Enterprise NAND and NVMe SSDs
BOTTLENECKThree manufacturers control NAND supply and endurance requirements restrict qualified enterprise SSD availability for AI training.
Flash SSDs in E1.S, E3.S, or U.2 form factors for training datasets, model checkpoints, and repositories. Speeds data loading to accelerators; endurance requirements differentiate enterprise from consumer grades. NAND pricing cyclical per oligopoly supply; QLC enterprise SSDs command premium for endurance management.
Flash-based solid-state drives in E1.S, E3.S, or U.2 form factors for AI training dataset staging, checkpoint storage, and model repositories.
Why the concentration exists
Enterprise NVMe SSDs use NAND flash memory organised as vertically stacked cells to deliver high-throughput, low-latency storage for data centers, AI training clusters and hyperscale environments. Supply is concentrated because only a handful of firms operate the multi-billion-dollar fabs needed to produce advanced 3D NAND at marketable cost. Each generation of node shrinks adds a layer count and raises yield risk, so qualification cycles for new products run long. The result is a tight oligopoly of integrated suppliers whose internal NAND availability directly governs how many SSDs they can ship.[15][6][5]
Capacity per drive is rising rapidly as stacking technology improves. SK hynix has begun mass-producing 321-layer 2Tb QLC devices and is developing a 244TB enterprise SSD by bundling 32 dies in one package through 32DP packaging. KIOXIA has demonstrated 2.5-inch enterprise SSDs up to 61.44TB and E3.S units up to 30.72TB on its eighth-generation BiCS FLASH. Samsung's own roadmap extends to 400-layer mass production now and a stated path toward 1,000 layers by 2030. Higher capacity per drive reduces the number of physical units needed per rack but also concentrates more value into each qualified vendor relationship.[8][10][11][2]
What the evidence shows
Samsung relies on own NAND for stable SSD supply; SK Group is second-largest.
techpowerup.comLead times for enterprise SSDs have extended from 8–12 weeks to 16–20 weeks or longer.
fusionww.comQualifying enterprise SSD vendors is a long and arduous process.
kingston.comWho supplies it
Five vertically integrated NAND producers set the structure of the enterprise SSD market. Samsung Electronics is the largest SSD manufacturer and produces its own NAND using proprietary V-NAND, while SK Group, comprising SK hynix and Solidigm, ranks second and focuses on high-capacity QLC products. Solidigm, a US subsidiary of SK hynix, sells enterprise PCIe Gen4 and Gen5 NVMe drives including the D7-P5520 and D7-PS1010, both offered at 7.68TB. KIOXIA, spun off from Toshiba and now owner of the LITE-ON SSD business it acquired in 2020, rounds out the top tier alongside Western Digital/SanDisk and Micron Technology.[1][6][15][19][12][14]
Who controls it
What it depends on, and what depends on it
What an enterprise SSD makes possible is the elimination of I/O starvation in GPU clusters. NVMe already delivers 3-7× the throughput of SATA and roughly 10-100× lower latency, and most AI training servers ship with 4-8 NVMe drives per node for performance. NVMe over Fabrics extends that performance across the data center using RDMA at latencies as low as 10-20 microseconds, or NVMe/TCP at 100-200 microseconds. Industry analyses describe 2026 AI storage as defined by unified namespaces, all-flash architectures, and disaggregated composable infrastructure.[18][16][17]
Where it sits in the stack
Takes in: 3D NAND wafers (L5), NVMe controller ASIC, DRAM cache
Sends on: High-throughput NVMe storage nodes for AI training infrastructure
What would break it
Supply tightness is the dominant fragility. Storage lead times have extended from 8-12 weeks to 16-20 weeks or longer, and analysts cited by resellers have projected price rises of 20% or more by the end of 2025. Phison's chief executive has stated that NAND prices more than doubled in six months, with a 1TB TLC chip moving from $4.80 in July 2025 to $10.70. All NAND production capacity for 2026 is reported as already sold out. Certain endurance classes, especially mixed-use enterprise SSDs, are noted as becoming harder to secure during tight supply periods.[4][9][3][7]
What to watch
Capacity additions and product launches are scheduled across the major suppliers through 2026 and 2030. Samsung is targeting mass production of 400-layer NAND with a stated path to 1,000 layers by 2030, and it began mass production of the PM1763 PCIe 6.0 enterprise SSD on July 8, 2026. SK hynix's 321-layer QLC is in mass production and a 244TB enterprise SSD is in development using 32-die packaging. KIOXIA demonstrated the CM9 Series PCIe 5.0 NVMe SSDs built with 8th-generation BiCS FLASH TLC on May 15, 2025, with 2.5-inch capacities up to 61.44TB and E3.S capacities up to 30.72TB.[2][13][8][10][11]
Sources
- techpowerup.com · 2026-03-13T00:00:00
- en.unibetter-ic.com · 2026-02-26T03:53:47
- comtek.co.uk · 2026-03-06T16:07:47
- fusionww.com · 2025-10-02T00:00:00
- kingston.com · 2021-07-31T00:05:22
- grokipedia.com · 2026-01-17T00:29:01
- hyssdpro.com · 2026-06-09T03:15:54
- tomshardware.com
- tomshardware.com · 10 November 2025
- americas.kioxia.com · May 15, 2025
- apac.kioxia.com · May 16, 2025
- en.wikipedia.org · 30 June 2020
- servethehome.com · 2026-07-08
- ssdwiki.com · 2021
- ssdwiki.com · 2024
- datacore.com
- starwindsoftware.com · 2026
- softwareseni.com
- solidigm.com
Full scorecard, owner shares, supply edges and the full tracked roster are in the desk letter.
GET THE BRIEFING