CASTROL ON
That redesign is being driven by the pace at which AI chips are evolving. Server generations are turning over every few months. Power densities are rising in step. And the infrastructure layer – cooling, piping, power – must keep pace with hardware that was designed with performance, not thermal compatibility, as its primary constraint.
Stranded power and the economics of heat density The commercial stakes of poorly executed cooling strategies are high. Peter frames the issue not in engineering terms but in revenue terms, which reflects a broader shift in how operators are beginning to think about thermal management.“ Think about AI as an accelerator,” he explains.“ You want to make sure it runs at full capacity. If you have 30MW running on AI, and your cooling is not up to the standard required for those 30MW, you’ re only going to generate 15MW. What’ s going to happen to the other 15MW? It’ s going to be stranded power.”
Stranded power – capacity that exists on paper but cannot be utilised because the cooling infrastructure cannot handle the load – is emerging as one of the defining problems in highdensity AI deployments. For operators selling compute capacity, this efficiency issue evolves into a direct drag on revenue.“ Once you’ re able to cool those usable watts, it’ s revenue for you and it’ s cash for the operators and the end customer,” Peter says.
“ We are no longer talking about individual designs – we’ re talking about a systemlevel redesign, a structural redesign of the data centre”
Peter Huang Global President of Data Center and Thermal Management Castrol
The bottlenecks extend beyond stranded power. Peter identifies scalability – the ability to match the newest rack densities – as the first challenge, followed by water supply and closed-loop design readiness for infrastructure at scale, and then execution reliability. That last point carries particular weight. Liquid cooling introduces failure modes that air cooling does not, and the consequences of those failures in a live AI environment are severe.“ We have seen so many situations on the ground,” Peter says.“ We have seen people, without the proper execution and on-site services, where things are leaking. When the water and the liquid cooling start leaking, it damages the expensive chips and the servers, and it becomes a mess.”
6 June 2026