10 PB in 2RU: Dell PowerEdge R7725xd with 40 Kioxia LC9 SSDs

Dell's PowerEdge R7725xd server now ships with 40 Kioxia LC9 245.76TB NVMe SSDs in E3.L form factor, yielding 9.8 PB raw capacity in just 2 rack units. The box runs on AMD EPYC 9005 processors and supports up to five 400 Gbps NICs — enough to saturate the storage for AI training pipelines, data lake ingestion, or backup workloads.

Kioxia's LC9 is a QLC (quad-level cell) NVMe SSD that pushes density by storing 4 bits per cell. Each drive hits 245.76 TB in the E3.L form factor, which is essentially a larger, thicker M.2-like card designed for high-capacity enterprise storage. Dell slots 40 of these into the R7725xd, hitting 9.8 PB. A full rack of 20 such servers would deliver 196 PB.

Why this matters for AI infrastructure

Arun Narayanan, SVP Compute and Networking at Dell, said: "The Dell PowerEdge R7725xd combined with Kioxia's high-capacity enterprise SSDs delivers the storage density and power efficiency our customers need to scale AI infrastructure without sacrificing performance."

Neville Ichhaporia, SVP and GM of SSD business at Kioxia America, added: "Customers can deploy massive ingestion streams, scale data lakes effortlessly, and handle large backups in a fraction of the footprint, improving TCO to new levels."

Competitors and roadmap

Kioxia isn't alone in the 256 TB-class SSD race. Micron has the 6600 ION, Sandisk the UltraQLC SN670, SK Hynix the AIN D, and its Solidigm subsidiary also fields competing drives. Scality revealed it's working on support for a future Samsung nearline-class SSD that could reach 1 PB per drive.

Technical details

  • Drive: Kioxia LC9, QLC NAND, NVMe, E3.L form factor, 245.76 TB
  • Server: Dell PowerEdge R7725xd, 2RU, AMD EPYC 9005
  • Maximum capacity: 40 drives → 9.8 PB raw
  • Networking: Up to 5x 400 Gbps NICs
  • Rack density: 20 servers → 196 PB

Developer implications

If you're building data pipelines for AI training, this density means you can colocate more data with compute, reducing network hops. For example, a training job that previously required 10 racks of HDD-based storage can now fit in one rack of these servers. The 400 Gbps NICs ensure you can feed GPUs without bottlenecking.

For database engineers, the E3.L form factor offers better thermal management than U.2, allowing sustained write performance under load. However, QLC drives have lower endurance than TLC — Kioxia rates the LC9 for 0.3 DWPD (drive writes per day), so they're best for read-heavy or cold storage workloads.

How to configure for AI workloads

To maximize throughput, stripe data across all 40 drives using Linux MD RAID 0 or ZFS. Example command for RAID 0 with 40 NVMe namespaces:

mdadm --create /dev/md0 --level=0 --raid-devices=40 /dev/nvme0n1 /dev/nvme1n1 ... /dev/nvme39n1

Then format with XFS for large file support:

mkfs.xfs -f /dev/md0
mount /dev/md0 /mnt/storage

For ZFS, use ashift=12 and recordsize=1M to match QLC write patterns:

zpool create -o ashift=12 tank /dev/nvme[0-39]n1
zfs set recordsize=1M tank

The 1 PB SSD horizon

Kioxia's LC9 is not the endgame. Samsung and others are developing SSDs that could reach 1 PB per drive using similar QLC or PLC (penta-level cell) technology. Scality's work with Samsung hints at nearline-class drives that could replace tape libraries. If 1 PB SSDs become real, a single 2RU server could hold 40 PB — enough for the entire Wikipedia archive 400 times over.

Bottom line

Dell and Kioxia have delivered the densest NVMe storage server available today. For developers managing AI data lakes or large-scale backups, this hardware changes capacity planning. Start evaluating QLC endurance and workload patterns now — the density is worth the trade-offs.