Kioxia's New SSD Hits 10 Million IOPS for AI Infrastructure

Imagine a storage drive that can feed a hungry GPU farm without a hiccup. Kioxia's latest arrival does just that: a GP1 series SSD engineered to deliver up to 10 million random IOPS and native GPU-direct access, built on PCIe 6.0 and NVMe 2.2.

Built to keep GPUs busy

GP1 is the first public member of Kioxia's GP family, a product line the company sketched out earlier this year. At its core sits second-generation XL-FLASH, an SLC-type flash optimized for low latency, paired with a new in-house controller. Short bursts. Long streams. The drive hits up to 10 million input/output operations per second for random reads, a metric that cuts straight to the heart of AI workloads where thousands of parallel requests must be handed to processors without delay.

Why does IOPS matter so much? Because training and inference pipelines are ruthless about latency. If storage can’t supply fresh data fast enough, powerful GPUs sit idle. High IOPS reduces those idle cycles and raises effective throughput without forcing datacenter operators to buy more high-cost HBM memory.

That last point is crucial. The GP1 line is explicitly aimed at modern AI architectures that mix high-bandwidth memory with a flash-backed expansion layer. Think of XL-FLASH as a larger, cheaper attic for datasets that can’t all live inside HBM. The trick is keeping the path from that attic to the GPU wide and fast enough that performance doesn’t fall off a cliff.

Kioxia also pushed endurance. The announced durability reaches up to 50 drive writes per day for a two-year warranty model at 3 TB capacity, which translates to roughly 109.5 petabytes written over that period. For context, some competitors tout higher single-number totals under different NAND choices; Kioxia’s approach trades raw write density for the latency and durability benefits of SLC-style XL-FLASH.

Form factors matter for deployment. GP1 will ship in E3.S and E1.S variants, in 9.5 mm and 15 mm thicknesses. Certain E3.S and the 9.5 mm E1.S models support liquid cooling, while all SKUs remain compatible with conventional air-cooled chassis. That flexibility makes these drives practical both for dense AI servers and for mixed datacenter racks.

There’s also a line of sight to much higher ambition. Kioxia’s architecture is designed for evolution; the team has publicly set a roadmap target of 100 million IOPS for future generations. Reaching that figure would require bold shifts in memory design and likely new categories beyond today’s SSD constructs, a direction Nvidia has been nudging suppliers toward for next-generation AI infrastructure.

Selected customers will get evaluation samples before the end of 2026, ahead of a public showing at the FMS storage conference in California. If you run or design GPU-heavy clusters, GP1 is a product to watch. It’s not just about raw numbers. It’s a reminder that storage can be the throttle or the accelerator in AI systems, depending on how it’s engineered.