SSP, Strategic Supply Partners Live inventory
The desk
GPU procurementNodes and clusters, landed to your rack ● Live inventoryVerified lines, indicative pricing
By platform
B300 and GB300Current-generation Blackwell H200 and H100New stock and fleet rotation All platformsThe full catalogue
Rest of the stack
Memory, servers and storageThe kit around the accelerators Networking and opticsTransceivers, switching, cabling
Capacity without capex
GPU-as-a-serviceReserved, committed spend, on demand Bare-metal leasingDedicated clusters on a term
Decide first
Cloud vs bare metalWhere each one actually wins Reserved vs on demandThe break-even, calculated Cluster cost calculatorThe whole bill, not the GPU line
Space and power
Colocation and powerSingle racks to 100 MW Above 40 kW per rackWhere the shortlist collapses
Where
United States Europe Asia Pacific Middle East and Africa All regions →
Supply side
Have capacity to fill?Tell us what you hold
Analysis
InsightsWhat things cost, with the numbers Sizing guidesMemory, nodes, power, cost ComparisonsBuy, lease or rent
Start here
How an engagement worksThe first week, held to dates Find your fitThree questions, the right desk Glossary
Talk to the desk
Blog · From the desk

The network nobody prices: optics, fabric and the quiet markup

Ask a team their GPU budget and you get a number instantly. Ask their optics budget and you get a pause. That pause is one of the most expensive silences in AI infrastructure.

Strategic Supply Partners26 August 20265 min read
The cluster network: fabric and optics as the unpriced layer between nodes

A cluster is a network with computers attached

Distributed training is a conversation between nodes, and the conversation runs over a fabric: switches, network cards, and a forest of transceivers and cables connecting all of it. Get the fabric wrong and fifty expensive nodes behave like fifty separate computers, which is a remarkable way to waste a cluster.

Yet in budget after budget, the network appears as a single hopeful line, if it appears at all. The GPUs got weeks of negotiation. The layer that makes them a cluster got an estimate.

Where the money hides

Two places. The first is quantity: a serious multi-node fabric consumes optics by the hundreds. Counting ports across a real topology produces numbers that startle first-time builders, and each of those ports needs a transceiver at each end and a cable between.

The second is the markup. Identical-spec transceivers vary in price several-fold depending on the channel they come through, and OEM list pricing on optics is one of the steepest markups in the entire industry. It survives because optics are bought late, in a hurry, by people who just want the cluster to work. Late, hurried and captive is the worst possible negotiating position, and the channel knows it.

The takeaway

Price the fabric with the cluster, not after it. Optics bought early and buyer-side are one of the easiest large savings in an AI build.

The questions to ask per layer

  • Fabric choice: InfiniBand or high-speed Ethernet? Training at scale still leans InfiniBand; modern Ethernet closes the gap for many workloads at better economics. The answer is your workload's, not the vendor's.
  • Topology: what does the job actually need? Over-built fabrics burn money; under-built ones burn GPU hours, which cost more.
  • Optics sourcing: verified compatible third-party where support terms allow, OEM where they genuinely do not. Knowing which is which is worth real money.
  • Between sites: if the workload spans locations, the interconnect between buildings is its own discipline with its own lead times. Plan it with the cluster.

The simple discipline

None of this is exotic. It is one sheet, priced early: every port, every transceiver, every cable, every link between sites, alongside the nodes they serve. The teams that do it save enough on the network to notice. The teams that do not, fund the markup.

What the sheet looks like

The fix costs one page. List the nodes, the ports per node, the switch layer, and count honestly: every port, transceiver and cable in the topology, plus spares, because optics fail and a cluster waiting on a transceiver shipment is burning GPU-hours to save pocket change. Price that list two ways: through the OEM channel, and through verified compatible third-party where support terms allow it. The gap between those two columns is usually the moment a team starts taking the network seriously.

Timing matters as much as sourcing. Optics have their own lead times, and a fabric ordered after the nodes arrive is a cluster standing idle in its most expensive month. Order the network with the compute, stage the delivery so the fabric is ready when the first rack lands, and commissioning stops being the phase where projects go quiet and schedules die.

None of this requires heroics. It requires refusing to treat the network as a detail, for one afternoon, early. The market charges an extraordinary price for skipping that afternoon, and it charges it in the two currencies that hurt most: money and time on the critical path.

Straight answers

Asked first, answered straight.

How much should I budget for optics in a GPU cluster?

More than feels reasonable at first pass. A serious multi-node fabric consumes transceivers by the hundreds, since every fabric port needs one at each end, plus spares. It is the most commonly underestimated line in AI infrastructure.

Why are transceivers so expensive?

High-speed optics are supply-constrained in their own right, and demand tracks the same AI build-out driving GPU scarcity. Bought late and in a hurry, they are bought at the worst possible price.

Does the network affect training performance?

Substantially, for multi-node training. An under-specified fabric leaves expensive accelerators waiting on each other, which shows up as poor utilisation rather than as an obvious network problem.

Working through this on a real requirement? Twenty minutes with the desk, no pitch, no quote at the end of it. We run your numbers, not ours.

Talk to the desk
Keep reading

The worked version, fully priced →

The networking desk →

Think in nodes →

Compute sourcing →

Talk to the desk

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.

Acknowledged within one hour, first sourcing pass within one business day.
✓

Sent. We are on it.

Your enquiry has landed with the desk. Acknowledged within one hour.