In the AI gold rush, data centre infrastructure vendors are selling shovels
As Computex in Taiwan wraps, the prominent presence of infrastructure vendors on the floors is noted, with a preponderance of liquid cooling technology on display. It is, of course, one thing demonstrating a PC or even a rack chilled by neon-lit plumbing, and quite another reticulating water supply and piping throughout a data centre. Seeing the hardware necessary for the latter is both fascinating and raises questions for the challenges involved in, among other things, retrofitting existing facilities with cooling systems capable of handling the growing heat generation associated with the high-performance computing sitting behind AI.
Schneider Electric Secure Power Division VP Pankaj Sharma sat down with Techday and explained that yes, the mechanical components of a Cooling Distribution Unit (that's the pumps, heat exchangers, and other physical, mechanically engineered bits) are relatively straightforward to build. At scale is where the magic happens. "Precision in managing temperature variability through firmware and software is what sets high-quality systems apart," he said. "And in AI data centres, where racks loaded with for example NVIDIA GPUs can cost millions, even slight temperature deviations can damage hardware or reduce efficiency, leading to significant financial losses."
He explained that dynamically controlling coolant temperature and flow to maintain optimal conditions for high-density compute is paramount. "This requires firmware to monitor and adjust parameters in real-time, ensuring stability and efficiency. For example, maintaining water temperatures within a tight range (as little as ±1°C) is critical for liquid-cooled systems supporting AI workloads."
Now, data centres were always expressly designed to keep water out, because water doesn't mix well with computers. That's a simple first principle now turned on its head as increased heat generation leaves no choice but to pipe it in, along with additives like ethylene glycol, good old antifreeze. Again, Sharma summons the challenge of scale and says delivering at scale is what sets Schnieder Electric apart from the multitude of cooling systems prominently in neon at Computex.
"Many companies can produce mechanical cooling systems, but few have the expertise to integrate advanced software/firmware for precise thermal management," he said. "This is where companies like Motivair, with a decade of experience in supercomputer data centres, stand out."
While it had progressed some work on its own-developed liquid cooling solutions, Sharma said Schneider Electric's recent acquisition of Motivair is an effective leapfrog over competitors, as the company comes with decades of experience in delivering air and liquid cooling in demanding data centre environments. Cray Supercomputer was mentioned – now there's a blast from the past.
Sharma explained that Motivair's chops extend to material science for pipelines and connectors capable of resisting corrosion, avoiding leaks, and degradation over time which could knock out data centre infrastructure. "Scaling liquid cooling across diverse global environments (different climates, power grids, and facility designs) demands expertise and most competitors lack this depth of experience," he said.
In fact, as AI data centres flourish like mushrooms on a global scale, Sharma agreed that the probability of a high-profile mechanical – or hydraulic – failure owing to insufficient heat management is quite likely. That's because the challenges are expressly mechanical, even to the extent of floor loading: redesigned brownfield data centres supporting higher-density computing must also support literally denser racks. Water is heavy, necessitating more robust racks, and even stronger floors.
What is certain is that in the AI goldrush, and the emergence of the data centre as the 'AI factory', there will be winners and losers. That's a consistent feature of every stampede in the direction of certain riches. On the Witwatersrand of AI, companies like Scheider Electric might not be digging out the lumps of yellow stuff directly, but without their expertise and tooling, nobody else can either.
Donovan Jackson is attending Computex as the guest of Schneider Electric. AI must live somewhere, and among other things, Schneider Electric makes data centre infrastructure.
