Micron's Memory Squeeze Runs Deeper Than Anyone Expected

The 2027 peak narrative just died.

Micron reported its fourth-quarter results and instantly rewrote the memory market’s timeline. Operating profit hit $43.75 billion on $54.23 billion in revenue, delivering an operating margin of 80.7 percent — up 48.4 percentage points from a year ago. Full-year revenue reached $133.19 billion with operating profit of $99.34 billion. Both are company records. But the numbers that matter most aren’t in the quarterly box; they’re in what CEO Sanjay Mehrotra said about the next two years.

The market has been pricing memory stocks on the assumption that 2027 marks the cyclical peak and prices roll over after. Mehrotra dismissed that directly. Demand will exceed supply through 2027 and 2028, he said, adding that it is unclear when equilibrium will return. More than 75 percent of Micron’s FY2027 shipment volume is already contractually reserved. Discussions with customers have moved past that into 2028 allocations.

This is not a cautious forecast dressed up as optimism. It is a supply-constrained market talking about locked-in revenue at prices that have surged dramatically.

Why the shortage runs deeper than expected

The core bottleneck is HBM — high-bandwidth memory — and the structural trade-offs it creates for everything else.

HBM production consumes significantly more wafers than conventional DRAM per equivalent bit output. As Micron shifts its product mix from HBM3E to HBM4 and HBM4E, the wafer intensity increases further. Every additional HBM unit displaces capacity that could have gone to standard server and desktop DRAM. The net effect is a constriction across the entire memory spectrum, not just in the premium segment where HBM lives.

Micron expects world DRAM bit shipment growth to slow from the mid-20 percent range this year to the early-20 percent range in both 2027 and 2028. NAND should grow faster at roughly the mid-20s annually, but the company still considers even that insufficient to meet demand. The constraint is physical: fabs cannot be built fast enough, and the construction timeline alone runs well beyond 2027.

Micron plans roughly $25 billion in capital expenditure in the first half of FY2027 after subtracting expected government subsidies, with a larger second half. More than half of the incremental spend goes toward new fab construction. The cleanrooms from that spending will not be operational until late 2028 at the earliest. “Building fabs does not instantly translate into bit production,” CFO Murphy said.

The long-term contract revolution

Perhaps the most structural shift is in how memory is now sold. Micron’s strategic customer agreements have grown from 16 to 26 in a single quarter. These contracts cover more than 35 percent of expected revenue through 2030, with some extending to 2031. They use take-or-pay terms — customers must either take the reserved volume or pay for it.

The remaining performance obligation calculated at minimum contracted pricing stands at approximately $150 billion. Financial agreements including customer deposits have risen to $32 billion. Mehrotra described this as a fundamental change in the industry’s character, replacing the old boom-and-bust cycle with multi-year revenue visibility that makes investment planning far more predictable.

Contract prices themselves are trending higher. Negotiations now reflect current market conditions and future price expectations on both sides — customers locking in supply, suppliers capturing upside.

What this means for Samsung and SK Hynix

Micron’s results land at a moment when Samsung and SK Hynix are about to report their own third-quarter earnings. The South Korean giants share the same supply constraints and the same HBM-driven demand surge, but their financial exposure differs in important ways.

The critical variable for both is HBM contract pricing. This year’s HBM prices were locked in during 2024 negotiations. Micron confirmed that its 2027 HBM pricing was significantly elevated and that new contract rates will take effect from early next year, narrowing the profitability gap between HBM and conventional DRAM. Samsung and SK Hynix face the same renegotiation calendar. How much of the current spot-market premium gets reflected in their forward contracts will determine whether their margin expansion outpaces Micron’s or falls short.

Standard DRAM and NAND pricing dynamics also remain favorable. Micron’s Q4 average selling prices rose roughly 10 to 20 percent quarter-on-quarter across both product categories while shipment volumes grew only in the single-digit to low-double-digit range. Price is still doing the heavy lifting. That pattern is likely to hold for its Korean competitors through at least the first half of next year.

Nvidia’s memory compression paradox

Reports that Nvidia is reducing HBM capacity per GPU in its next-generation Rubin Ultra design — and cutting LPDDR5X memory on CPU modules by half — could look like a demand warning. Micron’s COO Manish Pandita does not read it that way.

His interpretation is the opposite: customers are maximizing the number of accelerators they can ship given constrained memory supply. If memory availability relaxes, those per-chip allocations can expand. The net effect on total demand is neutral to positive because more GPUs still get built. Either way, the industry is not seeing a demand cliff — it is seeing a reallocation of scarce memory across a growing number of AI systems.

Pandita also flagged a secondary demand driver that most Western coverage has underweighted. Agentic AI — systems that perform tasks on behalf of users rather than simply responding to prompts — is increasing memory requirements per server across CPU workloads, not just GPU workloads. Lower-power DRAM, standard DDR, and SSD all see higher per-unit consumption. Micron projects server shipments to grow in the high-single to low-double-digit range over the next two years, driven partly by this shift.

What happens next

The immediate implication is earnings revision risk for Samsung and SK Hynix. If Micron’s 80 percent margin holds as a floor rather than a peak, Korean memory stocks have not yet repriced for the scenario they are now describing. The longer implication is structural: memory is no longer a commodity cycle that resets every 18 to 24 months. Multi-year supply contracts, construction timelines that run past 2028, and AI-driven demand that shows no signs of saturating are compressing the downside window.

For the broader AI hardware supply chain, the takeaway is simpler. Memory remains the binding constraint. Every new data center build, every GPU upgrade cycle, every agentic AI deployment pushes against the same finite wafer capacity. The shortage is not a temporary dislocation. It is the new operating environment — and the companies that understand that will price accordingly.