Micron's Q4 2026 Earnings and the Contract at the Top of the Memory Cycle

Micron reported fiscal Q4 2026 after the close on September 30: $54.2B of revenue, an 87.0% non-GAAP gross margin, $44.0B of operating cash flow in a single quarter, and a guide for $61.5B next quarter. A year earlier the same company did $11.3B at a 45.7% margin. The shares closed the regular session flat and slipped in after-hours trading. SK hynix did the same thing in July, falling around 11% intraday on the best quarter in its history.

That reaction is the market's thesis, and it is a coherent one: memory is a commodity, a commodity at a record margin is at its peak, and the three makers trade as one tape whose only question is when the glut arrives. This article disagrees with the last part of that sentence, not the first. The glut is coming — the capacity is already being built. What the print shows is that the next downturn will not hit the memory makers evenly, and that the variable that sorts them is the contract written at the top, not the technology lead.

The curve flattening at the top is the first thing a cycle bear points to, and it is fair: Q4 added about two points after Q2 and Q3 added ten and eighteen. But the flattening is a ceiling on margin, not on revenue. Micron's pricing rose in the high teens for DRAM and around 30% for NAND in the quarter, bits still grew, and management guided revenue up another $7B while holding margin roughly flat. What changed in the print is less the size of the numbers than what management said is constraining them.

Read down the last three rows. Management said it has no line of sight to supply and demand balancing, that 2027 and 2028 will be tighter than 2026, and that the capex it is spending is going into cleanrooms first and tools second. That is a statement about the industry, not about Micron: the bottleneck on how much memory the world can ship is floor space that takes two to three years to build. The next meaningful wave of new DRAM wafers, from Micron's Idaho and Japan fabs and SK hynix's Yongin cluster, lands between early 2027 and late 2028.

For the last three years the scarcity rent in the AI build sat with the accelerator. NVIDIA set the price, and memory was a line on its bill of materials that three suppliers competed to fill. Micron's print is the clearest evidence yet that the rent has moved. Most of next year's HBM is sold at significantly higher prices, customers are prepaying tens of billions of dollars for allocation, and management says those customers are optimising to ship more units, not to buy less memory per unit.

NVIDIA guiding its gross margin down a point for the October quarter, while the memory it buys reprices upward for 2027, is the first place the rent shows up on the other side of the invoice. We infer the link rather than read it off NVIDIA's release, which does not attribute the guide. NVIDIA can absorb a point; its network effects and embedding are why. AMD, starting from a far lower margin and shipping its most memory-dense rack yet, has less room, and Broadcom's custom accelerators carry the same HBM stacks whoever's logo is on the package. For the buyers, the memory shortage is a test of pricing power they did not have to prove while memory was cheap.

The third group is the way out of the shortage. Every new cleanroom Micron, Samsung and SK hynix finish has to be filled with scanners and inspection tools, and DRAM's move down-node puts more EUV layers and more process-control steps into every wafer. The construction-before-equipment sequencing in Micron's capex means the tool orders come after the buildings, in 2027 and 2028, which is exactly when the memory makers' own pricing is most at risk. ASML and KLA are paid to end the shortage regardless of which memory maker wins it.

Memory has no system of record and no network effects, and the matrix says so. The framework has always scored the memory makers as scale businesses with good IP and no lock-in — which is precisely why a record memory margin is priced as a peak. The one pillar that has moved is transaction embedding, and it has moved for Micron specifically: the stock page upgraded it to strong on the Strategic Customer Agreements, while Samsung sits at intact on the industry-wide shift to multi-year HBM contracts.

The distinction matters because of what those contracts are. An SCA is typically five years, non-cancellable, with committed volume, a cash deposit, and in three cases out of four a price floor and ceiling. The quarter took the count from 16 to 26, pushed coverage past a third of revenue through 2030, added the first extensions into 2031, and lifted customer commitments to $32B. A memory maker whose next-year output is three-quarters committed, with deposits it keeps if the customer walks, does not have a commodity trough. It has a shallower one: the floor holds on the contracted share, and only the uncontracted remainder takes the spot price down.

The market reads the same contracts the other way. Part of the case against SK hynix in July was that long-term agreements capped how much of the spot-price surge it could capture. That is true, and it is the trade the contracts make on purpose: give up some of the top to raise the bottom. A market that prices memory makers only on the top of the cycle will always read that trade as a cost. A framework that asks what a business keeps through the cycle reads it as the first real switching cost memory has ever had.

The bear case deserves its strongest form, because the history is on its side. Every memory upcycle has produced long-term supply agreements, and every downturn has tested them. When a customer is holding committed volume at a price well above spot, the supplier faces a choice between enforcing the contract and keeping the customer for the next cycle — and a supplier with three customers who matter usually chooses the customer. Take-or-pay language is only as strong as the supplier's willingness to take a hyperscaler to court. Nobody has yet seen an SCA survive a falling market.

The numbers also look smaller against the business than the headline suggests. $32B of commitments is less than one quarter of revenue at the guided run-rate. More than 35% coverage through 2030 means close to two-thirds is not covered, and a quarter of the agreements price periodically rather than against a floor. The ceilings cut both ways: in a tightening market they already hand some of the upside back, and the three-quarters priced inside a band may simply be priced at today's band, which sits far above where a normal cycle clears.

Then there is the supply response. Low-20s percent industry DRAM bit growth in each of 2027 and 2028 is a lot of new memory. Micron, SK hynix and Samsung are all building at once, Samsung holds the largest capacity and the least reason to defend price over share, and Chinese DRAM makers are adding capacity outside the export-control perimeter. Micron itself says its capital intensity is at historic lows relative to revenue — which is what the top of a capex cycle has always looked like, a moment before spending catches up with pricing. And a customer paying an 87% gross margin on a component has every incentive to engineer it out: compression, tiered memory, and custom HBM that puts the buyer's spec, not the supplier's, at the centre. Micron's custom HBM4E work with NVIDIA is embedding, but it is also NVIDIA defining what gets built.

Finally, the market may simply be right about valuation even if it is wrong about mechanism. Micron's own stock page already underwrites a digestion year before FY2030 rather than compounding from this base, and a peak-margin multiple on peak earnings is how memory stocks have always been priced. The shares not moving on a beat-and-raise does not have to mean investors missed the contracts. It can mean they read them, noticed the deposits cover less than one quarter, and decided a cycle with a slightly higher floor is still a cycle.

All of that is correct, and none of it is tested yet. This article's claim is narrower than "this time is different": it is that the next trough will separate the makers who contracted from the makers who did not, and that the separation will be visible in gross margin within a few quarters of the first price decline. If the contracts buckle the moment spot prices fall, the bear case wins outright and the embedding upgrade on the stock page was wrong.

• The memory makers. Micron is the better-built memory business after this print, because it has disclosed the contract coverage that makes a trough shallower. That makes it the memory name to hold through a cycle, not a reason to stop treating it as one. Samsung's memory exposure sits inside a phone and foundry conglomerate and reads as a cleaner spot-price bet on the same shortage.
• The buyers. The shortage is a pricing-power test for the accelerator vendors through 2028. NVIDIA's gross margin is the gauge to watch: a slow drift while HBM reprices is the moat working; a step down is the rent moving. AMD has the least margin to give, and Broadcom passes more of the cost to the hyperscaler that designs the chip.
• The way out. ASML and KLA get paid to build the capacity that ends the shortage, and their memory orders arrive after the cleanrooms do. They are the cohort's cleanest exposure to the memory build that does not depend on memory prices holding.