Micron Sold Most of Next Year's Memory Before It Was Built. The Real Story Is Why 2028 Won't Bring Relief.

The numbers that got the headlines

Micron's fiscal fourth quarter didn't just beat guidance. It rewrote what a memory company looks like. Quarterly revenue hit $54.25 billion, against $11.3 billion a year earlier. Full-year revenue landed at $133.2 billion, up from $37.4 billion. Net income rose 895 percent to $85 billion. DRAM revenue for the year crossed $100 billion for the first time.

The margins are the part that stops you. Micron's cloud memory unit, which sells HBM, posted a fourth-quarter gross margin of 83 percent, up from 59 percent a year earlier. The core data center unit, which sells more conventional DRAM, hit 90 percent, up from 41 percent. A business that was fighting for its life two years ago is now more profitable than most software companies. Data center SSD revenue reached nearly $10 billion for the quarter, more than 10 times the year-ago figure.

Those are the numbers that moved the stock. They are not the ones that should move your supply plan.

Two sentences that matter more

On the earnings call, CEO Sanjay Mehrotra said the quiet part out loud. "In calendar 2027 as well as 2028, we see demand exceeding supply," he told analysts. "In fact, we see greater tightness in the industry in 2027 and in 2028 versus 2026. Overall, supply-demand environment is only getting tighter."

Then came the line that belongs pinned above every hardware procurement desk: "We do not have line of sight to when supply and demand will return to balance."

He also made clear where next year's pricing is heading. Micron has already sold most of the memory it will make next year, and customers will pay "much higher prices" than they paid in 2026, according to The Register's read of the call.

Micron is not alone in this. SK hynix CEO Kwak Noh-jung called 2027 "the most challenging year on the supply side in the industry's history" in an interview with Reuters, as relayed in subsequent coverage of the memory market. Samsung said its HBM capacity for the year was already booked and it was taking advance orders for 2027 output. When three competitors who normally spend their time fighting each other for share all tell the same story, it stops being positioning.

Why new fabs don't fix a shortage quickly

The intuitive fix is capacity. Build more fabs, make more memory, watch prices fall. Micron is doing exactly that. It plans to spend around $25 billion in the first half of fiscal 2027 alone, and its guidance implies full-year capital spending above $40 billion, with most of the increase going to construction.

The catch is timing, and Micron spelled it out. New fabs do not produce at scale when the concrete sets. "Production from new DRAM and NAND fabrication facilities takes time to ramp and gradually becomes more meaningful starting a few quarters after initial output." That is corporate language for a simple fact: memory from a fab that starts output in late 2028 arrives in volume sometime in 2029.

Look at the timeline the company laid out. Its ID1 fab in Idaho starts wafer output in mid-2027, ID2 in late 2028. The Japan DRAM expansion starts initial output in late 2028. The new Singapore NAND facility begins output in the second half of 2028. The New York fab, the one Micron celebrated with a concrete pour last quarter, doesn't expect initial wafer output until calendar 2030.

Then there is the HBM effect. High-bandwidth memory, the kind that sits next to an AI accelerator, consumes far more wafer capacity per bit than ordinary DRAM. Morgan Stanley estimates that for the same effective bit output, HBM's wafer-consumption penalty rises from about 2.5 times today to about 3 times by 2028, and that HBM's share of advanced DRAM wafers climbs from 6 percent in 2023 to 34 percent in 2028. Every wafer that goes to HBM is a wafer that does not become the DDR module in a server or a laptop. AI demand isn't just eating memory. It's eating the capacity that would have made cheaper memory.

Micron's own roadmap makes the same point from the other end. It is working with NVIDIA on the industry's first custom HBM4E implementation, called NV-HBM, for next-generation GPUs and NVLink platforms. That is a premium product competing for the same wafers as the commodity part of the market.

The part that changes how you buy

Here is the structural shift, and it matters more to a hardware company than any single price forecast.

Micron has signed 26 strategic customer agreements, multi-year take-or-pay deals it expects to cover more than 35 percent of its revenue through 2030, with several extending into 2031. Customer financial commitments under those deals have grown to $32 billion, the vast majority of it cash deposits. Three quarters of the contracted revenue has a defined pricing framework, mostly bands with floor and ceiling prices. Management said more than 75 percent of fiscal 2027 output is already committed once those agreements and other customer commitments are counted.

Memory is moving from a commodity you buy on a quarterly cycle to infrastructure you contract years ahead. If you want guaranteed allocation, you sign. If you sign, you accept a price band set in a seller's market.

The other side of the trade

None of this means memory prices stay high forever. It means the case for them staying high is stronger than the case for them falling soon, and both cases are being argued with real numbers.

UBS raised its Micron price target to $1,625 and said the industry stays short of DRAM at least until the second quarter of 2028, with HBM average prices up roughly 76 percent year over year in fiscal 2027. Morgan Stanley points at 2028 from the other direction. In a report covered in late August, it estimated China's CXMT expanding from roughly 180,000 DRAM wafers a month in 2025 to about 300,000 this year, 500,000 by 2028 and 800,000 by 2031. On those numbers, CXMT could pass Micron as the world's third-largest DRAM supplier around 2028.

The constraints on that scenario are real. Export controls keep the most advanced lithography out of Chinese fabs, so CXMT's effective supply share runs below its capacity share, roughly 10 percent against 15 percent, according to Mawer's Shan Rui Yeo. As of a late-August review, CXMT had booked out its output through the end of 2027 anyway. But the underlying economics of memory haven't changed. When capacity arrives at the same moment demand cools, shortages disappear faster than anyone expects. That is the risk sitting on the far side of 2028, and it is the reason memory companies still trade at a few times forward earnings.

Downstream, the squeeze is already visible. Raspberry Pi reported record first-half revenue of $256.9 million, up 90 percent, partly because it stockpiled memory in 2025 before prices moved. CEO Eben Upton said the decision to build that inventory "allowed us to maintain product availability at a time when smaller competitors have struggled to secure allocation." Then the arithmetic caught up. The average cost of memory in Raspberry Pi's inventory went from $3.60 per GB at the end of 2025 to $13.30 per GB at the end of June. Same component, nearly four times the price.

What this means for hardware teams

The lesson isn't that memory is expensive. Hardware teams have lived that cycle before. The lesson is that the mechanism changed. A spot market that cleared every quarter has become a contract market with multi-year commitments, supplier concentration, and a supply response measured in years rather than months. Companies that treated memory as a procurement line item and companies that treated it as a strategic constraint are heading for very different margins through 2027.

If your product's bill of materials depends on DRAM or NAND, the questions worth answering now are not about today's price. They are about allocation. Do you have supply locked beyond your current purchase orders? What happens to your unit economics if you're buying at spot while a competitor is buying under a floor-and-ceiling band? And what does your roadmap look like if memory stays tight, and expensive, through 2028?

The supply chain complexity behind a hardware product in the middle of an AI-driven memory crunch is the kind of problem that doesn't show up in a spec sheet. At DMC, we work with hardware companies navigating exactly these constraints, from sourcing strategy and cost modeling to production ramp planning when demand outruns the fab. If you're stress-testing your hardware roadmap against a memory market that isn't easing, let's talk.