Auto Provision: deploy a GPU when it comes into stock

For a GPU machine that is not in stock yet, an Auto Provision order does the waiting for you. You choose the GPU, the machine types you accept and the most you will pay per hour, and DeployGPU deploys the machine as soon as one comes into stock at or below that price. It is built for the machines where GPU availability, not price, is the hard part: an 8x B300 or 8x H200 VM can sell out between two refreshes of a marketplace page. Auto Provision watches Verda, Hyperstack and QuantaCloud.

When the table above says that no B300 VM is available, that is the case Auto Provision is for: B300 stock comes and goes. The B300 page explains which models need one.

Why GPU availability needs an order, not a refresh button

The marketplace lists what is in stock at the moment you load it. Large machines on the newest GPUs come and go faster than that: an 8-GPU node can appear and be taken by someone else before you look again. An Auto Provision order does the looking for you, and it deploys with settings you chose in advance, so nothing waits for you to fill in a form.

Before you place an order, the page shows what you would be waiting for. For each machine type it lists the price now, the price range over the last 30 days, how often the machine came into stock in that time, how long it usually stayed and when it was last in stock, with the range you can expect to pay per hour.

How to place an order in 4 steps

The page is Auto Provision in the dashboard sidebar, under Services.

1. Choose a GPU: the family, the model where a family has several, and the count, 1, 2, 4 or 8 GPUs.
2. Pick the providers and machine types you accept, up to 20 across Verda, Hyperstack and QuantaCloud. The page labels Verda and Hyperstack "Live · checked every second" and QuantaCloud "Live · checked every minute".
3. Set the order: an OS image for each provider, one or more SSH keys, a disk of 100, 200, 500, 1,000 or 2,000 GB per machine, and a maximum price per hour for the whole machine, fees included. A start time and an end time are optional.
4. Press Start auto-provision.

If one of your machine types is in stock already, the page lists it with a Deploy now button that uses the same settings. If someone else gets that machine first, the order keeps waiting for the next one. Without an end time an order waits until it is filled or you cancel it, and with one it expires if nothing was deployed by then.

Open Auto Provision in the dashboard to see the current quote for the GPU you need. The Auto Provision docs go through every field.

What happens when a machine comes into stock

DeployGPU checks two things before it deploys: that the machine's hourly price, fees included, is at or below your maximum, and that your balance covers 1 hour at that price, the same rule as any launch. When both hold, it sends the create request to the provider and the order shows Deploying. The result is an ordinary DeployGPU deployment. It appears under Deployments, you connect over SSH with the keys on the order, and it bills by the hour like any other instance.

Two things can go wrong at that moment, and neither costs you anything. If the provider runs out before the create goes through, the order goes back to waiting. If your balance is short, that machine is skipped, the order keeps waiting, and you get an email saying a machine came into stock but your balance was short, at most once a day per order.

What an order costs

Placing an order is free, and so is waiting. You do not need credits to place one, although the order summary warns you when your balance is below one hour at your maximum price. Money moves only when a machine is deployed: the first hour is reserved from your balance, as for a launch from the marketplace, and billing then follows the rules on the pricing page, including the per-second refund when you stop. Each open order can deploy its own machine, and each machine is billed separately.

Order states and what pauses an order

An order pauses when one of its SSH keys is deleted, when a deploy does not finish within about 35 minutes or the machine stops before it is ready, or when the last machine type on it has no OS image left. Resume asks for the SSH keys and OS image again and puts the order at the back of the queue. A low balance, a price above your maximum, provider maintenance and a machine someone else got first do not pause an order: it keeps waiting.

Limits to know before you order

• 5 open orders per account, counting waiting, deploying and paused ones. One order deploys one machine.
• Verda, Hyperstack and QuantaCloud only. Massed Compute and Vast.ai are not part of Auto Provision.
• A plain GPU VM. App templates and volumes cannot be added to an order.
• The dashboard only. The API does not place or manage orders.
• An end time stops the order, not the machine. A machine that was deployed keeps running until you stop it.

Auto Provision can only deploy what a provider lists. It cannot make a machine appear, and a machine can sell out before your order reaches it. For more than 8 GPUs, a fixed term or a multi-node cluster, use the cluster request form, which a person answers.

The first order in production

The first order placed after Auto Provision went live on 27 October 2026 was for an 8x B300 VM. DeployGPU sent the create request 1.7 seconds after it saw the machine come into stock, and the VM was running about 3 minutes later. That is one order, not a promise of speed: how fast an order fills depends on the provider and on who else wants the same machine.

The machines people order this way are the ones the frontier models page and the Kimi K3 page call for: 8x B300, B200 and H200 nodes. The multi-GPU page covers what an 8-GPU VM gives you, and the H200 page lists the H200 VMs you can launch right now.