Micron is selling every tier of AI memory at once

Micron (NASDAQ: MU) came to COMPUTEX arguing that the ceiling on AI performance has moved off the processor and onto memory, both bandwidth and capacity, and it brought a part for every piece of that argument. Sumit Sadana, the chief business officer, put numbers behind it, with AI context lengths growing roughly 30x a year and memory per server doubling in three years.

The most consequential part is HBM4, where Micron says its 36GB 12H stacks are in high-volume production for Nvidia's (NASDAQ: NVDA) Vera Rubin, running past 11 Gb/s per pin for more than 2.8 TB/s, about 2.3x the bandwidth of HBM3E at over 20% better power efficiency. The entire 2026 HBM4 output is already committed under binding contracts, and a denser 48GB 16H part is sampling. SK Hynix (KRX: 000660) and Samsung (KRX: 005930) also build HBM4 for Vera Rubin, so Micron is one of three sources into that platform.

The more contested fight is SOCAMM2, the compact LPDDR module for servers. Micron leads on capacity at 256GB per module, against SK Hynix's 192GB part that reached mass production in May, with Samsung pushing its own design. The pitch is roughly a third of the power and a third the board area of an RDIMM at 9.6 Gbps, with time-to-first-token in real-time inference cut by more than 80%.

Then the storage tier. The 9650 is Micron's first commercially available PCIe Gen6 SSD. The 6600 ION reaches 245TB and stands in for a wall of HDDs at 82% less rack space and half the power. A client part, the 4600, loads a 13-billion-parameter model in under a second.

Across those four lines, Micron now has a current part at every tier of the AI stack.

For anyone watching consumer drives, Gen6 SSDs like the 9650 are shipping but going to the data center first, and every wafer Micron commits to HBM4 and SOCAMM2 is DRAM it is not selling into the consumer market while memory prices are already high.

Drafted with AI assistance against parallel reporting.

Sources:

• Micron, “Micron Powers AI Everywhere at COMPUTEX 2026” (press release via GlobeNewswire, June 1, 2026): HBM4, 256GB SOCAMM2, 9650 Gen6 SSD, 6600 ION 245TB, 4600 client SSD, GDDR7, DDR5 RDIMM, Sadana strategy quote
• DigiTimes, “Micron unveils AI memory and storage roadmap at COMPUTEX 2026” (June 2, 2026, paywalled stub: headline and lead only)
• Tom's Hardware, “Micron enters high-volume production of HBM4 for Nvidia Vera Rubin” (Q1 2026): >2.8 TB/s, 2.3x HBM3E bandwidth, >20% power efficiency, 48GB 16H sampling, 2026 capacity committed
• StorageReview / VideoCardz, “Micron 192GB SOCAMM2” coverage (May 2026): 1-gamma LPDDR5X, 9.6 Gbps, one-third power and footprint vs RDIMM, time-to-first-token reduction
• SK hynix newsroom, “SK hynix Begins Mass Production of 192GB SOCAMM2” (May 2026)
• TrendForce, “SOCAMM War Heats up: Micron Ships 256GB SOCAMM2 Samples” (March 2026): Micron topping industry capacity at 256GB