomes a problem through repetition, a percent or two a year, until the margin is gone.

So here is how you check this yourself on any dividend stock, and it is simple. Find the growth rate of the dividend over the last five years, and find the growth rate of free cash flow over the same period. If free cash flow is growing at least as fast as the dividend, the cushion is holding or widening, and the dividend is being earned. If the dividend is consistently outgrowing the cash, the company is writing checks its business is not yet backing at the same pace, and you are watching a cushion deflate. It does not mean sell. It means watch, and know the number that would turn a watch into a worry. For IBM, that number is free cash flow. If October shows it holding at or above roughly $12 billion, the dividend’s cushion has room and the streak is funded by the business. If it slips toward $10 billion while the dividend keeps climbing, the gap Dave flagged starts to bite, and the safest dividend on my holdings becomes one I have to question. That is the whole discipline of dividend safety in one comparison: is the cash growing as fast as the promise.

What would truly change my mind

Here is the part that keeps this honest. It is easy to say you will grade a holding strictly. It is hard to say, in advance, what would make you let it go. So: if IBM cuts guidance a second time in the same year, if software and Red Hat growth both fall toward their break levels, or if free cash flow drops far enough to put the dividend’s coverage in question, then the thesis I have held since 2020 is weakening in a way that price alone cannot excuse, and I have to treat it as a business in trouble rather than a winner having an off year. I am not predicting that. I am telling you the conditions under which I would change my mind, because an investor who cannot name those conditions is not holding a position, they are married to it.

The through-line from yesterday

Part one was about how a low price forgave a mistake. Part two is about not needing that forgiveness again through carelessness. In 2020 the margin of safety was enormous, ten times earnings, a 6 percent covered yield, and it made me hard to hurt. Today the price is not that, the easy protection is gone, and what replaces it is discipline: a bar set in public, a dividend watched by someone without my attachment, and a written list of what would make me sell. That is how you hold a long-term winner without becoming its hostage.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker, including every bar we set and every grade we publish.

The stock has been good to me for five years. That is exactly why I have to be hardest on it now. In October, IBM reports, and I will grade it against the five lines above, and find out whether the company I own on its history is still the company I would own on its merits.

Not investment advice. The subscriber decides.

In March 2020 I bought a 109-year-old company at about ten times earnings, an eleven-year low, paying a dividend yield above 6 percent that was well covered, with a fresh cloud strategy and a working AI. One of my reasons for buying, the AI, later collapsed. The investment still returned more than 200 percent, and the reason is the most important idea in value investing: a price low enough is a margin of safety wide enough to forgive being wrong.

Disclosure: the author has held IBM since March 2020 at an average cost of $113.57 and reinvests the dividend. Independent dividend analysis contributed by Dave Ahern of Dividend School, printed unedited. This is analysis, not a recommendation.

This is the first of two pieces on a company I have owned for five years. This one is the story, and the lesson inside it. The second, tomorrow, is the discipline: the bar IBM has to clear to keep its place in the portfolio. You cannot understand the second without the first, so let me start where I started, at the bottom of a crash, with a stock almost nobody wanted.

What IBM looked like in March 2020

Cast your mind back. The COVID crash was tearing through the market, and IBM, already an unloved stock, fell to its lowest point since 2009, an eleven-year low. On March 23, the worst day of that crash, I started buying at $93.49.

Here is what I was buying, in the language of value. IBM’s price had fallen to roughly ten times its earnings. To put that in perspective, over the previous decade IBM had traded at an average of about twenty-six times earnings, so I was buying it at well under half its own normal valuation. The dividend yield, the annual dividend divided by the share price, had climbed above 6 percent, because when a price falls far enough, the fixed dividend becomes a larger and larger share of it. And that dividend was not a fragile promise. IBM had raised it for twenty-four straight years, and it was consuming less than half of the company’s free cash flow, the cash left after running and investing in the business, and about 61 percent of earnings. A well-covered 6 percent yield on a 109-year-old company at ten times earnings. That is not a lottery ticket. That is a value.

The reasons I bought, and I want to be honest that there were several

I did not buy IBM for one reason, and being straight about that is the whole point of this piece.

The business case had real parts. IBM had just closed its $34 billion acquisition of Red Hat in July 2019, and it was building a hybrid-cloud strategy, the business of helping big companies run their computing across their own data centers and multiple public clouds at once, with Red Hat’s OpenShift platform as the engine. Not public cloud, IBM had already lost that race to Amazon and Microsoft, but hybrid cloud, a specific corner where IBM was positioned to lead. Alongside that was Watson, IBM’s artificial intelligence, and it was not a slide in a pitch deck, it was a working system being used in real applications, including cancer research, and its potential looked enormous.

So the case was: a deeply cheap price, a fortress dividend, a credible new cloud strategy, and a working AI with a huge runway. Several independent reasons to own it, bought at a price that demanded almost nothing go right.

The part I got wrong

Now the honest part. The Watson story, specifically Watson in healthcare, the piece I found most exciting, became one of the most expensive disappointments in IBM’s modern history. The company had poured roughly five billion dollars and thousands of people into Watson Health, the marquee hospital partnerships faltered, and in 2022 IBM sold those assets to a private equity firm for around one billion dollars, a fraction of what went in.

One of the reasons I bought IBM did not just underperform. It failed outright.

Why it worked anyway

And the investment still returned more than 200 percent. Sit with how those two facts coexist, because the space between them is where value investing really lives.

It worked for two reasons, and they are the same reason wearing two hats. First, the parts I got right, the cheapness, the dividend, the hybrid-cloud repositioning, were the durable ones, and they carried the whole. Second, and underneath everything, the price. I paid about ten times earnings for a profitable, cash-generating business. At that price, I was not depending on Watson to save me. I was being paid a well-covered 6 percent every year just to wait, on a business with several other ways to win. So when one of those ways collapsed, it cost me a piece of the upside, not the investment. The margin of safety absorbed the mistake.

That phrase, margin of safety, is the oldest idea in value investing, and it means buying so far below what a business is worth that you can be wrong about part of your thesis and still come out fine. It is insurance you buy with a low price. If I had paid a premium in 2020 and leaned the entire case on Watson, its failure would have been a wound. Because I paid a crash price for a business with multiple engines and a fat, covered dividend, the failure was a footnote.

An independent look at the dividend that anchored it

Because the dividend was so central to why this worked, and because I have reinvested it every quarter for five years and am therefore too attached to judge it cleanly, I asked someone who is not. Dave Ahern, who writes Dividend School, gave me his honest read, printed as he sees it.

His verdict was that IBM’s payout looks fairly safe. The scary-looking number is the earnings payout ratio, which has run above 100 percent in four of the last ten years, meaning IBM paid out more in dividends than it booked in reported profit in those years. Dave calls that not awesome, but not a killer, and here is why: he weights free cash flow over earnings, because a company like IBM carries heavy non-cash charges that make reported profit understate the actual cash coming in. On that truer measure, IBM’s free-cash-flow payout has stayed well below 70 percent, his threshold for concern. The cash covers the dividend even when earnings optically did not. He flagged one thing to watch, dividend growth is running slightly ahead of free-cash-flow growth, which slowly narrows the cushion over time. That is the honest kind of caution I wanted and could not have given myself. My thanks to him for the outside eyes.

The point for this piece is that the dividend was not decoration on the value case. It was a load-bearing part of it, then and now. A 6 percent yield you can trust is a huge share of the margin of safety, because it pays you to be patient while the rest of the thesis plays out, or does not.

The lesson, stated plainly

Here is what I want you to take from this, because it is the truest thing I know about investing, and it runs against how most people think about being right.

You do not have to get the story completely right to do well. You have to get the price right enough that being partly wrong cannot ruin you. I was wrong about the single most exciting reason I bought IBM. In most framings of investing, being wrong about your thesis means you lose. But value investing is not built on being right about the future. It is built on paying so little for the present that the future has many ways to reward you and few ways to hurt you. The margin of safety is not a nice-to-have. It is the whole mechanism by which ordinary investors survive their own mistakes, and everyone makes them.

Buy a good business at a punishing price and you need everything to go right. Buy a good business at a genuine discount, with a covered dividend paying you to wait, and you can be wrong about a great deal and still win. That is not a trick. It is the entire discipline, in one holding.

Tomorrow, the harder half

None of this tells me whether IBM still deserves its place in my portfolio today. That is a different question, a forward-looking one, and it does not get the gentle treatment just because the position has been good to me. The stock recently fell 25 percent in a day on a real disclosure, and it reports again in October. Tomorrow, I set the bar it has to clear, in public, before the numbers, and I grade it no more gently for owning it. The value that forgave my mistake in 2020 does not entitle IBM to anything now. Today it earns its place or it does not, and that is part two.

Run any company through the Stock Story Firewall, free, at firewall.readthelongview.com. Subscribers get the Watch cards, the Evidence tool, and the full Research Tracker, on the companies we own and the ones we do not.

I was half wrong about why I bought IBM, and at ten times earnings, with a 6 percent dividend paying me to wait, it barely mattered. Remember that the next time someone tells you investing is about being right. It is about being wrong safely, and the price is what makes the safety.

Not investment advice. The subscriber decides.