With the development of AI, more pressure is being put on data centres and how they are cooled, and many are turning to hybrid cooling options to do this, which presents opportunities for resellers.
Data centres have been a part of the working environment for many years but are in the spotlight more than ever thanks to AI putting greater demands on them. It also means how data centres are cooled is more of a priority than ever.
“Traditional enterprise environments were often designed for more predictable workloads and lower rack densities, while many AI workloads, especially GPU-based training and inference, place far greater pressure on power and cooling,” says Karl Mendez, managing director of CWCS.
“In my experience, many existing data centres were not designed for sustained higher-density deployments at scale. Some can be adapted, but operators now need to think much more carefully about power, heat removal, resilience and future growth.
“What is changing most is not just peak demand, but the consistency and intensity of that demand. AI workloads tend to run hotter, for longer and with less tolerance for fluctuation, which exposes limitations in legacy infrastructure more quickly. That is why we are seeing a shift from incremental upgrades to more fundamental infrastructure planning, with operators designing for higher baseline densities and more flexible power and cooling strategies.”
Jason Beckett, head of architecture at Hitachi Vantara, adds that traditional environments were built for steady, lower-density compute. “Whereas AI introduces highly concentrated demand driven by GPU clusters, parallel processing and unpredictable load patterns,” he says.
“The impact is most visible at server rack level. Standard racks historically operate at around 5–15kW, but AI deployments can exceed 40kW and in some cases go far higher, putting pressure on power delivery, cooling and physical infrastructure.
“This creates a misbalance. Many facilities hit maximum rack power limits before they can fully populate AI hardware, forcing underutilisation or redesign.
“Cooling is another constraint, as traditional air-based systems struggle with the heat generated, accelerating the shift to liquid cooling.
“As a result, operators are having to rethink layouts, retrofitting where possible, or building new AI-specific environments designed for higher density, resilience and energy demand from the outset.”
Adhum Carter Wolde-Lule, a director at Prism Power Group, agrees, adding: “the industry is in a genuine retrofit scramble right now, and new builds are having to make completely different design decisions compared to what was standard five or 10 years ago.”
Advantages of hybrid cooling
With data centres requiring greater cooling power, hybrid solutions are becoming increasingly popular. “Hybrid cooling combines air and liquid cooling technologies to address varying heat loads across the data centre,” says Slawomir Dziedziula, senior director of application engineering EMEA – IT Systems at Vertiv.
“Liquid cooling is applied directly at the chip level to manage the highest density heat sources, while air cooling continues to manage lower density areas,” he explains. “This approach improves overall thermal efficiency, supports higher return air temperatures, and reduces energy consumption compared to air only systems. Hybrid cooling also provides flexibility, allowing data centres to support mixed workloads and gradually adapt as computing densities increase.”
Flexibility is another advantage, Adhum adds. “Hybrid cooling, which typically means combining traditional air cooling with some form of liquid cooling whether that’s direct-to-chip or rear-door heat exchangers, lets operators handle mixed workloads without locking themselves into one approach,” he explains.
“Air cooling on its own simply can’t move the heat that modern AI hardware generates. Pure liquid cooling is very efficient but expensive to deploy at scale and brings its own operational complexity. Hybrid sits in the middle: air where you need it, liquid where the density demands it, and the outcome is better PUE, lower energy cost and the ability to run GPU-heavy workloads without tearing the facility apart.”
Retrofit difficulties
Hybrid cooling can be retrofitted into existing data centres – but this can present challenges. “It needs to be planned properly,” says Karl. “It is not simply a case of adding new cooling hardware. Operators also need to consider power design, floor layout, heat rejection, resilience, maintenance access and the impact on live customer infrastructure. In live environments, operational risk is often the most important factor, particularly where customers are running business-critical systems.
“Any changes need to be carefully phased to avoid disruption, with clear segregation of works and a staged deployment approach where needed. There are also structural and mechanical considerations that can limit what is achievable in older facilities, so feasibility assessments are essential. In many cases, hybrid cooling can extend the life and capability of an existing data centre quite effectively, but it needs to be approached as an integrated infrastructure change rather than a standalone upgrade.”
Rich Day, mechanical engineering manager at Mercury Power, agrees that when it comes to installing hybrid cooling into an existing data centre every site has its own challenges.
“Retrofitting evaporative cooling into an existing data centre can be straightforward on one site and a complete headache on another,” he says. “Often, the cooling itself is not the biggest issue. You also have to look at available power, space, existing infrastructure and in some cases, the structure of the building itself.
“If the data centre is spread across multiple floors, for example, you must ask whether the building can take the weight and load of the new AI equipment and any associated plant.
“It all comes down to what the site can realistically support, from a cooling perspective and a wider infrastructure point of view.”
Reseller conversations
For resellers, hybrid cooling presents opportunities. “Resellers and MSPs should focus on outcomes, not just the cooling technology itself,” says Karl.
“Customers want to know whether the environment can support the workloads they plan to run, whether it will remain efficient as demand grows, and whether it gives them a realistic path to scale. It is important to translate what can be a technical topic into clear business implications, including performance stability, cost predictability, energy efficiency and long-term flexibility.
“Hybrid cooling should be positioned as an enabler of performance, resilience and future capacity. There is also a growing advisory and commercial opportunity here. As more businesses look at how their infrastructure needs to evolve for modern AI workloads, resellers that can offer genuine consultative support, not just hardware, are best placed to add value and build longer-term customer relationships.”
Jason says resellers should also emphasise scalability. “Hybrid approaches allow infrastructures to evolve in stages, avoiding large-scale redesign while still supporting future growth.
“Efficiency and cost control are equally important. By applying liquid cooling only where it is needed, organisations can manage energy use and operational spend more effectively than with a full transition.
“Future-proofing is another key message. As compute density continues to rise, hybrid models provide a pathway to adapt without repeated disruption.”
Discussions should also focus on flexibility, scalability and protection of existing investments, adds Slawomir. “Hybrid cooling enables higher density AI workloads without forcing immediate, largescale infrastructure overhauls,” he says. “It supports phased deployment, improves energy efficiency and provides a clear path toward future cooling strategies as AI requirements continue to evolve.”
Increasing demand
With AI use increasingly, it is expected that hybrid cooling for data centres will continue to grow in popularity. “Hybrid cooling is fast becoming the default design standard for any data centre handling AI workloads,” says Shaheed Salie, senior technical manager at FiveNines Group.
“The global data centre sector is in an infrastructure investment supercycle — JLL forecasts up to $3 trillion in investment by 2030 and a doubling of global capacity. A significant proportion of that will include hybrid cooling as a core specification.
“In Europe, the EU Energy Efficiency Directive and growing sustainability pressure from hyperscalers mean that operators who cannot demonstrate efficient cooling will struggle to win contracts or secure planning approvals.
“Also, AI inference – the process of running live AI applications – is expected to overtake AI training as the dominant workload by 2027, according to JLL. Inference requires geographically distributed, lower-latency facilities, meaning hybrid cooling will need to be deployed at scale across secondary and edge markets too, not just in hyperscale campuses.”
Shaheed also notes that the talent and expertise required to design, install and commission hybrid cooling systems is already in short supply. “Demand will significantly outpace available skills in the near term,” he says.
Adnum adds that the compute density of AI hardware isn’t plateauing. “Each generation of GPUs is drawing more power and generating more heat, and air cooling as a standalone solution becomes less viable with every product cycle,” he says. “My expectation is that hybrid cooling becomes the default specification for new data centre builds within the next two to three years and retrofit demand will keep growing as operators try to keep existing facilities relevant rather than writing them off.”
Slawomir agrees that demand for hybrid cooling will intensify as AI adoption accelerates. “AI deployments are expanding from pilot projects to full production scale, increasing the need for scalable and adaptable thermal solutions,” he adds. “Hybrid cooling supports this transition by accommodating rising rack densities while maintaining operational continuity and efficiency across diverse computing architectures.”
