Kioxia Goes Liquid-Cooling on SSDs, Signaling a Quiet Shift in AI Server Design

Storage Is Now Part of the Thermal Equation

Kioxia announced its NX1 series of NVMe SSDs in late July 2026 and began sampling them to select hyper-scale customers shortly after. The specs — PCIe 5.0, E1.S form factor, capacities from 1.92 TB to 15.36 TB, sequential write speeds up 38 percent over the prior XD8 generation — are worth noting. The bigger signal is that this is the first Kioxia SSD to support direct liquid cooling.

That detail matters more than most English-language headlines covering the announcement will suggest.

For years, Japanese memory and storage vendors competed on a single axis: raw performance per dollar. Capacity went up. Latency went down. Prices fell. The cooling strategy for an AI server was still defined by a handful of hot chips — the GPU, maybe the CPU — while the dozens of SSDs sitting on the same motherboard were treated as background infrastructure. If a fan pushed enough air, everything survived.

That era is ending.

Why an SSD Needs a Cold Plate

The logic behind liquid-cooled SSDs is straightforward once you trace the path of an AI workload.

A modern AI server stacks multiple GPUs, each drawing hundreds of watts and generating substantial heat. To manage that heat, system designers are increasingly turning to direct liquid cooling — a cold plate bonded to the chip, with coolant flowing through it. When GPUs and CPUs are cooled this way, the need for massive airflow through the chassis drops dramatically. Fans move less air. The server interior becomes a constrained thermal environment in a different sense: you have solved the big-heat-source problem, but the smaller heat sources — SSDs, power delivery components, memory modules — still generate waste heat, and now they are no longer sitting in a fast-moving stream of ambient air.

You have to cool them some other way.

Solidigm, Intel’s storage spin-off, anticipated this shift. It announced a PCIe 5.0 E1.S SSD with cold-plate liquid cooling at Computex 2025. Kioxia’s move now places a second major vendor inside the same design paradigm. The difference between the two approaches — direct liquid cooling versus conventional forced-air — is not marginal. It changes how the entire server board is routed, how the enclosure is sealed, how the cooling loop is plumbed, and ultimately how many drives can be packed into a single rack unit.

The Real Contest Is System Architecture

What makes this moment quietly significant is that the boundary between chip vendor and systems engineer is dissolving.

Kioxia did not simply design a faster SSD and hope the customer’s cooling solution would handle it. The NX1 was engineered for a specific thermal environment — one where the storage device sits inside a liquid-cooled chassis and must transfer heat through a cold plate rather than rely on convective airflow. That requires a different PCB layout, different thermal interface materials, potentially different NAND wear management strategies, and a different relationship with the server OEM that integrates it.

This is not a product decision. It is a systems decision. And it is a decision that Japanese memory vendors are now forced to make at the component level, rather than leaving it entirely to the system integrator.

The same shift is happening across the industry. AMD, NVIDIA, and Intel are all designing their next-generation data-center silicon with liquid-cooling thermal envelopes in mind. The question is no longer whether an AI server can be air-cooled — it is how much of the chassis can be optimized for liquid contact, and which components are included in that optimization.

SSDs, it turns out, are now on that list.

What This Means for the Supply Chain

The immediate effect is modest. Kioxia is sampling to a limited set of hyper-scale customers. The 9.5-millimeter E1.S model supports both liquid and air cooling; the 15-millimeter version is air-only. Not every buyer will adopt the liquid-cooled variant immediately. Server designs are slow to change, and hyperscalers are still rolling out air-cooled AI racks at scale.

But the direction is set. As liquid cooling moves from niche to standard in high-density AI deployments, every component inside the chassis must be re-evaluated for thermal compatibility. Storage is one of the largest categories of components by count — an AI server may hold a dozen or more NVMe drives — so ignoring SSD thermal management becomes a liability as rack density climbs.

For Kioxia, the implication is clear: competing on performance alone is no longer sufficient. The company that can deliver a drive that works efficiently inside a liquid-cooled thermal envelope, while still offering strong IOPS and power efficiency, gains access to a segment of the market that is growing faster than any other in infrastructure. The 38 percent improvement in sequential write speed over the XD8 is useful. The fact that the drive can shed its own heat without disrupting the server’s cooling loop is the moat.

The Unseen Constraint Is Not Power — It Is Volume

Most coverage of AI server design focuses on power consumption. How many megawatts does a rack draw? What is the PUE? These are real questions. But the less visible constraint is physical volume inside the chassis.

Cold plates take up space. Coolant lines take up space. Pump capacity is limited. Every component that needs active cooling inside a liquid-cooled server must be mapped against the available cold-plate footprint and plumbing topology. An SSD that can cool itself through a direct cold plate is easier to integrate than one that requires a dedicated airflow path or an auxiliary fan.

This is why the E1.S form factor matters. It is a relatively compact drive size designed for high-density storage applications, and it is precisely the kind of form factor that aligns with liquid-cooled server architectures. Kioxia’s decision to offer both a 9.5 mm and a 15 mm variant suggests the company is positioning the product for two different integration strategies — one optimized for direct cooling, the other for environments where air remains sufficient.

The 15 mm model may turn out to be the bridge product that keeps legacy air-cooled designs viable as the market transitions. The 9.5 mm liquid-cooled model is the one that defines the next generation.

A Structural Change Disguised as a Product Launch

The FMS conference in San Jose, where Kioxia showcased the NX1 series, is a niche event by most measures. But the trajectory it reveals is not niche. It is the direction of the entire AI infrastructure stack.

Storage was the last major component class to enter the liquid-cooling conversation. That it is happening now — with a Japanese memory vendor leading the charge rather than following an American systems house — reflects a deeper change in how these products are conceived. The days when a storage company could ship a drive and assume the server OEM would solve the thermal problem are over. Thermal design is now part of the product specification, and the vendors who treat it as an afterthought will lose ground to those who bake it in from the start.

Kioxia’sNX1 series is not a dramatic departure in performance. It is a quiet admission that the rules of the game have changed, and that the company intends to write the next version.