Liquid Cooling Is Moving Beyond the GPU

Until recently, the liquid-cooling discussion in AI data centers focused almost entirely on GPUs and CPUs.

That made sense.

Accelerators were the largest heat sources, and Direct-to-Chip cooling offered the fastest way to remove tens or hundreds of kilowatts from increasingly dense AI racks.

But the architecture is changing.

As servers become more densely integrated and fans begin to disappear, components that previously depended on chassis airflow must find another way to reject heat.

Enterprise SSDs are one of the clearest examples.

The storage device is no longer thermally independent from the rest of the server.

In a fully liquid-cooled AI rack, the SSD also has to become part of the liquid-cooling architecture.

NVIDIA Vera Rubin Changes the Thermal Baseline

NVIDIA describes the Rubin generation as its first AI infrastructure generation designed around 100% liquid cooling, with chips and networking components operating in closed-loop liquid-cooled systems without conventional server fans.

NVIDIA is also promoting hot-liquid cooling with coolant temperatures as high as approximately 45°C to reduce facility cooling energy and water consumption.

That changes the design assumptions for every component inside the AI server.

The traditional model was:

GPU/CPU Liquid Cooling + Air-Cooled Memory, Networking and Storage

The emerging model is moving closer to:

GPU + CPU + Networking + Storage → Integrated Liquid-Cooled Rack

Vera Rubin therefore represents more than a new GPU generation.

It is accelerating a transition toward an entirely new thermal architecture for the AI factory.

Storage Is Becoming Part of the AI Compute Path

AI storage is also changing functionally.

Enterprise SSDs are no longer used only as passive long-term storage devices.

They increasingly feed massive datasets into GPUs, support checkpointing, inference databases, vector databases, retrieval systems and AI model loading.

NVIDIA’s new STX architecture reflects this change by bringing Vera, Rubin and BlueField-4 into a common AI-native data infrastructure platform.

The closer storage moves to the GPU data path, the more important storage bandwidth, latency, power efficiency and thermal management become.

This means that SSD cooling is no longer a secondary mechanical issue.

It is becoming part of AI system performance.

Samsung PM1763: PCIe 6.0 Meets Direct Liquid Cooling

Samsung Electronics made this transition explicit in July 2026 when it announced mass production of its PM1763 enterprise SSD.

The PM1763 is based on PCIe 6.0 and Samsung’s ninth-generation V-NAND with a new 4 nm controller.

Samsung specifies sequential read speeds of up to approximately 28.4 GB/s and sequential write speeds of up to 21 GB/s.

But one of the most interesting features is thermal.

Samsung states that the PM1763 is optimized for liquid-cooled AI server environments and supports Direct-to-Chip cooling using a cold plate attached directly to the device.

This is a significant development.

An SSD is now being engineered from the beginning with cold-plate cooling in mind rather than treating liquid cooling as an external server modification.

Why PCIe 6.0 Makes SSD Cooling More Important

PCIe 6.0 provides roughly twice the interface bandwidth of PCIe 5.0.

But higher bandwidth normally brings higher signal-processing requirements and increasing thermal density.

At the same time, AI systems demand sustained rather than short-duration performance.

Thermal throttling becomes unacceptable when the SSD is feeding expensive GPU resources.

If the storage subsystem slows down because of temperature, GPU utilization may also fall.

That creates an economic argument for cooling the SSD properly.

The objective is no longer simply to keep the SSD below its maximum temperature.

The objective is to maintain predictable storage performance continuously.

Solidigm Started the Transition Earlier

On the SK side, one of the most important developments comes through Solidigm, the enterprise SSD company originating from Intel’s NAND business and acquired by SK hynix.

Solidigm worked with NVIDIA to develop one of the industry’s first directly liquid-cooled enterprise SSD solutions.

Its D7-PS1010 E1.S platform uses a cold plate designed for liquid-cooled, fanless AI servers.

The most interesting engineering feature is that a single-sided cold plate can cool both sides of the SSD.

This solves an important packaging problem.

An enterprise SSD contains heat-generating components on both sides, but conventional dual-sided cooling can make mechanical design and hot swapping difficult.

Solidigm’s approach transfers heat from both sides while maintaining a single cold-plate interface.

Hot Swap Is Critical

Cooling performance alone is not enough for a data center SSD.

Enterprise operators expect drives to be replaceable while systems remain operational.

This is where liquid cooling creates a mechanical challenge.

If a cold plate permanently traps an SSD inside the cooling assembly, serviceability becomes much worse.

Solidigm therefore designed the liquid-cooled D7-PS1010 around hot-swappable operation.

The company says its E1.S cold-plate configuration can dissipate more than 25 watts while maintaining the serviceability expected from enterprise storage.

This may become an important design principle for future liquid-cooled storage:

Cooling Performance + Hot Swap + High Density

All three are required.

The End of the Server Fan Changes Everything

The move to liquid-cooled SSDs is partly caused by the success of GPU liquid cooling itself.

When GPUs and CPUs are cooled directly by liquid, large internal server fans become less necessary.

That sounds positive.

But those same fans historically provided airflow across:

SSDs

NICs

DIMMs

Power Components

Voltage Regulators

When the fans disappear, secondary components lose their traditional cooling mechanism.

The industry therefore faces a second stage of liquid cooling.

Stage one was cooling the hottest processors.

Stage two is cooling everything that depended on the airflow generated for those processors.

SSD liquid cooling is a direct result of that transition.

SK hynix Is Also Bringing Cooling Into Memory Packaging

The thermal transition is not limited to Solidigm SSDs.

In May 2026, SK hynix announced its iHBM technology for next-generation High Bandwidth Memory.

The concept integrates cooling elements directly into the HBM package to create an additional thermal path and reduce thermal resistance.

SK hynix says the approach can reduce HBM thermal resistance by approximately 30%.

This is not the same technology as SSD cold-plate cooling.

But the strategic direction is similar.

Cooling is moving closer to the semiconductor itself.

Instead of designing the semiconductor first and solving cooling afterward, future memory and storage products increasingly need thermal architecture built into the package or device from the beginning.

Vera Rubin Makes This More Important for Korea

NVIDIA has identified SK hynix, Samsung and Micron as HBM4 suppliers within the Vera Rubin ecosystem.

NVIDIA and SK Group also expanded their strategic partnership in July 2026, including plans for large Vera Rubin-based AI factory infrastructure and long-term next-generation AI memory cooperation.

This puts Korea’s two major memory manufacturers directly inside the next phase of AI infrastructure development.

Samsung is advancing PCIe 6.0 liquid-cooling-ready enterprise SSDs.

SK hynix is developing new HBM thermal technology.

Its Solidigm business is already commercializing liquid-cooled enterprise SSD architecture developed with NVIDIA.

The convergence is becoming increasingly clear.

SSD Cooling May Become Another D2C Branch

Today’s AI rack liquid loop is often shown as:

CDU → Rack Manifold → Hose → QD → GPU/CPU Cold Plate

Future systems may add another branch:

CDU → Rack Manifold → SSD Cold Plate

This sounds simple, but it creates new engineering requirements.

An SSD cooling branch needs:

Compact Cold Plates

Low-Profile Hose Routing

Low Pressure Drop

Quick Disconnects

Blind-Mate or Hot-Swap Interfaces

Leak-Free Serviceability

Thermal Interface Materials

The mechanical system must fit around increasingly dense E1.S and E3.S storage devices without blocking service access.

The SSD Cold Plate Will Be Different From the GPU Cold Plate

A GPU cold plate can justify relatively large hoses and significant coolant flow because the GPU may dissipate hundreds or even thousands of watts.

An SSD is very different.

Individual power is much lower.

But an AI server may contain many drives packed closely together.

The problem therefore becomes one of density.

Instead of cooling one 1,000-watt device, the system may need to cool many 20- to 40-watt storage devices efficiently.

The SSD cooling system must therefore emphasize:

Very low pressure drop

Small manifolds

Thin cold plates

Simple mechanical connection

High device density

Fast replacement

That can create a new component market distinct from conventional GPU cooling.

Samsung and Solidigm Are Taking Different Paths to the Same Destination

Samsung’s PM1763 represents the next generation of performance-oriented enterprise storage.

It combines PCIe 6.0, high NAND density, improved controller efficiency and compatibility with cold-plate liquid cooling.

Solidigm’s approach focuses heavily on solving the mechanical integration problem for fanless AI servers.

Its single-sided cold plate, dual-side thermal extraction and hot-swap architecture demonstrate how the SSD itself can be redesigned around liquid cooling.

The two approaches are different.

But both point in the same direction:

enterprise SSD cooling is becoming part of AI server design.

DATAAD View: Liquid Cooling Will Expand Across the Entire PCB

DATAAD believes SSD cooling is part of a much larger transition.

The industry started with the GPU.

Cooling is now spreading to CPUs, networking silicon, HBM, SSDs and power-delivery components.

Eventually, the engineering question may no longer be:

Which chip needs liquid cooling?

It may become:

How should we thermally manage the entire server board?

That could lead to new architectures combining multiple cold plates, thermal bridges, integrated manifolds and eventually full-board liquid-cooling structures.

The traditional boundary between electronics design and cooling design is disappearing.

The Next Opportunity Is the “Small Liquid Connection”

GPU cooling created demand for large rack manifolds, CDUs, hoses and quick disconnects.

SSD and memory cooling may create demand for something different.

Much smaller connections will be needed.

Potential requirements include:

DN2–DN4 class fluid interfaces

Very small flexible hoses

Miniature blind-mate couplings

High-density manifolds

Low insertion forces

Robot-friendly service interfaces

When dozens of memory and storage devices require cooling inside one rack, connection density becomes as important as flow capacity.

This may create another important engineering market around liquid-cooled AI infrastructure.

From GPU Cooling to Full-System Cooling

NVIDIA Vera Rubin is accelerating the shift from component-level cooling toward system-level cooling.

Samsung’s PM1763 demonstrates that PCIe 6.0 enterprise SSDs are already being designed for liquid-cooled AI servers.

Solidigm demonstrates that liquid-cooled SSDs can retain critical data center requirements such as hot swapping.

SK hynix’s iHBM demonstrates that thermal management is also moving directly into advanced memory packaging.

These developments suggest that AI data center cooling is entering a new stage.

The first generation was:

Liquid-Cooled GPU

The next generation is becoming:

Liquid-Cooled AI Server

And ultimately the market may move toward:

Liquid-Cooled AI Factory

DATAAD Outlook

As Vera Rubin and future AI platforms increase compute, memory and storage density, cooling technology will spread beyond the traditional GPU cold plate.

Enterprise SSDs are likely to become an important part of this transition because AI infrastructure increasingly depends on sustained high-speed access to enormous datasets.

Samsung and SK hynix/Solidigm show that Korea’s memory industry is already moving in this direction.

The next major competition may therefore not be limited to NAND capacity or SSD bandwidth.

It may also involve:

Who can deliver the fastest storage, at the highest density, with the most efficient thermal architecture?

In the AI era, even storage is becoming a liquid-cooling technology.