Complete servers, not loose accelerators. Each one is a machine that goes in a rack, draws serious power and holds the GPUs that do the mathematics.
The accelerators and the servers that hold them — the part that actually does the mathematics.
Why it mattersThis determines which models you can run, and how quickly.
No prices are shown. These are configured and quoted against a requirement, and nothing here can be bought at a checkout.
An eight-accelerator AMD Instinct training server in an 8U air-cooled chassis.
An eight-GPU NVIDIA Blackwell Ultra training server, air-cooled.
A 4U two-socket server taking up to ten double-wide GPUs, sized for inference.
A Grace Hopper server where the CPU and GPU share one pool of memory.
A CPU-optimised 2U server built on the NVIDIA Vera processor for agentic AI.
A 2U dual-socket EPYC workhorse that also takes GPUs for moderate inference.
A 1U single-socket server for light inference and general-purpose work.
A compact edge-hardened 1U server for AI inference outside the datacentre.