Advanced Cooling Technologies for AI Data Centres

AI-driven demand is accelerating the pace of change in AI data centre cooling and data centre design.  As rack densities rise and traditional approaches are pushed to their limits, developers are being faced with increasingly complex decisions around performance, efficiency and delivery risk.

How are AI workloads changing heat density at rack level, and what does that mean for cooling design?

AI workloads are driving a fundamental increase in rack level heat density due to a widespread adoption of high-performance GPUs and accelerators. These require significantly more power than traditional CPU-based servers.

As a result, rack densities are increasing from traditional levels of 5–15 kW/rack to 30 kW+, with many AI deployments now reaching 50–100 kW/rack and beyond.

This increase in rack density has a direct impact on data centre cooling design. At these levels, air-cooled systems become increasingly impractical and can struggle to maintain operational temperatures inside data halls.

To resolve this challenge, data centre designs are increasingly adopting liquid-cooling solutions, particularly direct-to-chip cooling, which can remove heat far more efficiently than air.

In practice, many facilities are adopting hybrid cooling strategies, combining air-cooled systems for lower density loads and liquid cooling for high density AI workloads.

At what density do air-based systems stop being viable?

There is no single fixed density at which air-based systems definitively stop being viable. Suitability depends on multiple factors including maintaining design flexibility, energy restrictions, climate conditions, specific site constraints as well as CapEx and OpEx priorities.

In data centre applications, air based cooling typically becomes impractical when power densities exceed ~15–25 kW/rack. With typical row configurations of 24–40 racks, this equates to a design benchmark of ~600 kW/row.

This is generally the tipping point where the technical corridor length required to accommodate air cooled plant (at these densities – fan wall units) begins to approach or exceed the available white space, reducing the overall spatial efficiency.

As row density increases, air flow rates increase. This results in:

• increased air velocity through racks
• increased pressure losses
• increased fan energy
• increased noise levels
• difficulties maintaining uniform air distribution

Direct air handling units (DAHUs) can be used as an alternative to fan wall units, however, their suitability is highly dependent on local climatic conditions and they are subject to similar limitations as rack densities increase.

The key point is that while air cooled systems are almost always technically achievable, they are not always the most efficient solution. As rack densities rise, liquid cooling solutions become increasingly effective due to their higher heat transfer capability and reduced reliance on large air volumes.

How are cooling strategies selected on live projects?

The primary driver in selecting between direct-to-chip, immersion and hybrid cooling systems is rack density, as this determines the fundamental heat rejection requirement and whether air based systems remain viable.

However, on a live project, a wider set of factors influence the decision:

• technology maturity and operator preference
• operation and maintenance considerations
• flexibility and future proofing
• space and infrastructure constraints
• CapEx, OpEx and programme

Direct-to-chip liquid cooling is currently the most widely adopted liquid cooling solution at scale. Immersion cooling, while offering higher heat transfer performance, is less mature and typically reserved for more specialised applications.

Hybrid solutions allow a mix of air and liquid cooled racks in the same hall. This allows operators to flex to evolving tenant requirements and reduces the risk of overcommitting to a single technology while it is still in its infancy.

What are the main risks or constraints when delivering and operating liquid-cooled environments?

The introduction of liquid cooling in data centre designs presents a new range of risks to designers and operators. These risks must be carefully managed across the project lifecycle.

Leakage

Leakage is one of the most significant concerns. If distribution systems or components are not installed and maintained correctly, coolant can escape and potentially damage critical equipment.

Installation

Liquid cooling systems require careful installation. Mistakes during installation can increase the risk of leaks, and installers and operators need to be upskilled to ensure correct procedures are followed.

Operational complexity

Liquid cooling introduces additional operational requirements, including coolant management, water treatment and maintenance of cooling distribution units and distribution networks. There is currently limited field experience with these systems, and we expect designs to continue evolving as more high-density data centres enter the market.

Standardisation and integration

There is limited standardisation across vendors, which can create integration challenges and potential vendor lock-in.

Cost and supply chain

Liquid systems can carry higher upfront costs and rely on more specialised supply chains, which requires careful management to limit impacts on procurement and programme.

Liquid cooling is no longer a future concept, it is becoming a critical component in delivering high-performance, scalable data centre infrastructure.

If you’re exploring how these technologies could shape your next project, our team can support you in evaluating, designing and delivering the right strategy. Get in touch to discuss your requirements.

Copy article link and share