Sell Spare Data-Center GPU Capacity Without Losing Control

Nobody sizes a GPU cluster for the average day. It's sized for the training run before a release, the month-end batch, the customer who might scale up. The rest of the time a share of it sits idle, and idle accelerators are expensive to own.

Selling that slack has usually meant handing it over: long-term contracts, a reseller's control plane, capacity you can't get back on short notice. Spot nodes on ParalonCloud work the other way round. The capacity stays yours, and you decide minute by minute whether the network can use it.

The model

• You choose which GPUs join. On a multi-GPU server, start the agent with --gpus device=… to expose only some cards. The rest stay invisible to the network and free for your own jobs. See Add a Node.
• Each node is spot. It rents at the platform default price for its GPUs minus a spot discount, and you keep 80% of what it bills.
• One button per node: Pause / Resume. Available, the network can rent it or run inference on it. Paused, nothing new arrives, a rental gets two minutes' notice and stops, and inference completes the requests in hand within two minutes, then stops.
• Fleet controls. Pause or release all spot nodes at once, or a selected group: one rack, one model of card, one site.
• Reclaiming is free. Spot rentals don't count toward reputation, so taking capacity back for your own peaks is never held against you.

There's no contract and no minimum term. Capacity you mark Available is used if there's demand for it; capacity you pause is yours again within two minutes.

What runs on your hardware

Two kinds of work, both in containers:

• Rentals. A developer rents the GPU and gets a Jupyter environment with root inside the container. Rentals are isolated from the host's private networks: a container can't reach addresses on your internal LAN.
• Inference. ParalonCloud runs open models on the node and serves them through its OpenAI-compatible API. The node is paid for the work it does.

Spot nodes are never given long render jobs, so nothing on them assumes it can run for hours.

Connecting a server

1. In the Dashboard, Add node → GPU → your OS → Advanced options → Spot.
2. Run the install command on the server. Add --gpus device=0,1 (or the indices you want) to the docker run line to limit which GPUs the node uses.
3. Repeat per node. Each node is verified by measurement (memory bandwidth and compute) before it serves anything, so every card is listed with what it actually delivers.

Requirements: Linux with Docker and the NVIDIA Container Toolkit, current NVIDIA drivers, a stable connection of at least 50 Mbps per node, and outbound port 7844 for the tunnel. No inbound ports, no public IP. See Hardware Requirements and Prerequisites.

Spot or regular for your capacity?

Use spot for capacity you may need back at short notice. If you decide to commit some cards to the network for the long term, run those as regular nodes: they rent at the full price, and a record of clean rentals builds the reputation renters choose by. Many operators will want both: a regular core and a spot fringe that breathes with internal demand.

Pricing

What each GPU model rents for is on the Pricing page. Spot nodes rent at that price minus the current spot discount, shown on each node card in your dashboard. You know your own power and facility costs; set them against the listed price per GPU before deciding how much capacity to make available.

Next

• Spot Nodes — Pause / Resume and fleet controls, in full.
• Add a Node — the install command and how to limit which GPUs a node uses.
• Payouts — how earnings are withdrawn.