Liquid Cooling Solves One Problem – and Reveals Another
AI data centers are being redesigned around a simple physical reality: modern GPUs and CPUs now dissipate heat at levels that air cooling can no longer manage efficiently. As processor power climbs past the kilowatt mark, liquid cooling has become a practical requirement. Cold plates and manifolds now sit at the center of next-generation server architectures, carrying thermal energy away from flagship chips with far greater efficiency than fans ever could.
At first glance, this transition appears unambiguously positive. GPU and CPU temperatures stabilize, performance ceilings rise, and the thermal margin required for advanced AI workloads becomes achievable. But as with many large engineering shifts, the change has second-order effects that only emerge once systems are deployed at scale.
AI server rack unit featuring low-power data center subsystems, including memory modules, SSDs, retimers, voltage regulators, and optical transceivers.
One of those effects is subtle but increasingly consequential: as liquid cooling is introduced, system fans are reduced or removed altogether. The fans that once moved large volumes of air across the entire server are no longer required to cool liquid-cooled processors. Space constraints from plumbing, power budgets, and cost pressures accelerate their removal. What remains is a server that is highly optimized for its largest heat sources – and increasingly inhospitable to everything else.
The Disappearance of “Free” Airflow
For decades, air cooling provided more than just targeted thermal management. Large system fans created a broad, continuous airflow path from the front of the server to the back. That airflow cooled CPUs and GPUs, while also actively managing the thermal needs of memory modules, SSDs, retimers, voltage regulators, and optical transceivers through shared system airflow.
In deployed systems, many of these components are assigned to thermal zones, with fan speeds dynamically ramping based on their temperature, ensuring adequate cooling even without dedicated heatsinks or cold plates.
Liquid cooling changes that dynamic. It is precise by design. Heat is removed exactly where cold plates are attached and nowhere else. As fan arrays disappear, the secondary components that once relied on incidental airflow are left in increasingly stagnant thermal environments.
These components were never designed with full liquid cooling in mind. Many are hot-swappable. Others sit in dense board-level assemblies with limited surface area and no practical way to attach cold plates. Extending liquid cooling to each of them adds cost, complexity, leak risk, and flow-balancing challenges that quickly outweigh the benefit.
The result is a growing class of what engineers now informally call “left-behind” components: parts that operate at relatively low power compared to GPUs, but that are thermally sensitive and increasingly exposed.
When Small Temperature Rises Create System-Level Limits
In isolation, a few extra degrees on a memory module or retimer might seem inconsequential. In practice, these components define hard operational limits. Memory devices begin to throttle as they approach temperature thresholds. SSDs reduce write speeds to protect data integrity. Retimers and voltage regulators lose efficiency and reliability as junction temperatures rise. Optical transceivers experience accelerated aging and signal degradation.
Unlike GPUs, these parts fail quietly. Performance drops incrementally. Latency increases. Error rates climb. Mean time to failure shortens. At the system level, these effects aggregate into reduced throughput, lower uptime, and higher maintenance costs.
Operators often respond in the most direct way available: by increasing fan speeds on the few fans that remain. That response restores airflow, but at a steep energy cost. Fan power scales nonlinearly with speed; a modest increase in RPM can produce a disproportionate rise in power consumption. At hyperscale, small changes compound quickly, turning secondary cooling into a meaningful contributor to operating expense.
This creates a paradox. Liquid cooling is adopted to improve efficiency and unlock performance, yet its introduction can force operators into energy-intensive mitigation strategies to protect components that liquid cooling does not reach.
Optical Transceivers: A Case Study in Thermal Blind Spots
Optical transceivers illustrate this problem particularly well. In modern AI servers, the high-power digital signal processing sections of transceivers are increasingly liquid-cooled inside the chassis. But the optical subassemblies – the portion that extends outside the server cage and handles the electrical-to-optical transition – remain exposed.
OSFP transceiver featuring µCooling fan-on-a-chip
As fans are reduced, these external sections lose the airflow they once relied on. Yet they still dissipate several watts of heat in a confined space. Simulations conducted by xMEMS engineers show that targeted airflow applied directly to these optical sections can reduce temperatures by nearly 10 degrees Celsius – a meaningful improvement for reliability and power efficiency, and one that avoids reintroducing high-power fans.
The broader lesson extends beyond transceivers. As server architectures evolve, cooling blind spots emerge wherever liquid cooling stops and airflow disappears.
Why Extending Liquid Cooling Isn’t the Answer
It is tempting to treat these issues as temporary growing pains that will be resolved by extending liquid cooling to more components. In practice, that approach faces steep barriers. Liquid systems require flat interfaces, precise pressure control, leak detection, and serviceability considerations that conflict with the design of many secondary components.
Hot-swappable devices such as memory modules and SSDs are poorly suited to fixed plumbing. Board-level components like retimers and voltage regulators lack the mechanical real estate for cold plates. Each additional cooling branch increases system complexity and risk.
From a systems perspective, the challenge is no longer about removing kilowatts of heat from a single chip. It is about maintaining thermal balance across dozens of smaller components without undermining the efficiency gains that liquid cooling delivers.
Targeted Airflow as a Complementary Strategy
This is where localized, solid-state airflow becomes relevant. Rather than attempting to recreate system-wide air cooling, micro-scale airflow can be applied precisely where it is needed. Small, solid-state micro-coolers can deliver directed airflow to specific hotspots – retimers in dense assemblies, memory banks, SSD controllers, or the exposed portions of optical transceivers.
The power required for this kind of localized cooling is modest compared to ramping large fans. The airflow is confined, controllable, and predictable. In effect, it restores the thermal function that system fans once provided, without reintroducing their energy and reliability drawbacks.
xMEMS has been working with server and component vendors to explore how solid-state micro-cooling can be integrated at both the component and system level. In some cases, micro-cooling is embedded directly into subcomponents, giving suppliers control over thermal performance independent of the server’s overall cooling design. In others, airflow is distributed through small manifolds to cool densely packed regions efficiently.
Cooling the System, Not Just the Chip
The evolution of AI data centers mirrors earlier shifts in computing. When processors first outpaced power supplies, system design adapted. When memory became a bottleneck, architectures changed. Cooling is undergoing a similar transition.
As liquid cooling expands, system fans are often reduced or eliminated, leaving fewer fans to maintain thermal margins for components that still require air cooling. Those remaining fans must ramp to higher RPM, which drives a nonlinear jump in power draw, noise, and mechanical stress. Localized micro-cooling delivers targeted airflow to these air-dependent components, easing thermal bottlenecks and reducing the need for aggressive fan ramping. In hybrid systems, micro-cooling power is modest versus the energy penalty of high-RPM fans, while providing confined, repeatable airflow.
As AI infrastructure continues to scale, the most impactful thermal innovations will focus less on cooling the biggest chips and more on maintaining balance across the entire server. Solving the hidden cooling bottleneck is not about replacing liquid cooling – it is about completing it.