An existing data center may have enough total power and cooling capacity for an AI deployment while lacking the local infrastructure required at the planned racks. High-density computing concentrates electrical demand, heat and physical weight in a much smaller area than conventional enterprise IT.
A practical assessment follows the proposed hardware from the rack back through the power and cooling systems. Maintenance conditions, future expansion and operating responsibilities also determine whether the retrofit is viable.
Start With the Workload and Rack Configuration
“AI-ready” does not describe a single infrastructure requirement. A small inference deployment may run within conventional rack densities. A large training cluster can require much higher power per rack, low-latency networking and direct liquid cooling.
The proposed server configuration establishes the starting point. Core design inputs include expected electrical load, cooling requirements, rack dimensions, weight and network topology. Equipment-vendor data also shows how power demand may change during operation.
Training and inference systems can produce different power profiles even when they use similar accelerators. Large clusters may create rapid, correlated changes in electrical demand. The hardware, workload and power-management settings determine the size and timing of those changes.
Infrastructure sized only for the first few racks can reach its limit as soon as the cluster grows. The initial design should therefore include a realistic expansion stage.
Trace Power Capacity to the Intended Racks
A megawatt of spare capacity at facility level does not guarantee that a particular room or row can support several high-density racks.
The complete power path includes the utility supply or onsite generation, switchgear, UPS, downstream distribution, busway or panelboards, rack PDUs and server connections. Each stage has its own rating and redundancy limits. Spare capacity may exist upstream while the local UPS module, busway or branch circuits remain sized for lower-density IT.
The location of the available capacity matters as much as the total. Bringing additional power into another room or row may require new distribution equipment, longer cable routes and a planned shutdown. Capacity during normal operation may also differ from the amount available during maintenance or after the loss of one power path.
Schneider Electric identifies density, synchronicity and variability as three power challenges when existing data centers are upgraded for AI clusters. Equipment-vendor power profiles help assess the UPS response to rapid load changes. The same review covers downstream capacity, protective settings and the remaining redundancy margin through the complete power path.
Temporary monitoring establishes the existing baseline and reveals local constraints. Rack-level and upstream measurements show current peak demand and available margin, while the proposed AI hardware defines the future design load.
Cooling Becomes a Heat-Path Decision
The cooling design shows how heat moves from the chips, through the rack and cooling system, and out of the building.
Air cooling may remain suitable when the selected servers and rack density stay within the tested capacity of the room. Higher-density deployments may use rear-door heat exchangers or direct-to-chip liquid cooling. The available heat-rejection path often determines which architecture is practical.
| Cooling approach | How heat leaves the rack | Main facility requirement | Common retrofit context |
| Enhanced air cooling | Heat enters the data hall air | Sufficient airflow and room cooling at the rack location | Lower-density AI deployments within the room’s tested limit |
| Rear-door heat exchanger | Heat is captured at the rear of the rack | Facility or secondary water loop, rear clearance and compatible rack airflow | Local density increases where server-level liquid connections are unavailable |
| Direct-to-chip with liquid-to-air CDU | Captured heat returns to the data hall through the CDU | Secondary liquid loop and room cooling sized for the CDU output and residual server heat | Retrofits without an accessible facility water loop |
| Direct-to-chip with liquid-to-liquid CDU | Captured heat transfers to facility water | Compatible facility and technology cooling loops with adequate heat rejection | Higher-density or multi-rack deployments with a suitable facility water system |
A liquid-to-liquid design contains two separate cooling systems. The facility water system, or FWS, removes heat from the CDU. The technology cooling system, or TCS, circulates coolant through the servers and cold plates. The CDU transfers heat between the two loops and can also provide pumping, temperature control and separation between their water-quality requirements.
The facility loop and TCS have separate operating requirements for temperature, pressure, flow, water quality and material compatibility. A facility water connection alone does not establish that the server and CDU can operate correctly.
The TCS supply temperature is normally maintained above the room dew point, with an engineering margin. Designs that allow surfaces below the dew point require specific insulation, sealing and condensation controls.
Direct-to-chip cooling also leaves part of the server heat in the room. Cold plates typically remove heat from processors and accelerators, while memory, storage, power supplies, networking and other components remain air-cooled.
The expected liquid heat-capture ratio should come from server-vendor data or validated thermal testing. An 80 kW rack still releases 20 kW into the room when the liquid loop captures 75% of the heat. The percentage in this example is illustrative; the actual capture ratio depends on the server design.
Several racks with a similar profile can create a substantial residual air load. Room airflow, humidity control and the remaining air-cooling capacity continue to affect operating conditions after liquid cooling is installed.
Plan for Physical Installation and Operations
AI racks may be deeper and heavier than the equipment they replace. The slab and raised-floor structure must support the installed load. The delivery route must accommodate the equipment dimensions and transport load.
Larger power feeds, dense network cabling, coolant manifolds, hoses and CDUs compete for the available overhead and underfloor space. A coordinated layout prevents one system from restricting access to another.
Rear-door heat exchangers require enough clearance for the door to open and for technicians to reach rack PDUs, cables and server components. The final arrangement also preserves access to valves, power connections and isolation points. A rack may fit into the row while leaving too little space for commissioning or maintenance.
Operational procedures change as the infrastructure becomes more integrated. Facilities teams may monitor coolant temperature, pressure, flow and leak alarms. IT technicians require a defined method for connecting and removing liquid-cooled servers. The controls platform can show how rack conditions relate to CDU operation, facility water and the remaining air cooling.
Failure response forms part of the operating design. A pump fault, facility-water interruption or leak alarm may trigger workload throttling, automatic isolation or a controlled shutdown. The server’s thermal tolerance and the redundancy built into the cooling system determine the appropriate response.
Power alarms also require operating context. A temporary increase in cluster load may reflect normal workload behaviour while consuming the remaining redundancy margin in an older UPS system. Shared thresholds and an agreed escalation process help facilities and IT teams interpret the same event consistently.
When Does a Retrofit Make Sense?
A retrofit can reuse existing land, buildings, utility capacity and operating infrastructure. It may also allow an organisation to deploy AI capacity sooner than a new construction project.
The strongest candidates have suitable upstream power, a practical distribution path to the planned racks, adequate heat rejection, sufficient structural capacity and space for new equipment.
A site with spare capacity in the wrong location may require extensive electrical and mechanical work. Colocation or a new AI-focused facility may offer a clearer expansion path when the retrofit affects several major systems.
The site decision influences rack architecture, cooling method, utility requirements and the long-term cost of expansion. Resolving that decision before detailed equipment procurement reduces the risk of selecting hardware that the facility cannot support efficiently.
What a Pilot Can Confirm
A 2025 Equinix HK1 liquid-cooling pilot provides an example of an operating data center adding AI infrastructure.
The deployment combined Dell liquid-cooled servers with Schneider Electric’s in-rack CDU. The pilot used a 50 kW in-rack, direct-to-chip liquid-cooling system inside the existing HK1 facility.
The joint announcement reports a PUE of 1.2 or lower with liquid cooling, compared with 1.4–1.5 for the referenced CRAC-based arrangement. It also reports potential annual energy savings above 2,000 MWh if liquid cooling is implemented across the site.
These figures apply to the assessed HK1 scenario. PUE is a site-level metric shaped by the measurement boundary, IT utilisation, climate, cooling architecture and reporting period.
The project combined liquid-cooled servers, an in-rack CDU, existing facility infrastructure and a site energy audit within one pilot deployment.
A representative test uses the planned server configuration and a workload close to the expected production profile. Measurements can include:
- rack power;
- coolant supply and return temperatures;
- flow and differential pressure;
- CDU power;
- room inlet and exhaust air temperatures;
- air-side airflow used to estimate the residual cooling load.
Collecting facility and rack measurements at the same time shows how the complete heat path performs under load.
The commissioning plan also covers the response to pump faults, loss of facility water, leak alarms and power interruptions. Controlled tests, component-level checks or simulation can be selected according to the site’s risk controls.
Commissioning records state the expected response to each event and identify the responsible team. The results inform the capacity model, identify the next constraint and show whether the operating team is ready for a larger deployment.
Five Questions Before Design Approval
- What hardware and workload will run in the first deployment and at the planned expansion stage?
- Can the complete power path support the racks during normal operation, planned maintenance and the defined failure condition?
- Where will the server heat go, including the portion that remains air-cooled?
- Can the rack location support the weight, dimensions, cabling, coolant distribution and maintenance access?
- Who will operate and maintain the connected power, cooling and monitoring systems?
Frequently Asked Questions
Can an existing data center support AI workloads?
Yes, when the facility has a suitable power path, cooling architecture, rack space and operating model. The assessment focuses on the intended rack location and workload because total site capacity does not reveal every local constraint.
Does every AI data center require liquid cooling?
No. Smaller inference deployments and lower-density AI servers may operate with air cooling. Liquid cooling becomes more relevant as rack density rises or when the selected servers require direct liquid connections.
Why can a data center have spare power but still be unable to support AI racks?
Available power may sit upstream or in another data hall. The UPS, busway, panel, rack PDU or branch circuit serving the planned location may have a lower limit. Redundancy requirements can reduce the usable capacity further.
Does direct-to-chip cooling replace room air conditioning?
Usually no. Components outside the cold plates still release heat into the room. The residual air-cooling requirement is determined by the server’s liquid heat-capture ratio and the other equipment installed in the rack.
