Tool · Indicative, editable, honest

What does a GPU cluster actually cost?

On-demand cloud, reserved lease, or buy: the same cluster, three structures, one comparison. Defaults are indicative market figures; edit any of them. This is the shape of the decision, not a quote.

On-demand cloudpay only for used hours, at the flexibility premium

$0

Reserved bare-metal leasecommitted rate, all hours of the term, facility included

$0

Buy and colocatehardware + hosting + power over the horizon

$0

The gapbetween the cheapest and dearest structure

$0

Want this modelled on your real numbers?

Send the requirement and we will come back with verified availability, current pricing and the structure that actually fits. No pitch, no quote at the end of it unless you ask for one.

We attach the configuration you just modelled so you do not have to retype it.

Sent. We are on it.

Your configuration has landed with the desk. Acknowledged within one hour, first sourcing pass within one business day.

Indicative only. Excludes staffing, spares, networking build, import and residual value. Reserved lease modelled as the committed rate across all hours (that is what committed means); on-demand across used hours only. Real pricing moves with term, volume and facility. Run it on verified numbers.

Read the result properly

What the comparison is really telling you.

A worked example with real numbers: a 4,096-GPU B300 cluster, and what the space costs.

Within one hourAcknowledged, every enquiry, every time
You contract directWith the supplier, never through us
No cost to buyersCompensation comes from the supply side, disclosed
Global coverageSourced wherever the position actually fits

At low utilisation or short horizons, on-demand wins and should: you are renting flexibility you genuinely use. Somewhere in the middle, the reserved lease overtakes it, usually far earlier than teams expect. Owning wins at high sustained utilisation over long horizons, if you can carry the capital and the operations. The full arguments live in cloud versus bare metal and buy versus lease.

Tell us what you need Find your fitAcknowledged within one hour, first sourcing pass within one business day.
Rack elevation drawings and a GPU module on a planning desk
Access policy

We turn work away.

A desk that never declines anything is selling something. Ours is not. Stated up front, so nobody spends a week finding out.

Open to

The party running the workload

Whoever's workload occupies the capacity and whose balance sheet stands behind the term, or an advisor they name in writing.

Structure

One layer, never two

You contract directly with the OEM, ODM, distributor, cloud or facility. We are not in the chain and we do not add one.

Compliance

KYC before any introduction

Run first, not after the commercial terms. It is why suppliers quote our buyers real numbers instead of screening quotes.

Declined

Positions assembled to flip

Brokers and resellers without a nameable end user do not get a file. If a requirement is wrong for the desk, we say so in the first reply.

Straight answers

Asked first, answered straight.

Where do the default numbers come from?

Published market ranges for on-demand rates, our live inventory for hardware pricing, and typical colocation and power figures. Every field is editable, every output is indicative, and none of it is a quote: real numbers depend on term, volume and facility, which is the conversation the desk exists for.

Why does utilisation change the answer so much?

Because on-demand pricing charges for flexibility. Run the cluster hard and you are paying that premium on every hour; run it rarely and the premium is worth it. Utilisation is the single biggest lever in the whole comparison.

What does the buy figure leave out?

Deliberately, the long tail: staffing, spares, refresh risk and residual value. It covers hardware, hosting and power, which is enough to see the shape. The full model is a working session, not a widget.

Can SSP verify my numbers?

Yes, that is the point. Bring your utilisation and horizon to a twenty-minute call and we run the comparison on verified pricing rather than defaults.