Lium vs Vast.ai: where should you rent a GPU?
Compare Lium and Vast.ai on GPU prices, spending controls for coding agents, host checks and failed-start protections, then bring your existing job to Lium.
Lium teamupdated 7 min read
comparison
vast-ai
h100
h200
pricing
Lium offers competitive GPU prices and lets you run the whole rental from your terminal: choose a machine, run your container and remove the pod. Claude or Codex can handle those steps too, using its own API key with a spending budget you control.
You also get recurring host checks and billing protections if a rental fails. If you're comparing Lium with Vast.ai for training or inference, these are the reasons we'd put Lium first.
Why developers pick Lium over Vast.ai
- Competitive GPU prices. A single H100 SXM was available for $1.39 an hour on Lium in our October 17 snapshot, below Vast's $2.40 market median.
- Built for agents. Give Claude or Codex its own API key, spending budget and access to only the pods it creates. It can rent, run and remove pods through the CLI.
- A pod that never starts costs nothing. If the node was at fault, Lium tries another node with the same GPU at the same or a lower price.
- Recurring host checks. Lium's validator tests every connected node every 15 minutes. You can also filter by reliability, uptime and download speed.
- No surprise price rises. On a Secure node a raise can't reach your pod for 21 days and is capped at 25%.
GPU pricing
This pricing snapshot is from 17 October 2026. Lium prices come from the public node feed at 22:08 UTC and show the lowest listed price on a Secure node you could rent then. Vast prices are the market medians in Vast's public pricing feed, last refreshed at 21:30 UTC. The table compares minimum prices with medians, so check current offers for the configuration you need. Prices are per GPU per hour, before storage and transfer.
Lium lists the H100 SXM as "H100 80GB HBM3". Compare the available machines' CPU, RAM, storage, location and host requirements as well as the GPU price. For today's prices run
lium ls --gpu H100 or open the H100 page.What a week of inference would cost
Say you serve a model from one RTX PRO 6000 for a week, around the clock. That is 168 hours.
At the rates above, the compute estimate is $1.29 × 168 = $216.72 on Lium, compared with $1.47 × 168 = $246.96 at Vast's median. That leaves $30.24 for another run.
A shorter example is 10 hours on one H100 SXM: $13.90 on Lium and $24.00 at Vast's median. These estimates cover a fixed number of GPU-hours; actual job runtime depends on the machine and workload. Add storage and any applicable transfer charges on either platform.
Built for Agents, Claude and Codex.
Give each agent its own API key, permissions and daily, monthly or lifetime budget. An evaluation agent could have a $20 daily allowance and access to only the pods it creates, while a training agent works under a separate key and budget.
Lium enforces those budgets on its servers. A new rental that would exceed a key's budget is refused, and its running pods are removed when spending reaches a budget. The agent cannot raise its own allowance. Keep outputs on a persistent volume or copy them out before removal.
pip install lium.io installs the CLI and Python SDK. An agent can also create an account with one API call, without an email address. The account still needs funding before it can rent.Add the rental, execution and cleanup steps to
CLAUDE.md or AGENTS.md. Our Claude Code walkthrough shows the instructions and a session where the pod was ready 26 seconds after lium up.What Lium adds to the bill (and what it doesn't)
A pod on Lium is billed per second at its listed price, with no minimum and no rounding. Volumes and backups are billed on their own line, by the data you actually store rather than the size you reserve; a volume is about $0.036 per GB-month.
Transfer charges are another part of the bill to check. Vast's billing docs describe host-set charges for data sent to and from an instance. Lium does not meter a pod's own internet traffic, but reads from network volumes are charged as transfer.
Two rules in the billing docs are worth knowing when hosts set their own prices. A cut reaches your running pod right away. A raise on a Secure node can't reach it sooner than 21 days after you rented, comes with seven days' notice by e-mail, and is capped at 25% and once a month.
Who checks the hosts
Lium's validator tests every connected node every 15 minutes. Each node is listed as Secure or Spot, and Browse Pods lets you filter by the provider's reliability score, uptime and measured download speed.
Vast also offers verified hosts and a Secure Cloud tier. On Lium, recurring checks are backed by specific billing and recovery rules when a machine does fail:
- A pod that never reaches RUNNING is not charged. If the node was at fault, Lium tries another node with the same GPU at the same or a lower price. A successful start includes provisioning time in its bill.
- If the provider is at fault for force-closing your Secure pod, you receive a credit for the interruption. The broken-pod guide explains retries and credits.
- If a node stops reporting, its pods are charged nothing once it is marked inactive, and what they were charged since its last report is credited back.
- A provider who needs the machine for maintenance has to announce it at least 24 hours ahead, and the window lasts at most an hour.
When something does go wrong on the host, a rented pod can open a private Discord support channel with its provider and Lium support when the required connections are configured.
For private workloads, Lium has Confidential Computing machines with Intel TDX isolation for VM memory and disk. The private models guide explains the required attestation and encryption steps; GPU memory protection needs a separate check.
Renting from the terminal
Filter for the GPU memory and count you need, then choose a node from the results:
bash
pip install lium.io
lium init
lium ls --min-vram 96 --count 1 --tier secure # cheapest first
lium up <node> --count 1 --name serve --ttl 7d --volume new:name=weights -y --no-ssh
lium exec serve "nvidia-smi -L"
lium rm serve -y--ttl 7d removes the pod at the end of the week even if you forget, and the new volume mounts at /mnt, so weights you keep there outlive the pod. The vLLM guide walks through serving a model end to end.Moving from Vast
Bring your Docker image with
lium up --image, then configure its environment variables, ports and startup command. Update scripts that call the vastai CLI to use lium or the Python SDK.Move your datasets and weights into the new pod, and attach a volume at
/mnt for checkpoints and outputs that must survive removal. The migration guide walks through images, data transfers, running a job and getting the results back. You pay by card, stablecoins or TAO.FAQ
Is Lium cheaper than Vast.ai?
Our lowest listed Secure-node prices were below Vast's market medians for every GPU in the table on 17 October 2026. A single H100 SXM was $1.39 an hour on Lium against Vast's $2.40 median. This compares minimum prices with medians; check current offers for the configuration you need.
Does Vast.ai charge for bandwidth?
Yes. Vast's billing docs charge for every byte sent to or received from an instance, at a per-TB price each host sets, and storage keeps billing while an instance is stopped.
How does Lium check its hosts?
Lium's validator tests every connected node every 15 minutes. You can filter nodes by reliability score, uptime and download speed, and you are not charged for a pod that never reaches RUNNING.
Can I rent a whole 8x H100 machine on Lium?
Yes. 8x H100 SXM nodes with NVLink were listed from $2.09 per GPU-hour on 17 October 2026.
See every GPU's live price on the pricing page, or start with your first hour on Lium.
Run it on Lium
See H100 pricing and availability before you start.
