For roughly two years, the data center industry has treated liquid cooling as the answer to a single question: how do you keep GPUs from melting? That question has been answered. Direct-to-chip liquid cooling now handles 120 kW racks full of NVIDIA Blackwell silicon. The fans are gone. The thermal problem, supposedly, is solved.
Except it isn't. Because while everyone was busy plumbing coolant lines to GPUs and CPUs, they forgot about storage.
Every NVMe SSD ever designed for a data center assumed one thing: airflow. The controller and NAND flash sit on both sides of a thermally insulated PCB. Heat dissipates through convection, pushed along by server fans that move air across drive bays in a predictable pattern. The entire thermal model of enterprise storage depends on that moving air.
Now strip the fans out. That is exactly what happens inside NVIDIA's GB200 NVL72 system, where 72 GPUs and 36 CPUs sit in a single rack drawing 132 kW, with 115 kW handled by liquid cooling and 17 kW by residual air. There are no fans in the rack. The storage drives tucked into those few remaining NVMe caddies? They are sitting in still air, generating heat with no designed path for removing it.
This is not a theoretical edge case. It is the default architecture for frontier AI compute in 2026.
For a long time, the math worked. A PCIe Gen 4 SSD pulls maybe 20 to 25 watts. In a traditional server with multiple fans running, that heat gets swept away without anyone thinking about it. The storage sits downstream of the CPU cooler in the airflow path, catches some residual breeze, and stays well below its 70 to 75 degree Celsius throttling threshold.
But the thermal envelope is shifting fast. PCIe Gen 5 enterprise SSDs already push closer to 30 watts at the platform level. PCIe Gen 6, arriving in volume through 2026, could push 40 to 60 watts per drive. That is real heat. In a fanless chassis, there is no convection loop to carry it away. The drives throttle. Performance collapses. And in a system where GPU utilization depends on storage feeding data fast enough, throttled SSDs become a chokepoint for the entire AI training or inference pipeline.
Roger Corell, Senior Director of AI and Leadership Marketing at Solidigm, put it bluntly: "Power and cooling are the top challenges facing AI data centers." He was not talking about GPUs. He was talking about the drives.
Right now, the industry's default response to this problem is what you might call a hybrid cooling architecture. Liquid loops handle the GPUs and CPUs. A small amount of residual airflow, maybe from CRAC units on the data center floor or from supplemental fans bolted onto specific components, handles everything else. Storage, networking, power delivery, voltage regulators. The leftovers.
This approach works until you try to increase density. Because every watt of air-cooled overhead is a watt you cannot spend on compute. Every cubic centimeter reserved for an airflow channel is space that could hold another drive or another NVLink connection. The hybrid model does not scale. It creates a ceiling on rack density that has nothing to do with the capability of the processors and everything to do with the thermal management of the peripheral components surrounding them.
The constraint on AI infrastructure is no longer compute performance. It is system-level thermal design. And storage is the weakest link in the chain.
In September 2025, Solidigm did something no enterprise storage company had done before. They shipped the D7-PS1010 E1.S, the first enterprise SSD with an integrated cold plate that connects directly to a server's liquid cooling loop. Greg Matson, Solidigm's SVP and Head of Products, called it "the world's first single-sided Cold-Plate solution that cools both sides" of the drive.
The technical trick is elegant. The 9.5mm E1.S form factor leaves just enough room for a cold plate assembly that draws heat from the controller and NAND on both sides of the PCB through a single thermal interface. The drive plugs into the same coolant distribution unit already serving the GPUs. No fans. No airflow dependency. No hybrid tax.
Capacities start at 3.84 TB and 7.68 TB on PCIe 5.0. Supermicro's Vik Malyala confirmed that the drives work inside NVIDIA HGX B300 liquid-cooled systems. James Zhao, the senior principal SSD analyst at Omdia, framed the product as a signal that liquid cooling would become standard for enterprise storage, not just an experiment.
If Solidigm answered the component-level question, Wiwynn answered the system-level one. Their white paper on GPU-initiated, liquid-cooled, ultra-high-density storage describes a 100% fanless chassis housing 96 E3.S NVMe drives, all liquid cooled, with the GPUs themselves initiating storage I/O through NVIDIA's SCADA architecture.
The numbers are striking. By eliminating the CPU from the storage control path and allowing over 100,000 concurrent GPU threads to issue their own data requests, the system approaches a theoretical ceiling of 100 million IOPS. Physical footprint drops by up to 75%. Data center OpEx falls 40 to 50%. The fanless design eliminates mechanical vibration entirely, which stabilizes tail latencies and extends drive life.
This is what a fully integrated thermal architecture looks like. No orphaned components. No hybrid cooling. Every element in the rack connected to the same liquid loop, managed by the same thermal telemetry, and optimized as a single system.
The idea of liquid-cooling storage is not brand new. Back in 2022, Iceotope and Meta published a study on precision immersion cooling for high-density storage JBODs. They submerged 72 hard drives in dielectric fluid and found the temperature variance across all drives was just 3 degrees Celsius regardless of position. System-level cooling power dropped below 5% of total power consumption. Failure rates improved because every drive ran at the same steady-state temperature instead of the thermal gradient you get in an air-cooled enclosure.
That was HDDs. The principle translates directly to SSDs, but the urgency is far greater because NVMe drives generate more concentrated heat per unit volume and throttle more aggressively when thermal limits are breached.
If you are designing or specifying cooling infrastructure for AI deployments, the storage thermal problem changes your job scope. You can no longer treat the CDU as something that serves the GPU tray and call it done. The coolant distribution architecture needs to reach every component in the rack that generates meaningful heat, and in a Gen 6 world, that includes every SSD.
Form factor selection matters more than it used to. The EDSFF standard exists specifically because the old 2.5-inch U.2 drive was designed for an airflow-dependent world. E1.S and E3.S were built with thermal headroom and serviceability in mind, but even they assumed some air movement. The next generation of these form factors will need to assume zero airflow as the baseline.
Cold plate integration, per-drive thermal telemetry, multi-zone leak detection for storage bays. These are not luxury features. They are requirements for any system targeting rack densities above 80 kW.
The AI infrastructure buildout is a trillion-dollar global project, and the industry has spent most of that money thinking about two things: GPU supply and power delivery. Cooling got attention only when racks crossed the 40 kW threshold and air could not keep up. Storage got almost no attention at all.
That era is ending. As VentureBeat reported, the shift to liquid-cooled AI systems has exposed a fundamental design assumption baked into every storage device on the market: that some amount of air will always be moving. In the racks being deployed today and planned for tomorrow, that assumption is wrong. The companies that recognize this, Solidigm, Wiwynn, Iceotope, and the ODMs building next-generation NVIDIA platforms, are redesigning storage from the thermal interface up. Everyone else is shipping drives into environments that will slowly cook them.