Table of Contents
Joe Capes, CEO of LiquidStack, told RCR Wireless News that he expects single-phase direct-to-chip cooling to dominate the next three to four years
As AI workloads push data centers to new thermal and power limits, liquid cooling is no longer optional, according to Joe Capes, CEO of LiquidStack, a provider of advanced cooling technologies.
“Liquid cooling is a means of removing heat or providing thermal management for data center components,” Capes told RCR Wireless News. “What’s really happened in the last two years is with the advent of AI scale-up and much higher power from GPUs and CPUs, we’re seeing the adoption of direct-to-chip liquid cooling most prominently.”
This method, he explained, circulates a cooled liquid directly over processors through a cold plate, efficiently capturing and removing heat. Though the technology dates back to IBM’s mainframes in the 1950s, Capes said its broad adoption today is driven by AI’s escalating power demands.
“The three most pressing pain points are power, power, and power,” Capes emphasized. AI data centers, he noted, are “very power intensive,” creating bottlenecks in mature markets like Northern Virginia, where energy and grid constraints are slowing expansion.
This is pushing AI infrastructure growth into new regions, including parts of Southeast Asia, the Middle East, and South America, where power availability remains higher. But scaling is also constrained by limited supply of key infrastructure components such as switchgear, transformers, backup generators, and coolant distribution units (CDUs)—critical for deploying direct-to-chip systems.
According to Capes, the notion that liquid cooling is complex or costly is outdated. “The biggest misconceptions are that liquid cooling is difficult or expensive to deploy,” he said. “Using liquid cooling can be 40% to 50% more efficient than air cooling for AI workloads.”
Still, the transition poses a learning curve. Many operators, he said, are implementing liquid cooling “for the first time or the first time at scale,” making vendor experience and global after-sales support essential for success.
Looking ahead, Capes expects single-phase direct-to-chip cooling to dominate the next three to four years. As chip power levels exceed 1,500 watts per processor, data centers will need to shift toward two-phase cooling and even hybrid solutions combining immersion and direct-to-chip systems.
“The expectation is we’re going to see one- to two-megawatt IT racks in the future,” he said. “To support those heat loads, we will need to move to two-phase.”
Environmental sustainability is another priority. Capes urged the chemical industry to accelerate development of PFAS-free refrigerants and environmentally safe fluids, warning that current compounds “were developed to last a long time” and can persist in the environment.
LiquidStack previously announced its GigaModular platform, a 10-megawatt modular CDU system launching globally in the last quarter of the year. The design enables a “pay-as-you-grow” model, allowing operators to scale AI cooling capacity as demand rises.
“We think that having the flexibility of deploying CDUs in a modular form factor, and at a scale that supports higher power densities, will be really helpful in solving customer challenges,” Capes said.
As the AI race accelerates, Capes believes the industry’s ability to balance performance, power, and sustainability will determine who leads in the next generation of data center design.
