By Eamonn Ryan
This article is derived from a presentation delivered by Michael Young, application engineer – thermal management at Vertiv and a regular contributor to RACA Journal, at FRIGAIR 2025.
Michael Young, application engineer – thermal management at Vertiv, at FRIGAIR 2025. Image by © RACA Journal
Young’s insights, shared on the topic of ‘Design of data centres for optimal energy efficiency’, explores the impact of artificial intelligence on data centre operations and the evolving strategies for effective thermal management.
The pervasive integration of Artificial Intelligence (AI) into our daily lives, from Gmail’s smart filters toYouTube’s recommendations and the transformative capabilities of ChatGPT, is driving an unprecedented surge in data centre energy consumption. This exponential growth presents a significant challenge for thermal management, demanding innovative cooling solutions and a re-evaluation of traditional data centre design.
AI-driven applications require immense computing power, directly translating into higher energy consumption. Projections indicate a staggering 5.3-fold increase in data centre energy consumption by 2032, with a threefold increase expected by 2030 alone. To put this into perspective, data centres are now being designed with capacities exceeding 40 MW, rivalling the power demands of entire industrial sectors.
This escalation is evident at the individual server level.Historically, a data centre rack consumed as little as 3 kW. With the advent of AI, high-density computing is pushing this figure dramatically, with racks now demanding 50 to 150 kW. This substantial increase in power density creates an equally formidable challenge in dissipating the generated heat, as electrical energy consumed by IT equipment is almost entirely converted into heat within the data centre environment.
The shift to liquid cooling: managing high-density heat
The sheer heat generated by AI-intensive racks makes traditional air cooling insufficient. Air simply lacks the thermal capacity and practical space to effectively cool such high-density environments. This necessitates a shift towards liquid cooling solutions.
Liquid cooling works by directly transferring heat from high- processing chips, such as GPUs (Graphics Processing Units) – the backbone of AI computations – to circulating cold fluid. This fluid flows across a cold plate attached to the chip’s back, efficiently absorbing the heat and carrying it away from the server.
A typical liquid cooling system comprises two main loops:
- Technology loop: This loop directly cools the IT servers. Hot fluid from the servers flows into a Cooling Displacement Unit (CDU), where it exchanges heat with colder water from the facility loop via a plate heat exchanger
- Facilities loop: This loop supplies chilled water to the CDU and is responsible for ultimately rejecting the heat to the outside environment, often through chillers or dry coolers.
Effective liquid cooling requires sophisticated control. Sensors monitor the supply water temperature to the racks, adjusting a three-way valve to regulate water flow back to the chiller. A pressure transducer then senses changes in water flow rate and adjusts the speed of variable-speed pumps. This dynamic control ensures optimal cooling precisely when and where it’s needed, adapting to varying computational demands (e.g., peak loads during events like Black Friday).
Designing for uptime: redundancy in data centres
Data centre downtime is incredibly costly and to mitigate this risk, redundancy is a non-negotiable principle in data centre design.This means implementing N+1 redundancy, where if two pumps are required for operation, a third is provided as a backup. Similarly, if 600 kW of cooling capacity are needed, two 600 kW cooling units might be installed to ensure continuous operation even if one fails. Each cooling unit within the system typically contains multiple pumps, a plate heat exchanger, expansion tanks, and comprehensive monitoring sensors and valves.
What could the data centres of the future look like?
Data centres that implement AI can have two types of construction, namely an existing data centre that implements AI and a new data centre.
With an existing data centre, low density server racks may still employ the use of air-cooled cooling units that operate on relatively lower water temperatures which may be in the region of 20°C to 30°C. To utilise the principle of free cooling, a separate water loop may be implemented to provide cooling for the AI servers whereby the water in this loop is at a higher temperature. Therefore, within this arrangement, we may possibly see data centres that contain two separate water loops, one for the low-density servers and one for AI servers.
This drive for high-temperature water management is a significant industry trend because it dramatically enhances the potential for free cooling. If a data centre’s AI cooling loop runs at 43°C, it will almost always operate in free cooling mode in climates like Johannesburg, where ambient temperatures rarely reach such high extremes.
In essence, the data centre of the future will be characterised by sophisticated thermal management, hybrid cooling approaches, and a relentless focus on energy efficiency driven by the ever- increasing demands of AI and advanced computing.
