From welcome email to nvidia-smi

1. 1 · Connect

   Provisioning finishes in under 10 minutes; the hostname and IPv4 address appear on the server page in your console. The SSH key from your order is already installed for root — password login is disabled unless you chose the one-time root password.

   ssh root@203.0.113.10Copy
2. 2 · Check the GPUs

   Every card you ordered must be listed, at 0 % utilisation, with the driver and CUDA versions of your image.

   nvidia-smiCopy
3. 3 · Run something on them

   This proves the container runtime sees the GPUs. From here the machine is yours: pull images, mount storage, install frameworks.

   docker run --rm --gpus all \ nvidia/cuda:12.6.0-base-ubuntu24.04 nvidia-smiCopy

Nothing else is installed on the machine — no agent, no monitoring daemon inside your OS. We watch the hardware from outside.

Older toolkits install side by side — conda, uv or a container all work. The ML and inference images add JupyterLab on port 8888, bound to localhost: tunnel it rather than opening the port.

When a plan ends, drives are cryptographically wiped before the hardware goes back into the pool.

A stateful firewall runs at the network edge with every inbound port closed except SSH (22). Ask support to open a port, or keep the service private and reach it through an SSH tunnel. For public inference endpoints, put a reverse proxy with TLS in front — never expose a framework's dev server.

Private VLANs between your own servers in the same region are available on request, at no charge.

Order 8 GPUs of a data-center model and you get a single NVLink node rather than eight separate machines — cheaper than eight singles, and with full peer-to-peer bandwidth between the cards. Check the topology before you tune anything:

NCCL is installed and tuned for the fabric, so a standard torchrun --standalone --nproc_per_node=8 works as-is. Node-level specs and prices are on the clusters page.

• What we watch — GPU ECC errors, temperatures and clock throttling, drive SMART data, link state and power. From outside your OS: no agent is installed on the machine.
• What we never touch — your data, your processes, your logs. We do not log into a running server without your written consent.
• Planned maintenance — announced at least 7 days ahead and scheduled with you.
• Hardware failure — 4-hour replacement target, credits under the SLA.

1. You fund the balanceFrom Billing: pick an amount and a coin — BTC, ETH or USDT (ERC-20 / TRC-20) — and we mint a deposit address for that payment only. Minimum $50.00. No card, no KYC.
2. The chain confirmsYour balance is credited on the first on-chain confirmation, day or night, and any invoice it covers settles itself.
3. Renewals charge the balanceThe renewal invoice is issued 7 days ahead and charged on the renewal day when auto-renew is on. We email you if the balance will not cover it.

An invoice left unpaid 3 days past its due date suspends the server — data intact — and it resumes automatically as soon as the invoice is settled. Auto-renew can be switched off per server.

Building a tool, or asking an AI assistant? The catalogue, prices and stock are published as JSON at /pricing.json, and a plain-text summary of this site for language models at /llms.txt.

Open a ticket from the Support section of your console — tickets are reserved for customers with an active server. Pick the server it is about and, for anything hardware, paste the output of nvidia-smi -q. That single detail is what turns a two-day thread into a same-hour fix.