Memory families
Six Memory Families, Three That Matter for AI
DRAM, HBM and NAND carry nearly all memory revenue and all of the AI load. SRAM lives inside logic chips; NOR, MRAM and FRAM serve boot code and safety-critical state in cars and industrial gear.
| Family | Products | Process | Leading makers | Where it goes |
|---|---|---|---|---|
| DRAM | DDR5, LPDDR5X, GDDR7, DDR4 | 1-alpha to 1-gamma DRAM; EUV on a few layers | Samsung, SK hynix, Micron, CXMT, Nanya, Winbond | Working memory in every server, PC, phone and car |
| HBM | HBM3E, HBM4 | DRAM core dies on a logic base die, stacked with TSVs | SK hynix, Samsung, Micron | AI accelerators: NVIDIA, AMD, Google, AWS |
| NAND flash | TLC and QLC 3D NAND | 218 to 321 layers in volume; etch-limited | Samsung, SK hynix and Solidigm, Micron, Kioxia, SanDisk, YMTC | SSDs, phones, AI training and inference storage |
| SRAM | Embedded cache; standalone parts | Built on the host logic node; standalone at 28 to 130nm | TSMC and Samsung (embedded); Renesas, Infineon, ISSI, GSI Technology | CPU, GPU and NPU caches; network buffers |
| NOR flash | SPI NOR (Winbond W25Q, Macronix MX25) | 45 to 130nm planar | Winbond, Macronix, GigaDevice, Infineon, Micron | Boot and firmware in cars, MCUs and IoT |
| MRAM, FRAM, ReRAM | Everspin MRAM; embedded MRAM at GlobalFoundries 22FDX and Samsung 28nm; Fujitsu FRAM; Weebit ReRAM | Back-end-of-line integration at 22 to 28nm | Everspin, GlobalFoundries, Samsung, TSMC, Fujitsu, Weebit Nano | Automotive MCUs, industrial state retention; Intel Optane (PCM) ended in 2022 |
Makers
Who Holds DRAM, HBM and NAND
Samsung, SK hynix and Micron are the only makers at scale in all three AI memories. Every one of them is an integrated device maker: memory has no foundry model, so capacity decisions in Suwon, Icheon and Boise reach world supply within months.
| Maker (HQ) | DRAM | HBM | NAND | Position |
|---|---|---|---|---|
| Samsung (South Korea) | 39.4% | 33% | 29.3% | Largest DRAM and NAND maker; HBM4 in mass production since February 2026 |
| SK hynix (South Korea) | 24.9% | 50% | 18.1% | HBM leader; NAND includes Solidigm |
| Micron (United States) | 23.3% | 18% | 15.1% | Only US-headquartered maker; DRAM output in Manassas, Virginia; Boise and Clay fabs building |
| Kioxia (Japan) | - | - | 13.6% | NAND only; shares Yokkaichi and Kitakami fabs with SanDisk |
| SanDisk (United States) | - | - | 11.4% | NAND only; spun off from Western Digital in February 2025 |
| CXMT (China) | 9.5% | Sampling | - | Commodity DDR4, DDR5 and LPDDR5; HBM3 sampling |
| YMTC (China) | - | - | ~14% of bits | Third by NAND bits; on the US Entity List |
DRAM and NAND shares: TrendForce (SK hynix, Micron and SanDisk NAND shares derived from its revenue figures). HBM shares: Counterpoint. YMTC bit share: Tom's Hardware reporting.
The HBM diversion
HBM Takes About Three Times the Wafer per Bit
Each HBM stack needs 12 to 16 known-good DRAM dies, through-silicon vias on every die and a stacking step that loses yield at each layer. Micron has put HBM3E at about three times the wafer supply of DDR5 for the same bits. When the three leaders move wafers to HBM for AI accelerators, DDR5 for servers and LPDDR5X for phones and cars tighten, even with no fab running short.
HBM then has to be joined to the GPU in TSMC CoWoS packaging, a separate line with its own capacity limit; see CoWoS.
Record
Page Record
- Covers
- Memory families, makers and shares, and the HBM wafer diversion
- Sources
- TrendForce DRAM and NAND revenue rankings (August and September 2026); Counterpoint HBM shares (September 2026); Micron investor statements; company releases
- Last verified
- October 2026