256TB in a Single SSD Is Real, and It's Coming to Consumers

TL;DR

Kioxia's new LC9 QLC drive packs 256TB into a single E3 form factor SSD, enabling petabyte-scale AI storage in just five drives.

Key Points

• 1.Kioxia's LC9 drive delivers 256TB in a single E3 form factor SSD. Five of these drives fill a half-rack with 1.23 petabytes of flash storage, using PCIe 5.0 and QLC NAND manufactured in Japan.
• 2.A separate GP1 SLC drive targets GPU-direct storage with 512-byte access granularity. It achieves 10.5 million IOPS today and has a roadmap to reach 100 million IOPS by 2027, with 8–16TB capacity in an E1.S form factor.
• 3.Nvidia's new storage architecture integrates these drives across tiered memory layers. Using Bluefield DPU NICs (800Gb/s), Spectrum-X switching, and Nvidia Dynamo, the system routes data across HBM, LPDDR5, local NVMe, and disaggregated QLC storage to serve hundreds of thousands of AI inference users.
• 4.GPU-direct NVMe access eliminates CPU overhead for AI workloads like graph neural networks and fraud detection. Direct GPU-to-NVMe copies keep latency in the 10–15 microsecond range via RDMA, making large-context applications viable without expanding VRAM.
• 5.Kioxia's open-source Isaac technology offloads RAG vector databases to NVMe storage, supporting 4.8 billion vectors at 200 queries/second with 50-microsecond latency. This delivers up to 4x performance over blind sharding and achieves 90–99% recall, significantly reducing AI hallucinations.