lium.io

Lium (formerly Celium) is a decentralized GPU rental marketplace operating as Subnet 51 (SN51) on the Bittensor network. The platform connects hardware providers who contribute GPU resources with developers and researchers who need computational power for AI training, inference, and other GPU-intensive workloads. Lium is fully permissionless and open to anyone, with no KYC required. Anyone can add compute, and anyone can rent it.

The platform offers a web interface at lium.io where users can browse available GPU pods, select machines based on GPU type, performance, and price, and spin up instances in seconds. Available hardware on the network includes a range of GPUs from consumer-grade to data center-class cards, such as NVIDIA B200, RTX PRO 6000 Blackwell, RTX 6000 Ada, A100, RTX 5090, RTX 4090, and RTX 3090. Users can also interact with the platform through a CLI tool (pip install lium.io) and a Python SDK, making it possible to launch, manage, and scale GPU compute directly from the terminal. Additional features include templates for common frameworks, persistent volumes, backups, and a referral program.

Lium positions itself as a cost-competitive alternative to centralized cloud providers. Validators on the network programmatically verify hardware specifications, bandwidth, and uptime for every machine, ensuring renters receive exactly what they pay for. The platform uses on-chain smart contracts to manage collateral and slashing mechanisms, automatically penalizing miners if a rented GPU goes offline or underperforms. Payments can be made in stablecoins or TAO.

Lium operates as a peer-to-peer marketplace with three key roles: miners (GPU providers), validators, and renters. Miners register their GPU resources on the network and make them available for rental. Validators securely connect to miner machines to verify hardware specifications and performance, maintaining the network's integrity and scoring miners based on the quality and reliability of their hardware. Renters browse available pods through the web interface or CLI, select a machine, and deploy their workloads.

Each available machine is listed as a pod with detailed specifications including GPU type and count, CPU, memory, storage, network speeds, uptime history, and geographic location. Users can filter pods by GPU type, location, pricing, and other parameters. The platform supports GPU splitting, allowing multiple users to share a single multi-GPU machine when the provider enables it. Once a pod is rented, users can SSH in, copy files, execute commands, and monitor their workloads through the dashboard or CLI.

Miners earn rewards through a combination of TAO emissions and direct rental fees from users. The incentive structure rewards machines based on uptime and performance, with validators continuously monitoring and scoring the fleet. The on-chain collateral system ensures that providers have skin in the game and face financial consequences for downtime or performance issues.