Lium vs RunPod: why developers and AI agents rent their H100s on Lium
Lium beat RunPod on price for every GPU in our 17 October 2026 table, charges nothing for a pod that fails to start, and lets Claude Code or Codex rent a GPU on its own. Here is where Lium wins, with the numbers.
Lium teamupdated 8 min read
comparison
runpod
h100
a100
pricing
ai-agents
You have a training run that needs eight H100s for a day, and a RunPod tab open with the price already in your head. Before you click Deploy, it is worth five minutes to see what the same day costs on Lium, and what it looks like when Claude Code or Codex does the renting for you.
We checked both price lists on the evening of 17 October 2026. On every GPU in the table below, Lium's cheapest Secure node came in under RunPod's Community Cloud and well under its Secure Cloud. A full 8x H100 node for 24 hours came to $401 on Lium against $766 on RunPod Secure Cloud.
Lium is also built so that a script or a coding agent can rent a GPU, use it and hand it back without a person clicking anything. If you already let an agent run your experiments, that part matters as much as the price.
Why Lium beats RunPod
- Lower prices on every GPU in our table. A single H100 SXM was $1.39 an hour on Lium, about half of RunPod Community's $2.69 and a third of Secure Cloud's $3.99.
- A smaller balance to start. Lium deploys with 15 minutes of the pod's cost in your balance; RunPod asks for an hour of credit first.
- A pod that fails to start costs nothing. If the node was at fault, Lium tries another node with the same GPU at the same or a lower price.
- No surprise price rises. On a Secure node a provider's raise can't reach your pod for 21 days, needs seven days' notice and is capped at 25%.
- Built for agents. Claude Code or Codex gets its own API key with a daily budget the server enforces, and can even sign up with one API call.
- Every host is checked for you. Lium tests every node every 15 minutes, and
--verify-gpusconfirms you got the GPUs you pay for.
Prices on 17 October 2026
Lium prices come from the public node feed at 22:08 UTC, and each one is the lowest ask on a Secure node you could rent at that moment. Providers set their own prices on Lium, so these move. RunPod prices come from runpod.io/pricing at 22:05 UTC. All prices are per GPU per hour.
Lium lists the H100 SXM as "H100 80GB HBM3". The 8-GPU node is rented whole, so its price for the machine is $16.72 an hour. A single H100 SXM at $1.39 is about half of RunPod's Community price. For today's numbers, open the H100 page or run
lium ls --gpu H100.What a day of GPU time costs
Start with the job from the top of this post: 24 hours on eight H100 SXMs with NVLink.
- Lium: $2.09 × 8 GPUs × 24 h = $401.28
- RunPod Community: $2.69 × 8 × 24 = $516.48
- RunPod Secure: $3.99 × 8 × 24 = $766.08
These are compute-only estimates before storage and transfer. The RunPod totals use its published per-GPU rates, rather than a quote for a matched eight-GPU NVLink node.
That is $364.80 less than Secure Cloud for one day. Repeat the run four times while you tune a learning rate and the savings, $1,459.20, pay for three more full days on Lium.
A smaller case is a week of A100 work, say 8 hours a day for five days while you iterate on a data pipeline. That is 40 GPU-hours: $26.00 on Lium, $47.60 on RunPod Community and $71.60 on Secure.
Getting started costs less too. For your first pod, Lium needs a balance that covers 15 minutes of the pod, which is $4.18 for the 8x H100 node. RunPod's docs ask for at least one hour of credit before an on-demand pod starts, so $31.92 for eight Secure H100s. These are deployment balance requirements, separate from payment minimums.
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 the 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. Save outputs to a persistent volume or copy them out before removal.
RunPod also supports coding agents through MCP. On Lium, separate budgets and pod access let you decide how much work to delegate to each agent.
Add a short rules block to
CLAUDE.md or AGENTS.md so the agent knows how to rent, run and clean up. This one is from our Claude Code walkthrough:markdown
## GPU compute (Lium)
Rent GPUs with the `lium` CLI. The API key is configured; never print it.
- Always rent with a name and a deadline: `lium up <huid> --name <name> --ttl 1h -y --json`.
- Run work with `lium exec <name> "<command>"`.
- When done, or if anything fails: `lium rm <name> -y`, then `lium ps` to confirm.
- If a rent is refused with API_KEY_BUDGET_EXCEEDED, stop and tell me.In that run the pod was ready 26 seconds after
lium up, and the whole session cost $0.006. An agent can even create its own account with one API call and no e-mail address. The docs have an MCP server at docs.lium.io/mcp and an llms.txt, so the agent looks up flags on its own instead of guessing.How billing works on Lium
Pods are billed per second at the listed price, with no minimum and no rounding. The billing page has the exact rules, and a few of them save real money.
A pod that never reaches RUNNING is not charged. If it fails to start because of its node, Lium tries another node with the same GPU at the same or a lower price. A successful start is billed from the deploy request, including provisioning time. If a provider force-closes a pod on a Secure node, you are credited, subject to review of who was at fault. The broken-pod guide explains retries and credits.
A provider can't spring a price rise on a running pod either. On a Secure node a raise reaches your pod no sooner than 21 days after you rented it, with seven days' notice by e-mail, at most once a month and by at most 25%. A price cut reaches your pod right away.
And
lium up --ttl 24h schedules the pod's removal the moment it is rented, so a run you forget about stops billing on its own.How Lium checks its hosts
Lium is a marketplace of independent providers, and the checking is done for you. Lium's validator tests every connected node every 15 minutes. Each node is marked Secure or Spot, and you can filter by the provider's reliability score, uptime and download speed. If a node stops reporting, its pods stop being charged once it is marked inactive, and what they were charged since its last report is credited back.
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
The whole flow fits in a few commands. This is the 8x H100 job from above:
bash
pip install lium.io
lium init # approve in the browser, key saved
lium ls --gpu H100 --count 8 --tier secure # cheapest first
lium up <node> --count 8 --name sweep --ttl 24h --verify-gpus -y --no-ssh
lium exec sweep "nvidia-smi -L"
lium ssh sweep # when you want a shell
lium rm sweep -y--verify-gpus counts the GPUs nvidia-smi sees inside the pod and compares them with what you are billed for. Your own Docker image runs with --image, and --dockerfile builds one on the node.Storage and payment
Volumes mount at
/mnt, outlive the pod and can move to another pod. They are billed on the data you actually store, about $0.036 per GB-month, not on the size you set aside. Reads from a network volume incur transfer charges; volume billing explains the rates. You top up with a card, stablecoins or TAO.Moving from RunPod
Bring your Docker image with
lium up --image, then configure its environment variables, ports and startup command. Update scripts that call runpodctl to use the lium CLI or Python SDK.Move your data 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 bringing an image, transferring data, running a job and getting the results back.FAQ
Is Lium cheaper than RunPod for H100s?
Yes. On 17 October 2026 a single H100 SXM on Lium was $1.39 an hour and a GPU on an 8x H100 node $2.09, against $2.69 on RunPod Community Cloud and $3.99 on Secure Cloud.
Does Lium bill by the second?
Yes, with no minimum. To start your first pod, you need a balance covering 15 minutes of its price; existing pods add to that requirement.
Can Claude Code or Codex rent GPUs on Lium by themselves?
Yes. Give the agent a scoped API key with a daily budget and a short rules block, and it rents, runs and removes pods through the
lium CLI.What happens if my pod fails to start?
If the pod never reaches RUNNING, you pay nothing. If the node was at fault, Lium tries another node with the same GPU at the same or a lower price.
Compare live prices on the pricing page, or follow your first hour on Lium to rent your first GPU from the terminal.
Run it on Lium
See H100 pricing and availability before you start.
