Hybrid Data Centre Cooling Emerges As The Near-Term Solution For AI Factories

Compute-intensive AI factories are scaling up AI training and fine-tuning, with expected rack densities exceeding 150kW – well beyond the capacity that shaped conventional air-cooled facilities. The growth of the liquid cooling market, which approached $3 billion in revenue in 2025 and is forecast to reach around $7 billion by 2029, clearly demonstrates the start of a shift from air to liquid cooling. This also reflects a broader reality: AI factories are becoming a major driver of compute density and cooling demand.

Why does that matter? Large AI training workloads behave differently than conventional enterprise workloads. At scale, they run in tightly synchronized cycles, alternating between compute-heavy and communication-heavy phases. This creates sharp swings in power draw and heat output across the cluster rather than a smooth, predictable load. Microsoft’s research on large-scale training infrastructure shows these swings become more pronounced as jobs scale across tens of thousands of GPUs, meaning cooling systems increasingly need to manage both high baseline density and short, intense thermal peaks.

For many operators, the practical decision is not whether to choose air or liquid cooling in absolute terms, but how to optimize cooling costs and operations simultaneously. To address this, organizations are deploying hybrid air-liquid cooling systems. Given that direct-to-chip liquid cooling can remove around 70 to 80% of rack heat at source, leaving air to handle the remaining load more economically, a hybrid system can both remove the required heat loads and optimize energy cost. With cooling still accounting for up to 40% of a data centre’s electricity use, this matters. Teams can target liquid cooling where thermal intensity is highest while avoiding the cost and complexity of redesigning the whole facility around peak conditions from day one.

Wide-scale adoption of data centre liquid cooling is limited by how difficult the technology is to scale with confidence. Uptime Institute’s 2025 Cooling Systems Survey highlights the practical issues operators are prioritizing as liquid cooling moves from the pilot phase into broader rollout. Analysis of the report reveals:

• Retrofit complexity remains a major hurdle. Nearly half of operators highlight retrofit practicality as a key consideration, reinforcing how difficult it can be to introduce liquid cooling into existing facilities without wider changes to layout, controls and supporting infrastructure. For this CAPEX-heavy investment, careful modular construction planning is a must to ensure that it can fulfil cost and performance requirements at scale, without compromising cascaded data centre systems.
• The investment case is not always straightforward. The survey suggests operators are still balancing the promise of lower operating costs against the upfront cost and disruption of deployment, which means the commercial case remains highly site-specific even where the technical case is clear. As economies of scale develop further, and cooling technology becomes more widely deployed, the operational and strategic payoffs will become clearer.
• Operational maturity is becoming a differentiator. Around a third of respondents cite both maintenance and reliability as concerns because liquid-cooled environments demand different maintenance routines, supplier coordination and failure-response processes than air-only facilities. This, in turn, raises the bar for in-house capabilities and outsourced service providers. Organizations must be invested in upskilling personnel to effectively implement liquid cooling at scale and mitigate operational risks.

Liquid cooling is becoming a core component of AI factory design, but the most practical near-term path is hybrid — supporting rising GPU density, managing variable training loads and building internal capability before committing to liquid-first at scale.

For more on data centre infrastructure strategy, see our reports on unified data centre management platforms and data centre software trends.