Q&A with Molex: data centre infrastructure demands

Power and cooling requirements 

The number one biggest trend shaping data centre infrastructure is AI, Smith said. As a result of AI, the kinds of questions that are being asked are, “How do you build data centres faster? How do you build data centres that can handle the power requirements, the cooling requirements?” 

The environmental impact of running data centres is also being thought about, alongside construction shortages, because of the sheer demand being placed on people to build data centres.

It is difficult to quantify at what stage we are in the AI boom - whether demand for data centres might calm down, or even amp up - but Smith said his customers were still providing signals that they want to continue to grow their business, so the boom seems to show no signs of slowing down.

AI has “unique requirements” over other applications that use data centre infrastructure, such as data storage and Cloud services. 

“One is higher power density,” Smith explained. “You’re often over 30 kilowatts per rack due to the intense computational needs of an AI workload. This necessitates very advanced cooling solutions. There are folks working on different types of liquid cooling, whether it’s direct to chip liquid cooling, whether it’s immersion cooling. There’s a lot of advancement in cooling technologies.”

Sometimes cooling can interact with the connectivity infrastructure, he said, and compete for space, which has to be thought about too. “You need to put all of this equipment really close together, and you’ve got to make it very scalable and flexible to accommodate the evolving nature of those AI applications.” 

Fibre optics 

Molex’s area of expertise is in connectors and optical connectivity solutions, including fibre optics, which Smith referred to as “the plumbing in your house,” with regards to how fibre optics are used to connect everything up. Molex’s fibre optics solutions are being shaped by customer demands - and there were a few examples Smith pinpointed to demonstrate how they’re being adapted for data centres used to run AI applications.

One example of this is in looking at the materials of connectors: “If you’re going for an immersion cooling solution, you’ve got to make sure that any transmission equipment inside that cooling tank is compatible with the liquids inside.

“So there’s been some material advances to make sure that things like fibre jacketing and connectors are compatible with those cooling liquids.”

The second example is making sure liquid in cooling tanks don’t escape. Being cognisant of this problem, Molex launched VaporConnect, which connect optical transceivers inside a cooling tank to cabling outside. “It provides that vapour barrier and provides the connectivity in and out of a tank.” 

“Another is the evolution of small form factor connectors,” Smith added. “There’s this new range of multi-fibre connectors that have come onto the market that can give you up to three times the density of historical connectors, while also reducing the space it needs and the amount of cleaning you need to do.” 

In his final remarks, Smith said that any business investing in AI infrastructure must understand their specific needs, their workflow, talent constraints, and the challenges they aim to solve with using AI. Making use of advancements happening in technologies ranging from cooling solutions to fibre optics is also going to be key.