GPUs for training
Rent an 8-GPU node for a training run from one command, keep your data on a volume that outlives the pod, back up checkpoints on a schedule, and remove the node when the run ends. Billing is per second, with no minimum rental.
H100: from $3.75 per GPU-hour, 24 free right now. Live H100 prices
Rent an 8x H100 node, start a training run, back up checkpoints, remove the node
pip install lium.io export LIUM_API_KEY=sk_... lium up --gpu H100 --count 8 --name train --ttl 12h -v new:name=data -y --no-ssh # volume at /mnt lium rsync train ./project /root/project lium bk set train --path /root/checkpoints --every 6h --keep 7d -y lium exec train "cd /root/project && nohup setsid torchrun --nproc_per_node=8 train.py > /root/train.log 2>&1 < /dev/null &" lium exec train "tail -n 20 /root/train.log" lium rm train -y # after the run has finished and its last checkpoint is backed up
Why train on Lium
You pay per second from the moment the pod is running until you remove it, so a run that ends early stops costing money right away. A rent with --ttl removes the pod when the time is up.
Live prices for every model are public: lium.io/gpu/h100, lium.io/gpu/h200 and lium.io/gpu/b200 show the current price and how many GPUs are free.
The CLI, Python SDK and REST API do everything the dashboard does, so a training pipeline or an AI agent can rent, run and clean up on its own.
Multi-GPU nodes
lium ls --gpu H100 --count 8 lists only 8-GPU nodes. Before a long run, check how the GPUs are wired with nvidia-smi topo -m; docs.lium.io/pod-users/multi-gpu-checks covers the checks to make in the first minutes.
For a job bigger than one machine, docs.lium.io/pod-users/clusters explains renting several whole nodes on one InfiniBand or RoCE fabric.
Data and checkpoints
A volume is storage that outlives the pod, mounted at /mnt. Use it for datasets and weights, and attach it to the next pod with lium up -v.
Backups copy a folder under /root to storage on your account on a schedule, and are kept after the pod is gone. New pods are backed up by default; lium bk set picks the folder, frequency and retention, and lium up --restore-backup starts a new pod with that data.
Need a single H100 by API instead? lium.io/cheap-h100-api shows how.
Frequently asked questions
What happens to my run if the pod restarts?
A pod can restart without any action from you, and the training process stops. Make the job resumable from its last checkpoint and keep a backup schedule on the checkpoint folder.
How is a pod billed?
Every pod has an hourly price and is billed per second, with no minimum rental period. Billing stops when the pod is removed. A pod rented with --ttl removes itself when the time is up.
Can an AI agent do this without a browser?
Yes. Signup, API keys, renting, running commands and removal all work from the CLI or the REST API with no browser and no e-mail confirmation. docs.lium.io/developers/agents is the one-page guide.
For AI agents — CLI install, agent skill, headless signup: llms-full.txt