GPU capacity for fine-tuning and post-training.
For teams adapting existing models: the footprint is smaller than most expect, and the desk prices it honestly. Fine-tuning sits between training and inference in every dimension: smaller than a pre-training run, heavier than serving, and far more bursty than either. That shape rewards a different buying structure.
What decides this requirement.
LoRA and full fine-tuning are different purchases.
Full fine-tuning updates every parameter and needs memory for weights, gradients and optimiser state, which puts it close to pre-training requirements per unit of model. Parameter-efficient approaches such as LoRA update a small fraction and cut the memory footprint dramatically, frequently bringing a model that needed a multi-node cluster onto a single node.
That difference is worth establishing before anyone quotes hardware, because it can change the requirement by an order of magnitude. We ask about method before we ask about budget.
Quantised and mixed-precision approaches shift the arithmetic again. The right answer is workload-specific and worth ten minutes of conversation rather than a guess.
Reserve the floor, burst the peaks.
Fine-tuning is intensive for days then quiet for weeks. Paying reserved rates for peak capacity that sits idle is waste; paying on-demand rates for a continuous baseline is also waste. The correct answer for most teams is both.
We size the reserved floor against what you genuinely run continuously, and leave the peaks to on-demand or GPUaaS capacity. The structures are on the leasing desk and GPUaaS page.
This is the clearest case on the whole site for not buying one structure. Teams that force fine-tuning into a single commitment shape almost always overpay on one side of it.
Start from what is verified.
Current verified lines with quantities, lead times and indicative pricing are public on the live inventory. Anything not listed becomes a sourcing requirement with a first pass inside one business day.
Asked first, answered straight.
It depends entirely on method. Parameter-efficient approaches like LoRA can bring a model that would otherwise need multiple nodes onto a single one. Full fine-tuning is much closer to pre-training in its requirements.
For most fine-tuning teams, both. Reserve what you run continuously, burst the peaks. Forcing bursty demand into one commitment shape means overpaying on one side of it.
More than inference, less than full pre-training, and heavily dependent on method and precision. Tell us the base model and the approach and we will size it properly.
Yes. Single-node and small multi-node requirements are entirely normal and we source them on the same terms as larger deployments.
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.
Sent. We are on it.
Your enquiry has landed with the desk. Acknowledged within one hour.