The Night Before Micron
The Night Before Micron
It's the night before Micron reports and I'm doing what I always do the night before a memory company reports a potentially cycle-defining number: sitting here reciting the old catechism to myself, which goes something like: cycles end, cycles always end, this is a cycle, therefore it ends. Simple syllogism. Worked in 2018. Worked in 2022 when MU fell 65% from peak to trough. Worked every time someone showed up at a conference with a new framework for why the DRAM supply-demand relationship had been permanently altered by some demand shock that would dwarf all historical precedent.
They're saying it again. Louder this time.
The numbers coming out tomorrow are going to be — there is no other word for it — insane. Consensus is $34.5 billion in quarterly revenue and roughly $19.72 in earnings per share. For one quarter. A company that generated $37 billion in revenue for its entire fiscal 2025 is now being asked to print $34 billion in three months. Gross margin guidance is 81%. Eighty-one. A DRAM manufacturer. The same category of business that investors spent thirty years treating like a slag heap — lumpy, capital-intensive, savage when the cycle turns, impossible to value — is apparently now a premium infrastructure franchise worthy of a multiple somewhere between Nvidia and God.
The stock is up somewhere in the neighborhood of 360% this year. It crossed a trillion dollar market cap in late May. It trades around $1,230 today. UBS, Deutsche Bank, Cantor Fitzgerald, TD Cowen — they have all slapped $1,500 targets on the name, which is above where it's currently trading, which means this is the rare situation where the sell-side is still playing catch-up to a stock they helped pump. Cantor Fitzgerald more than doubled its target from $700 to $1,500, calling this AI-driven memory cycle early to middle innings.
Early innings. Let me sit with that.
Here is what is genuinely different and I want to be honest about it because my instinct is always to be the person who spots the top, and being the person who spots the top is expensive when you're three years early.
What's different is this: HBM is not DRAM. High-bandwidth memory stacks dies vertically, connects them through microscopic silicon vias, and delivers bandwidth that standard DDR architecture cannot approach. Every AI accelerator Nvidia ships — Blackwell, Vera Rubin — requires HBM. Not prefers. Requires. Micron's entire 2026 HBM supply is fully contracted. Goldman Sachs pegged the 2026 DRAM supply-demand gap at 4.9%, the most severe shortage in 15 years. Sanjay Mehrotra has said publicly that Micron can fill only half to two-thirds of key customer demand in the medium term. They are not sandbagging. There are only three companies on the planet that make HBM at scale: Micron, SK Hynix, and Samsung. That is the entire market. The hyperscalers are locked in. There is no spot market to turn to. The capacity simply does not exist.
So the bull is not irrational. The bull is saying: this is not oversupply masquerading as cycle; this is a genuine oligopoly running hard against a physical constraint. The pricing reflects scarcity, not sentiment. DRAM average selling prices were up mid-60% sequentially in the most recent quarter, NAND up high-70%. Those are numbers you see once in a generation of the commodity cycle, if that.
And then the multi-year contracts. Micron has begun shifting from spot-price memory sales to multi-year supply agreements at pre-agreed pricing. That is not a commodity business model. That is infrastructure pricing. Semiconductor infrastructure, locked pricing, $725 billion in annual hyperscaler capex underpinning demand — the re-rating argument writes itself.
But here is the thing that keeps me up.
New fabs come online in 2027. Micron's own new York campus. SK Hynix's expansion. Samsung's recovery — they were briefly behind on HBM yields but they will catch up because Samsung has never, in fifty years of semiconductor history, stayed behind. If AI model efficiency improves faster than compute demand grows, the demand side softens just as new supply peaks. That is the classic memory bust scenario, dressed up in AI language. The cycle. The same one. Wearing a different coat.
And the valuation already prices zero probability that this happens. Zero. If you buy MU at $1,230, you are buying a company that the market has decided will never again face a normal downcycle. You are paying, effectively, for the permanent repeal of a thirty-year industrial pattern because the demand driver this time is machines that need memory the way people need oxygen. Maybe that is true. The hyperscalers have not blinked on capex. Big tech has signalled that AI capital expenditure is set to rise past $700 billion in 2026 from $400 billion in 2025, and Micron's results will either validate or complicate that trajectory. Every server that gets built pulls memory demand with it. The number is compounding faster than capacity.
The problem is not the demand. The problem is that every great memory bull market has carried its own version of this argument, and none of them repealed the cycle. The 2021 bull market said cloud adoption was so structural that DRAM demand was now secular. Correct, and yet: 65% drawdown in 2022 when supply caught up.
Tomorrow's number almost doesn't matter, which is the most unsettling thing I can say about it. If Micron beats by the expected 15-20% — which the whisper numbers suggest — the stock probably goes up another 5%. If they miss, it craters. But either way, the structural question doesn't get answered tomorrow. It gets answered in 2027 and 2028 when the new capacity ramps, when we find out whether AI model efficiency curves ate into per-server memory demand, when we find out whether three oligopolists all running maximum capacity simultaneously can avoid the prisoner's dilemma that has destroyed every memory cycle before this one.
The bear case for MU is not that the AI boom is fake. The bear case is that three companies will build enough capacity to satisfy a boom that may or may not plateau at the same moment the new fabs come online, and that when the price of memory falls, it falls fast, and that the stock at $1,230 prices none of that probability.
The bull case is simpler: the world is building an infrastructure that needs HBM the way power grids need copper, and there is no substitute.
Both of those things can be true at the same time. They just can't both be true at the same valuation.
I'll be watching the HBM revenue mix, the forward ASP commentary, and — the number that almost nobody is tracking loudly enough — the first sign of any customer pushing back on multi-year pricing. That's the canary. Not the earnings. The contracts.
Ninety-one minutes until the Asian session opens.