How Dell and Kioxia Built a 9.8 PB All‑Flash Server in Just 2U

At its core, the system is a storage‑optimized version of Dell’s PowerEdge R7725xd, a 2U rack server designed for data‑intensive workloads. The configuration demonstrated with Kioxia includes:

• 40 × Kioxia LC9 Series NVMe SSDs
• 245.76 TB capacity per drive
• ≈9.8 PB raw capacity total

Multiplying 40 drives by 245.76 TB results in 9,830.4 TB of raw flash capacity, which is why the system is described as a 9.8 PB configuration.

The platform itself can be powered by dual AMD EPYC 9005‑series processors, providing extremely high CPU core counts and PCIe bandwidth for storage‑heavy workloads.

How the system achieves such extreme density

The breakthrough comes from combining three architectural factors.

1. Ultra‑high‑capacity NVMe SSDs

Kioxia’s LC9 Series drives deliver up to 245.76 TB per SSD, among the highest capacities currently announced for enterprise NVMe storage. These drives use PCIe 5.0 and are designed specifically for data‑center and AI workloads.

They rely on advanced 3D QLC NAND architectures that allow very large capacities while maintaining enterprise‑class endurance and performance characteristics.

2. E3.L enterprise form factor

The drives are available in the EDSFF E3.L form factor, which is optimized for data‑center density and cooling. The longer E3.L module allows more NAND packages and higher power envelopes than traditional 2.5‑inch drives, enabling extremely large capacities in a single device.

3. High PCIe lane availability

The PowerEdge R7725xd platform provides massive I/O bandwidth thanks to modern server architectures built around high‑lane‑count CPUs and PCIe Gen5 connectivity. These systems are designed so that NVMe drives can connect directly to CPU lanes, minimizing bottlenecks and enabling large parallel storage arrays inside a single node.

Together, these design choices allow dozens of extremely large NVMe drives to operate concurrently inside a compact 2U chassis.

Networking and data movement capabilities

Dense storage is only useful if the system can move data in and out quickly. The R7725xd platform supports high‑speed networking options that allow the server to export large datasets efficiently.

For example, configurations can support multiple high‑bandwidth network interfaces—including 100 GbE and 400 GbE options—and up to five 400‑Gbps NICs in some setups, enabling extremely high aggregate bandwidth for AI clusters or storage fabrics.

Such networking capacity is essential for:

• Feeding GPU clusters during AI training
• Serving massive object‑storage datasets
• Moving training data, embeddings, and vector databases across clusters

Without high‑speed networking, a multi‑petabyte server would quickly become an I/O bottleneck rather than a storage accelerator.

Why this matters for AI and hyperscale data centers

Large‑scale AI workloads require storing enormous datasets: training corpora, checkpoints, embeddings, vector databases, and logs. High‑density flash servers reduce infrastructure overhead in several ways.

Smaller data‑center footprint

Packing 9.8 PB of flash into a single 2U system dramatically reduces the number of servers required to reach multi‑petabyte storage capacity. Fewer systems mean less rack space, cabling, and supporting infrastructure.

Potential energy and cooling savings

High‑capacity SSDs can replace many smaller drives—or even large numbers of HDDs—consolidating storage into fewer devices. This can improve energy efficiency and reduce cooling demand in large data centers, though exact savings depend on deployment design.

Better utilization of expensive AI hardware

AI clusters often contain extremely costly GPUs. If storage cannot deliver data fast enough, GPUs sit idle. Ultra‑dense NVMe storage arrays are designed to prevent those bottlenecks by providing very high parallel I/O throughput.

The emerging 245–256 TB enterprise SSD race

The configuration also reflects a broader industry trend: the rapid emergence of 245 TB‑class and 256 TB‑class enterprise SSDs.

Multiple vendors are developing drives in this capacity tier, targeting AI infrastructure and hyperscale cloud environments. For example, Micron has introduced a 245 TB version of its 6600 ION enterprise SSD, highlighting the growing competition in ultra‑high‑capacity flash storage.

Market demand is rising quickly. According to TrendForce, revenue for the world’s top enterprise SSD vendors increased over 50% quarter‑over‑quarter in late 2025, driven largely by expanding AI infrastructure deployments.

At the same time, the global SSD market is projected to grow from $36.33 billion in 2026 to $76.41 billion by 2031, reflecting the increasing importance of flash storage in modern computing systems.

What this milestone signals for the future

The Dell–Kioxia 9.8 PB server is less about a single product and more about a direction for infrastructure design. As NAND density increases and PCIe bandwidth expands, single servers are evolving into multi‑petabyte storage nodes capable of feeding entire AI clusters.

If the industry continues scaling SSD capacities toward 256 TB and beyond, it becomes plausible that future 2U systems could exceed 10 PB or even tens of petabytes of flash within a single chassis.

For hyperscalers, AI labs, and large enterprises, that shift could fundamentally reshape how storage clusters—and even entire data centers—are designed.