In the high-stakes world of modern AI factory design, storage has often been relegated to the role of "youngest sibling." While GPUs hog both the spotlight and the budget, storage is often treated as a low-priority afterthought, a necessary but "set-it-and-forget-it" component.
As we move into 2026 and beyond, this significant oversight can have serious performance and efficiency consequences. Your AI infrastructure is only as resilient as its weakest thermal link. Right now, many forward-looking organizations are realizing that how storage is treated becomes a missing ingredient in the AI systems cooling menu. Here's why liquid-cooled storage is the secret to scalable AI performance.
The thermal wall
We are currently hitting what data center architects call the "thermal wall." As GPUs get more powerful to support increasingly demanding AI training and inference workloads, they generate unprecedented amounts of heat. Today's superchips routinely exceed 1,400W+ per GPU. For a full rack, that's the equivalent of running more than a hundred full-sized microwave ovens in the space of a single refrigerator. Now, 2,000W+ chips are on the horizon. At this scale, relying on traditional air cooling simply does not work.
That's why we've already begun to see a massive shift to liquid cooling for GPU systems. Whether it's direct-to-chip (DTC) or immersion cooling (where hardware is literally submerged in a special fluid), liquid is up to 3,000 times more efficient at removing heat than air. However, in the rush to upgrade the "brains" of the operation, many people are forgetting the storage portion: the local SSDs in the GPU servers that feed data to those hungry GPUs.
The hybrid cooling trap
Modern GPUs by necessity have largely transitioned to liquid cooling. By comparison, SSDs draw significantly less power, leaving data center architects with the option to rely on traditional spinning fans to cool the storage subsystem. This combination of liquid and air is known as a "hybrid" cooling approach, and it creates a massive efficiency headache. You end up running and maintaining two different cooling systems at once, which is both expensive and wasteful.
The benefits of choosing fully fan-less liquid cooling for storage are compelling. Real-world results from High Performance Cooling (HPC) infrastructure leader DUG Technology show that adopting a full immersion cooling can slash overall system power consumption by over 50%. A huge part of that comes from simply getting rid of fans, which can eat up to 25% of a system's total power. Power Usage Effectiveness (PUE) is a data center's efficiency score. A 1.0 score means all energy is being directly used by IT equipment, but anything above that translates to overhead and wasted energy. By bringing storage into the more efficient liquid loop, organizations have achieved PUE ratings as low as 1.02 — far better than the industry average of 1.55.
Liquid cooling as a requirement
It turns out that liquid cooling isn't just about removing heat. It can also extend the life of your hardware. AI workloads are often bursty, leading to wild heating and cooling temperature swings. Thermal cycling stress describes the cumulative stress on hardware components that can lead to increased failure risk. Companies like Hypertec have pioneered "immersion-born" server designs that thrive in these thermally demanding operating environments. The results are startling: some servers in these liquid setups still run smoothly after nine years of operation. That's nearly double the typical five-year lifespan expected from air-cooled hardware.
We are quickly approaching a point where liquid cooling isn't a choice but rather a requirement. NVIDIA's new-in-2026 Vera Rubin NVL72 rack-level design mandates liquid cooling for every component in the system. That means that even if your storage devices could be cooled sufficiently with air, fans are no longer an option. NVIDIA's planned 2027 rack designs will draw four times the amount of power in the same space, so air is a dead end. In anticipation of this coming industry challenge, the Solidigm team, in collaboration with NVIDIA, led the effort to introduce the first liquid-cooled, fully hot-swap serviceable eSSD for Direct Attach Storage (DAS).
Moving past the limitations of air
True scalability comes from a unified strategy where storage and computation work together in the same liquid-cooled ecosystem. Collaborations between Solidigm, DUG, and Hypertec have already proven that this is possible at scale. For instance, using high-performance NVMe storage, a single 1U liquid-cooled storage node can now hold up to 320TB, with the potential to scale to a staggering 1.9 petabytes (PB).
This kind of density doesn't just save space — it keeps valuable data near revenue-generating compute engines, empowering real-time AI decision-making.
It's time to stop treating storage as an afterthought and start seeing it as a strategic asset. By thinking liquid first across your entire stack, you can move past the limitations of air and build an AI future that is as sustainable as it is powerful.
<iframe width="560" height="315" src="https://www.youtube.com/embed/avD8U30NcGY?feature=oembed" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen title="Immersion Cooling — Built for the Next Wave of Data Storage"></iframe>
This post was created by Solidigm with Insider Studios.
The post Why liquid-cooled storage is the secret to scalable AI performance appeared first on Business Insider
