Ten accelerators across four NVIDIA generations. Each one is a dedicated physical card — nothing here is virtualised or time-sliced between customers.
Largest memory footprint we offer. Built for frontier-scale training.
The workhorse for multi-node LLM training. NVLink across all 8 GPUs.
Hopper performance at a lower hourly rate for single-node work.
Proven, widely supported, and the best price-per-VRAM on the fleet.
Entry point into datacentre-class Ampere for mid-size workloads.
Ada-generation inference and rendering with hardware video encode.
48GB of ECC memory at a workstation price point.
Best raw throughput per dollar on the fleet for single-GPU jobs.
The cheapest way to get 24GB of VRAM under a CUDA runtime.
Legacy Volta capacity for pipelines pinned to older CUDA versions.