Trane Technologies has expanded its data centre thermal management system portfolio with two new 250-megawatt (MW) high-density AI factory reference designs, which are based on the NVIDIA DSX AI Factory platform.
The platform delivers high energy efficiency and zero water consumption cooling architectures.
The new reference designs build upon Trane Technologies’ industry-first thermal management system reference design for gigawatt-scale AI factories and add technologies from recently acquired companies LiquidStack and Stellar Energy.
The new 250 MW blueprints deliver zero water consumption cooling architectures that maximise Token per Watt and reclaim significant electrical capacity for compute workloads. They also introduce Trane Technologies’ first reference design featuring Stellar Energy modular cooling plants, which help reduce time on the job site.
Additionally, the LiquidStack GigaModular Coolant Distribution Unit (CDU) platform integrated across both reference designs has qualified as an NVIDIA DSX Ready, meeting applicable NVIDIA DSX reference design requirements for liquid-cooled AI infrastructure.
“When Trane Technologies introduced the industry’s first comprehensive thermal management reference design for gigawatt-scale AI factories, we set a new standard for holistic system efficiency across compute, facilities and climate systems,” according to Mauro J. Atalla, senior vice president and CTO at Trane Technologies.
“With our latest designs, we are incorporating new technologies and improvements that help data centre developers overcome skilled labour shortages, compress deployment timelines and reclaim critical power for AI factories.”
NVIDIA president of AI Infrastructure, Vladimir Troy, said efficiently scaling gigawatt-scale AI factories demands tightly coordinated power, cooling and compute infrastructure that operates as one unified system.
“Trane Technologies’ new cooling reference designs incorporate LiquidStack’s GigaModular CDU, which has received the NVIDIA DSX Ready qualification,” he said.
“These new modular reference designs reduce integration risk, accelerate deployment and help operators maximize useful AI compute within available power, cooling and grid constraints.”