Micron's $100 billion in locked contracts changes the memory cycle math

Micron just printed the most important earnings call in semiconductor history

Sanjay Mehrotra announced a $25 billion data center revenue quarter on Tuesday night. Let that context sit for a moment: a single quarter, a single segment, $25 billion. Annualized, that's a run rate above $100 billion from data centers alone. For a company that, four years ago, was grinding through a commoditized memory downcycle with sub-20% gross margins, this is a different business.

The headline numbers were extraordinary: 84.9% gross margin, $25.11 earnings per share, $41.5 billion in revenue up 346% year-over-year. But the number that actually matters came buried in the prepared remarks: $100 billion in contractual floor-price revenue locked across sixteen strategic customer agreements, with $22 billion in cash deposits already sitting on Micron's (NASDAQ:MU) balance sheet from customers who haven't taken delivery yet.

Think about what that means structurally. Micron's customers, the hyperscalers and AI infrastructure builders, are paying Micron in advance. They are depositing cash to reserve future HBM capacity because they cannot risk being caught short. The floor prices on those contracts are set above any gross margin Micron has ever printed in a prior upcycle. This is not a cyclical memory company anymore. The business model has changed.

Mehrotra has been telegraphing this shift for two years, and the skeptics, including me at points, looked at the history of DRAM cycles and said: it always comes back down. Oversupply arrives, ASPs collapse, margins crater. That is how memory has worked since the 1980s.

Here is why this cycle may be different, and I want to be precise about the word "may" because the history is real. HBM is not standard DRAM. The conversion ratio from standard DRAM wafers to HBM is roughly 3:1, meaning each HBM unit consumes three times the wafer capacity of equivalent standard memory. Every AI accelerator Nvidia ships requires HBM stacks. Every new GPU generation requires more HBM than the last. So as AI accelerator volumes climb, HBM demand climbs faster than anyone initially modeled, while simultaneously consuming supply that would otherwise address the standard DRAM market.

On top of that, Mehrotra named the supply constraints explicitly: greenfield fab lead times measured in years, EUV tool scarcity, HBM packaging complexity, and skilled-trades shortages at the new Idaho and New York fabs. These are not Micron-specific problems. Samsung and SK Hynix face the same constraints. The tightness Micron is pricing into contracts through 2030 reflects industry capacity, not one company's production schedule.

For ASML (NASDAQ:ASML), for advanced packaging players, and for every equipment supplier on Micron's build-out, this call was a demand confirmation worth more than any analyst note. Micron just told the market it is pulling in clean room capacity, that its Singapore packaging facility contributes to HBM volume from the first half of 2027, and that customers are willing to pay deposits years in advance to secure allocation.

The bear case is real and worth stating clearly: if AI capex hits a wall, if hyperscaler spending stalls, if a recession forces IT budget cuts, those SCA floor prices get tested and the deposit structure faces legal scrutiny rather than smooth execution. AI infrastructure spending has never been tested in a genuine downturn.

But if the bear case doesn't materialize in the next 18 months, Micron is sitting on a contractual revenue base that makes it look less like a cyclical memory supplier and more like a toll road on AI compute. That re-rating hasn't fully happened yet.

✦

Quick Takes

Three more stories worth knowing about

Nike's recovery is real in one country, not four

Elliot Hill's Camp Nou metaphor, scaffolding, cranes, limited capacity, was more honest than he probably intended. Nike's (NYSE:NKE) North America wholesale is genuinely turning: 11% growth, February sell-through positive across all channels for the first time in two years. But Greater China is guided down approximately 20% in Q4 and the cleanup runs through all of FY27. EMEA partners are sitting on elevated inventory and running promotions. The Q&A filler rate jumping from 0.1 per minute in prepared remarks to 3.2 per minute when analysts pressed on EMEA told you everything. For Foot Locker (NYSE:FL) and Dick's Sporting Goods (NYSE:DKS), the wholesale rebuild is real in North America. Don't read it across to Europe.

SpaceX is public, and Rocket Lab gets the spillover bid

The largest IPO in history closed this month at a $1.75 trillion SpaceX valuation. The capital chasing space exposure now has very few places to go on public markets. Rocket Lab (NASDAQ:RKLB) is the most direct beneficiary: it is the only other company currently flying an operational orbital rocket. Institutions that want launch exposure and missed the SpaceX offering will find RKLB is the closest available proxy. AST SpaceMobile (NASDAQ:ASTS) and Intuitive Machines (NASDAQ:LUNR) catch the same sector re-rating trade. None of this is fundamental analysis of those companies in isolation. It is straightforward capital flow logic: a landmark public comp reprices the whole category.

General Mills told you the lower-income consumer is still hurting

Jeff Harmoning spent most of the General Mills (NYSE:GIS) call arguing the worst is behind them, but the actual Q4 organic growth guidance shifted from 'positive' to 'improved trends,' with Kofi Bruce admitting the improvement comes from retailer inventory reversals and trade timing, not consumer demand. The detail on Totino's, customers resisting the value-perception change on the box, is a live read-through for any packaged food company selling at a premium to private label. The lower-income consumer is still trading into promotions. The Brazil divestiture is genuinely clean news. Everything else on that call was careful language around a difficult underlying picture.

✦

Overlooked Stock

CLS

Celestica Inc.

Celestica (NYSE:CLS) is a Canadian contract manufacturer that almost nobody outside semiconductor supply chain circles talks about, which is exactly why it scores a 93 here. The company builds the physical infrastructure layer of the AI boom: compute platforms, rack-scale systems, advanced hardware for hyperscalers. It has a strategic collaboration with AMD to develop Helios, a rack-scale AI platform, which puts it directly inside the hardware design cycle rather than just assembling finished specs. After Micron's call this week confirmed multi-year data center build commitments with cash deposits attached, the demand picture for companies sitting one layer up the AI hardware stack looks increasingly durable. Celestica is not a household name. Its market cap is a fraction of the companies it supplies. That gap between role and recognition is where the interesting setups tend to live. Worth pulling the filings.

The Contrarian Take

The SCA structure at Micron makes the old memory cycle thesis obsolete

The standard bear case on Micron has always been the same: memory is a commodity, cycles turn, margins collapse. It has been right many times. The argument is that AI capex is cyclical too, and when it slows, those contract floors get tested.

But the SCA structure changes the math in a way that cycle analysis doesn't capture. Customers depositing $22 billion in cash in advance are not doing that for a commodity they can source elsewhere. They are paying for guaranteed HBM allocation they cannot get from Samsung or SK Hynix fast enough to meet their GPU production schedules. The deposit is an admission of dependency.

The bear case requires believing either that AI accelerator demand collapses before 2030, or that Samsung and SK Hynix ramp HBM capacity fast enough to break the pricing floor despite identical fab lead times and EUV constraints. Both are possible. Neither looks probable on an 18-month horizon. The cycle thesis isn't wrong forever. It just may be wrong long enough to matter.