GPU colocation readiness checklist

Inventory the equipment

List each server, switch, storage unit and other device. Include manufacturer and model, quantity, rack units, width and depth, service clearances, weight and the vendor’s power and environmental requirements. Note whether the equipment is customer-owned, supplied by an integrator or still being selected.

Use the exact hardware documentation for dimensions, weight, inlet conditions and power. Do not estimate a rack plan from GPU count alone: systems built around the same accelerator can have different chassis, power feeds, cooling and service clearances.

• Example: NVIDIA DGX B300 physical and operational specifications

State the load and how you estimated it

Provide the expected IT load in kilowatts for each rack and for the full deployment. Show the source and whether the figure is a vendor maximum, measured load or planning estimate. Include expected growth and any start-up or peak conditions that affect sizing.

Ask the operator to distinguish equipment IT load from facility electrical capacity and from power that is currently available for a specific deployment. Request the relevant evidence and the date it applies. A site plan, utility connection, energized equipment and contractable capacity describe different stages of readiness; ask which stage a stated value represents.

Document cooling and facility interfaces

Describe each system’s heat-removal method and connections: air-cooled, direct-to-chip liquid, rear-door heat exchanger or another design. For liquid-cooled equipment, provide the server-side requirements and identify what the rack or customer supplies, such as manifolds or coolant distribution equipment. Ask the facility engineer to confirm interface compatibility, operating limits, water quality requirements where relevant, leak detection and responsibilities for connection and maintenance.

Do not treat “liquid cooling” as one universal interface. ASHRAE’s data-center framework distinguishes cooling-system arrangements and describes liquid-cooling loops; the equipment and facility boundaries need to be reviewed together.

If water use matters to your procurement, ask what boundary and metric are being described: on-site operational consumption, cooling-loop fill and maintenance, or water used in electricity generation are not interchangeable measures. Cooling towers can consume water through evaporation and other losses, while closed loops and facility cooling systems have different designs.

• ASHRAE AI Data Center Energy Performance Framework
• U.S. Department of Energy: cooling-water efficiency for data centers

Specify power and resilience requirements

State the required voltage, phase, connector or busway needs from the equipment documentation. Identify whether the design requires A/B feeds, dual power supplies, UPS coverage or generator-backed power. Ask how metering works, what redundancy the proposed service includes and how maintenance or a component failure affects the available feed. Do not infer facility resilience from a label; ask for the exact design and the contractual scope.

Plan connectivity and access

List carriers, fiber paths, cross-connects, bandwidth and routing needs. Explain whether you need private connectivity to cloud providers, corporate networks or other facilities. Identify physical access, remote-hands, escort, security and change-control requirements. Ask what is included in the service and what requires a separate order or third party.

Confirm readiness, installation and acceptance

Give the requested ready-for-install and service dates, deployment sequence and expansion plan. Ask the operator to state, for the specific site and date, whether the relevant capacity is planned, energized, commissioned or contractable. Confirm what evidence supports the status and when it was reviewed.

Agree on delivery access, staging, receiving, installation, cabling, commissioning and acceptance responsibilities. Define which party supplies and tests servers, racks, PDUs, coolant distribution, network links and monitoring. Record acceptance criteria, issue handling, service start and escalation contacts in the project documents.