Best GPU cloud providers in 2026
We compared on-demand H100 SXM, B200 and B300 prices across Lium, RunPod, Vast.ai, Lambda, CoreWeave, Nebius and Together AI. Lium had the lowest H100 SXM and B200 price for a whole 8-GPU node.
Lium team6 min read
comparison
gpu-pricing
h100
b200
b300
Lium had the lowest H100 SXM and B200 rental rates among the seven GPU clouds we checked on 22 October: $2.10 and $5.50 per GPU-hour, respectively, for whole eight-GPU nodes. You can rent through the CLI or API, connect over SSH and set a removal time or a spending budget for an agent.
Compare the total cost for the GPU count you need, then check how you'll connect, automate jobs and stop the rental. For multi-GPU training, the interconnect and other node resources matter too. The tables below show prices and rental sizes, followed by Lium's tools for agents and automation.
Why Lium
- Lowest H100 SXM and B200 price. $2.10 per GPU-hour for an H100 SXM and $5.50 for a B200, both for a whole 8-GPU node on the secure tier. Every other cloud in our H100 SXM and B200 tables was higher.
- B300 from one GPU. $8.00 per GPU-hour on the secure tier, below RunPod Secure Cloud, Nebius and Together AI.
- Per-second billing with a stop built in.
lium up --ttl 6hremoves the pod after six hours.--budget 12.50removes it once $12.50 is spent. - Made for agents. Claude Code and Codex can sign up with one API call, rent on their own key with a daily budget, and read the docs over MCP.
- SSH access to your GPU pod. Pick the Secure tier or the cheaper Spot tier.
How we priced it
Every price is on-demand, with no commitment, in USD per GPU-hour, before tax, storage and traffic. Last checked: 22 Oct 2026, 12:04 UTC. Lium prices are real asks from the Lium nodes feed. Providers set them, so they move during the day. Competitor prices are from each cloud's public pricing page. Preemptible and interruptible prices are left out, except the Lium spot B300 row, which is marked as spot. Treat the tables as a dated snapshot and follow the links before you rent.
On Lium, a secure node is one where the provider is penalized if your rental is interrupted. A spot node can be reclaimed at any time.
For one cloud at a time, read Lium vs RunPod and Lium vs Vast.ai. For serving a model, read Best GPU cloud for open-source LLM inference.
H100 SXM 80GB
Lium was the cheapest H100 SXM at the time of writing: $2.10 per GPU-hour for a whole 8-GPU node, $16.80 an hour for the node. Lium's next H100 SXM price, $3.75, was still below the $3.99 that Lambda, Together AI and RunPod Secure Cloud charge.
B200
Lium was the cheapest B200 too: $5.50 per GPU-hour for a whole 8-GPU secure node, $44.00 an hour for the node. A single B200 on Lium was $6.00, below Vast.ai, Lambda and RunPod Secure Cloud for one GPU; RunPod Community Cloud was $5.98.
B300
Lium's secure tier makes the B300 available from one GPU at $8.00 per GPU-hour, below RunPod Secure Cloud, Nebius and Together AI. For jobs that can resume from checkpoints, a full eight-GPU spot node was $6.98 per GPU-hour. The provider can reclaim a spot node, so choose the tier to match your workload.
Built for Claude Code and Codex
An agent can do the whole rental on Lium without a person in the loop:
- Signup by API call. One request creates an account, no e-mail address needed.
- A budget per key. Give the agent its own API key with a daily budget. Lium refuses a rent past it with
402 API_KEY_BUDGET_EXCEEDED, so the limit doesn't depend on the agent behaving. - A stop on every pod.
lium up --ttl 6hor--budget 12.50(CLI 0.27.2 or newer) ends the pod on time or on spend. - CLI, SDK and REST API.
pip install lium.iogives theliumCLI and the Python SDK. The REST API and the nodes feed work from any language. - Docs the agent can read. The hosted docs MCP at docs.lium.io/mcp signs in with Lium, and lium.io/llms.txt describes the workflow.
Step-by-step setups: Give Claude Code a GPU and Give Codex a GPU.
Rent an H100 on Lium
Install the CLI and sign in once:
bash
pip install lium.io
lium initList secure-tier nodes with eight H100s at up to $2.10 per GPU-hour:
bash
lium ls --gpu H100 --tier secure --count 8 --max-price 2.10Check that the node you choose has H100 SXM 80GB GPUs and that all eight are available. If no offer matches, check the H100 listings for current availability and pricing.
After confirming row 1 matches, rent all eight GPUs and schedule removal after six hours. At the snapshot rate, the whole node costs $16.80 per hour; confirm the listed rate before renting:
bash
lium up 1 --count 8 --name h100-job --ttl 6h -yWhen the job finishes early, remove the pod so billing stops:
bash
lium rm h100-jobChoose your GPU on Lium
Lium's combination of competitive rates, SSH access and agent spending controls makes it a strong choice for training, inference and automated experiments. Browse the live GPU listings, choose the GPU count your workload needs and set the pod's removal time when you rent.
Run it on Lium
Get new posts by e-mail: Subscribe to the Lium newsletter
