AI Data Centers and the Race for Cooling Efficiency
As AI workloads drive higher rack densities and intensify demands on data center infrastructure, cooling efficiency is becoming a critical factor in operational performance, energy consumption, and long-term investment decisions. Shivakumar Vaidyanathan brings a multidisciplinary perspective to this challenge, combining expertise in mechanical engineering, MEP systems, sustainability, and data center services. Drawing on his experience in global data center services at Kyndryl and his work across commercial and industrial cooling environments, he offers practical insight into designing, operating, and optimizing cooling infrastructure. His expertise in AI/ML-based equipment performance modelling and IIoT also provides a valuable lens on how intelligent technologies can improve energy efficiency and support more sustainable operations. The article examines the key decisions shaping AI-ready data center cooling, including when air or liquid cooling is better suited to different rack densities, operating conditions, and site requirements. Mr. Vaidyanathan explores the practical challenges that can undermine performance and increase costs, from hydraulic complexity and evolving GPU requirements to inadequate sensor infrastructure and unreliable operational data. He also discusses how AI/ML-driven optimization can improve cooling performance, maintenance planning, and energy use, alongside strategies for reducing water consumption and meeting sustainability goals across diverse operating environments. For data center operators, infrastructure leaders, and investors, the article offers a grounded perspective on evaluating cooling technologies beyond headline efficiency claims. It highlights the importance of adaptable infrastructure, sound engineering decisions, and measurable operational outcomes in determining whether cooling investments can support future AI workloads without creating costly retrofits or unsustainable resource demands.
Q1. From your experience with data center cooling systems, what factors have determined whether liquid cooling delivers better economics than traditional air cooling?
There isn’t a single factor that determines when liquid cooling becomes more economical than air cooling; it’s really about how rack power density, climate, and utilization all come together.
Rack Power Density
In my experience, rack density stands out as the clearest indicator. If you’re running racks below about 15-20 kW each, traditional air cooling with hot and cold aisle containment is still cost-effective. But once you go beyond 40-50 kW per rack, which is now typical for GPU platforms like NVIDIA’s GB200/B300 and the upcoming Vera Rubin generation, air cooling starts to struggle. The sheer volume of airflow required at those densities means you end up paying a lot more for fan power, floor space, and cooling hardware, which quickly wipes out any cost advantage of air. At that point, liquid cooling (whether it’s direct-to-chip cold plates or immersion) becomes the clear winner because it can remove heat right at the source and does so with much lower energy overhead.
Climate
Climate is a bigger deal than most people realize. In places that are hot, humid, …
Create an account to continue reading
Create AccountAlready have an account? Sign in
Need an expert in this space?
Talk to an Industry Expert
Knowledge Ridge connects decision-makers with carefully vetted subject matter experts for one-on-one calls, research sprints, and advisory engagements — across 11 sectors and 163 sub-industries globally.
Comments
No comments yet. Be the first to comment!