ZutaCore has launched a two-phase direct-to-chip cooling system designed to help hyperscalers manage the rising thermal demands of AI infrastructure without the need for facility water. Announced at the OCP APAC Summit, the company’s new sidecar solution integrates with legacy air-cooled data centres. It enables up to 240kW of cooling capacity using a closed-loop, waterless design.
The system supports either two 120kW racks or four 60kW racks, addressing a growing challenge for operators struggling to deploy next-generation GPUs in environments never built to accommodate such heat density. With many of the world’s largest hyperscale facilities still relying on air cooling, the sidecar offers a transitional approach for expanding AI compute capacity without major retrofits.
“Many of the world’s largest hyperscale data centres rely solely on air cooling, which makes it extremely difficult to support the thermal demands of next-generation AI GPUs,” says My Truong, Chief Technology Officer at ZutaCore. “Our sidecar solution addresses this challenge directly, enabling more efficient AI adoption within existing infrastructure.”
Cooling constraints meet AI acceleration
As demand for AI workloads intensifies, thermal management has become one of the primary limiting factors in scaling compute infrastructure. Operators are increasingly forced to contend with the mismatch between traditional air-cooled facilities and the extreme heat output of modern GPU clusters, often reaching 60kW or more per rack.
Liquid cooling is widely seen as the most viable path forward, but the shift is complicated by the cost and disruption associated with installing facility water systems. The new ZutaCore solution avoids that hurdle entirely by using a sealed, water-free system that mounts alongside standard racks and leverages a two-phase heat exchange process.
According to Dell’Oro Group, demand for direct liquid cooling is set to more than double between 2025 and 2029. Research Director Alex Cordovil says that hybrid architectures will be key to this growth. “Liquid-to-air technologies offer significant advantages for existing facilities aiming to deploy AI workloads with minimal disruption and incremental investment,” says Cordovil.
The sidecar’s design reflects this incremental logic. By enabling cooling upgrades without removing existing air-cooled systems, it helps operators extend the life and value of their infrastructure while introducing the efficiencies needed to support modern AI workloads.
Driving efficiency through design
The system’s two-phase cooling approach delivers significant gains in energy efficiency. While single-phase systems typically require 1.5 litres of coolant per minute per kilowatt of heat, ZutaCore’s solution operates at just 0.3 litres per minute per kW, cutting pumping power requirements by 80 per cent.
It also replaces traditional copper fin-and-tube radiators with an aluminium microchannel condenser, reducing weight and increasing heat exchange efficiency by up to 40 per cent. Key components such as pumps and fans are hot-swappable and configured with N+1 redundancy, ensuring continuous operation and simplifying maintenance.
Smart controls monitor flow, temperature, and pressure automatically and integrate with existing data centre infrastructure management (DCIM) tools for remote oversight and automation. Together, these features offer operators a way to future-proof their environments against the exponential power demands of AI training and inference workloads.
Rather than advocating wholesale change, the sidecar presents a pragmatic middle ground—one that acknowledges the urgency of AI deployment while recognising the long tail of legacy infrastructure still powering much of the digital economy.
As model complexity and training scale continue to rise, solutions that can bridge the gap between what was built and what is now required will shape the trajectory of AI infrastructure over the next decade.
