01
Training / fine-tuning
A bounded window to train or adapt a model — often with multi-node topology, checkpoint storage, and a hard stop date.
Typical when a research or product team already knows the model family and needs capacity for a run, not a forever fleet.
What we need
- Approximate model size, precision, and whether the run is single-node or multi-node
- Checkpoint and storage expectations
- The window: start, duration, and what slips if capacity slips
Scope notes
- Publish a training SKU, a cluster size, or a tokens-per-second number on this site
- Promise interconnect performance that has not been named in the order
02
Live inference
A serving path for models that must answer in a product — latency and concurrency as goals, not as a public SLA.
Typical when the constraint is a serving pattern (shared, dedicated, or isolated) rather than a one-off training burst.
What we need
- Model family, context length, and whether batch or interactive traffic dominates
- Latency or concurrency targets stated as goals
- Who operates the runtime after handoff
Scope notes
- Offer a public inference endpoint or a self-serve API under the Platinum brand
- Treat a latency goal mentioned on this page as a contractual SLA
03
Visual & media compute
GPU time for generation, rendering, transcode, or other visual pipelines — often bursty, often storage-heavy.
Typical when the job is framed by resolution, codec, batch size, or interactive vs offline throughput.
What we need
- Interactive vs batch, and any resolution / codec constraints already known
- Whether the pipeline is a short burst or a standing render path
- Whether NVIDIA RTX PRO 6000 Blackwell Server Edition is in scope, or datacentre-class accelerators only
Scope notes
- Publish render-farm capacity or a media product catalogue
- Guarantee codec or throughput numbers that belong in a measured engagement
04
Isolated deployments
When tenancy, network boundary, or operator model is the hard requirement — not an afterthought to a GPU name.
Typical for regulated data, customer-operated runtimes, or any case where shared tenancy is not acceptable until proven otherwise.
What we need
- Which controls are mandatory: network, access, tenancy, key custody
- Whether the customer, a supplier, or both must operate the runtime
- Whether EU/EEA (or another) residency is a must — stated, not inferred from Irish registration
Scope notes
- Treat ‘private’ as a marketing badge without listing controls in the SOW
- Imply that Ireland registration equals Ireland processing