Weekly recap cover, week ending September 18, 2026A 3D NAND word line staircase with channel holes through it, the lead topic of the week ending September 18, 2026; below it, one drawing for each topic the week covered: NAND flash (1), DRAM and HBM (1), Standards (1), Market and earnings (4), Research papers (3).THE WEEK IN MEMORY AND STORAGEWEEK ENDINGSeptember 182026LEAD · NAND FLASH11143

Week endingFriday, September 18, 2026Covers September 12 to September 18, 2026 · 10 items

SK hynix and Intel Talks, Solidigm Eyes a US NAND Fab, and CXMT Looks at Flash

Fab geography dominated the week. Reuters reported that SK hynix is talking with Intel about making memory in the United States and that its Solidigm unit is weighing a US NAND plant; SK hynix said nothing has been decided on the Intel talks. Reuters also reported that CXMT plans a NAND research line, which would move the DRAM maker toward YMTC's turf.

Micron showed a very dense server DDR5 module, PCI-SIG posted new base specification revisions, and Kioxia said it is preparing a US share listing after a Bloomberg report. Research this week kept returning to high bandwidth flash and SSD backed local inference.

Top stories

ReportNAND flash

Reuters: CXMT is preparing a NAND research line in Beijing (opens in a new tab)

Reuters, citing sources, says the DRAM maker plans a research and development production line for NAND at its new Beijing plant. CXMT had not commented. This is a report, not a company announcement.

TrendForce, citing Reuters ·

Reported by TrendForce, citing Reuters; the company has not confirmed it yet.

The rest of the week

Market and earnings

Research papers

  1. Study sizes DRAM and flash for a trillion parameter model served from high bandwidth flash (opens in a new tab)

    The authors explore how much DRAM bandwidth and host bandwidth an appliance needs once model weights move to high bandwidth flash. Results are analytical and the work is a preprint.

    arXiv ·

  2. SSD-LLaMA runs trillion parameter experts from an SSD on a consumer PC (opens in a new tab)

    The system streams mixture of experts weights from an SSD through RAM and GPU memory with an I/O pipeline built for expert delivery. The authors report faster decoding than baselines on a single consumer GPU. It is a preprint.

    arXiv ·

  3. HBFlex manages KV cache reads, writes and reclaim for an all flash memory system (opens in a new tab)

    The paper proposes a memory system built only from high bandwidth flash that balances KV placement, schedules write back around compute and packs data by lifetime to cut garbage collection. Results come from trace driven simulation.

    arXiv ·