Micron: the memory 'peak' question now has two answers

Micron has climbed more than seven times off its 52-week low to sit around $945, and it has retraced about a quarter from its $1,255 high. That pullback is where the debate is: is this the top of the memory supercycle, the way every memory cycle before has ended? For three decades the answer has been a reliable yes — prices boom, the makers overbuild, supply floods back in, prices crash, and the whole thing resets. Memory is the most dependable cycle in the semiconductor business, and most of the anxiety about Micron at a $1.07 trillion market cap comes from that history.

The problem with that frame is that it treats memory as one product. The numbers from Micron's latest quarter suggest the industry has quietly split into two economically different businesses — and the suppliers are behaving very differently in each.

This isn't a demand boom. It's a price story.

Micron's third quarter of fiscal 2026 (ended May 28) brought in $41.46 billion in revenue — up 346% from a year earlier and up 74% in a single quarter — on gross margins of 84.6% under GAAP accounting, versus 37.7% a year ago. No memory maker has sustained gross margins anywhere near that level. The instinct is to read that as a demand explosion. The volume data says otherwise.

From Micron's reported results, the split is stark:

	DRAM	NAND
Q3 FY2026 revenue	$31.3B (76% of total)	$9.9B (24% of total)
Revenue vs. a year earlier	+343%	+361%
Bit shipments (volume)	low-single-digit growth	mid-single-digit growth
Average selling price	up in the low-60s% range	up in the mid-80s% range

DRAM bit shipments — the actual number of memory chips sold — rose only in the low single digits, and NAND in the mid single digits. Nearly all of the 346% revenue jump came from prices, not volume. In plain terms, Micron sold roughly the same number of bits as it did a year earlier and charged far more for each one.

That is the signature of a supply-constrained market, not a demand surge. The constraint is reallocation: advanced AI memory called HBM — the high-bandwidth chips stacked directly onto the biggest AI accelerators — is eating the wafer capacity that used to make ordinary memory. One HBM wafer displaces two or more conventional DRAM wafers, so even holding total output flat, the industry produces fewer commodity bits. And the new supply that would fix it isn't close: the large new fabs Micron, SK Hynix, and Samsung are building have two- to three-year lead times, with meaningful volume arriving around 2028. Micron's CEO has said the company has no "line of sight" on when supply will catch up to demand, and that data-center customers want roughly 50% more memory than it can ship.

The bigger change isn't the price. It's the contract.

Prices explain one record quarter. They don't explain the structure of what Micron is building. Alongside the results, the company signed 16 multi-year "take-or-pay" agreements — contracts in which a customer commits to a minimum volume of purchases and pays even if it doesn't take the chips. Fourteen of them represent about $100 billion in committed minimum revenue, running through 2030, backed by roughly $22 billion in customer deposits and letters of credit. Management says it expects half or more of future revenue to sit under agreements like these.

This is the part that doesn't fit the "peak is near" script. In the commodity half of memory, the supplier is a price-taker: it builds into a shortage and eats the crash when supply returns. Under a take-or-pay contract with a price floor, that asymmetry is removed. The floor keeps gross margins above historical peaks through the weak half of any future cycle. That is not a promise that prices never fall — the agreements also carry price ceilings at current levels for a meaningful share of revenue, capping the upside — but it does change who bears the downside risk.

The two-market split is now visible in the P&L. The strategic segment — HBM and server memory — is capacity-constrained, high-margin (its Core Data Center unit ran at 87% gross margin in the quarter), and increasingly contract-protected. The commodity segment — the memory in laptops and phones — is still a price-taker and still exposed to the classic cycle. "Is the peak near?" gets two different answers depending on which half of the company you are valuing.

The market is still pricing one of them.

Micron carries a $1.07 trillion market cap at about 21 times trailing earnings, with the next quarter guided to $50 billion in revenue and roughly $31 per share — up from the prior quarter, not down. That multiple sits far below the AI infrastructure names that are the customers buying its chips: about 60 times for Broadcom and 116 for AMD. The gap is the whole debate compressed into a number. The market is not paying for these earnings as if they are durable; it is paying for them as if they are the top of a cycle.

So the question an investor actually needs to track isn't "when is the peak." It is whether the big 2028 wave of new wafer capacity — Samsung adding HBM capacity this year, and the new Micron, SK Hynix, and Samsung greenfield fabs reaching volume in 2028 — lands into a market that has also contracted. If AI demand holds and the $100 billion of contracts keeps filling in, the floor holds and a 21 multiple looks like a cyclical trap. If AI capex cools exactly as the new capacity comes online, the ceiling and the floor meet at the downside, and the commodity half re-prices the way it always has. The contracts de-risk the crash; they don't remove the clock.