by Datura AI
Decentralized GPU marketplace on Bittensor with no KYC and hourly billing.
Lium suits short, price-sensitive GPU jobs: you pay by the second at a provider's listed rate, can pick a cheaper spot pod that may be reclaimed, and can script everything through a command-line tool built for AI agents. It publishes no uptime guarantee, so steady or regulated workloads belong on a contracted cloud instead.
Renting an H100 for about $2.24/hour with zero KYC and a one-line CLI command.
TAO staking and slashing create real economic accountability without KYC, and the open-source subnet code lets anyone audit or run a validator, but the moat is thin since Vast.ai and io.net offer similar peer-to-peer pricing without requiring a crypto wallet.
Main trade-off
No formal uptime SLA or compliance certification, so regulated or mission-critical teams have no contractual guarantee if a provider's hardware drops offline.
| Plan | GPU | Price |
|---|---|---|
| RTX A4000GPU · Cheapest live ask for light inference and notebooks | 1× NVIDIA RTX A4000 | $0.12/hr |
| Tesla V100GPU · Older-generation card for small training runs | 1× NVIDIA Tesla V100 | $0.17/hr |
| RTX 3090GPU · Budget inference and small fine-tuning jobs | 1× NVIDIA RTX 3090, 24 GB VRAM | $0.25/hr |
| L40SGPU · Inference and rendering workloads | 1× NVIDIA L40S, 48 GB VRAM | $0.38/hr |
| RTX A6000GPU · Mid-size fine-tuning and image generation | 1× NVIDIA RTX A6000, 48 GB VRAM | $0.42/hr |
| RTX 4090GPU · Consumer-flagship inference and training | 1× NVIDIA RTX 4090 | $0.65/hr |
| RTX 5090GPU · Consumer-flagship inference and training | 1× NVIDIA RTX 5090, 32 GB VRAM | $0.70/hr |
| A100 SXM4 (reference price)GPU · No free pod on 2026-10-27; reference price shown | 1× NVIDIA A100-SXM4-80GB, 80 GB VRAM | $0.70/hr |
| RTX PRO 6000 Blackwell (Server Edition)GPU · High-VRAM inference | 1× NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB VRAM | $1.75/hr |
| H100 80GBGPU · Large model training and fine-tuning | 1× NVIDIA H100 80GB HBM3, 80 GB VRAM | $1.79/hr |
| H200 NVLGPU · Large-context LLM training and inference | 1× NVIDIA H200 NVL | $3.50/hr |
| H200GPU · Large-scale LLM workloads | 1× NVIDIA H200, 141 GB VRAM | $3.55/hr |
| B300 SXM6GPU · Frontier-scale training; pods come as 1, 4 or 8 GPUs | 1× NVIDIA B300 SXM6 | $4/hr |
| B200GPU · Frontier-scale training and inference | 1× NVIDIA B200 | $6.10/hr |
Billed per second from a prepaid balance, with no minimum rental and no subscription. Providers set their own prices, so the figures here are the lowest live ask per GPU-hour on Lium's pricing page, read on 2026-10-27: RTX A4000 $0.12, Tesla V100 $0.17, RTX 3090 $0.25, L40S $0.38, RTX A6000 $0.42, RTX 4090 $0.65, RTX 5090 $0.70, H100 80GB $1.79, H200 $3.55, B300 SXM6 $4.00 and B200 $6.10. The highest live asks reached $12.90 per GPU for a B300 SXM6, and an 8-GPU pod peaked at $206.40 per hour on the browse page. A pod needs a balance covering 15 minutes of itself plus 15 minutes of every pod already running. The listed price is the whole pod, local disk included; persistent volumes are charged separately per GB-hour plus data read out, at cost. Top up by card or crypto, and crypto, TAO and alpha top-ups earn a 2.5% bonus. For scale, Crusoe publishes $4.50 per GPU-hour for an H100 and CoreWeave lists an 8-GPU HGX H100 node at $49.24 per hour (Hokai records, October 2026).
Lium is a peer-to-peer GPU marketplace operated by Datura AI Corp. On 27 October 2026 its pricing page tracked 94 GPU models, with 17 rentable at that moment across 267 GPUs, and every pod is billed per second with no minimum rental or contract.
Per-second billing
Every pod is billed by the second with no minimum rental, no reservation and no subscription, drawn from a prepaid balance.
Agent-first CLI and SDK
The lium CLI and Python SDK (0.20.0, released October 2026) start, run and remove a pod in one command, with a time limit flag such as --ttl 1h that auto-removes the pod after an hour.
Signup without an email
Create an account with one POST to the fingerprint signup endpoint, no mailbox needed, and the response carries an API key.
Spot and secure tiers
The browse page filters pods by Secure or Spot; spot pods are cheaper but a provider can reclaim them.
Public price feed
GPU prices are served as JSON at lium.io/pricing.json and live nodes at lium.io/api/public/v1/nodes, readable with no account.
Per-node filters
Search pods by GPU type, count, country, reliability, VRAM, open ports, CUDA version, NVLink, confidential computing and GPU profiling.
Card and crypto top-ups
Pay by card through Stripe, in supported cryptocurrencies, or with TAO and Bittensor alpha; non-card top-ups earn a 2.5% bonus.
Persistent volumes
Storage volumes outlive a pod and are billed per GB-hour plus data read out, both at cost.
The nearest AI infrastructure services on Hokai for the same job.
| Service | Regions | Pricing | Compliance | Checked |
|---|---|---|---|---|
| Asia Pacific · Canada · Europe · US | Pay as you go · Hourly | -- | 27 Oct | |
| US · Europe · Asia Pacific | Free tier · Pay as you go · Hourly | SOC 2 · HIPAA · GDPR | 29 Oct | |
| Europe · Asia Pacific · US | Pay as you go · Hourly | SOC 2 · HIPAA | 23 Oct | |
| US · Europe · Canada | Pay as you go · Hourly | SOC 2 | 19 Oct |
Every pod is billed per second from a balance you top up in advance, and the provider sets the per-hour rate, so it moves with supply. At the time of this check the lowest live asks ran from $0.12 for an RTX A4000 through $1.79 for an H100 to $6.10 for a B200, while pods of the newest multi-GPU cards sit far higher. The pod price already includes local disk, while persistent volumes are charged on top. Spot pods cost less but can be reclaimed.
There is no free plan, but a new account receives $5 of credit once the emailed verification link is clicked. The credit must be spent on a rental within 14 days, is limited to one per person, and is withheld from signups that look automated or arrive after the day's credit budget is gone. Startups, researchers and open-source teams can email support@lium.io for a larger grant, which expires within 90 days.
For a contract and committed capacity, CoreWeave sells bare-metal H100 nodes by the node-hour and Crusoe publishes per-GPU rates and includes free egress. Baseten fits teams that want managed inference instead of raw pods, and Atlas Cloud suits anyone who only needs a hosted-model API. Choose any of them over Lium when you need support commitments or one named region.
On the numbers Hokai holds, Lium is cheaper per GPU-hour: its lowest H100 ask in this check was $1.79 against Crusoe's published $4.50 (read 19 October), and its L40S started at $0.38 against $1.50. Crusoe sells the steadier product, with free network egress, managed Kubernetes and committed-spend options, while Lium draws on independent providers who can run out of a given card. Pick Lium for bursty experiments and Crusoe for production runs.
Create an account on lium.io, or, if an agent has no mailbox, make one with a single call to the fingerprint signup endpoint. Top up a balance that covers 15 minutes of the pod you want, then install the CLI with pip install lium.io. A command such as lium up --gpu H100 --ttl 1h starts a pod that removes itself after an hour, and lium ssh opens a root shell on it.
Lium is a GPU rental marketplace operated by Datura AI Corp, a company incorporated in Saint Kitts and Nevis, and it launched in 2024 on Bittensor's Subnet 51. Independent providers list their own machines at their own price per GPU-hour and keep 95% of what renters pay. Renters launch Docker pods with root SSH access and direct GPU passthrough, billed per second with no minimum rental and no reservation.
The product changed shape in 2026. It now sells itself to AI agents as much as to people: the lium command-line tool and Python SDK (version 0.20.0, released in October 2026) let an agent start a pod, run a job and remove it inside a budget it sets, and an agent with no mailbox can create an account with a single API call. The pricing page tracks dozens of GPU models, yet only a minority have a free pod at any moment, and a spot tier offers cheaper pods that a provider can reclaim.
Lium suits ML engineers, indie developers and coding agents running short fine-tuning jobs, inference tests and one-off experiments, where per-second billing beats a monthly commitment. If you want contract-backed clusters, read the CoreWeave profile first, then check what Crusoe publishes. Teams after managed inference should weigh Baseten, while a hosted-model API with a short GPU menu is Atlas Cloud territory.
Skip Lium if you need an uptime SLA, a compliance certificate or one named data center; none of those appears on its pricing, about or home pages. Supply is peer-to-peer, so a specific card or country can have no free pod on a given day, and Lium bans access from US-embargoed countries and sanctioned parties under its export-control policy. Steady or regulated workloads sit better on a dedicated server from Hetzner.
Plans, regions, the uptime commitment and the compliance list on this page are read from Datura AI's own pages and documentation. The verdict is Hokai's reading of them, not the provider's claim.