LiquidStack has moved its GigaModular coolant distribution unit (CDU) into general availability, with the company stating the platform has completed multi-module integration and full-load testing for deployments up to 14 MW and earned ETL certification. The single-phase, direct-to-chip platform is built to meet NVIDIA Vera Rubin specifications and is already carrying early customer orders for AI, HPC, and hyperscale sites.
That 14 MW figure is the headline change. When LiquidStack first unveiled GigaModular in mid-2025, the platform topped out around 10 MW per deployment; the commercial release validates a higher ceiling assembled from multi-megawatt building blocks. For operators, the relevant detail isn't the peak number so much as how that capacity gets added.
Why CDUs are now a planning constraint
The coolant distribution unit has quietly become one of the gating items in AI buildouts. As rack power densities climb toward 600 kW by the end of 2027 — driven by Vera Rubin and GB300-class hardware — air cooling is off the table and direct-to-chip liquid loops are the default. The CDU is the box that pumps, filters, and regulates coolant between the facility water loop and the cold plates on the silicon, and its capacity has to be committed before the racks land.
That creates a familiar over-provisioning problem: size the CDU for full design density on day one and you sink capex into cooling that sits idle until the GPUs arrive in volume. GigaModular's pitch is to break that with multi-MW increments. "Purpose-built for high-density AI environments, GigaModular helps operators expand cooling capacity without requiring significant infrastructure redesigns or over-provisioning upfront," said Joe Capes, Vice President at Trane Technologies and General Manager of LiquidStack.
The numbers builders care about
LiquidStack puts the GigaModular at roughly 25% lower capex and floor-space requirements than traditional CDUs, and rates the platform to operate in ambient temperatures as high as 122°F (50°C) — a meaningful spec for sites in hot climates or with constrained mechanical cooling. The architecture supports N, N+1, and N+2 redundancy, uses a high-efficiency IE5 pump design, and is offered either skid-mounted with pre-installed piping or as separate cabinets for on-site assembly.
What changes
For infra teams scoping next-generation halls, a validated, ETL-certified 14 MW platform with incremental scaling shifts cooling from a fixed upfront bet to something closer to the modular power and networking buildouts already common in AI factories. It also signals where the thermal envelope is heading: when a CDU vendor is testing to 14 MW and citing 600 kW racks, the cooling subsystem is no longer an afterthought bolted onto the GPU order — it is part of the critical path, and the lead times now matter as much as the accelerators themselves.
