Table of Contents
Dell’Oro research director Alex Cordovil told RCRTech that the real bottleneck is the power delivery chain, and specifically the long lead-time equipment that sits upstream of the white space — transformers, medium-voltage switchgear, and on-site generation
In sum – what to know:
Power bottleneck – Dell’Oro says transformers, medium-voltage switchgear, and on-site generation—not liquid cooling—remain the biggest constraints slowing AI data center deployments.
Thermal evolution – Direct liquid cooling is reshaping data center architectures, while heat rejection technologies are emerging as a critical but often overlooked part of AI infrastructure design.
Geographic shift – Permitting delays and community opposition could increasingly steer AI data center investment toward regions that can provide faster access to power.
Demand for AI infrastructure continues to reshape the data center supply chain, but the industry’s biggest obstacle is no longer cooling technology—it’s getting enough power to facilities quickly enough, according to Dell’Oro Group.
The research firm’s latest Data Center Physical Infrastructure (DCPI) report found the market grew 28% year over year to $12 billion in the first quarter of 2026, marking a fifth consecutive quarter of more than 20% growth as hyperscalers, neoclouds and colocators continue expanding AI infrastructure.
While direct liquid cooling has become one of the defining technologies of AI-era data centers, Dell’Oro research director Alex Cordovil told RCRTech that the industry’s primary bottleneck lies further upstream.
“The bottleneck is the power delivery chain, and specifically the long lead-time equipment that sits upstream of the white space — transformers, medium-voltage switchgear, and on-site generation,” Cordovil said. “These are the components with multi-year lead times and supply chains that were never sized for a buildout of this pace.”
He added that although liquid cooling receives considerable attention, it is rarely the factor delaying projects. “Liquid cooling gets the attention, but you can stand up a cold plate loop far faster than you can energize a substation,” Cordovil added.
While power remains the primary constraint, Cordovil said thermal technologies are increasingly becoming a point of differentiation for infrastructure vendors.
According to Dell’Oro, thermal management was the fastest-growing DCPI segment during the quarter, expanding nearly 50% year over year as direct liquid cooling (DLC) continues transforming data center cooling architectures.
Cordovil believes another part of the cooling chain deserves greater attention. “Thermal management grew nearly 50% year-over-year in Q1 2026, and Direct Liquid Cooling is now reshaping cooling architecture end to end rather than sitting as a bolt-on.”
“The part of that chain the market still under-appreciates is heat rejection — chillers, dry coolers, adiabatic systems. Once you have moved the heat off the chip, you still have to move it out of the building, and that is where site-level physics like water availability and ambient conditions start to dictate design,” he added.
The research firm also highlighted accelerating demand for AI-oriented power distribution technologies. Busbar revenue grew in the low-30% range during the quarter, outperforming the broader market as higher rack densities and Open Compute Project (OCP)-style DC power distribution become more common.
Beyond technology, Dell’Oro also sees permitting and community opposition emerging as new risks for AI infrastructure development, particularly in the United States.
According to Cordovil, these issues are extending project timelines at a time when rapid deployment has become a competitive advantage.
“Permitting delays and local opposition have grown into a genuine watchout over the past few quarters, particularly in the United States, and they compound with unsettled compute architectures and giga-scale projects slipping their schedules,” said Cordovil.
Rather than cancelling projects outright, he said, these challenges increasingly influence where new AI infrastructure is built.
“Opposition often doesn’t kill a project outright. It lengthens or relocates it, and in a market where time-to-power is the scarcest currency, a two-year delay is functionally a rejection,” he added.
Cordovil expects North America to remain the world’s largest DCPI market, accounting for more than half of global manufacturer revenue, but suggested future capacity could increasingly favor locations able to provide faster access to electricity.
“The regions best positioned are those that can offer speed to energization — markets with generation headroom, cooperative utilities, and a willingness to accept flexible load commitments in exchange for faster interconnection.”
He added that if community opposition becomes significantly stronger, Asia-Pacific markets could emerge as some of the largest beneficiaries of future AI infrastructure investment.
