Questions buyers ask before they spend six figures

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 lands in the low-to-mid six figures in pounds sterling depending on CPU, memory, networking and support. Tell us your interests and we will quote your exact configuration.

Our delivery target for NVIDIA GPU-powered Supermicro systems is 4 to 5 weeks from order, against an industry standard of two months or more. That is made possible by a direct channel into Taiwan's manufacturing ecosystem. SuperClusters and GB200 racks are quoted case by case, and we confirm the lead time in writing before you commit.

Yes. SuperServer.ai supplies NVIDIA GPU-powered Supermicro systems directly, from single HGX nodes to rack-scale clusters, and delivers them to the UK and Europe in 4 to 5 weeks. Every order starts with a consultation so the configuration matches the workload.

No. SuperServer.ai is owned and operated by Karl Finance LLP and is independent of Supermicro and NVIDIA. We supply systems built on their technology through our own channel.

We will say so. Our cloud GPU price comparison page is there to help you decide, and the buy-versus-rent guide gives the honest rule of thumb. If your utilisation is low, renting is cheaper and we would rather tell you that than sell you a server you do not need.