How does multi-GPU node pricing work?
A multi-GPU node is a single machine with several GPUs, priced either as a per-GPU rate times the count or as a fixed rate for the whole node. Most providers scale close to linearly, so an 8-GPU node costs roughly eight times one GPU, plus the shared CPU, memory, and high-speed interconnect that come with it. The interconnect is what you are really paying extra for, since it lets the GPUs train a single model together.
For workloads that fit on one GPU, renting eight separate single-GPU instances can be cheaper than one 8-GPU node, because you skip the premium for the fast internal fabric. Choose a node when your job needs the GPUs to talk to each other quickly, as in large-model training or tensor-parallel inference.
That fabric matters. NVLink and NVSwitch inside an SXM node move data between GPUs far faster than PCIe or the network between separate instances, which is why big training runs insist on it. A node also gives you a single machine to manage instead of coordinating many instances.
When you compare node prices, divide by the GPU count to get an effective per-GPU rate, then check what interconnect and how much CPU, memory, and local storage are included. Two nodes at the same per-GPU price can differ a lot once you account for the fabric and the surrounding hardware.
Related questions
Numbers on this page come from today's verified snapshot. Full table on the homepage; method in the methodology.