Liquid cooling.
Designed into the build.

Specify the complete thermal path

Liquid cooling moves heat from compatible computing equipment into a coolant circuit. The practical design question is how that heat travels from the servers to the site’s heat-rejection equipment, and how the whole system behaves under load.

For a GPU container, the cooling plan needs to account for both the selected servers and the site. Server requirements, supply temperatures, flow, pressure and coolant compatibility must align with the equipment suppliers’ specifications. Any residual air-cooling demand also needs to be covered.

QIT’s offering is centered on liquid-cooled GPU infrastructure. The final cooling architecture, equipment selection and performance commitments are established through project engineering. We do not assume the same design will fit every GPU platform or climate.

How a QIT container is cooled

Every QIT container uses direct liquid cooling, with a water block on every GPU and CPU. A sealed closed loop recirculates 1.5 m³ of water continuously to dry coolers, so the container consumes no water and produces no wastewater. Temperature holds within 2 °C under sustained load, and the system runs quietly enough for sites close to users. The full container specification lists capacity, GPUs and dimensions.

Liquid cooled data center methods: direct-to-chip and immersion cooling systems

A liquid cooled data center removes most of the heat load with a liquid instead of air. Water and specialized coolants carry far more heat per unit volume than air, which is why liquid cooling supports the high density racks used for AI and high performance computing. In a traditional data center, air cooled rooms rely on CRAC units, raised floors and chilled water coils; those systems struggle as rack power rises.

Each cooling solution trades efficiency, serviceability and cost differently. Many GPU platforms now ship designed for direct-to-chip cooling, so confirm what the server vendor supports before choosing immersion. QIT selects the method during project engineering rather than assuming one approach fits every platform.

Energy efficiency, benefits and risks of liquid cooling

The main benefits of a liquid cooled data center come from moving heat with a fluid rather than supply air:

• Density: liquid can dissipate heat from racks that air alone cannot cool, so more computing power fits in the same space.
• Energy efficiency: fewer and smaller fans, and the option of warmer coolant temperatures, can reduce cooling energy. Higher efficiency must be shown at stated conditions, not assumed.
• Less noise and more stable temperatures at the server, which supports sustained performance under full load.
• Heat reuse: warm return water is easier to capture than warm exhaust air.

The risks are different from air cooling, not absent. Leaks, fluid quality, pump failures and loss of flow need monitoring and a tested response process. Immersion cooling technology adds its own questions: the cooling fluid, whether a synthetic dielectric or a mineral oil, must suit the servers, and each immersion cooling tank changes how technicians lift, drain and maintain equipment. Water cooling loops outside the container need treatment and freeze protection. A critical review of these items is part of thermal management, not an afterthought.

Modular data center cooling and immersion cooling containers

Modular data center cooling packages the thermal system with the IT space. In a container, the CDUs, pumps, piping and controls are installed and tested in the factory, and the unit connects on site to dry coolers, cooling towers or chillers that reject the heat.

Immersion cooling containers are one published example. GRC’s ICEtank is a 40 ft containerized immersion data center with eight 42U ICEraQ immersion tanks, rated by GRC at up to 480 kW. DUG Technology’s DUG Nomad uses 10, 20 and 40 ft shipping containers with single-phase immersion. Direct-to-chip designs reach high density in a container format too; Vertiv’s MegaMod CoolChip, for example, pairs direct-to-chip liquid cooling with supplemental air in a prefabricated module.

Whatever the method, the container is only part of the data center cooling system. The external heat rejection, its power demand, water use and noise belong in the same scope, along with energy efficiency targets stated at defined operating conditions.

Sources: GRC edge products; DUG Nomad; Vertiv, July 25, 2024. Vendor figures, reviewed November 2, 2026. Compare suppliers in the modular data center companies guide or read what is a modular data center.

Three interfaces to resolve early

01 — GPU platform to coolant system

Choose the compute platform and obtain its operating envelope. Confirm the server-side connections, coolant requirements, maintenance access and conditions for safe operation. A different rack-scale platform can change the supporting infrastructure.

02 — Distribution to facility loop

Define the coolant distribution equipment, isolation points, controls and monitoring. Record which party owns each interface and the procedures for filling, flushing, testing and servicing the system.

03 — Facility loop to heat rejection

Establish how the site removes heat and under what ambient conditions. The external equipment, available space, power and local constraints are part of the facility design, even when the compute arrives in a container.

NVIDIA’s DSX POD reference design also distinguishes factory integration from the prepared site’s supporting power, water, cooling and network infrastructure.

What to ask before choosing a system

• Which GPU and server platforms are included in the validated design?
• What are the full-load and partial-load thermal requirements?
• How are leaks, loss of flow and out-of-range temperatures detected and handled?
• Which components still need air cooling?
• What maintenance tasks can be performed while the system operates?
• Who supplies, commissions and maintains the external heat-rejection equipment?
• What operating conditions apply to efficiency or water-use claims?

These questions turn a broad cooling promise into a reviewable engineering scope. They also expose dependencies that could affect the delivery schedule.

Plan your QIT cooling scope

Send the proposed site, available electrical capacity and intended GPU workload. If you already have a server vendor or rack design, include its thermal specifications. If the platform is still open, describe the workload and the capacity you need.

We use those inputs to discuss the container configuration and the work needed outside it. The 15-hour installation requires a prepared site with power and cooling interfaces in place.

Liquid cooling at a Texas site

Establish the outdoor design conditions and the selected GPU platform’s coolant limits before choosing heat-rejection equipment. A closed server loop does not by itself establish zero site water use: check the external cooling design and its operating conditions.

Compare dry and evaporative heat rejection on space, electrical demand, water needs and performance at the design ambient temperature. Include residual air cooling and maintenance access in the same scope. Use the site checklist to connect the thermal design to the rest of the project.