ardelvo.
← All perspectives

Before you buy GPUs

A useful AI infrastructure plan starts with the facility and workload—not a model number on a server quote.

1. What work will the system perform?

Training, fine-tuning and inference place different demands on accelerators, memory, storage and network traffic. Define expected models, data volumes, concurrency and growth before selecting hardware.

2. How much rack power is actually available?

Nameplate facility capacity is not the same as usable capacity at a specific row or rack. Verify the path from utility and UPS through distribution, branch circuits and rack PDUs.

3. Can the current cooling system remove the heat?

Higher-density racks may exceed the practical limits of existing air cooling. Review supply temperatures, airflow, containment and the feasibility of rear-door or direct-to-chip liquid cooling.

4. Can data reach the accelerators fast enough?

Expensive compute sits idle when storage throughput, east-west traffic or data preparation cannot keep up. Size storage and fabric from the workload backward.

5. Who will operate the platform?

Assign ownership for firmware, drivers, scheduling, monitoring, incident response and capacity. A technically complete installation can still fail without a clear operating model.

6. What happens when demand changes?

Plan expansion units, floor space, power blocks, cooling connections and network ports before the first deployment. Scaling should be a designed step, not a second redesign.

Do not approve the equipment order until compute, power, cooling, network, storage and ownership have one shared design baseline.

Make the purchase decision explicit

Ask each responsible team to approve the assumptions that affect its work. A facilities review should identify capacity limits and any required changes. The platform team should define the workload and performance target. Procurement should separate equipment cost from installation, support and operating costs.

Compare more than one viable configuration using the same assumptions. A lower server price may come with extra networking, cooling or integration work. Record those costs alongside the quote so the comparison remains meaningful.

What the readiness review should produce

RecordDecision it supports
Workload profileWhat to size, benchmark and accept
Facility constraintsWhere equipment can be deployed
Equipment and connection planWhat is included in the purchase
Operating responsibilitiesWho owns incidents and maintenance
Unresolved issuesWhat must be closed before approval

A useful review does not need to be hundreds of pages. It needs to make the unknowns visible and assign each one an owner. Treat unverified capacity or missing operating resources as an open issue, not an assumption that installation will resolve.

Use this brief as a starting point for planning. Confirm equipment and facility requirements with the relevant suppliers and qualified specialists.

Discuss your project