Analysts expect the data center physical infrastructure (DCPI) market will grow at a 10% CAGR between 2022 and 2027, driven in part by vendor backlogs and fewer supply chain constraints. But artificial intelligence (AI) workloads, including generative AI, will primarily be responsible for the market's growth, Dell'Oro Group Research Director Lucas Beran told SDxCentral.

Due to the rise of data center planning and design changes meant to incorporate AI workloads, that 10% growth rate represents an upward revision from the analyst firm's initial five-year forecast. And while AI's impact is significant enough to warrant such revision, "this growth won’t materialize overnight," Beran noted. He expects data center designs created to support AI will begin materializing at scale in 2025, given vendors' existing backlogs from "waning but still present supply chain constraints."

Microsoft, for example, is seeing increased demand for its AI platform and plans to "accelerate investment in our cloud infrastructure," CFO Amy Hood noted during the hyperscaler's Q4 2023 earnings call. "The real focus here is being able to be aggressive in meeting the demand curve and focusing on the transition and growth," Hood told investors.

As far as the types of infrastructure Microsoft will direct investments toward, "it’s both on the data centers and the physical basis, plus CPUs, GPUs and networking equipment," she said.

Alphabet CFO Ruth Porat similarly highlighted Google Cloud's increase in data center server capex this last quarter, "which included a meaningful increase in our investments in AI compute," she told investors.

Even though AI is already contributing to "a broad proliferation" of high-performance computing (HPC) infrastructure, continued market growth will require investments in next-generation DCPI that supports higher energy and thermal management needs, Beran argued.

More specifically, AI infrastructure will necessitate shifts in rack power densities from less than 10 kilowatts (kW) per rack to 20 kW per rack – "which already isn't too uncommon today" – to 50 kW per rack and more than 100 kW per rack. "This means bringing more power," he said.

From a facility perspective, the data center technology doesn't necessarily need to change, but there will need to be more of it "to support new computing capacity" driven by AI, he said. "The ideal data center physical infrastructure solutions to support AI workloads exist today, but not at the scale or maturity that the data center industry would like to see," he explained.

A major infrastructure change organizations will need to invest in is higher power rated rack power density units (rPDUs). "This will also accelerate the transition from basic rPDUs to intelligent rPDUs, which provide the ability to remotely monitor and manage your rack power infrastructure and optimize operations and reliability," Beran said.

High power consumption at the rack level will also generate more heat than standard data center infrastructure. Currently, data center operators can support about 20 kW of power density per rack using air-cooling, but managing AI infrastructure "at scale with cost, efficiency and sustainability in mind will require transitioning from air to liquid cooling."

Despite the relatively small size of the liquid cooling market, it's "growing fast," Beran said. "The biggest inhibitor here is the lack of operational readiness for data center operators to incorporate liquid cooling in their data center designs. To overcome this, data center operators are deploying AI infrastructure in existing facilities, but with low rack utilization," he noted.

As an alternate solution, data center operators "may also retrofit existing facilities with air-assisted liquid cooling, such as a rear door heat exchanger or direct liquid cooling (DLC) with heat rejection at the back of the rack, into a hot aisle," he added.

To that point, DLC is beginning to emerge as a leading thermal management solution and boasts early momentum and ecosystem support. Form factor is the biggest characteristic driving DLC's early success because "it utilizes existing vertical rack infrastructure that end-users are familiar operating," Beran said. Server OEMs have also begun shipping "DLC-ready servers" that need fewer modifications than immersion cooling–ready servers.

"It’s also worth noting that DLC has been utilized in the HPC industry for a number of years, so there are already best practices learned there that can be applied to the growing number of DLC deployments," he added. "The transition from air to liquid cooling will be an evolution, not an overnight revolution. However, for early adopters, it will drive the best performance and economics for the coming wave of AI IT infrastructure."