In 2023 we wrote that AI and high-performance computing were pushing racks toward 75 to 100 kW, and that liquid cooling was becoming a requirement. That ceiling is gone. The NVIDIA GB300 NVL72 racks FEA is designing for today are rated for up to 140 kW each, and next-generation platforms are expected to go higher.
Liquid cooling is no longer a question of whether. The questions owners bring us now are practical ones: what water temperature, how much redundancy in the liquid path, where the pipe goes, and whether an existing building can take it at all.
Why air runs out
The physics have not changed. Air has a specific heat of about 0.24 BTU/lb-°F; water's is 1.0, roughly 4.17 times higher. Water is also about 832 times denser (62.4 lb/ft³ versus 0.075 lb/ft³). Multiply the two and a given volume of water carries nearly 3,500 times as much heat as the same volume of air.
Air systems compete by moving more air: bigger fans, bigger ducts, bigger coils, and server fans running faster and louder while drawing more UPS power. With good containment and controls, air is efficient in the 10 to 20 kW per rack range, and rear-door heat exchangers stretch it further. At 140 kW, air alone is not an option.
What a liquid-cooled hall actually looks like
Most of the AI capacity we are designing uses direct-to-chip cooling. Heat moves through a chain of loops, and each link is a design decision:
- Cold plates and rack manifolds pick up heat at the GPUs and CPUs.
- Coolant distribution units (CDUs), mounted in the rack or in the row, separate the server coolant from the building water and control its temperature, pressure and flow.
- The technology cooling system (TCS) is the secondary loop between the CDUs and the IT equipment.
- The facility water system serves the CDUs, often as a tempered chilled water loop, and connects to the plant and heat rejection.
Liquid does not take all of the heat. A direct-to-chip rack still rejects part of its load to air, and the network and storage racks around it, typically 20 to 40 kW, are air cooled. That is why our designs pair CDUs with rear-door heat exchangers, sometimes on different water systems. The room still needs an air strategy; it is just a much smaller one.
The design decisions that matter
Water temperature
Supply water must stay above the room dew point to avoid condensation inside the IT equipment, and within the server maker's limits. Within that window, warmer is better: it allows more economizer hours and less compressor time. On existing sites, a tempered loop blended up from central plant chilled water is often the fastest way to serve liquid-cooled racks without building a new plant.
Redundancy and maintainability of the liquid path
A CDU, a pump, a valve or a section of TCS pipe is now as critical as a UPS module. Every one of them needs a way to be serviced without dropping racks: redundant CDUs, isolation valves, and a header or loop arrangement that keeps cooling flowing while a section is out. Thermal ride-through at these densities is measured in seconds, so CDU pumps and controls generally need uninterrupted power.
Water quality, materials and leak management
The TCS is a closed loop with tight requirements for filtration, fluid chemistry and wetted materials, usually set by the server manufacturer. Flushing and cleaning before the IT load is connected, leak detection, and drip protection over electrical equipment belong in the design, not in the field.
Pipe routes, weight and power
Dense racks, filled piping and CDUs add floor load, so a structural review comes early. Overhead and under-floor routes each carry trade-offs for access and leak risk. On the electrical side, high-density racks call for larger feeds, 400 V class distribution or 480Y/277V to the rack, and multiple power shelves per rack.
What it means for the central plant
Liquid cooling does not shrink the plant: the heat still has to go somewhere, so the plant is sized for the full planned load. The gains come at the terminal end and in operating temperatures. Pipes take far less space than ducts and air handlers, which frees floor area for IT and power.
A plant built around warmer water can run in economizer mode for many more hours each year. In milder climates, dry coolers may carry the load year round. In the Northeast, a hydronic economizer with chillers that run only on peak days is a practical middle ground. Both reduce compressor energy, and closed-loop heat rejection cuts make-up water and chemical treatment.
Where we have done this
- 150 kW per rack, in 2015A 10 MW HPC conversion for a research client, with custom ASICs at 150 kW per rack cooled through custom heat exchangers on a tempered chilled water loop. Spare substation capacity and central plant chilled water were put to work instead of building new infrastructure. See the project
- A 40 MW AI GPU data centerBuilt for NVIDIA GB300 and Vera Rubin platforms, with rack-mounted CDUs and rear-door heat exchangers on a TCS loop, fed by a tempered chilled water loop and air-cooled chillers. The first 16 MW phase was completed in 2026. See the project
- A 1 MW AI pod inside a live enterprise data centerGB300 NVL72 racks designed for up to 140 kW each, with row-based CDUs and rear-door heat exchangers, in a facility that stays in production throughout. In design, 2026. See the project
Start with what you have
For many owners, the first question is not which CDU to buy but whether their existing building can support liquid-cooled racks at all. That answer comes from a capacity study of power, chilled water, structure and space, with a phasing plan that keeps the critical load online.