For an enterprise-owned data center, a new AI project can turn liquid cooling from a future consideration into a current planning decision. A university adding research computing or a pharmaceutical company deploying AI may need infrastructure its existing server rooms were never designed to support.
The decision is not simply whether to buy a liquid-cooled server. It is whether the facility can support the proposed equipment’s power and cooling requirements while protecting the workloads already running there.
AI and high-performance computing can concentrate power and heat into a few racks. Having spare cooling capacity somewhere in the building does not necessarily mean that capacity can reach the equipment that needs it. The rack, distribution systems and heat-rejection equipment need to work together.1
Direct-to-chip cooling removes heat at the processor. It brings a supporting system of coolant distribution units (CDUs), manifolds, connections, controls and monitoring. The heat still has to leave the building. Most arrangements connect to facility water; a liquid-to-air CDU avoids that, but only works if the room’s air cooling can absorb the heat it rejects.1
Rear-door heat exchangers capture heat from air leaving the rack. Used alongside air cooling, they can support rooms with mixed rack densities, but they cannot replace direct-to-chip cooling when the selected hardware requires it.1 In Supermicro’s 23 September 2026 announcement for NVIDIA Vera Rubin NVL72 racks, for example, rear-door heat exchangers are optional equipment to capture residual heat alongside direct liquid cooling.2
For some sites, a defined AI zone or contained pod may be a manageable starting point. Deploying rack by rack lets air and liquid cooling coexist during the transition, and gives the team a way to prove out installation, maintenance and operating procedures before committing to a wider rollout.3
Before committing to the platform, bring IT and facilities together around five decisions:
- Capacity: can the rack, room and facility support the proposed electrical load and cooling demand?
- Cooling architecture: what does the hardware require, where will its heat go, and what still needs air cooling?
- Installation: are piping routes, floor loading, clearances and service access workable without unacceptable disruption?
- Operations and resilience: how will coolant quality, leaks, alarms, maintenance and cooling failures be managed, and who owns each responsibility?4
- Cost and schedule: what is the complete cost and timeline, including infrastructure, commissioning, ongoing support and provision for expansion?
The objective is to provide the computing capacity the business needs while preserving uptime and a sensible path for growth. A small, carefully engineered deployment may be enough to get started. Where the building cannot support the proposed load economically, colocation or cloud capacity may deserve a place in the same evaluation.
Bringing the hardware requirements and facility capabilities together early gives your team more choices before the purchase becomes a commitment.
Sources
Checked 30 September 2026. All four are manufacturers' own publications.
- Vendor source. Vertiv · Liquid and immersion cooling options for data centers
- Vendor source. Supermicro · Now shipping NVIDIA Vera Rubin NVL72 racks, 23 September 2026
- Vendor source. Schneider Electric blog · Brownfield data center modernization for AI, 26 February 2026
- Vendor source. Schneider Electric · Managing direct liquid cooling deployment risks in AI data centres, 27 July 2026