Blog/What is a GPU node?
Definition

What is a GPU node?

By Samuel Seidel · Published September 9, 2026

A GPU node is a single machine, physical or virtual, with one or more GPUs attached, treated as one unit of compute. In a rental context, one GPU node is typically what you provision and get billed for: a discrete machine with its own CPU, memory, storage, and network connection, built around the GPU or GPUs that do the actual computation.

Node as the unit of allocation

"Node" is cluster terminology: a group of machines working together is a cluster, and each machine in it is a node. GPU rental borrows the word for one discrete, addressable machine, whether or not it belongs to a larger cluster. When a provider advertises a GPU node, it means one machine you get exclusive or shared access to, with a specific GPU model, memory capacity, and CPU, RAM and storage attached.

Dedicated vs shared nodes

A dedicated node is allocated entirely to one renter; nothing else runs on it while it's yours. A shared node splits its GPU or GPUs across several renters or workloads at once, by time-slicing, memory partitioning or virtualization. The difference matters for two reasons:

See dedicated vs shared GPUs for more on the isolation mechanics.

Single-node vs multi-node

Some workloads fit entirely within one node's GPU memory and compute; others need more than one node working together, connected over a network, to handle a model too large for a single machine or to parallelize training across more GPUs than one node has. Multi-node setups add networking overhead and coordination work that a single node avoids, so check whether a workload needs more than one node before you provision several. If a model fits in one node's memory, inference usually runs fine on that one node.

What defines a node's capacity for LLM inference

For serving language models specifically, a node's GPU memory sets the ceiling on how large a model (and how much context) it can hold, and memory bandwidth sets how fast it can generate tokens. CPU and system RAM matter less for the inference itself, though they matter for data preprocessing and general system overhead. A node's specification sheet is usually the fastest way to tell whether it can handle a given model before renting it.

Where GPUwerk fits

GPUwerk's unit of rental is a dedicated DGX Spark node: one machine with 128 GB of unified memory, allocated entirely to you, at $0.79/hour. For a model too large for one Spark, two linked nodes with 256 GB between them cost $1.79/hour. See pricing for current rates and node specifications.

Related pages

First top-up: pay $20, get $30 in credit

One dedicated node, no sharing, no noisy neighbors.

One dedicated Spark, $0.79/hour, deployed in minutes in Europe, North America or East Asia.

Create account