How liquid cooling enables AI data centres

A high-tech data centre cooling system shown as a network of glowing transparent pipes weaving through a dark server hall, with heat waves rising into a night sky where wind patterns subtly shape thei

As AI workloads push computing densities to new extremes, data centres across the UK, US, UAE, and Kazakhstan are re-evaluating their thermal strategies. The shift from air to liquid cooling is no longer optional — it’s a prerequisite for scalability, efficiency, and long-term operational control. But the change goes beyond simply swapping cooling methods. It represents a fundamental rethinking of how cooling systems integrate with hardware, power, and facility design.

Recent insights from Shell Global Solutions, validated through collaboration with NVIDIA and the Open Compute Project (OCP), highlight that the industry’s focus has moved from raw cooling performance to lifecycle reliability, material compatibility, and system-wide sustainability. Garry Hwang, associate technology manager at Shell, notes that operators now demand more than just heat transfer — they want cooling solutions that support long-term operations, reduce risk, and can be monitored and managed.

Fluids as system enablers, not commodities

Coolants are no longer seen merely as heat-transfer media. In modern direct liquid cooling (DLC) systems, they are integral to system stability. As data centres operate at higher temperatures to improve waste heat recovery and reduce reliance on primary cooling, coolant stability becomes critical. Thermal degradation, sludge formation, or phase separation can compromise heat transfer and damage sensitive components.

Material compatibility is equally important. Coolants must remain inert not only with metals but also with plastics, seals, and hoses, especially under elevated temperatures. Hwang warns that mixing incompatible fluids — even from different suppliers — can lead to precipitation or corrosion, undermining system integrity. This challenges the outdated view of coolant as a commodity, akin to gasoline.

Shell has responded by developing fluids validated under NVIDIA’s DSX Ready Coolant program, ensuring compatibility with high-performance AI hardware. The company also emphasizes real-world testing over lab simulations, noting that field deployments reveal issues around contamination, fluid aging, and maintenance that simulations often miss.

AI density demands holistic thermal design

AI and high-performance computing (HPC) racks now exceed traditional air-cooling limits, with some GPUs generating thermal loads that air cannot dissipate efficiently. Direct-to-chip cooling, where cold plates extract heat directly from processors, is becoming standard in high-density deployments.

This approach reduces reliance on high-speed fans, lowering both energy consumption and noise. However, liquid cooling does not eliminate heat — it relocates it. The heat is transferred via a secondary loop to a Coolant Distribution Unit (CDU), then to facility-scale chillers, dry coolers, or cooling towers. The efficiency of this entire thermal pathway, from chip to atmosphere, determines overall performance.

Thermal modelling using computational fluid dynamics (CFD) is now essential to prevent hot-air recirculation, especially in large AI campuses. Poorly positioned rooftop chillers, for instance, can create localized heating or reduce cooling efficiency if discharged air is drawn back into intakes. Site layout, wind patterns, and neighbouring structures all influence dispersion.

The noise-heat trade-off in sustainable design

Cooling infrastructure is also a major source of noise. Chillers, pumps, and backup generators contribute to sound levels that can affect nearby communities, particularly at night. Yet noise reduction measures like acoustic barriers can inadvertently restrict airflow or worsen heat recirculation.

CtrlS, a data centre operator, integrates acoustic planning from the outset, using enclosures and site layout to manage sound propagation without compromising cooling. The key is holistic design: evaluating combined acoustic impact, not just individual equipment specs, and considering both daytime and nighttime conditions.

High-voltage infrastructure adds complexity

Beyond cooling, the rise of AI data centres is reshaping electrical systems. Many new facilities require direct connections to high-voltage (HV) transmission lines — 66kV, 110kV, or higher — due to insufficient capacity in local distribution networks.

This introduces new commissioning challenges. HV systems demand specialist testing for transformers, switchgear, protection relays, and grid interfaces. Incorrect settings or installation flaws can lead to arc flash, equipment damage, or grid instability. Commissioning must begin at project inception, not as a final step, to ensure safety and reliability.

What it means

For data centre operators, liquid cooling is no longer a tactical upgrade — it’s a strategic enabler of AI scalability. The choice of coolant, compatibility with hardware, and integration with power and heat-rejection systems directly affect operational risk, efficiency, and regulatory compliance.

Businesses should treat cooling as a full lifecycle partnership, not a one-time procurement. Selecting a supplier with real-world validation, material science expertise, and integration capabilities — such as Shell’s collaboration with energy and CDU teams — can reduce downtime and extend system life.

Operators must also align thermal, acoustic, and electrical planning from day one. This includes engaging HV commissioning specialists early, using CFD for heat dispersion, and designing noise controls that don’t compromise airflow.

In markets like the UAE and Kazakhstan, where new AI campuses are being built, siting within industrial zones can ease environmental integration. In the UK and US, regulatory scrutiny of noise and heat discharge is increasing, making proactive design essential.

The goal, as Hwang puts it, is clear: cooling should not limit AI growth — it should enable it.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *