Capacity is reserved,
not queued.
Dedicated GPU capacity contracted against a defined term, a named site and a stated power budget. Contact our team for the latest availability.
Reservation
Three steps, no discovery phase.
We do not run a sales cycle before we run your model. The first technical conversation produces a configuration and a measurement.
01
Scope
Model class, parameter count, context length, concurrency and latency target. One call, written up as a configuration.
02
Measure
Your model on the proposed hardware. Latency, throughput and tokens per watt — measured, not interpolated from a datasheet.
03
Commit
A named site, a term, a power budget and a ramp schedule. Nodes are held once the reservation is signed.
Submit requirements.
Tell us the workload and the target date. We will come back with a reference configuration, a site recommendation and the availability position for that site.
If the right answer is a smaller tier, or a different site, that is what the reply will say.
Prefer email?
info@agatecap.com
Include the model, the concurrency and the date you need capacity.
Questions
Commercial, mostly.
One node for twelve months. Terms of 36 and 60 months carry the reserved-capacity pricing; shorter terms are priced at the on-demand rate.
Yes. Reserved nodes are allocated to one customer for the term. You are not sharing a scheduler with another tenant unless you contract for shared capacity at a lower price.
Within a platform and a country, yes, subject to availability at the receiving site. Across borders it depends on the data terms of both sites — we will not move customer data without a written instruction.
You renew, resize or release. Released nodes return to the published inventory, which is why the table above moves. There is no automatic rollover at a higher rate.
Not sure which tier you need?
Send the workload and we will size it. The answer is sometimes smaller than the request — and it is always measured on the hardware you would actually be buying.