Help center

1. Pick a GPU. Find GPUs lists 47 models with the hourly price of the whole node.
2. Configure it. Number of GPUs, spot or on-demand, region, template, optional persistent disk. The price updates as you go.
3. Add credit. Top-ups start at $40 and are paid in crypto. Your configuration is saved through payment and you land back on it.
4. Confirm. The order page shows the hourly rate and what it needs from your balance before anything starts.

No. The catalog, the configurator and every price are public. An account is required to order, to hold a balance and to manage keys.

About 60 seconds from a confirmed order in the median case. The instance page updates on its own and shows the SSH command as soon as the endpoint exists. Boot time is not billed.

A dedicated GPU passed through to a KVM virtual machine, a dedicated share of CPU and RAM, local NVMe scratch, an optional persistent network disk, SSH and Jupyter access, and a choice of 9 templates such as PyTorch, CUDA, vLLM and ComfyUI. You can also bring your own container image.

No. You add credit, launch, and stop whenever you want. Pay as you go — no contracts, no minimum commitment. Reserved capacity at a fixed price exists for teams that need guaranteed GPUs for a month or longer.

VRAM is the memory on each GPU. Model size, precision, context length and batch size all pull on it, and training adds gradients and optimizer state on top. Estimate before you choose: the Hugging Face memory estimator is a good starting point.

On multi-GPU nodes the catalog shows both per-GPU and total memory. Using the combined memory means your software has to shard the model across GPUs — tensor or pipeline parallelism — it does not happen by itself.

LLM inference: NVIDIA GPUs that vLLM supports (Turing or newer), 24 GB or more, vLLM template. Image generation: NVIDIA, 16 GB or more, ComfyUI. Training: NVIDIA, 48 GB or more, PyTorch. Each starts at one GPU and sorts by hourly price. They are shortcuts, not recommendations — adjust them to your model.

Check vLLM's hardware requirements before committing to that runtime.

The large price is the whole node: every GPU in it, plus the persistent disk you selected. CPU, system RAM and local NVMe are included, never billed separately. To compare two configurations fairly, keep the same GPU count and region.

They are relative to your current filters, not to the whole catalog. “Lowest price” marks the cheapest configuration matching what you asked for; “Most memory” marks the largest per-GPU VRAM among those results. Change a filter and both move.

Every tier sits at least 20% below the lowest price we observed for that same tier across the public market index (public price index (getdeploying.com), September 2026), and the rule is checked by the pricing pipeline, not by hand. Spot runs 30–53% under our own on-demand rate, so an H100 SXM is $0.55/h on spot against $0.89/h on-demand.

The same GPU and the same virtual machine as our on-demand tier, sold from capacity nobody has reserved. Because we can reclaim it, it is priced 30–53% below on-demand.

You get 2 minutes' notice through the instance metadata endpoint, a webhook and the console. The instance is then stopped, not deleted: the persistent disk, its data and the IP reservation stay attached.

You can relaunch on the next available spot capacity, or move the same disk to an on-demand instance. Compute is billed only while the instance runs, so the notice window and the stop cost you nothing.

It depends on the model and the region, and we publish the trailing 30-day reclaim rate for every model in the price table rather than asking you to guess. Most data-center models sit below 5% — fewer than one interruption per twenty instance-days. Consumer cards in high demand run higher, and the table says so.

Anything that can checkpoint or retry: training and fine-tuning with periodic checkpoints, batch inference, embedding jobs, rendering, hyperparameter sweeps, data processing, CI. Latency-sensitive production endpoints belong on on-demand, which is never reclaimed.

Yes. Both tiers share the same images, disks, templates and API. A common shape is a small on-demand baseline for serving, with spot capacity for everything that can wait a few minutes.

Per minute, from the moment the instance is reachable until you stop it. Every started minute costs the hourly price divided by 60 — no minimum duration, no rounding up to the hour. Boot time and reclaim events are not billed. Persistent storage is billed per GB-month for as long as the disk exists, including while the instance is stopped. There are no egress fees.

Open Balance & billing, pick an amount from $40 up, choose a currency and network, and you get the exact amount to send with an address and a QR code. The rate is locked for 30 minutes. We add no top-up fee: every dollar you pay lands on your balance. Your wallet's own network fee is charged on top and must not be deducted from the amount shown.

USDT (TRON · TRC-20), USDT (Ethereum · ERC-20), Bitcoin (BTC), Ethereum (ETH), Solana (SOL), Litecoin (LTC), Monero (XMR), TRON (TRX). Always send on the network shown on the payment screen — funds sent on another network cannot be recovered.

Before an order starts, your available balance has to cover one hour of that instance, counting the instances you already have running. It is a balance requirement, not a charge: nothing is deducted up front, and usage is still billed per minute. It exists so an instance never starts on a balance that cannot pay for it.

Yes. Set a monthly budget in Balance & billing. New orders that would exceed it — first hour included — are refused. It does not stop an instance that is already running, so treat it as a guard rail, not a kill switch.

Billing holds every payment, every credited top-up and every usage charge. Each completed top-up has a printable receipt, and the whole transaction ledger exports as CSV.

Local NVMe comes with the node, costs nothing extra and is fast. It is scratch space: it disappears with the instance. A persistent disk is a network volume you attach; it survives stops, reclaims and terminations, and you can move it to another instance. Checkpoints, datasets and anything you would hate to redo belong there.

$0.08 per GB-month, billed for as long as the disk exists — including while the instance is stopped, and including after it is terminated. A disk stops costing when you delete it, not when you stop using it.

Nothing. The instance stops; the persistent disk and its contents stay exactly as they were. Local NVMe scratch is lost, which is why checkpoints go on the persistent disk.

No. There are no egress fees.

Add your public key under Access keys — the private half never leaves your device. Once the instance is running, its page shows the exact ssh command to copy.

Run through this in order: the instance is Running, not still provisioning; the public key registered under Access keys is the one matching the private key you are offering (ssh -i if you keep several); you are connecting as the user shown on the instance page; and your private key file is not group-readable, which OpenSSH refuses. If it still fails, tell us with the instance reference.

Create one in Access keys. The secret is shown once, at creation — copy it then. Revoke a key any time; revocation is immediate. If a key leaks, revoke first, ask questions after.

Yes. Account settings walks you through it with any TOTP app in about a minute, and hands you eight single-use recovery codes. Turning it on signs out every other device.

Use one of your recovery codes to sign in, then set up the authenticator again from scratch. Each code works once. If you have neither the authenticator nor a code, open a request from a device that is still signed in.

Account settings lists every signed-in device with its browser, IP address and last activity. Sign out one device, or all the others, from that same list. Changing your password also signs out every other device.

Your password, your private SSH key, a wallet recovery phrase or seed, or a two-factor code. Nobody here needs any of them, and no genuine message from us will ever ask. Keep them out of support tickets too.

Crypto payments credit after the network confirms, which is minutes on most chains. The payment page keeps watching on its own — you can close it. If the quote expired after you sent the funds, do not send a second payment: we keep checking and credit it when it lands. If nothing has happened well past the usual confirmation time, tell us with the payment reference (sg-…) and we trace it on-chain.

The payment page shows what arrived and what is still missing, and the same address stays valid. Send the remainder — not the full amount again.

Funds sent on a network other than the one shown cannot be recovered. This is the one mistake with no undo, which is why the payment screen states the network twice and pairs it with the coin icon. If you are unsure before sending, stop and ask.

The page refreshes itself and switches to Running with connection details as soon as the machine answers. If it stays on Provisioning noticeably longer than usual, open a request with the instance reference (i-…) rather than ordering a second one — compute is not billed while it is not running.

Expected behaviour, not a failure: you had 2 minutes' notice and the instance was stopped, not deleted. Your persistent disk is untouched. Relaunch on spot, or move the disk to on-demand if the deadline matters more than the price. If it happens repeatedly on the same model, the reclaim rate in the price table will show it — pick another model or another region.