Rent a GPU cluster
Say how many machines and which GPU. We find the cheapest free ones, start them together, and put them on one private network — each reaches the others by name. Pay by the hour; each machine is paid only for the time it delivers.
| Machine | GPU | Memory | Region | Price / h |
|---|
What you get
One private network
The machines reach each other as node1, node2, … whatever network each host is on — nothing to open, nothing to configure.
Ready for distributed work
Every machine knows its place: NODE_RANK, NUM_NODES, MASTER_ADDR — the names PyTorch’s torchrun, Lightning and most launchers read. One command starts a job on all of them.
Run on all
Type a command once and it runs on every machine, with each one’s output side by side. Each machine also has its own terminal, files and web access.
Pay for what each delivers
Every machine is its own rental, held in escrow. One that stops early is refunded for itself; the others keep running and are paid for their time.
Questions about clusters
What can I run on a GPU cluster?
Training across machines with PyTorch (DDP or FSDP through torchrun), Ray or Dask; batch work split over many GPUs; or one model too big for a single card spread over several (llama.cpp RPC, vLLM with pipeline parallelism) — as long as it fits the cluster’s combined memory.
How fast is the network between the machines?
It runs through RentGPUPower, so it suits coordinating work and moving moderate amounts of data rather than a datacenter’s interconnect. Machines in one region are closer to each other: tick “All in one region” for training that talks a lot between machines.
Can the machines come from different hosts?
Yes. We pick the cheapest free machines that match what you asked for, whoever hosts them. Each host is paid for their own machine; you get one cluster.
What if one machine stops?
Its unused time comes back to your balance automatically, and the rest of the cluster keeps running. You can stop the whole cluster at any moment from its page; every machine settles on its own.
How do I pay?
From your balance, in one step, for every machine at once. Short of credit? Add the difference by card, PayPal or another way, and you come straight back to confirm.