Samsung used its GMIF2026 keynote in Shenzhen to show a 400-plus-layer NAND, two new server SSDs and a flash part designed to sit inside the processor package.
The number that will travel furthest from Samsung’s appearance at the Global Memory Innovation Forum in Shenzhen is 256TB, the top capacity of its new BM1773 server SSD. The idea worth watching, though, is Samsung zNAND-O, a flash design that Samsung wants to co-package directly with a processor so that local AI models stop leaning so hard on expensive DRAM.
Kevin Yoon, CVP and CTO of Samsung Memory China, laid it all out in a keynote titled “The Memory Revolution: Building the Foundation of Agentic AI”. According to Samsung’s own release, zNAND-O is in development in 4-layer and 8-layer versions and connects to an MPU over UCIe, the chiplet interconnect standard, inside the same package.
Samsung’s claims are big. It says zNAND-O offers ten times the bit density of DRAM, and seven times the read bandwidth and read energy efficiency of conventional NAND. In a simulation built around Moonshot AI’s one-trillion-parameter Kimi K2 model, Samsung says an NPU-optimized design using zNAND-O cut memory cost to a quarter at the same performance, and fit a model nearly twice as large.
Two drives, two jobs #
The SSDs are less speculative. The PM1763 is the performance part: a PCIe Gen6 server drive built on TLC V-NAND with an upgraded cooling design, sampling in the fourth quarter of 2026. The BM1773 goes the other way. It runs on V9 QLC, spans 16TB to 256TB, and Samsung calls it the world’s smallest 256TB SSD. “With these two SSDs, we will drive both the performance and capacity markets,” Yoon said.
SamMobile fills in more detail on the BM1773: an E3.S drive on PCIe 5.0 with a 16-channel controller and 2Tb QLC dies, rated for up to 4.5GB/s sequential reads and 50K random write IOPS. Endurance is listed at 0.6 drive writes per day, roughly double the total bytes written of Samsung’s previous QLC generation, and the drive is designed to keep running if a single die fails rather than dropping into read-only mode.
Underneath both stories sits V10 BV-NAND, which Samsung bills as the first V-NAND with more than 400 layers. The memory cells and the peripheral circuitry are built on separate wafers and then bonded vertically. Against V9, Samsung quotes 58% higher bit density, 33% faster I/O and 25% lower power.
What to keep in perspective #
Every figure here comes from Samsung, and none of it has been tested independently. zNAND-O also has no specification sheet, no customer and no date attached. Because it is co-packaged over UCIe, it is something a chip designer would build into an SoC, not a part you will ever drop into an M.2 slot. Whether it reaches laptops depends on Qualcomm, MediaTek, Apple or Nvidia deciding it is worth the package space.
The BM1773 is a capacity drive, not a fast one. 4.5GB/s is modest for a modern server SSD; the point is to put a huge amount of flash into a small slot and take ground from hard drives in AI data lakes.
It still matters to PC buyers. AI demand for memory is a big part of why DRAM and NAND prices climbed this year. If flash can carry some of the load that currently needs DRAM on local AI machines, that is one of the few developments that could ease the pressure, assuming zNAND-O leaves the simulator.