Renting GPU compute on market.tokenos.ai

market.tokenos.ai is for renting GPU compute by the minute to run your own workloads, such as model training, fine-tuning and inference. It is separate from deploying a node on compute.tokenos.ai.

How renting works

Create a TokenOS account at user.tokenos.ai and verify your identity with a government photo ID through our verification provider; a verification lasts up to 24 months. Then pick a GPU tier and a rental type, and pay with your TokenOS balance, a card or USDC on Solana. The full price is paid before a machine is assigned. An auto-allocator assigns an idle GPU that matches your tier and region, usually in under 60 seconds. Access is by SSH key only, with no password, and it ends when the rental ends. The operator sees the public key for your rental, not your email, wallet or payments.

Pricing tiers

• On-demand: you pay for the minutes you use. Stop early and the unused part is refunded, minus at least one minute. The minimum is 1 day, or 1 hour for hourly rentals.
• Spot: cheaper, but can end with 90 seconds' warning when someone is waiting for an on-demand rental of the same GPU class. The unused part is refunded.
• Reserved: a 7, 30 or 90 day commitment, paid up front and not refundable if you end it early. It is never cut short for on-demand demand, and it is a paid claim on operator uptime rather than on a specific machine.

On-demand rates are published in the fee schedule at market.tokenos.ai/pricing. Inference workloads on datacenter GPU tiers are billed 20 percent below the on-demand rate. Spot and reserved prices are shown at checkout.

Hardware, regions and tools

GPU classes include H100, H200, B200, B300, GB300, A100, L40S, RTX 4090, RTX 3090 and consumer NVIDIA cards. The class shown is capped at what the operator's own benchmark reported; that benchmark is self-reported, so check the hardware when your rental starts. Active regions are US West, US East, Europe (Frankfurt, Amsterdam) and Asia Pacific (Singapore, Tokyo). CUDA, PyTorch, TensorFlow, HuggingFace, vLLM, Ollama and Docker on Ubuntu are supported. A REST API with an OpenAPI specification is available for developers. AI agents cannot rent on the marketplace on their own; for AI inference, use TokenOS P2P.

Operator reputation

Each operator has a public reputation score made up of 60 percent uptime, 25 percent buyer ratings and 15 percent completed jobs. The formula is public, and rank cannot be bought.

The marketplace has its own terms, linked from its sign-up page. For account or billing questions, submit a ticket.

Last reviewed: November 5, 2026.