# AI Transforms Data Centers into Power and Cooling Plants — retained record

> Retained news record from news:datacenterknowledge.

Canonical: https://highsignal.app/data/records/e6cab527ae5017ce

[← sources](https://highsignal.app/data)


source: [news](https://highsignal.app/data/news) (news:datacenterknowledge)  published 2026-05-14 11:16 ingested 2026-05-15 08:50

[original source ↗](https://www.datacenterknowledge.com/build-design/ai-transforms-data-centers-into-power-and-cooling-plants)

## extract

AI Transforms Data Centers into Power and Cooling Plants
Dense GPU clusters are reshaping data center design and widening the gap between AI factories and legacy enterprise facilities.
AI infrastructure is upending how data centers are designed. As GPU clusters push rack densities far beyond historical norms, operators are devoting more campus area to cooling plants, electrical systems, and mechanical equipment than to server rows. The change is splitting the market into two infrastructure models: conventional enterprise facilities built for traditional workloads and purpose-built AI environments engineered for extreme compute density.
For decades, most facilities operated within relatively stable rack densities using air-cooled x86 servers. AI systems are breaking those assumptions. According to data shared by Vladimir Galabov, senior research director for cloud and data center research at Omdia, hyperscale deployments are moving sharply upward in density. He estimates that Meta’s average rack density rose from roughly 18 kW in 2022 to about 34 kW in 2025. Maximum rack density climbed to approximately 130 kW with the deployment of Nvidia Blackwell NVL72 systems.
Galabov expects that trajectory to continue as future AI systems push toward dramatically higher power envelopes.
“From where I stand, the industry is ready for a 600 kW rack,” Galabov said, noting that such figures would have been unthinkable five years ago.
AI Factories Rebalance the Data Center
The density increase is also changing how data centers allocate floor space.
Traditional enterprise and cloud facilities were largely built around “white space” – the areas occupied by IT equipment. AI factories invert that balance, according to Steven Carlini, Schneider Electric’s chief advocate for AI and data centers.
“AI factories have very limited white space, and the majority of the physical space is for the ‘grey space’ chillers, generators, transformers, switchgear, etc.,” Carlini said. Cooling plants expand, electrical systems scale up, and mechanical equipment occupies a larger portion of the footprint than in conventional environments.
Hyperscalers Pull Away from Legacy Enterprise Facilities
Hyperscalers and cloud providers are moving first, and many are choosing to build new AI-focused facilities rather than retrofit older environments.
“The top 20–30 buyers are the biggest AI deployers,” Galabov said, adding that many operators are likely to separate infrastructure fleets instead of modernizing large populations of legacy facilities.
The result is a growing separation between hyperscale AI campuses and legacy enterprise facilities.
“It’s extremely challenging to convert a legacy data center into an AI factory using the latest GPUs,” Carlini said. Operators must determine whether existing utility feeds, electrical systems, racks, and cooling architectures can support dense AI deployments. Many older facilities were never designed to meet the heat and power requirements associated with modern GPU clusters.
For enterprises, the economics and timelines may favor colocation or hosted AI infrastructure over internal upgrades.
“It’s early in the game,” Carlini said. “Many enterprises are running AI pilots or are just starting to deploy working models hosted in cloud and colo.”
Liquid Cooling Adds Another Infrastructure Layer
Rising rack power is also driving widespread adoption of liquid cooling.
Direct-to-chip cooling is becoming standard in many hyperscale AI deployments. But liquid cooling is not replacing air cooling outright. Operators are increasingly managing environments that require both systems simultaneously.
“The cooling complexity [is] driven by the fact that these AI factories have both air cooling and liquid cooling,” Carlini said.
That creates parallel cooling systems operating at different temperature ranges inside the same environment.
“You need two cooling architectures as the chillers run at different temperatures,” Carlini said. “You also need plumbing for the liquid cooling and ducting for the air cooling.”
The added complexity is increasing demand for software that can coordinate thermal and power performance across mixed infrastructure.
Cooling towers and dry coolers on rooftops are growing in both count and footprint as heat loads climb. (Image: Alamy)
Standardized AI Pods Emerge
As GPU clusters consume more power and cooling, custom deployments become more difficult to design and operate.
Vendors are responding with standardized AI pod architectures, reference designs, and prefabricated infrastructure modules to simplify the deployment of high-density racks.
Carlini pointed to growing standardization around IT room power distribution and rack infrastructure as deployments move toward next-generation GPU systems. “This is pushing the industry to standardized IT Pods as reference designs and or even prefabricated modules,” he said.
Galabov said that Open Compute Project specifications and Nvidia reference architectures are helping standardize approaches to rack power and liquid-cooling interfaces.
Vendors are also packaging cooling and rack infrastructure into integrated systems. Recent examples include collaborations between OptiCool and Belden on combined rack hardware and rear-door liquid cooling designs.
Five years ago, a 100 kW rack was considered extreme. Today, hyperscalers are planning infrastructure measured in multiples of that figure. Many enterprise facilities were built for conventional server workloads, not dense GPU clusters that demand far higher heat and power loads.

## retained text

No document body was retained for this record — only its metadata was stored.