All resourcesGPU Infrastructure Guide6 min read

GPU colocation: plan the deployment before you order the rack

Build a GPU colocation brief covering hardware, usable power, cooling, storage, connectivity, installation responsibilities, and acceptance checks.

By Amherst Systems · Published

GPU colocation means placing your GPU servers in a third-party data center while retaining responsibility for the equipment and the parts of the platform you operate. A suitable facility needs to support the specific hardware, usable power, cooling method, network connections, and operating model of your deployment. Available rack space alone does not establish that fit.

Amherst Systems helps teams turn those requirements into a facility shortlist and a coordinated deployment plan. Hardware installation, specialist cooling work, software configuration, and ongoing operations are assigned explicitly in the project scope.

Start with the workload and equipment list

Write down whether the environment will serve inference, training, rendering, or another compute workload. Include the initial capacity, growth plan, data location, availability requirements, and target service date. An inference service with external users may have different network priorities from a training environment moving large datasets between storage and compute.

Build a bill of materials that includes the complete server configuration, storage, switches, optics, cables, rack accessories, and management equipment. Confirm dimensions, weight, airflow direction, connectors, and vendor installation requirements. Identify what is already owned, what is on order, and what remains undecided.

Use the documentation for the exact system being purchased. GPU model names or advertised GPU power ratings are not a complete facility specification. The rest of the server and rack equipment must be included in the design.

Qualify usable power and cooling at the proposed location

Ask the provider to review the equipment schedule and confirm what the proposed cabinet or cage can support. Obtain the usable power allocation, circuit and PDU arrangement, connector requirements, and the cooling design that will serve that footprint.

For an A/B design, establish the intended behavior if one feed becomes unavailable. The plan must account for the load on the surviving path; two feed labels do not establish redundant capacity by themselves. Our rack power and redundancy guide explains the questions to take into that discussion.

Cooling requirements depend on the selected systems and configuration. Confirm whether the deployment uses air cooling, liquid cooling, or a combination, and which party supplies and supports each required component. Ask for written confirmation for the specific footprint, including any work that must be completed before equipment can be installed.

Treat cluster networking and external connectivity separately

Document the connections inside the deployment as well as those leaving the facility. The compute fabric, storage network, management access, internet service, and private transport can have different performance and resilience requirements.

Confirm switch ports, optics, cable types, handoff locations, routing ownership, and the testing each party will perform. Include remote administration and recovery access in the plan. A large internet circuit does not resolve an internal storage or cluster-network constraint.

Where availability matters, review shared building entrances, underlying carrier access, and other common dependencies. Use our redundant connectivity guide when comparing carrier proposals, then confirm route details with the providers serving the actual locations.

Compare proposals against one requirements brief

Send each candidate facility the same inputs so that differences are visible. A useful sourcing brief includes:

  • Target markets, latency needs, and preferred service date.
  • Equipment models, quantities, dimensions, weight, and delivery status.
  • Initial and planned rack count, power requirements, and cooling method.
  • Required carriers, internet capacity, private connections, and cross-connects.
  • Access, remote hands, receiving, staging, installation, and support requirements.
  • Budget, contract term, acceptance criteria, and expansion assumptions.

Separate recurring charges from setup and project charges. Check what each proposal includes for power, cooling, connectivity, cross-connects, remote hands, and any facility modifications. Identify who pays for equipment, installation, and future changes. Capacity and delivery timing need provider confirmation before a purchasing decision.

Our colocation selection guide provides a broader framework for comparing facility and commercial fit.

Define installation and acceptance before shipment

Name an owner for receiving, inventory, rack placement, power connections, cabling, configuration, and testing. Confirm access approvals and the sequence of facility, hardware, and carrier work. Where specialist installation is needed, agree on the responsible vendor and its prerequisites.

Set acceptance criteria around the actual workload. These may include inventory checks, vendor diagnostics, temperature and power observations under an agreed load, network and storage checks, remote access, and approved failover exercises. Coordinate tests with the facility and equipment vendors; acceptance work should not disrupt unrelated services.

The handover should include equipment records, rack and connection documentation, configurations, support contacts, monitoring ownership, and an escalation process. Use the data center deployment checklist to organize the project from order through operational handover.

Bring us the requirements you have

You do not need a final design to start. Share the workload, draft equipment list, preferred locations, budget range, and target date. Amherst Systems can help identify missing inputs, qualify facility options, source connectivity, and coordinate the agreed data center deployment services.

Technical references

The equipment vendor's current documentation and the facility's written design confirmation govern a specific installation. These primary sources provide context for the planning questions above:

Questions, answered

Related questions

Short answers to the questions this guide raises most often.

What should we send to get help with GPU colocation?

Start with the workload, equipment models and quantities, preferred locations, target date, budget range, and any known power, cooling, or connectivity requirements. Amherst Systems can help turn incomplete inputs into a sourcing brief.

Can any colocation cabinet support GPU servers?

Do not assume so. The provider needs to confirm the power, cooling, physical fit, and support requirements of the exact equipment in the proposed location. Available space is only one part of that review.

Does GPU colocation include managed GPU compute?

Not automatically. Colocation provides an agreed facility service for equipment. Hardware ownership, platform software, monitoring, maintenance, and user access depend on the contract and operating model; managed compute is a separate service scope.

Can Amherst Systems help after the facility is selected?

Yes. We can help plan equipment layout, power and network requirements, provider dependencies, installation responsibilities, acceptance checks, and handover. Physical installation and ongoing operations are agreed separately in the project scope.