Beyond chips: the power and cooling bottleneck in AI

In July, Vertiv unveiled its first manufacturing base in Southeast Asia, located in Senai, Johor. This plant is tasked with producing CoolChip liquid cooling distribution units (CDUs) and prefabricated SmartRun power systems – the kind of gear that keeps multi-GPU racks from overheating as they push toward 100 kW per rack. “At these levels, you physically can’t get enough power into a rack without moving to higher-voltage 800V DC designs,” explained Paul Churchill, Vertiv’s VP for Asia. Those designs are already emerging in the US and are expected to roll out in Asia soon. The Johor factory is not just a capacity play; it’s a direct answer to the supply chain stresses that have plagued global shipping.

A factory built for speed

The approximately 236,000 sq ft facility has been operational since the first quarter and has been shipping its XDU liquid-cooling line for several months. Full technical capacity is expected by mid-2027, with later phases adding new liquid-cooling lines and larger SmartRun systems. Vertiv says these prefabricated systems can cut on-site deployment time by up to 85%. Producing inside the region slashes lead times and avoids the freight disruptions that rocked the industry in recent years. The plant is targeting up to 500 skilled jobs, and Churchill noted that about 98% of the current workforce are local hires, a result he attributed to strong recruitment drives rather than the need to bring in outside labour.

Johor’s tightening market isn’t slowing demand

While Johor has carried the bulk of Malaysia’s data center buildout, state agencies are now becoming more selective, weighing new projects’ power and water consumption against other economic sectors. Of the 52 data center projects approved nationally, 17 are operational, with the rest moving through construction and approval. Vertiv, however, remains unfazed. Asked whether AI infrastructure spending is in a bubble, Churchill said customer demand remains “extremely healthy,” underpinned by a three-to-five-year investment horizon across Asia. The company is building for that demand and expects continued growth.

Why this matters for self-hosted AI

For organizations evaluating on-prem deployment of large language models, the news goes beyond industrial geography. Liquid cooling and high-voltage power are the silent enablers of enterprise-grade GPU clusters. Without infrastructure that can handle densities above 100 kW, running LLMs on your own hardware hits a physical wall long before it becomes affordable. A regional factory churning out CDUs and prefabricated power modules shortens lead times, cuts shipping risks, and bolsters technological sovereignty – a non-trivial advantage for teams that need to control latency and data residency. As Asia’s AI ambitions scale, the ability to cool and power those racks locally will be just as critical as the chips inside them.