Why liquid-cooled storage is the secret to scalable AI performance

In the rapidly evolving landscape of artificial intelligence, the significance of storage has often been overshadowed by the more prominent focus on GPUs. As we approach the year 2026, the neglect of storage solutions in AI infrastructure could lead to substantial drops in performance and efficiency. The effectiveness of AI systems hinges on the thermal management of their components, making it crucial to rethink how storage is integrated into the cooling strategies of these setups. Currently, many data center architects are confronting a challenge known as the "thermal wall." As GPUs evolve to handle increasingly complex AI training and inference tasks, they produce remarkable amounts of heat—often exceeding 1,400 watts per GPU. To put it in perspective, a full rack of these GPUs generates heat comparable to that of over a hundred microwave ovens confined within the space of a refrigerator. With the emergence of 2,000-watt chips on the horizon, traditional air cooling is rapidly becoming inadequate. This has triggered a significant shift towards liquid cooling solutions for GPU systems. Liquid cooling methods, including direct-to-chip (DTC) and immersion cooling—where components are submerged in a specialized fluid—are proving to be up to 3,000 times more effective at dissipating heat compared to air. However, in this rush to enhance GPU cooling, the storage aspect—particularly local SSDs in GPU servers—has been overlooked. While SSDs consume less power, relying solely on traditional air cooling for these components leads to inefficiencies, creating what is known as a hybrid cooling scenario that can be both costly and ineffective. Opting for a fully liquid-cooled storage solution offers numerous advantages. Research from DUG Technology, a leader in High Performance Cooling (HPC) infrastructure, indicates that implementing full immersion cooling can reduce overall system power consumption by over 50%. This reduction is largely attributed to the elimination of fans, which can consume up to 25% of a system's total power. Power Usage Effectiveness (PUE) is a vital metric for assessing data center efficiency; a score of 1.0 indicates that all energy is utilized by IT equipment, whereas higher scores reflect overhead and waste. By integrating storage into a liquid cooling system, companies have achieved impressive PUE ratings as low as 1.02, significantly better than the industry average of 1.55. Moreover, liquid cooling not only enhances thermal management but also extends hardware lifespans. AI workloads often exhibit erratic heating and cooling patterns, which can stress components and increase failure rates. Innovative designs from companies like Hypertec demonstrate that servers in liquid-cooled environments can operate reliably for up to nine years—almost double the lifespan of standard air-cooled systems. As the industry evolves, the necessity for liquid cooling is becoming undeniable. NVIDIA's upcoming Vera Rubin NVL72 rack-level design for 2026 mandates liquid cooling for all system components, indicating that even storage devices will no longer have the option of air cooling. Anticipating this shift, Solidigm has collaborated with NVIDIA to develop the first liquid-cooled, fully hot-swap serviceable eSSD for Direct Attach Storage (DAS). The future of scalable AI performance lies in a cohesive approach where storage and computation coexist in a liquid-cooled environment. Partnerships between Solidigm, DUG, and Hypertec have already showcased the potential for high-performance NVMe storage solutions, enabling a 1U liquid-cooled storage node to accommodate up to 320TB, with scalability up to an astonishing 1.9 petabytes (PB). This density not only optimizes space but also places crucial data in proximity to revenue-generating compute resources, facilitating real-time AI-driven decision-making. It's imperative to stop viewing storage as a secondary consideration and instead recognize it as a critical asset. By prioritizing liquid cooling across the entire infrastructure, organizations can transcend the limitations of air and pave the way for a sustainable and powerful AI future.