From GPU count to nodes, racks and megawatts.
The most useful twenty minutes in AI infrastructure planning is converting a GPU number into everything it actually implies. Almost nobody does it before their first supplier call.
What decides this requirement.
Five minutes with a calculator.
Start with GPUs and divide by eight: that is your node count for typical HGX platforms. Fifty nodes for four hundred GPUs, and immediately the conversation changes from a hardware question to a facility question.
Multiply node count by the real power draw of your chosen generation, then add cooling overhead. Current AI racks are measured in tens of kilowatts, and even modest node counts become megawatt-class conversations quickly. That number determines which buildings can host you.
Rack count follows from what density the facility actually supports today, which varies enormously. The same fifty nodes might occupy very different footprints in two halls, and the lower-density option usually costs more in space what it saves in cooling.
Optics, freight and commissioning.
Fabric ports multiply fast. Every node wants multiple high-speed connections, every connection wants a transceiver at both ends and a cable between, and spares matter because a cluster idle on a missing transceiver is burning far more than the part costs.
Freight at node weights is a real line item, insured and shock-monitored, and it has its own lead time that belongs on the critical path rather than beside it. For international deliveries the border adds more, covered on the import and IOR desk.
Then commissioning: racking, cabling, burn-in and acceptance. It is the phase where schedules quietly slip, and it goes better when the network arrived with the compute rather than after it.
Start from what is verified.
Current verified lines with quantities, lead times and indicative pricing are public on the live inventory. Anything not listed becomes a sourcing requirement with a first pass inside one business day.
Asked first, answered straight.
Around 125 nodes on typical eight-GPU HGX platforms. More importantly, that is a multi-megawatt power conversation and a facility shortlist rather than a purchase order.
Current AI densities run in the tens of kilowatts per rack and newer generations push higher. The binding question is what your target facility can actually deliver and cool today.
More than most first-time builders budget: every node has multiple fabric ports, each needing a transceiver at both ends, plus spares. It is one of the most commonly underestimated lines.
At node weights and values, yes. It has its own lead time and belongs on the critical path, and for international deliveries the border adds more time still.
Working through this on a real requirement?
Twenty minutes with the desk, no pitch and no quote at the end of it. Tell us roughly what you need and we will come back within one business day.
Sent. We are on it.
Your enquiry has landed with the desk. Acknowledged within one hour.