Micron just tried to delete the memory cycle

Micron Technology ($MU) reported Wednesday night, and the headline numbers are the kind you read twice. Revenue of $41.5 billion, up 346% from a year ago. A gross margin of 84.9%, a company record. A guide for next quarter of $50 billion at roughly 86% margins, well above what Wall Street expected. The stock jumped double digits after hours.

I want to walk you past those numbers, because the records are not the story. A memory company printing a cyclical peak is the most ordinary thing in the world. It has happened in every up-cycle for thirty years, and every time it has been followed by a glut, a price collapse, and years of nursing losses. If that is all this was, it would be a great quarter and a familiar one.

The story is what Micron signed alongside the numbers: sixteen long-term, take-or-pay supply contracts, with floor prices and customer cash put down up front, built to lock in the good part of the cycle before the bad part arrives. If those contracts hold up the way management describes them, then the thing I have always told you to respect about memory, that it is brutally cyclical, may be quietly turning into something structural. That is the shift worth your time.

This is part one of two. Today I lay out the machine that could break the cycle and the bear case against it. In part two I walk through the demand engines underneath it, including a robotics angle I have been waiting to see show up in a memory call, and how I read the whole thing for the portfolio.

Set the records down as context, then move on. Data center revenue alone crossed $25 billion in the quarter, an annualized run rate above $100 billion, with data center SSD revenue topping $5 billion and more than doubling from the prior quarter. All four of Micron's business units hit records, including the automotive unit, which quadrupled from a year ago as carmakers pulled orders forward ahead of the shortage. Free cash flow was a record $18.3 billion, and net debt is essentially gone.

These are the symptoms of a tight market, and tight markets always look like this near the top. The question that matters for a long-term holder is not how good the peak is. It is whether the next trough is as far down as the last one. That is what the contracts are designed to answer.

Micron signed sixteen strategic customer agreements, or SCAs, with four very large customers, three medium-sized ones, and a set of smaller automotive buyers. They run five years, calendar 2026 through 2030. They cover roughly 20% of Micron's DRAM volume and about a third of its NAND, and management expects that to climb toward half or more of total company revenue as the rest get signed.

Three features make these different from the loose, quarter-to-quarter handshakes memory has always run on.

First, they are take-or-pay. Think of it like a gym membership you signed for five years: whether or not you show up, you owe the monthly fee. The customer is contractually obligated to buy the committed volume, or pay for it anyway. On the call, Micron's chief business officer was unusually blunt that there is no provision for a customer to walk away.

Second, the largest deals have a price floor and a price ceiling. The price still gets negotiated each quarter inside that band, but it cannot fall below the floor no matter how soft the market gets. And on the contracts that carry these bands, the floor, in management's words, is set to deliver a gross margin "well above our peak quarterly margins in any past cycle." Not every SCA has a band; a modest slice are fixed-price or move with the market, so this floor applies to the larger banded agreements rather than all sixteen. Read that slowly. The worst case inside these contracts is engineered to beat the best case of every prior cycle. That is the sentence that turns a cyclical business into something closer to a structural one.

Third, the customers put cash down. Across the sixteen signed deals, Micron expects $22 billion in deposits and related commitments, roughly $18 billion of it in actual cash. That is customers posting collateral on their own demand. You do not wire eighteen billion dollars to a supplier you expect to abandon.

One honest accounting note so you read the headline correctly. Micron disclosed about $100 billion of "remaining performance obligations" tied to fourteen of these agreements. That figure is calculated at minimum committed volumes and minimum prices, so management was explicit that it is a deliberately conservative floor, not a revenue forecast, and that actual revenue should run well above it. I would treat the $100 billion the way management framed it: as the contractual floor under the next several years, not a target.

If contract mechanics make your eyes glaze, here is the version that matters, and you do not need a model for it. What these agreements change is the answer to two questions: how bad can it get, and how long does the good part last.

I compared an SCA above to a five-year gym membership you pay whether or not you show up. Stretch that one more step. The old memory business was a month-to-month gym, packed in good years and empty in bad ones, with the owner riding every boom and bust. The new one is a five-year corporate contract with a minimum monthly fee and a clause that says the company pays even if its employees never walk in. The owner can borrow and invest against that, because the revenue no longer vanishes the moment the mood turns.

Three things do the work. Volume: think of roughly 20% of Micron's DRAM and a third of its NAND as reserved seats in every quarter from 2026 through 2030, a share management expects to climb toward half of revenue. Price: on the contracts that carry bands, the floor is set to hold margins above any prior peak, so the worst case is built to beat the old best case. Cash: customers wired Micron close to $18 billion up front, which is not how buyers behave when they plan to walk.

So the practical read is this. Treat the $100 billion as a floor under the next few years, not a forecast of the peak. Assume the next downturn cuts shallower than the last one rather than wiping margins out the way past busts did. And watch behavior over words: if customers keep honoring these contracts and keep signing more through the first soft patch, the structural read earns more trust, and if they start trying to renegotiate, move Micron back toward cyclical in your head.

The contracts do more than steady pricing. They hand Micron a wall of cash and a reason to trust it. Free cash flow was a record $18.3 billion this quarter, and on the call one analyst walked through the math that next year's free cash flow could be large enough to buy back close to a tenth of the company, framing it around a figure north of $30 billion. Management did not endorse that specific number, so treat it as the analyst's math rather than guidance, but it did commit to returning 100% of excess cash to shareholders over time, mostly through buybacks, with the pace stepping up after December 9, the second anniversary of its CHIPS Act agreement. I would read that "tenth of the company" figure as a stress test of what this cash engine could do, not a base case: the only firm commitment is to return 100% of excess cash over time and to raise the return rate after December 9, not any fixed buyback number. The $18 billion of upfront customer cash matters here too, because it helps fund the buildout without forcing Micron to choose between investing in new fabs and returning capital. That is the financial version of the same structural story: a company that used to hoard cash to survive the next downturn is now talking about handing it back.

The bear reflex on any memory peak is correct by default: high prices invite new supply, new supply becomes a glut, the glut kills the price. So the real question is whether supply can respond fast this time. Management's answer, and the physics behind it, is that it cannot.

New memory capacity now requires greenfield fabs, entirely new plants built from bare ground, rather than the cheaper trick of upgrading an existing one. Those plants are slow. Micron's Idaho and Taiwan fabs do not produce meaningful output until calendar 2028, and even then management said it has no line of sight on when supply catches demand. Two structural forces make it harder. High-bandwidth memory uses far more wafer capacity per gigabyte than standard memory, by common industry estimates roughly three times, so every wafer that shifts to AI memory is a wafer pulled out of the consumer pool. And each new manufacturing node yields fewer additional bits than the last, so the old habit of squeezing more supply out of the same factory is fading.

The cleanest tell came when an analyst asked for Micron's long-term bit-demand growth rate, the single most important number for modeling the industry. Management declined, and the reason was revealing: shipment growth, they said, is "not really determined by demand anymore. It's actually more determined by the supply," because demand sits so far above what anyone can make. When a memory company stops forecasting demand because supply is the only thing that matters, that is the cycle inverting.

I would be doing you a disservice if I only made the bull case, so here is the other side, stated as fairly as I can.

The oldest objection is the strongest: the spending that creates today's shortage is the same spending that has always created tomorrow's glut. Some analysts, including at TechInsights, are already modeling an industry downturn by 2027, and others at Morningstar flagged the whole memory group as overvalued earlier this spring. The SCAs are Micron's answer, but they are unproven through a real downturn. Take-or-pay contracts are only as good as the customers' willingness to honor them when prices fall and lawyers get involved, and we have not seen that tested yet.

There is also a subtler cost to the contracts, and an analyst put his finger on it on the call. By locking customers into price bands with ceilings, Micron may be capping its own upside in exactly the moments when buyers are most desperate and would pay almost anything. Management's defense was essentially that durability is worth more than the occasional spike, which I find reasonable, but it is a real trade-off, not a free lunch. You are trading the top of the cycle for a higher floor.

And the structural supply story cuts both ways. China's CXMT and YMTC keep climbing the technology ladder. Most of their output stays inside China today, so Micron says it sees little of them in the open market, but they are exactly the kind of supply that could eventually answer the shortage on a timeline no one is modeling. The whole thesis rests on three disciplined makers staying three disciplined makers.

None of this dismisses what happened Wednesday. It is the set of things that have to stay true for the structural read to hold, and I will be watching every one of them.

Micron is a disclosed holding for me (almost as long as my NVDA position), and I’ve traded this stock along the way. A great quarter is not a reason to chase a stock that has already run hard this year, and I am not going to tell you what it is worth, because that is your call, not mine. What I will say is that the cyclical-versus-structural question is the one that should drive how you think about memory, and Wednesday moved the evidence toward structural.

In part two I get into the demand engines that have to keep showing up for any of this to hold, the robotics angle that gives the story a second decade, and the supply-chain names I watch one layer down from Micron. If the move from cyclical to structural is real, the companies that feed Micron get pulled along with it.

Stay disciplined - Koh

Disclosure: I have a position in $MU.

Disclaimer: Nothing in this newsletter constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed above. Always do your own research and consult a licensed financial advisor before making investment decisions.