Straight answers on GPU sizing, buy versus rent, H100 versus H200, what an 8-GPU H200 server costs, and delivery times to the UK and EU.
For 16-bit inference a 70B model needs about 140GB for weights plus KV cache. That is two H100 80GB, or one H200 141GB at a small batch. For serving thousands of users, plan on an 8-GPU node with tensor parallelism and quantisation to 8-bit, which halves the memory. Training or full fine-tuning a 70B model needs roughly 16 to 64 GPUs depending on how long you are willing to wait.
It depends on utilisation. Below about 30% utilisation, rent on-demand. Between 30% and 60%, reserved cloud or a hosted dedicated node is usually cheapest. Above 60% for 18 months or more, owning wins: an 8× H200 node pays back against on-demand cloud in roughly 12 to 16 months before power and colocation costs. Our quotes show all three side by side.
Same Hopper compute, different memory. H200 has 141GB of HBM3e at 4.8TB/s against the H100's 80GB at 3.35TB/s. For inference on large models the H200 is often 1.5 to 1.9 times faster because it is memory-bound, and it lets you fit a model on fewer GPUs. For compute-bound training the difference is small.
Prices are set by GPU allocation and move month to month, which is why nobody publishes a list price. At the time of writing an eight-GPU HGX H200 node from an authorised integrator lands in the low-to-mid six figures in pounds sterling depending on CPU, memory, networking and support. Request quotes and we return three comparable figures.
Single HGX H100/H200 nodes: typically four to eight weeks from order through the integrators we work with. SuperClusters and GB200 racks: quoted case by case and longer. We give you the real lead time from each supplier in the quote, not a marketing figure.
No. SuperServer.ai is an independent sourcing desk. We obtain quotes from authorised Supermicro integrators and cloud providers on your behalf and you buy directly from the one you choose. We are paid a referral fee by the supplier; the price you pay is the supplier's direct price.
No. SuperServer.ai is owned and operated by Lovell Group Ltd and is independent of Supermicro and NVIDIA. We source Supermicro and NVIDIA systems from authorised integrators on behalf of buyers.
Yes. Tick 'Cloud' or 'No preference' on the form. We compare on-demand and reserved pricing from specialist clouds and hyperscalers for your workload and region, and include data-residency and egress in the comparison.
Comparable pricing and lead times from authorised integrators, plus a cloud and hosted alternative so you can see the trade-off.