GPU colocation means placing hardware you own in a third-party data centre while coordinating space, power, cooling, connectivity and support. For GPU systems, the fit of the facility matters as much as the rack itself. A location should be assessed against the exact server and workload?not a generic ?GPU-ready? label.
Power: size for the whole system
Request the server manufacturer?s electrical specifications, input voltage, connector and circuit requirements, peak and typical draw, power-supply redundancy design and any power-capping requirements. Then have the facility confirm available capacity per rack, circuit topology, metering, redundancy and the process for outage notification.
Do not multiply the GPU?s board rating and treat the result as total rack demand. CPUs, memory, storage, network switches and conversion losses also consume power. The facility and hardware integrator should validate the system-level design.
Cooling: match heat removal to the rack
Nearly all electrical energy consumed by IT equipment ultimately becomes heat that must be removed. Ask how the site handles hot and cold aisles, containment, airflow paths, temperature monitoring and high-density racks. For liquid-cooled equipment, verify the facility?s supported coolant distribution, connection standards, leak detection, maintenance and responsibility boundaries before committing.
Keep cable routing and equipment installation consistent with the manufacturer?s airflow guidance. NVIDIA?s DGX H100/H200 safety documentation, for example, warns that airflow must not be obstructed and calls for chassis covers to remain installed during operation. Always follow the relevant manufacturer documentation for the actual model.
Connectivity: plan the workload path
List required carriers, bandwidth, cross-connects, public or private addressing, VPN or private connectivity, firewall policy, out-of-band management and latency needs. For distributed training, east-west fabric and topology can be critical; for inference, customer access patterns and service availability may matter more. Confirm which services the facility provides and which require separate partners.
Remote hands and maintenance
Clarify what ?remote hands? includes: visual checks, cable work, power cycling under an approved procedure, component swaps, receiving parts or other tasks. Agree access authorisation, maintenance windows, ticket response, evidence or photos, spare-parts handling and escalation. Hardware repair, software administration and data-centre work may be separate responsibilities.
Operations and reporting
Before deployment, define monitoring ownership, alert routing, incident severity, response contacts, planned maintenance, change management and recurring reports. Decide how temperature, power, connectivity and hardware-health events are communicated. Avoid relying on an informal promise; put service boundaries and exclusions in writing.
Questions to ask a prospective facility
- Can the exact server configuration be supported at the required rack density?
- What power capacity and redundancy are available, and how are they metered?
- What cooling method and environmental limits apply?
- Which carriers, cross-connects and bandwidth options are available?
- What remote-hands tasks and response processes are included?
- Who is responsible for installation, maintenance, spares and incident reports?
- What site, partner and regulatory checks are required before equipment moves?
There is no single colocation specification suitable for every GPU. PocketChina coordinates feasibility discussions around customer-owned hardware, infrastructure needs and local operations in China. Actual availability, power arrangements, price and service scope must be confirmed for each project.
