The infrastructure around the GPU.
GPU procurement is the entry point. These are the layers that decide whether a cluster lands on time, on budget and in a building that can actually run it.
Import, export & importer of record
Export screening, insured freight at node weights, customs classification and IOR where you cannot import yourself.
Networking & optics
InfiniBand and Ethernet fabrics, switching, transceivers and cabling, at the price the channel should charge.
Financing & leasing
Structures that make a deal fundable: take-or-pay paper, deposits and terms a credit committee will underwrite.
ITAD & buyback
Certified data destruction, secure decommissioning and buyback that turns retiring kit into credit against the refresh.
Data centre space & power
Megawatts, rack density, cooling and connectivity qualified against real secured capacity, not brochures.
Everything around the GPU still has to arrive.
Procurement conversations concentrate on the accelerator because it is the largest single line and the easiest one to compare. Meanwhile the cluster does not run without power distribution, cooling capacity, fabric switching, optics, storage, racks and a cleared route across a border. Each of those looks small beside the GPU number and together they routinely exceed it.
They also fail differently. A GPU that is late delays the project. Optics that are late, or an import that stalls at a border, delays the project while you are already paying for the facility. The unglamorous layers carry more schedule risk than the headline hardware does.
Sourced together, because they arrive together.
These pages set out the parts of the build that sit outside the accelerator itself: the networking and optics layer that consumes transceivers by the hundreds and is the most commonly underestimated line in AI infrastructure, and the import, export and Importer of Record work that determines whether hardware reaches your rack or sits at a border.
The desk sources these alongside the compute rather than after it, because splitting them across suppliers is how commissioning dates slip. If you want the arithmetic behind how much of each you need, that sits on the sizing guides.
Asked first, answered straight.
Fabric switching and optics, power distribution and cooling, storage fast enough to keep the accelerators fed, racks, and the freight and import path. Individually each looks minor beside the GPU line. Together they frequently cost more.
The entity that carries responsibility for an import, its duties and its compliance obligations. If you have no registered presence in the destination country you will need one, and arranging it late is a common cause of hardware stalling at a border.
Yes, and sourcing them together is deliberate. Splitting the build across suppliers is one of the most reliable ways to miss a commissioning date, because the parts arrive on unrelated schedules.
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.