Inference plus periodic fine tuning capacity, for organisations adapting open weight models on their own data. This is the tier you move up to on evidence, when retrieval alone measurably is not enough, and it is explicitly not a training cluster. The distinction is worth several times the price of the pod.
Designed for
Adapting open weight models on your own data as a planned, recurring workload: domain language, house style, task specialisation.
The full inference and retrieval estate runs here too. Fine tuning capacity sits alongside serving, not instead of it.
The testing loop that makes fine tuning worth doing: evaluation runs, comparisons against the base model, and the evidence that the adapted model is actually better.
Refreshing adapted models as your data changes, on a cadence you plan, with capacity sized for the cycle rather than idling for it.
Not designed for
Foundation model training. Sustained utilisation, specialist networking and data engineering at a scale this pod deliberately does not carry. If the plan is genuinely to train, that is a different conversation, and we will have it honestly rather than sell you the wrong tier.
Pure inference estates. If you are serving models and running retrieval with no recurring adaptation work, the Inference Pod does the job at meaningfully lower cost. Buying tier 2 for a tier 1 workload is exactly the oversizing this design line exists to prevent.
Regulated sovereign workloads. Where data residency and regulatory control are the design constraint, the Private AI Pod carries the security and residency wrap as part of the architecture.
Logical architecture
What is in the box
Enterprise rack servers from our Dell and HPE lines, with the memory and I/O headroom that data preparation and training jobs demand alongside serving.
Training capable GPUs in modest counts, sized for periodic fine tuning runs on open weight models, not for the sustained utilisation economics of a training cluster.
Fast flash for datasets, checkpoints and model versions, with capacity tiers for training data. Checkpoint write speed shapes how long every run takes.
Higher bandwidth east west networking for multi GPU jobs, plus an orchestration stack covering serving, training runs, experiment tracking and model versioning.
No part numbers and no prices on this page, deliberately. The specifics are produced in the sizing engagement, configured to order through our Dell, HPE and distribution relationships.
Sizing signals
The Fine Tuning Pod fits when the workload sounds like this:
Lifecycle
Every pod is wrapped by the IDEAL framework.
Sizing, procurement, deployment, adoption, then refresh and expansion managed proactively. Fine tuning cadence is reviewed as part of the lifecycle, so capacity tracks the work rather than the original guess.
Questions
More accelerator memory and faster storage than inference, but far less than most sizing conversations assume. Parameter efficient methods have made fine tuning open weight models practical on modest GPU counts, and for periodic runs the capacity question is about the cycle, how often and how large, not about sustained cluster utilisation. Model size, method and cadence set the number.
Try retrieval first. Most enterprise tasks are answered by serving a strong open model with retrieval over your own data, which is cheaper, faster to change and easier to govern. Fine tuning earns its place when the model must consistently speak your domain, follow your formats or perform a specialised task where retrieval measurably falls short. The honest test is an evaluation, not a preference.
Explicitly not. It carries training capable accelerators in modest counts for periodic adaptation of open weight models. Foundation model training is a different discipline with different economics, sustained utilisation, dedicated data engineering and specialist networking at a different scale. If a plan genuinely calls for that, we will say so plainly and help you size that decision on its own numbers.
On a planned cadence, typically driven by how quickly the underlying data or task changes: quarterly and monthly cycles are common, continuous retraining is rare. That cadence is a sizing input, because capacity that fine tunes one week a month should not be bought as if it trains every day. Expansion stays a costed option you exercise when the cadence genuinely rises.
No published price, deliberately. The number depends on model size, method, cadence and the serving estate alongside, so a published figure would be padded or wrong. The sizing conversation produces a validated bill of materials configured to order through our Dell, HPE and distribution relationships.
Bring the task and the evaluation evidence. We will tell you honestly whether fine tuning earns its place, and exactly what capacity the cadence needs.
Book a sizing conversation