Micron Stock and the $26.5 Billion Tell

The head of SK Hynix says the memory shortage will last beyond 2030. He has just raised $26.5 billion to build the factories that would end it. Micron, meanwhile, has quietly capped its own selling price for five years. Watch what they do.

Micron stock: A silicon wafer. Photograph by Windell Oskay, CC BY 2.0, via Wikimedia Commons.

Micron Technology · NASDAQ: MU · $979.30 · 11 July 2026 All figures as at this date. Sources: SEC filings, company earnings materials, and market data.

On Friday, two things happened, and almost everyone read them as the same thing.

In New York, SK Hynix rang the bell on the largest listing ever undertaken in the United States by a foreign company. It sold 177.9 million American depositary shares at $149 apiece and raised $26.5 billion, beating the record Alibaba set in 2014. The book was more than seven times oversubscribed. The shares closed the first day around $168, up roughly 13%.

That same day, in an interview after the debut, SK Hynix's chief executive Kwak Noh-Jung said the memory shortage would probably persist beyond 2030, and that 2027 would be the worst year for supply in the industry's history.

The market treated the second statement as the reason for the first. A shortage without end; a stampede to own a piece of it.

Now read them in the other order. SK Hynix told regulators what it intends to do with the money: build new production facilities. The largest capital raise a foreign company has ever executed on an American exchange is going into memory fabs. Which is to say that the single most authoritative voice on the durability of the shortage has just raised twenty-six and a half billion dollars to bring it to an end.

That is not a contradiction. It is the memory industry's oldest and most reliable mechanism, and it has never once failed to work.

What memory is, and why it suddenly matters

Two products. DRAM is the fast, forgetful memory a computer thinks in; NAND is the slower storage that remembers when the power goes off. Both are made in enormous quantities to common standards, and for most of their history both have been bought the way flour is bought: by the metric ton, from whoever is cheapest this month.

AI changed the volumes but not, on the evidence so far, the nature of the product. Training and running large models is bounded less by raw compute than by how fast and how much memory you can feed the processor. Data center operators have spent accordingly, and in doing so have drained the pool that everyone else — phones, cars, laptops — was also drinking from.

The result has been the most violent upswing in the industry's history. In its fiscal third quarter, reported on 24 June 2026, Micron did $41.5 billion of revenue, up 346% on the same quarter a year before. Gross margin was 84.9% on a non-GAAP basis (84.6% on GAAP). Operating margin was 81.2%. It guided the current quarter to $50.0 billion of revenue. The shares are $979.30 and the company is worth $1.11 trillion.

Those are real numbers, audited or company-guided, and any honest analysis has to start by taking them seriously.

The bull case, at its strongest

Do not caricature it. The argument for Micron in July 2026 is not “AI go up.” It is specific, and parts of it are strong.

Supply cannot answer quickly. Micron's own account is detailed and, as far as I can check, accurate: new memory capacity requires greenfield fabs, and those are gated by construction lead times, permitting, a shortage of skilled trades, and grid connections. Process nodes are getting harder. High-bandwidth memory consumes more wafer per usable bit with each generation, which cannibalizes ordinary DRAM supply. NAND cleanroom space is being converted to DRAM. None of that is spin. It is why Micron says it has no line of sight to when supply catches demand.

Micron is no longer purely a price-taker in the part that matters most. High-bandwidth memory is co-designed, qualified into a customer's platform, and hard to second-source mid-cycle. Micron held roughly 21% of the HBM market in the second quarter of 2026, moving past Samsung into second place behind SK Hynix. Its HBM4 12-high is ramping at twice the pace of the prior generation and has already passed $1 billion in revenue. Switching supplier on HBM is not like switching supplier on a DIMM.

And then the thing the bulls actually care about: the contracts. Micron has signed sixteen strategic customer agreements. They are take-or-pay. They typically run five years, from calendar 2026 to the end of 2030. They carry roughly $100 billion of remaining performance obligations, and customers — Nvidia among them, per Reuters — have committed $22 billion in deposits and related financial commitments. Crucially, the agreements carry a price floor, and Micron says that floor “enables a very robust gross margin for Micron, well above our peak quarterly margins in any past cycle.”

Sit with that last claim, because it is the load-bearing one. Micron's previous best quarterly gross margin, at the top of the 2018 supercycle, was 61%. If a contractual floor holds a substantial slice of revenue above that level through 2030, then the catastrophe of fiscal 2023 — when Micron's gross margin was minus 9.1% and it lost $5.83 billion — becomes arithmetically impossible to repeat. That would be a genuine change in the shape of the downside, not a story about it.

The agreements will significantly enhance the durability and predictability of Micron's strong financial performance. — Sanjay Mehrotra, Chairman, President and CEO, Micron Technology

That is the case. It deserves better than a sneer.

The complication

Chart of Micron DRAM and NAND selling-price changes, with the memory price move dwarfing the change in bits shipped.

Start with where the record quarter came from. Micron's DRAM revenue rose 67% sequentially. DRAM bit shipments rose in the low-single-digit percentage range. DRAM prices rose in the low-60s. In NAND, revenue nearly doubled; bit shipments rose in the mid-single digits; prices rose in the mid-80s.

Micron did not sell meaningfully more memory last quarter. It sold almost exactly as much memory, for very much more money.

That is not a criticism. It is a description, and it is the description of a commodity in shortage. A business with pricing power raises price and keeps its customers because they cannot get what it sells anywhere else. A business in a shortage raises price because there is not enough of the thing, and every one of its competitors is raising price by an almost identical amount at the same moment. The two look identical on an income statement for exactly as long as the shortage lasts.

Profit comparison

Three years ago, Micron sold memory for less than it cost to make it. Gross margin was negative. Not thin — negative. Today it is 84.9%. Nothing about the company's factories, patents, engineers or customer relationships swung ninety-four percentage points in eleven quarters. The price of a commodity did.

In a business selling a commodity-type product, it's impossible to be a lot smarter than your dumbest competitor. — Warren Buffett, 1990 letter to Berkshire Hathaway shareholders

He was writing about airlines. The sentence has never needed rewriting for memory.

The tell

Which brings us back to the contracts, and to the detail almost every write-up has skipped.

Micron's prepared remarks state that its largest agreements “generally have a ceiling price for existing products at the current CQ2 market price, and a floor price through the term of the agreement.” When all the planned deals are done, agreements with fixed prices or ceilings at or close to today's prices are expected to cover about 40% of company revenue.

Read that again. At the peak of the most extreme pricing environment the memory industry has ever seen, with a competitor's chief executive on the record saying the shortage will outlast the decade, Micron has agreed to cap what it can charge its largest customers — at today's price — until the end of 2030.

You do not sell five years of upside cheaply in a market you believe is going to stay this good. You sell it because you want the floor, and the floor has to be paid for. The ceiling is the premium. What Micron has bought is insurance, and the price of the policy tells you what the company privately thinks the risk is.

There is a version of the bull case that says: the floor is worth more than the ceiling costs, so this is a good trade. That may well be true. But notice what it concedes. A business that needs to contract away half a decade of pricing in order to make its earnings “durable and predictable” is a business conceding that, left alone, its earnings are neither. Nvidia has not done this. TSMC has not done this. ASML has not done this. The companies with moats do not have to buy stability. They have it.

Three further things sit underneath the headline:

The coverage is partial. The sixteen agreements represent roughly 20% of Micron's DRAM volume and about a third of its NAND volume over the term. DRAM is 76% of revenue. So four-fifths of the bits in the great majority of the business remain exactly what they were: a spot commodity, priced daily, by everyone.

Take-or-pay is binding, and renegotiable. Reuters, reporting the deals, noted that the memory industry has tried long-term contracts before and that past attempts failed, because memory was a commodity and buyers could always swap suppliers and squeeze. Ben Barringer of Quilter Cheviot put the risk to Reuters directly: “The bear case is that these contracts only hold while supply remains tight. If demand softens and the market turns, there is a risk they are renegotiated or abandoned.” A customer holding a five-year obligation to buy chips at 2026 ceiling prices, in a market where 2029 spot is a third of that, does not sue. It calls. And a supplier that wants to keep Nvidia as a partner for the next decade takes the call.

And the $22 billion is not earnings. Micron's CFO was careful, even if the coverage was not: approximately $18 billion of it is customer deposits, they appear in financing cash flow rather than free cash flow, and they “will be returned to customers over time, towards the latter half of the agreement term.” It is a genuine show of commitment. It is also, in accounting terms, closer to a loan from your customer than a payment by one.

What it costs

The question is not the multiple. It is which E is real.

Micron trades on roughly 22 times its trailing earnings and roughly 6.5 times fiscal 2027 consensus. Both numbers are true. They are answers to different questions.

Here is the arithmetic in plain terms. At $979.30 a share, with $44.31 of earnings over the last twelve months, you are buying about 4.5 cents of annual earnings for every dollar you put in. That is an earnings yield of 4.5%. A ten-year US Treasury paid 4.56% at Friday's close. On the last twelve months of the greatest boom in the history of this industry, Micron pays you no more than a government bond — and it pays you with a share whose price moves more than twice as much as the market's.

Run it on fiscal 2027 consensus instead — $149.64 of earnings per share — and the same $979.30 buys 15.3 cents on the dollar. Now the stock is extraordinarily cheap. (Fiscal 2026 ends in August, so 2027 is the first full year the contracts govern; it is the year the bull case actually rests on.)

Run it on fiscal 2023, when Micron actually lost $5.34 a share, and the yield is negative.

Same company. Same price. Three answers, spanning everything from a wipeout to a gift. The multiple is not the question. Which E is real is the question, and no amount of staring at a P/E ratio will answer it.

It is worth doing the arithmetic properly, because the spread is the finding.

Every number here is defensible. That is the problem.

Fifty-four dollars to two thousand six hundred and ninety-four. That is not a failure to analyze. That is the analysis. Every one of those numbers is defensible, each rests on a different assumption about the durability of Micron's earnings, and nothing in the historical record tells you which assumption is right. Anyone publishing a fair value to the dollar on this company is not being more rigorous than that range. They are simply concealing which assumption is carrying the weight.

So let us do the arithmetic properly, on every basis a reasonable person might choose, at multiples drawn from Micron's own history rather than invented for the occasion.

Table valuing Micron on five-year-average, last-year, trailing, and forward EPS at a range of multiples — a spread from roughly $29 to $2,694 a share..

Look at where the current price sits. At $979.30, Micron is defensible in exactly one box: fiscal 2027 consensus earnings, at a modest eight times. Every other cell in that table says the shares are worth a fraction of what they cost. The entire investment case rests on one number — an analyst forecast, for a year that has not happened, in a business whose earnings have swung from minus $5.34 to plus $44.31 a share inside three years.

That is worth stating plainly, because it is the thing the bull case never quite says out loud. You are not buying a cheap stock. You are buying a forecast.

And this is where the published analysis gets genuinely careless. It is common to see Micron valued on a “normalized” or “mid-cycle” earnings figure of around $63 a share — a sound method, correctly applied to a cyclical business. But Micron's actual five-year average earnings are $3.61 a share. A mid-cycle figure seventeen times the real mid-cycle is not a normalization. It is the peak, wearing a normalized label. Apply a through-the-cycle multiple to it and you have built the single most expensive error in cyclical investing: you have assumed the cycle away, using the machinery designed to account for it.

Peter Lynch's rule for cyclicals, in One Up on Wall Street, was that a low price-to-earnings ratio is not a bargain signal but a warning: it means the market is applying a small multiple to earnings it does not expect to last. Memory stocks look cheapest at the top and dearest at the bottom, and they do so with a regularity that ought to embarrass anyone who screens on the ratio.

The professionals covering the stock know this, and their published numbers say so more clearly than their ratings do. Forty-five analysts carry an average price target of $1,486 on Micron. The lowest is $361. The highest is $2,200. That is a six-fold spread on a single, highly liquid, exhaustively covered large-cap. It is not a forecast. It is an admission that the range of outcomes here is enormous, published underneath a consensus rating of “Strong Buy.”

The market itself is less sure than the ratings suggest. Micron touched $1,255 on 25 June, the day after those results. It closed Friday at $979.30 — about 22% lower, in a fortnight, while the SK Hynix listing was being celebrated.

The cycle, in one table.

What would make me wrong

I want to state this properly, because a bear case that cannot describe its own defeat is just a mood.

If AI demand growth is structural rather than a build-out spike, and if supply really cannot respond in scale before 2029, then the floor prices hold, the contracts are honored, and Micron earns something like $130–150 a share for several consecutive years. At $979.30, that is under seven times earnings. You could halve those earnings and still be paying about thirteen times for a business with net cash, an investment-grade rating and the best product cycle it has ever had. That is not an expensive stock. That is a cheap one, and I would have missed it.

The HBM point is also real, and it is the strongest thing the bulls have. If high-bandwidth memory keeps taking share of the DRAM mix, the blended business becomes progressively less commodity-like over time, because HBM genuinely is qualified, co-designed and difficult to swap. A slow migration from commodity to component is exactly how a business stops being a commodity business.

And the balance sheet means the question is about return, not survival. Micron holds $26 billion of cash against $6.4 billion of debt, was upgraded to BBB+ this year by all three agencies, and carries a current ratio of 3.4. Whatever the next few years do to its earnings, Micron will be there at the end of them. That has not always been true of memory companies.

The bear case is not that Micron is a bad company. It is that Micron is a cyclical company being valued, and increasingly narrated, as though it has stopped being one — on the strength of contracts whose own structure suggests management is not certain of that either.

The verdict

The two questions.

Wonderful? No. A wonderful business does not have a gross margin that travels from minus 9.1% to plus 84.9% in three years without changing anything it does. It does not generate a record quarter almost entirely from price while shipping the same number of bits. And it does not need to sign away five years of pricing to persuade the market that its earnings will last. Micron is a well-run, well-financed, technically excellent company operating in an industry that has never permitted anyone to be wonderful for long. Those are different things, and the difference is the whole discipline. At a fair price? Not knowably — which, for the purposes of committing capital, has to be treated as a no. The price is defensible on one set of earnings, demanding on another, and catastrophic on a third, and I have no honest basis for choosing between them. The professionals do not either; that is what a $361-to-$2,200 range of targets means. A margin of safety cannot be built on an earnings figure whose durability has not yet been demonstrated.

The call: avoid. Not short — the supply constraints are real, the shortage may well run for years, and betting against a commodity in a genuine squeeze is a good way to be right eventually and broke first. Simply avoid. There is no way to size this position such that I am comfortable both if Kwak Noh-Jung is right and if he is wrong, and when that is true of a stock, the answer is to own something else.

This has a cost, and I will name it. If the bulls are right, I will watch a very large gain go past. I would rather do that than pretend to an ability to forecast the price of DRAM in 2029, because that — beneath the AI narrative, the record listing and the take-or-pay contracts — is the only forecast that actually matters here.

What would change my mind

A verdict that cannot be falsified is not analysis; it is a mood. So here is what I am watching, and the number that would move me.

Checkable. I will report on these, including when they go against me.

What this means for you

If you own Micron, the useful question is not “should I sell?” but “which number am I relying on?” If your reason for holding is that it trades on seven times earnings, you are relying on fiscal 2027 consensus. Write that number down. Then ask what happens to your thesis when Micron's Idaho fab starts producing wafers in mid-2027, SK Hynix's $26.5 billion of new Korean capacity lands behind it, and Samsung's does too.

More generally, the transferable lesson is about insurance. When a company starts paying to make its own earnings more predictable, that is information. Ask what the policy costs, and ask what it implies the company believes about the risk. Micron paid for its floor with a ceiling on the best prices it has ever seen. That was a rational trade — and it is the clearest statement anyone has made this year about how long these prices are expected to last.

Questions a reasonable person would ask

If the shortage is real, why does it matter that it is a commodity?

Because a shortage sets the price, and the price sets the profit. When the shortage ends — through new fabs, or a pause in AI spending, or both — the price falls to whatever it costs the least efficient producer to make a bit, and so does the profit. In a business with a moat, the profit survives the shortage. In memory, the profit is the shortage. What is a “take-or-pay” contract with a floor and a ceiling?

The customer promises to buy a set volume each year or pay for it anyway. The floor is the lowest price Micron can be pushed to; the ceiling is the highest it can charge. In plain terms, Micron has agreed not to profit from prices going higher, in exchange for protection if they collapse. It has swapped a lottery ticket for an insurance policy.

Micron trades at under seven times earnings. Isn't that cheap by definition?

Only if the earnings are durable. A price-to-earnings ratio is a price divided by a number that may or may not recur. For a cyclical business at the top of its cycle, that number is at its most flattering and its least repeatable, which is why memory stocks routinely look cheapest just before they fall. The low multiple is the market telling you it does not believe the E.

Could Micron simply be a different company now?

It could. If high-bandwidth memory keeps growing as a share of the mix, and if it stays genuinely hard to second-source, the business becomes structurally less commodity-like over time. That is the strongest bull argument, and it is testable: watch HBM's share of DRAM revenue, and watch whether the price floors in the contracts are ever actually tested. We will know a great deal more by 2028. The point of not owning it today is that you can still buy it then.

This article is general commentary and information, not financial advice, and does not take account of your objectives, financial situation or needs. It is not a recommendation to buy or sell any security. Do your own research or speak to a licensed adviser. The author holds no position in Micron, SK Hynix, Samsung Electronics or Nvidia. Prices and facts change; this piece will not be updated.

Related reading

• What Buffett Meant by “A Wonderful Company at a Fair Price” — the cornerstone framework this analysis applies.

• The Sky Is No Longer the Limit: Is SpaceX Stock Wonderful at a Fair Price? — the same two questions, a different answer.

• The Fed Under Kevin Warsh — why the discount rate you compare this against is going up, not down.

Stay with it

If this was useful, subscribe — free, one considered piece at a time, no hype and no price targets I cannot defend.

The full research archive lives at wonderfulatafairprice.com.