Subscribe to Our Digest
Artificial intelligence is transforming data centres at a rapid pace, but the computational power driving breakthroughs in machine learning comes with a critical challenge: unprecedented heat generation. Modern GPU clusters produce thermal loads that legacy cooling systems simply cannot manage, forcing the industry to reimagine thermal management from the ground up.
The numbers tell a compelling story. While traditional servers might generate 5-10 kilowatts per rack, AI-optimized configurations now routinely exceed 100kW, with cutting-edge deployments approaching 140kw. NVIDIA’s H100 and H200 GPUs each dissipate 700 watts of heat, while the upcoming B200 will reach 1,000-1,200 watts per chip. The company’s GB200 NVL72 system—combining 72 GPUs and 36 CPUs in a single rack—demands an extraordinary 140kW of cooling capacity.
These extreme densities reveal the fundamental limitations of air cooling, which has served as the industry standard for decades. Heat simply cannot be removed quickly enough using air alone, prompting facilities to embrace liquid and immersion cooling technologies that were once considered exotic but are rapidly becoming essential.
Air cooling relies on fans, computer room air conditioning units, and carefully designed hot/cold aisle containment to manage temperatures. The approach works well for moderate densities but faces insurmountable physics at AI-scale heat loads.
The main issue is efficiency. Air conducts heat poorly compared to liquids—approximately 3,000 times less effectively. This means massive volumes of air must be constantly circulated to remove heat, consuming enormous amounts of energy. Data centres that rely solely on-air cooling typically have Power Usage Effectiveness (PUE) ratios between 1.6 and 2.0, meaning that for every watt consumed by computing equipment, an additional 0.6 to 1.0 watts are needed for cooling.
Practical limitations arise around 15-20kW per rack. Beyond this threshold, the air volume required becomes impractical, and cooling efficiency declines sharply. Since modern AI racks routinely surpass 50kW, air cooling alone cannot satisfy the demand.
That said, air cooling retains advantages for specific scenarios. It requires minimal initial investment, leverages widely available expertise, and is suitable for legacy infrastructure handling traditional workloads. For facilities operating at less than 20kW per rack, air cooling remains practical and cost-effective. However, for AI-intensive computing, it represents an increasingly outdated method.
Liquid cooling has emerged as the preferred solution for modern AI data centres, using water or specialised coolants to absorb and transport heat with far greater efficiency than air. The most common method, direct-to-chip cooling, circulates coolant through cold plates mounted directly on heat-generating components such as GPUs, CPUs, memory modules, and power delivery systems.
These cold plates contain intricate microchannels through which coolant flows, absorbing heat via conduction before being pumped to a Coolant Distribution Unit (CDU). The CDU transfers thermal energy to the facility’s chilled water system, which ultimately dissipates the heat outside the building. This design enables rack densities exceeding 200 kilowatts while significantly improving energy efficiency.
Leading infrastructure providers have developed sophisticated solutions for liquid cooling deployments. Vertiv’s Liebert XDU series offers capacities from 70kW to 1,368kW, while Schneider Electric (through its Motivair acquisition) provides comprehensive portfolios, including CDUs, dynamic cold plates, and rear door heat exchangers specifically engineered for AI workloads.
The performance improvements are significant. Research carried out by NVIDIA and Vertiv showed that implementing direct-to-chip cooling for 75% of a facility’s thermal load resulted in an 18% decrease in facility power consumption and more than a 10% reduction in overall data centre power. Optimized liquid cooling systems achieve PUE ratios between 1.3 and 1.4—a significant improvement over air-cooled equivalents.
Energy savings of 40-80% compared to traditional air cooling are consistently reported. Additionally, liquid cooling reduces server fan requirements by up to 80%, creating a compounding efficiency benefit. Modern cold plate designs can keep GPUs 35°C cooler than air-cooled alternatives, extending hardware lifespan whilst enabling sustained maximum performance during prolonged AI training runs.
Rear door heat exchangers provide a transitional approach for existing facilities. These devices attach to the back of server racks, capturing hot exhaust air and transferring its thermal energy to chilled water before the air enters the data centre environment. Unlike direct-to-chip cooling, rear door exchangers do not require modifications to servers themselves, making deployment relatively straightforward. They effectively remove 60-70% of server heat, making them suitable for moderate high-density applications that surpass traditional air cooling capacity but may not need full immersion capabilities.
Immersion cooling is the most advanced thermal management technology, submerging entire servers in dielectric fluid that directly absorbs heat from all components. This innovative method eliminates server fans completely, allowing for the most compact and energy-efficient data centre designs possible.
Two primary methods. Single-phase immersion cooling circulates dielectric fluid that remains liquid throughout the thermal cycle, with pumps moving the heated fluid to external heat exchangers. Two-phase immersion uses engineered fluids with low boiling points; as components generate heat, the surrounding fluid boils, carrying thermal energy away through vapour that then condenses and returns to the system. The two-phase method requires no pumps, which reduces complexity and energy consumption.
Industry pioneers including Green Revolution Cooling, Submer Technologies, LiquidStack, Iceotope, and Asperitas have developed comprehensive immersion solutions for AI and HPC applications. Submer’s SmartPod platforms support rack densities up to 140kW with PUE ratios between 1.03 and 1.1—far superior to the global data centre average of 1.56. LiquidStack’s two-phase DataTank technology has achieved cooling energy reductions exceeding 90% in production deployments.
The efficiency advantages extend beyond just energy. Immersion cooling eliminates the need for traditional CRAC units, hot/cold aisle containment, raised floors, and extensive air distribution infrastructure, allowing space savings of approximately two-thirds compared to similar air-cooled setups. Water use is nearly eliminated in many cases, as closed-loop systems operate without the need for evaporative cooling. Microsoft, for example, deployed closed-loop chip-level liquid cooling, which eliminated evaporative water use, resulting in a reduction of more than 125 million litres of water consumption per facility annually.
The uniform thermal contact provided by immersion ensures consistent temperature distribution across all components, eliminating hotspots and enabling reliable operation at peak computational loads. This thermal consistency supports sustained high-performance computing essential for multi-week AI training runs.
However, immersion cooling is the most complex and capital-intensive method. Facilities need specialised immersion tanks, dielectric fluid systems, and often require complete architectural redesigns. Server hardware must either be compatible with immersion or be retrofitted for fluid submersion, and maintenance procedures differ significantly from traditional methods. This technology is most cost-effective in new-build facilities designed for immersion from the start, where high-density requirements and sustainability goals make the investment worthwhile.
Each cooling technology serves distinct use cases based on power density requirements, infrastructure constraints and operational priorities.
Air cooling remains suitable for legacy data centres and facilities that predominantly house traditional server workloads below 20kW per rack. It provides the lowest initial costs and easiest deployment, but cannot efficiently manage modern AI densities.
Liquid cooling has become the industry standard for AI and HPC workloads, supporting rack densities from 20kW to over 200kW. It delivers substantial efficiency improvements with moderate infrastructure investment and can be deployed incrementally, allowing facilities to transition progressively as workload densities increase.
Immersion cooling suits extremely dense applications exceeding 50kW per rack, where maximum performance and sustainability are priorities. It provides unmatched thermal performance and energy efficiency but requires the highest capital investment and the most significant infrastructure changes.
Leading facilities increasingly adopt hybrid methods, combining air or liquid cooling for legacy and moderate heat loads with immersion for state-of-the-art racks. AI-driven analytics and real-time sensor networks further optimize these systems, dynamically adjusting cooling mechanisms to match fluctuating computational loads while reducing energy consumption.
The data centre cooling market is undergoing rapid transformation, with nearly 40% of facilities now adopting liquid cooling technologies. The liquid cooling market is expected to grow at over 20% annually, reaching a value of £7 billion by 2028.
New data centre builds consistently achieve PUEs of 1.3 or better, demonstrating that facilities with advanced cooling from the start significantly outperform legacy infrastructure. However, the industry-wide average PUE has plateaued at approximately 1.56, as nearly half of all facilities are over 11 years old and struggle to adopt newer technologies easily. Further efficiency improvements will require not only innovation but also significant infrastructure modernization across the existing facilities base.
As AI workload densities continue to rise—with next-generation GPU clusters potentially exceeding 1 megawatt per rack—the limitations of traditional cooling methods become increasingly clear. Liquid and immersion cooling technologies that once seemed exotic just years ago are quickly becoming standard for AI-intensive environments.
The thermal challenges posed by artificial intelligence demand fundamental changes to data centre infrastructure. Air cooling, while still viable for specific applications, cannot meet the demands of modern GPU-accelerated computing. Liquid cooling has emerged as the industry standard, delivering substantial efficiency improvements and enabling rack densities that air alone cannot achieve. Immersion cooling is the ultimate solution for ultra-high-density applications, offering unmatched performance at the expense of increased investment.
The future belongs to hybrid approaches optimized by intelligent control systems, with facilities deploying multiple cooling technologies tailored to different workload types. As the industry faces both rising computational demands and increasing environmental pressures, advanced thermal management is no longer optional—it is essential infrastructure for the AI-driven data centre landscape.