AI research wiki
hardware · 2026-10-01 · Tier 1

Micron's Quarter: Revenue Up Nearly 5x, HBM Locked for 2027, and NAND Rides the KV Cache

Micron's Quarter: Revenue Up Nearly 5x, HBM Locked for 2027, and NAND Rides the KV Cache

Sources: The Information · @StockSavvyShay on HBM · @StockSavvyShay on NAND Raw: raw/rss/2026-09-30-the-information-revenue-quintupled-at-ai-memory-maker-micron.md and the 2026-10-01 X feed capture

TL;DR

Micron reported revenue of $54.2B for the quarter ending 2026-09-03, nearly 5x a year earlier, on demand for HBM (high-bandwidth memory, the stacked DRAM next to GPU dies). Gross margin nearly doubled to 86.8% on price increases. Micron said HBM had been lower margin than conventional DRAM this cycle, but most 2027 HBM supply is already locked at much higher prices, closing that gap. It also announced NVHBM, which it calls the first custom HBM4E implementation, for Nvidia's next GPUs and NVLink Fusion. The surprise line: NAND revenue up nearly 8x to $14B in 18 months, with data-center SSDs about 71% of it.

Why it matters here

  • The KV cache is now a storage market. The memory-hierarchy page has tracked KV spilling down tiers: SGLang HiSparse fetching top-k misses from host DRAM (09-29), KVCMAS sharing caches across agents (09-30). Micron's NAND growth is the demand-side signal that the tier below DRAM (SSD-backed KV and context stores for long agent sessions) is being bought at scale.
  • HBM pricing power moves to the supplier. Locked 2027 prices plus a custom HBM4E part for Nvidia means memory, not compute, sets a growing share of accelerator cost. Pair with SK hynix validating HBM5 on CoWoS (09-30 digest).
  • Context for the capex debate: the same day, The Information reported banks (SocGen, SMBC, MUFG) pulling back from data-center loans and CleanSpark conceding terms to finance a Meta site. Memory makers are pricing in demand that the debt market is starting to question.

Links