Servers for AI

Rackmount systems built to hold multiple datacenter GPUs, from single-accelerator 1U machines to 8-GPU HGX platforms.

408 products from 17 vendors, from $449 to $795,000 (median $31,488). Most common accelerators: H200, A100, B200, H100, B300.

What to compare

The number that matters is total GPU memory, not core count: a model must fit in VRAM before anything else is relevant. Check whether the listed price includes the GPUs — many vendors quote a base chassis — and whether the interconnect is NVLink or PCIe, which decides how well training scales past one card.

Who buys these

Teams training or serving models at a scale that has outgrown a workstation, and anyone comparing the real cost of owning against renting.