Summary: Everyone debates the merits of direct-to-chip versus immersion, but the real bottleneck in liquid cooling retrofits is the facility’s existing plumbing and power architecture. Vendor thermal specs mean nothing if your raised floor cannot hold the weight or your chilled water plant cannot hit the required setpoints.

Adding liquid cooling to a rack makes sense as AI workloads grow. However, finding room for the pipework, supplying enough power, and installing it all while the neighboring racks stay online takes more work.

Retrofits start here. Before choosing between direct-to-chip cooling, rear-door heat exchangers, or immersion, you need to understand what the building can support and what will have to change.

Start with the capacity you can actually use

An empty rack doesn’t necessarily mean you have room for installing a high-density load. The hall may have spare cooling capacity, but the electrical feed serving that row may already be close to its limit.

Trace the proposed load through the rack PDU, branch circuits, upstream distribution, and UPS systems. Check what those systems can support during a failure and under normal operation.

Power monitoring and load measurements are needed here. AKCP’s contactless current sensors can monitor suitable existing cables without disconnecting the load. Combined with a Power Train view in Quicklime DCIM, those readings help identify where spare capacity may be available.

Decide where the heat will go

Liquid cooling moves heat away from the servers, but you still need somewhere to reject it.

With a liquid-to-liquid coolant distribution unit, or CDU, a heat exchanger separates the server coolant loop from the facility water circuit. That separation lets each side operate with its own fluid requirements and pressures. The facility circuit must still provide the temperature and flow needed to remove the new heat load.

A water connection isn’t the only option. A liquid-to-air CDU can support a smaller deployment without extending facility water to the racks, provided the room’s air cooling can handle the heat it releases.

Size the CDU for the conditions it will actually see. The stated capacity figure is less useful than its performance at your available water temperature, required coolant temperature, and flow rate. Include pump redundancy, filtration, service access, and a workable route for expansion. Planned isolation valves and connection points can make the next phase considerably easier.

Keep watching the air-cooled racks

A retrofit can leave air and liquid cooling operating together for years. Even within a direct-to-chip rack, components outside the liquid loop may still need air cooling. Get the expected split from the server supplier rather than assuming the entire rack load has moved to liquid.

The room may also behave differently after the change. Some racks release less heat into the aisle, a liquid-to-air CDU adds heat elsewhere, and new pipework under a raised floor can obstruct the air supply. Existing cooling controls may need adjustment.

Start with temperature measurements before the installation so you have something useful to compare afterward. AKCP’s Cabinet Thermal Map Sensors measure temperatures at the top, middle, and bottom of the rack, helping operators spot changes that a room-level reading can miss. In facilities with raised-floor plenums or containment, differential pressure monitoring can also help assess changes in air delivery.

Don’t reduce air cooling simply because the liquid loop has started running. First confirm that the remaining air-cooled equipment stays within its inlet temperature limits across the operating conditions you expect.

Design leak detection around the pipe route

Liquid cooling doesn’t make nearby electrical equipment leak-tolerant, even in a purpose-built hall. In a retrofit, the extra challenge is fitting new pipework around existing equipment and cable routes.

Look at where a leak could start and where the fluid would travel. Pay close attention to CDU connections, rack manifolds, valves, and server hose connections. Also consider low points along the route and locations where leaked coolant could reach neighboring racks or power distribution.

A floor sensor alone may detect a leak only after coolant has passed through the rack. Use detection close to likely leak points, with additional coverage along the route the liquid could take. Check that the selected sensors detect the actual coolant formulation; performance with tap water doesn’t establish performance with every cooling fluid.

Monitor the loop as well as the floor. Supply and return temperatures, flow, pressure, and pump status help explain what is happening. A pressure drop can have several causes, so it should prompt investigation rather than automatically being treated as proof of a leak.

Quicklime DCIM can integrate supported CDU data with AKCP leak, temperature, and power readings through protocols such as Modbus or SNMP. Operators can then review related alarms and trends in one place. Protective shutdown and isolation sequences should remain part of the cooling and server control design, with their behavior tested during commissioning.

Commission the first racks before expanding

A phased installation gives the team a chance to prove the design before repeating it across the hall. The first rack or row should be a controlled commissioning stage, with clear acceptance criteria.

That includes pressure testing and flushing the pipework, checking coolant quality, confirming flow at the racks, and testing alarms and failover. Follow the equipment suppliers’ procedures for each step. Check condensation control as well: exposed cold surfaces below the local dew point can collect moisture without any pipe or hose leaking.

Test the operating procedures alongside the hardware. If a pump fails or a leak alarm activates, the shift team needs to know which equipment is affected, how to isolate it, and when the servers must reduce load or shut down.

After the new racks are online, compare their measured power and cooling demand with the design assumptions. Track facility energy use and PUE, but read them alongside IT load and the work being delivered. A change in PUE alone doesn’t tell you how much energy the retrofit has saved.

Know what the next phase will require

The first installation should leave you with more than a working liquid loop. It should establish how much capacity remains, how the neighboring racks have been affected, and what is needed to add the next group of servers.

Ask for those answers before committing to the rollout. A well-planned retrofit gives the operations team a system they can maintain and expand while keeping the rest of the hall running.

Frequently Asked Questions (FAQ)

Yes, provided the facility can support the proposed installation. The assessment should cover electrical capacity, heat rejection, floor loading, pipe routes, and maintenance access. Spare rack space alone doesn’t tell you whether the building can handle the new load.

A phased retrofit can limit disruption by converting selected racks or rows while other areas remain online. However, connections to shared power or cooling infrastructure may still require planned outages. Those dependencies need to be identified before installation begins.

Not always. A liquid-to-air CDU transfers heat from the server coolant into the room’s air, avoiding a facility water connection at the rack. The existing air cooling must have enough capacity to remove that heat from the building.

Usually, some air cooling is still needed. Cold plates cool selected components, while power supplies, storage, and other parts may continue to release heat into the room. Ask the server supplier how much heat the liquid loop captures and size the remaining air cooling accordingly.

The choice depends on the servers, available infrastructure, and how the equipment will be maintained. Rear-door heat exchangers can serve compatible air-cooled racks; direct-to-chip cooling requires servers designed for it. Immersion introduces different equipment handling and space requirements. Assess the complete installation before choosing a method.

Place detection close to likely leak points, including CDU connections, manifolds, valves, and hose connections. Add coverage where escaped coolant could collect or reach electrical equipment. Confirm that the sensors work with the actual coolant formulation, rather than assuming every water sensor will detect it reliably.

Monitor coolant supply and return temperatures, flow, pressure, pump status, and leak alarms. Keep watching rack inlet temperatures and electrical load as well. Quicklime DCIM can bring supported equipment data and AKCP sensor readings together, helping operators investigate changes across both cooling systems.

No. Savings depend on the cooling design, operating temperatures, climate, and workload. Compare measured energy use before and after the retrofit under comparable conditions. PUE is useful to track, but it cannot show the full efficiency benefit on its own.