Micron's $50 Billion Quarter Meets an AI Stack That Moves Its Cache

Micron reports fiscal Q4 after the close on Wednesday, 30 September 2026, guided to $50 billion of revenue↗ plus or minus $1 billion at an 86% gross margin ¹²³. Last quarter printed $41.46 billion at an 84.6% gross margin ²⁷. The stock has more than tripled this year, up over 230% ¹, a run priced for another round of big estimate hikes.

Consensus sits above the guide at $51.24 billion↗, up 353%, the fastest on record ⁴. Morgan Stanley stays Overweight at a $1,200 target and still expects estimates to rise, just "less so than prior quarters↗" ¹⁵. Fiscal 2027 profit estimates jumped↗ from $95.80 to $154.70 in three months, then rose just 1.2% in the month through early August ⁶. For holders, the size of the next hike is the trade.

Frontier models push cache into host memory

None of that math prices what the newest frontier serving stack does to memory. Z.ai's GLM-5.x models, 744 billion parameters with 40 billion active per token, use sparse attention to read less of the cache at each step ⁸. The catch: picking which tokens to read requires the full context to sit in HBM first, so the capacity bottleneck survives ⁸.

SGLang's HiSparse layer offloads cache entries to host DRAM, and Z.ai says GLM-5.3 adds local storage as a further tier ⁸. Measured on a B200 serving GLM-5.2↗, doubling concurrency from 8 to 16 requests cut GPU-memory token reuse from 90.3% to 54.8% while host-memory reuse rose from 6.0% to 40.3%, with the overall hit rate above 95% ⁸. Agents, which recycle long histories and write little, are the workload doing this ⁸.

Sparse attention does not shrink the memory bill. It forwards the bill to a cheaper tier.

Conventional DRAM delivers the volume while HBM keeps the margin

Micron sells both ends of that ladder, and its growth already lives at the bottom. Its June-quarter DRAM revenue rose 65.5%↗ from the prior quarter to $36.0 billion, the fastest of the big three, while its share of the HBM market slid from 21% to 18% ⁹.

The premium stays with HBM. SK Hynix holds about half that market and runs a 76% operating margin, and each HBM4 wafer eats roughly three DDR5 wafers of capacity, part of why conventional DRAM prices jumped 58% to 63% in a quarter ⁹.

Micron's own investor site pitches agentic AI under a white paper↗ titled "Agentic AI in Data Centers: The Critical Role of High-Bandwidth Memory" ¹⁰; the serving measurements put agentic traffic one tier down.

Long-term purchase pacts cannot resolve where future margins settle

The bull case has receipts: 16 strategic customer agreements worth about $100 billion↗ through 2030, with roughly $22 billion of customer cash and other commitments ²³. It also has arithmetic: $100 billion is about two quarters of revenue at the guided pace ². Synovus sees supply tight through fiscal 2027 ¹, and SK Hynix is weighing a Solidigm IPO↗ that values its NAND unit at up to $150 billion ¹¹.

Even the Overweight camp concedes there is "no real way to prove that near term" ¹. The beat will take care of itself; at an 86% margin, scarcity is not the question. Which memory the agentic era pays a premium for is. The test on Wednesday's call: whether Micron's HBM share keeps sliding while conventional DRAM does the growing, and whether management's DRAM pricing talk credits scarcity or substitution ⁹.