The first of the things I got wrong this year. This one I can at least write down.
Two people valued the same company on the same day, using the same method. One arrived at about $980 a share. The other, working from a different but entirely defensible set of assumptions, arrived at about $206.
Neither of them made an arithmetic error. The entire gap is assumptions.
I know this because I had spent the previous week producing a third number for the same company, and mine sat comfortably between theirs. I was rather pleased with it. It had a pessimistic case, an optimistic case, and a current market price sitting politely in the middle, which is what a valuation is supposed to look like.
Then someone pointed out that I had answered a question I hadn't asked.
First, a bakery
Everything below rests on one calculation, and if you have never met it, the rest of the post is noise. So here it is, with small numbers, and no spreadsheet.
Imagine a bakery on your street is for sale. It hands its owner $10,000 a year in cash, and there is no reason to think that stops.
What is it worth? Not $10,000 — you get that every year. But not infinity either, because a dollar arriving in 2046 is worth less to you than a dollar arriving on Friday. You could have done something with the Friday dollar in the meantime, and the 2046 one might never turn up at all.
So you shrink each future year before you add it up. Say you want 10% a year to make owning a bakery worth the bother. Then next year's $10,000 is worth $9,091 to you today, the year after's is worth $8,264, the one after that $7,513, and so on, each year a little smaller than the last. That 10% is called the discount rate, and shrinking future money by it is called discounting.
Add up that whole shrinking tail, out to forever, and it comes to a clean number: $100,000. That is a discounted cash flow — a DCF. Ten thousand a year, at 10%, is worth a hundred thousand. That is the entire idea. The models get baroque; the idea doesn't.
Now watch what happens when I change one belief. Suppose I decide the street is getting busier and the bakery's profit will grow 3% a year, forever. Same shop, same ovens, same $10,000 today. New value: $142,857.
I did not learn anything about the bakery. I changed one assumption in my own head, and the answer moved 43%. Hold on to that, because it is the whole problem.
Now run it backwards
The owner wants $200,000 for the bakery.
The forward way to think about this is to build your own number and compare. I've just shown you how fragile that is — nudge your growth assumption and you can make $100,000 or $142,857 or almost anything you like, all of it defensible, all of it a description of your own mood.
The backwards way asks a different question, and it is a much better question: what would have to be true for $200,000 to be the right price?
Run the same arithmetic in reverse and it falls out in one line. At $200,000, with $10,000 of profit and my 10% requirement, the growth rate baked into the asking price is 5% a year, forever.
Notice what just happened. I am no longer arguing about a valuation. I am holding a single claim, with a number in it, about a bakery on a street I can walk down. Has it grown 5% a year? Is the neighbourhood filling up or emptying out? Is there a second bakery opening in the spring?
I still might be wrong. But I am now wrong about something specific, in public, where the street can correct me. That is the difference between the two directions, and everything else in this post is that same move applied to a company far too large for me to walk past.
Two directions
Those two ways of pointing the same arithmetic are not variations on a theme. They are opposites, and almost every investing book ever written teaches only one of them.
The first goes company → value → compare to price. You estimate what the business will earn, discount it back, arrive at intrinsic value, and hold that up against the quote. This is the direction every book on my shelf teaches. Graham built the discipline of it, Fisher added the qualitative layer, Buffett made it famous.
The second goes price → expectations. You take the current price as a given — not as a verdict, as data — and work backwards: what future does this price already assume? What growth, what margin, sustained for how many years, is baked into what people are paying right now?
The first direction produces a number you can fall in love with. The second produces a claim you can argue with.
And you do not need to do any of the arithmetic to start. Here is the crudest possible version, and it works on a phone in two minutes. Take the company's market capitalisation — the price of one share times the number of shares, which is what the whole business is being sold for. Take last year's free cash flow — the cash left over after the business has paid for everything it needs to keep running. Divide the first by the second. If the answer is 40, the price is asking you to accept forty years of today's cash flow — which is only sane if you believe there is going to be considerably more of it later. That single ratio is a reverse DCF with the roof torn off. It is not precise. It is directionally honest, which for most decisions is the part that matters.
The two missing shelves
Here is what unsettled me. I counted my library by what it actually teaches, and the split was total: roughly forty books on how to find an investment, and none on what the price already assumes, or on how much of it to own.
Two enormous gaps, invisible to me for years because nothing on the shelf pointed at them. Everything I own I chose with the tools from shelf one. Everything I own I sized with instinct.
This post is about the first of those two gaps. The second one — the one about size — is harder by a distance, and I have been putting off writing it for a reason that will be obvious by the end of this one.
The trap in a cyclical
The company I was valuing is a memory-chip maker. I am not going to name it; principles travel further than tickers, and I own it, which is exactly the situation in which my opinion is worth least. All figures below are from its most recently reported quarter, as at September 2026.
Revenue of $41.5 billion. Gross margin — what is left of each dollar of sales once you subtract the cost of physically making the thing — of 84.6% on a GAAP basis — 84.9% on the company's adjusted basis, so this is not an accounting artefact — with the next quarter guided to roughly 86% on $50 billion of revenue.
That margin is the whole problem, and I did not see it.
Here is the history. This same company's previous all-time cycle-peak gross margin, in 2018, was 58.9%. Today's number sits twenty-six points above the best it has ever managed at a top. And every prior memory cycle ended the same way. Margins fell from 59% to 27% between 2016 and 2019, and the stock lost 56%. Operating margins went negative in the 2020–2023 cycle and the stock halved. In 1996–97, DRAM prices fell 51% and then another 65%; memory stocks lost between 60% and 80%.
Now look at what my model said. My pessimistic case was about 12–14% below the current price.
For a cyclical business at an unprecedented margin peak, a 14% decline is not a pessimistic case. It is a mild-disappointment case. The mildest outcome in the historical record is minus fifty percent.
The mechanism of the error is the classic one: a DCF projects forward from wherever you start it. Start a cyclical at peak margins, let them fade gently and politely, and the model hands back a large number — a number that describes your starting point rather than the business.
And here is the part that stung, because the answer was already on my shelf. At the current price the stock trades at roughly 6.6 times next year's expected earnings. That is the P/E ratio: how many years of profit you are paying for. Six and a half years looks absurdly cheap — most large companies ask for twenty or more. Lynch wrote the warning decades ago: for a cyclical, a low P/E is a top signal, not a bargain, because the denominator is peak earnings. A high P/E can mean the trough. I owned the book. I had read the book. I walked into it anyway.
One objection to my own confession
Before I go further, the honest counter-argument, because a confession you don't stress-test is just a different kind of performance.
Intrinsic value should fall less than trough earnings. Most of a DCF's value sits past the years you actually forecast, in the lump the model uses to stand for "and then it carries on" — the same forever-tail that made the bakery worth $100,000; a two-year collapse in margins that later recovers genuinely does not destroy 50% of a business's worth. So a −14% value case is not automatically absurd.
What was absurd was what I did with it. I produced a number about value and then behaved as though it were a number about price. Between here and any eventual recovery there is a price path, and the price path is the part I actually have to live through — with my own nerve, my own timeline and my own capacity to keep holding while the screen is red. A valuation that ignores the path is a description of a destination given to someone who has to walk there.
"But what if the cycle is broken?"
This was my honest objection, and I still think it is the strongest argument on the table.
The company has signed sixteen long-term customer agreements running through the end of 2030: take-or-pay, and not cancellable. They carry roughly $100 billion in minimum contracted revenue. Customers have put up $22 billion in deposits and letters of credit — about $18 billion of it cash, paid up front.
That is genuinely new. Memory has never been sold this way. In every previous downcycle the customer simply cancelled the order; that was the mechanism. Here they cannot, and their money is already on the table.
So I asked: doesn't that change the valuation? And I realised I had not calculated it at all. I had assumed it.
Four things came out of actually doing the arithmetic.
First, divide the headline by the base. One hundred billion dollars is an enormous number until you set it against the company's own scale. Next quarter's guidance is $50 billion — call it $200 billion a year. Spread across five years, the contracts guarantee about $20 billion a year. Ten percent of the current run rate. A cushion, not a transformation.
Second, the contracts have a ceiling. This is the part almost nobody mentions. The largest agreements are priced within a band: a floor, yes, but also a price ceiling set at calendar Q2 2026 market levels. If memory prices keep climbing, the company does not capture the upside on that revenue. Once all the planned agreements are in place, fixed or capped pricing is expected to cover roughly 40% of revenue.
A structural change that reduces variance is not automatically a bullish thesis.
Less volatile cash flow deserves a lower discount rate, which raises value. Capped peak earnings lower the optimistic case. The two effects pull in opposite directions, and which one wins depends entirely on numbers I had never put in a spreadsheet. "The cycle is broken" is not a reason the stock is worth more. It is a reason the distribution is narrower — at both ends.
Third, contracts allocate supply; they don't absorb it. Industry DRAM bit shipments are growing in the low-to-mid twenties percent this year. The company's own capital spending is around $27 billion and rising. Its two largest competitors are expanding, and a Chinese entrant is arriving. The contracts cover about 20% of DRAM volume. The other 80% is sold at the going rate on the open market — the spot price — and spot is what sets the price for everyone, including the customer deciding whether to take delivery cheerfully or grudgingly. That is the mechanism that broke every previous cycle, and the contracts do not touch it.
Fourth — and this one cuts against me. In a real downcycle, revenue does not politely stay at $200 billion. It halves. That same $20 billion floor is then twenty percent of revenue, not ten, and it is the highest-quality fifth of it. My denominator flattered the argument I had just talked myself into.
Which is its own lesson, and a more useful one than the first three: divide by the base in the scenario you are testing, not the base you can see today. I caught myself being sloppily bullish and replaced it, within about an hour, with being sloppily bearish. The direction changed. The sloppiness didn't.
The mistake I made about markets
A second error, unrelated, which I'll keep short because it deserves its own post.
I had argued that since the US market persistently trades at higher multiples than XETRA in Frankfurt, and XETRA higher than the BIST in Istanbul — for real reasons: liquidity, disclosure, enforcement, currency, the investor base, index flows — there ought to be a "fair premium" you add on top of a DCF value for a US listing.
There should not, and the error has a name: double counting. The market premium is already inside the DCF. It lives in the discount rate; that is how it enters. Discount the same cash flows at 8% instead of 12% and you get a much higher value, and the gap between those two numbers is the premium I was proposing to add again.
Underneath that sits the more useful confusion. A lower cost of capital is not a gift; it is the price of admission. It raises today's value of a given stream and lowers the return the buyer should expect from here. If US-listed equities are priced to return less than Istanbul-listed ones, that is the compensation structure working exactly as designed — not a discount waiting to be claimed.
I originally wrote a neat closing argument here: that a premium applied to every stock in a market cannot change a relative ranking, so it cannot change a decision. That is true within a market, and it is false across markets — which is precisely the comparison I had been making. So: the correction needs its own correction. The double-counting objection is what actually kills the idea. The level-shift argument doesn't, and I'm leaving my error visible rather than quietly deleting it.
Where the instinct does earn its keep is somewhere else entirely. The market premium is not a constant. It is a variable, and it mean-reverts. Buying a US-listed company while the US premium sits near a historical high is a different bet from buying the same company when the premium is average. Which means something uncomfortable about my own portfolio: nearly all of my equity is listed in one market. I am not only long these companies. I am long the premium of the market they trade in — a position I never chose, sized, or priced.
How much of the gap between markets is payment for real risk and how much is simply fashion is a question I have started pulling at, and I do not much like where the thread is going. Another time — once I have finished being wrong about it in private.
So what did I do?
Nothing. And I want to be straight about that rather than let an essay end on a flourish.
I have not sold, not trimmed, not added. My sell rule is that I exit when the thesis is broken and the position is at a loss, and neither is true. Nothing in the analysis above breaks the thesis; it changes what I think the thesis is.
But "my rule didn't fire" is a description of a rule, not a decision. The honest position is that I don't yet know what the right size for this is, at this point in this cycle, with this much of my net worth in one market. That is not modesty. That is the second missing shelf, and I am standing in front of it empty-handed.
The claim I am actually holding
When I finally ran the model backwards, the result was not what I expected. Even at a normalised gross margin far above anything the industry has ever sustained, the model would not reach the market price.
Which means the expectation embedded in this price is not about the level of margins at all. It is about their duration, and about growth. The market is not saying "these margins are permanent." It is saying "this business keeps getting bigger from here."
That is a far more specific claim than the one I thought I was buying. And it is falsifiable, which is the only property that makes a thesis worth holding.
One correction to something I nearly wrote. I was going to say that reverse DCF is immune to the $980-versus-$206 problem, because it takes the price as given. That is not true, and it is worth saying plainly. A reverse DCF still needs a discount rate, a view on how quickly today's margins drift back to normal, and an assumption about the forever-tail. Change my bakery's 10% to 8% and the growth rate the price implies drops from 5% to 3%; change it to 12% and it jumps to 7%. Same asking price, three different stories about the street. The arbitrariness does not vanish; it relocates.
What reverse DCF actually does is smaller and better. It corners the disagreement. Instead of two people trading whole valuations that each contain a dozen buried judgements, you end up arguing about one variable, out in the open, with a number attached. It does not make you right. It makes you checkable.
A view you cannot state precisely enough to be proven wrong is not a view. It is a mood.
The record, as of September 2026
So that this is checkable rather than merely well-written, here is what I am on the hook for. I intend to come back to this box.
- Position: held, not trimmed, not added. Sized by instinct, which is the thing I'm trying to fix.
- What the price assumes: not that these margins persist, but that the business is materially larger in five years than it is now.
- What would prove me wrong about the risk: a full cycle turn in which the drawdown stays inside 25%, or contracted revenue rising well past 40% of the total.
- What would prove me right: spot pricing rolling over while contracted volume stays flat, and margins reverting toward the high fifties.
- Next review: the quarter after next, whatever the price has done in between.
In short
- A forward DCF describes your starting point. Start a cyclical at a record margin and the model will hand you back your own optimism, formatted as arithmetic.
- For a cyclical, a low P/E is a warning, not a bargain — the denominator is peak earnings.
- A structural change that reduces variance cuts both tails. It is not automatically bullish.
- Divide by the base in the scenario you are testing, not the base you can see today.
- A market premium is already inside your discount rate. A lower cost of capital is the price of admission, not a discount.
- Reverse DCF doesn't remove assumptions. It corners them — which is what makes a view checkable.
Try this
Take one thing you own. Find its market capitalisation and its last twelve months of free cash flow — both are on any free finance site, no login required — and divide the first by the second. That number is how many years of today's cash you are being asked to pay for. It is the bakery question, asked about something you already own.
Now write one sentence, with a number and a date in it: "This price assumes ____ grows ____% a year through 20__."
Then write the second sentence: "I will know I was wrong if, by ____, I see ____."
If you cannot finish the second sentence, you do not have a thesis. You have a mood, and it is costing you money.
And one more, thirty seconds long, that I should have done first: look up your company's worst peak-to-trough drawdown in its own history, and ask whether your bad case is anywhere near it. Mine wasn't. It was off by a factor of four.
There are two more of these waiting. One I owe you. The other I am still losing an argument with myself about, which is usually the sign that it is the one worth reading. They will come when they are honest.
Disclosure: I own shares in the company discussed, unnamed throughout. Nothing here is a recommendation to buy or sell anything. All figures are from public company reporting as at September 2026 and were accurate when I wrote this, which is a different thing from being accurate when you read it.
0 Comments