What is Lium?
Lium is a GPU rental marketplace run by Datura AI Corp, a company incorporated in Saint Kitts and Nevis. Lium owns little of the hardware: independent providers connect their machines, and Lium checks the GPUs, scores their performance, meters usage and pays providers a share of rental fees plus rewards from Bittensor Subnet 51, a crypto network. You rent a pod, a container with root access, SSH and optional Jupyter, billed per second with no minimum.
On 14 November 2026 its public price feed listed 2,170 GPUs, 1,804 of them rented, from consumer cards such as the RTX 4090 and 5090 to H100, H200, B200 and B300 data-centre GPUs. You can rent through the web app, the lium.io CLI and Python SDK, a REST API, or a hosted MCP server that lets an AI agent check live prices and rent or remove pods.
Lium's prices now appear in the ReviewRamble GPU calculator at /gpu-calculator, next to the other clouds we track.
Is Lium worth it?
Renting a GPU for an evening of fine-tuning usually means choosing between a big cloud that wants a quota request and a smaller one that bills a full hour whether you use it or not.
Lium is a marketplace. Independent hosts list their machines, and you rent a slice as a pod with root access, SSH and optional Jupyter, billed per second. Hosts set the prices, so they move. On 14 November 2026 Lium's live feed showed an RTX 4090 from $0.30 an hour, an H100 from $1.85, an H200 from $3.80 and a B200 from $5, with 1,804 of 2,170 listed GPUs already rented. One CLI command starts a pod with a timer that deletes it, and the hosted MCP server lets an agent such as Claude Code look up prices and rent or remove pods within a spending budget set on its API key.
Reliability is published rather than promised. Lium's own figures over 14 days show 97 to 98% of single H100, H200, RTX 4090 and RTX 5090 launches reaching a running pod, with a median under 20 seconds to SSH but a slow tail of more than two minutes on H100s. Spot machines are cheaper and can be taken back with no compensation; Secure hosts are penalised for interrupting you. If a host's fault kills a rental, Lium credits that pod's fees for the day of the failure and the day before. Planned maintenance can take a machine away for up to 6 hours, announced at least 24 hours ahead, and you are not charged for it.
The catch: on a standard machine the host can read your files, environment variables and process memory. Lium says so plainly and tells you to keep API keys, SSH keys, database passwords and proprietary weights off those machines; its defence is a rule, in force since 19 August 2026, that permanently bans any host caught looking. Only nodes labelled CVM run your pod in an Intel TDX confidential VM whose memory is encrypted in hardware. Billing has edges too: top-ups are non-refundable, an unused balance is not returned when you close the account, pods have no stop button, anything not on a Volume disappears when a pod is deleted, and when your balance hits zero pods are stopped after 2 minutes. Disputes go to the courts of Saint Kitts and Nevis unless your own country's law says otherwise.
Our take: Lium is a good home for work you would run on a stranger's computer without worry, such as open-weight models, public datasets and batch renders. Customer data and keys belong on a CVM node or a conventional cloud, and the GPU calculator shows what that safety costs per hour.
Where it earns its place
A PhD student fine-tuning an open-weight model on a public dataset
Get a big GPU for one evening without a cloud quota request
Rent a single H100 from about $1.85 an hour with lium up --ttl so the pod deletes itself, and keep checkpoints on a Volume at about $0.036 per GB-month, since everything else on the pod goes when it is deleted.
A two-person startup whose developers work in Claude Code
Let the agent rent a GPU for a benchmark and clean up afterwards
Connect the hosted MCP server with an API key that has a daily USD budget; the agent searches live prices, rents a pod, runs the job and removes it. Keep the agent's own API keys off non-CVM nodes, where the host could read them.
A health-tech team training on patient records
Use cheap GPUs without exposing sensitive data to a stranger
Only CVM-labelled nodes encrypt pod memory in hardware; on any other node the host can read the data. Either verify a CVM node's TDX attestation before releasing keys, or keep this work on a cloud with a data-processing agreement.
Before you commit
- Setup
- Sign up with email, Google, GitHub or Discord, top up by card or crypto, add an SSH key and rent; Lium's figures show a median of about 12 to 20 seconds from request to SSH on popular single GPUs.
- Integrations
- Web app, the lium.io CLI and Python SDK on PyPI, a REST API with an OpenAPI spec, a hosted MCP server at docs.lium.io/mcp, and public price feeds at lium.io/pricing.json.
- Data and privacy
- Lium says it does not read the contents of pods or volumes, but on non-CVM nodes the host can read files, environment variables and memory, and a host caught doing so is banned for good. CVM nodes run pods in Intel TDX confidential VMs. Account data is kept while the account exists; processors include Stripe and Amazon Web Services.
- Cost at scale
- An 8-GPU H100 node with NVLink was listed at $23.20 an hour ($2.90 per GPU), about $16,900 for a month left running; volumes add about $0.036 per GB-month plus egress at AWS rates on every read.
- Limits
- To rent, your balance must cover 15 minutes of the new pod and of every pod already running; availability changes by the minute, and some listed GPU models had none free.
What’s good
- Per-second billing with no minimum, from about $0.30 an hour for an RTX 4090
- Launch success and time to SSH published for each GPU type
- CLI, Python SDK, REST API and hosted MCP server, with per-key spending budgets
- Confidential (Intel TDX) nodes with attestation you can verify
- Range from consumer cards to B200 and B300
What’s not
- On standard nodes the host can read your files, environment and memory
- Top-ups and unused balances are non-refundable
- Spot pods can be reclaimed without compensation
- Operator incorporated in Saint Kitts and Nevis, whose courts hear disputes
- Pods cannot be stopped; data off a Volume is lost on delete
Best for
Developers and researchers running open models or public data who want cheap per-second GPUs and can keep secrets off the machine.
Lium alternatives
Lium: common questions
- Is Lium worth it?
- Lium rents big GPUs by the second for less than most clouds, as long as nothing on the machine is secret: on most nodes the host can read your pod. We rate it 3.8 out of 5.
- How much does Lium cost?
- Billed per second with no minimum, at prices each host sets. Lowest available prices on 14 November 2026, per GPU-hour: RTX 4090 $0.30; RTX 5090 $0.49; H100 $1.85; H200 $3.80; B200 $5; B300 $5. Volumes $0.00005 per GB-hour (about $0.036 per GB-month) plus egress. Prepaid by card or crypto; top-ups are non-refundable.
- What are the downsides of Lium?
- On standard nodes the host can read your files, environment and memory; top-ups and unused balances are non-refundable; spot pods can be reclaimed without compensation; operator incorporated in Saint Kitts and Nevis, whose courts hear disputes; pods cannot be stopped; data off a Volume is lost on delete.
- Who is Lium best for?
- Developers and researchers running open models or public data who want cheap per-second GPUs and can keep secrets off the machine.
- How much does Lium cost for a team or in production?
- An 8-GPU H100 node with NVLink was listed at $23.20 an hour ($2.90 per GPU), about $16,900 for a month left running; volumes add about $0.036 per GB-month plus egress at AWS rates on every read.
- Is Lium safe for company data?
- Lium says it does not read the contents of pods or volumes, but on non-CVM nodes the host can read files, environment variables and memory, and a host caught doing so is banned for good. CVM nodes run pods in Intel TDX confidential VMs. Account data is kept while the account exists; processors include Stripe and Amazon Web Services.
- What are the best Lium alternatives?
- The closest alternatives we have reviewed are Runpod, Modal and Baseten.
Answers checked Nov 14, 2026.