← Blog

Guide

GPU VMs Billed by the Minute: H100, H200, MI300X, L40S

machine0 GPU VMs are full virtual machines with a dedicated GPU, billed per minute with no commitment, from a 1x RTX 4000 Ada at $0.836/hr to 8x H200 at $39.336/hr. They run in seven GPU datacenters (Toronto, New York, Atlanta, Frankfurt, London, Amsterdam and Singapore), come with a NixOS image that has NVIDIA drivers, CUDA and GPU containers ready, and can be suspended between runs so you pay only for image storage while idle. One H100 for 3 hours costs about $14.55.

## GPU sizes and prices

Every GPU VM is a full KVM virtual machine with the GPU attached, its own disk, a public static IP and HTTPS. Prices are per hour, billed per minute, the same in every region that offers the size. The monthly column assumes 730 hours.

SizeGPUVRAMvCPURAMDisk$/hour~$/month
gpu-4000ada-11x RTX 4000 Ada20 GB832 GB500 GB$0.836$610
gpu-l40s-11x L40S48 GB864 GB500 GB$1.727$1,261
gpu-6000ada-11x RTX 6000 Ada48 GB864 GB500 GB$1.727$1,261
gpu-mi300x-11x MI300X192 GB20240 GB720 GB$2.849$2,080
gpu-h100-11x H10080 GB20240 GB720 GB$4.851$3,541
gpu-h200-11x H200141 GB24240 GB720 GB$4.917$3,589
gpu-mi300x-88x MI300X1,536 GB1601,920 GB2 TB$22.792$16,638
gpu-h100-88x H100640 GB1601,920 GB2 TB$38.808$28,330
gpu-h200-88x H2001,128 GB1921,920 GB2 TB$39.336$28,715

The monthly figure is what an always-on machine costs. Most GPU work is not always on, which is where per-minute billing and suspend matter.

## Regions

A region is a specific datacenter, and a GPU VM, its disks and its snapshots live exactly where you create it. GPU sizes are available in seven datacenters:

  • tor1 (Toronto)
  • nyc2 (New York)
  • atl1 (Atlanta)
  • fra1 (Frankfurt)
  • lon1 (London)
  • ams3 (Amsterdam)
  • sgp1 (Singapore)

nyc3 (New York) and sfo3 (San Francisco) have CPU sizes only, so pass a GPU datacenter with --region. Availability per size changes; machine0 sizes shows it. Pick the datacenter closest to your data or your users; the price is the same everywhere.

## Creating a GPU VM

$ machine0 new train --size gpu-h100-1 --region tor1 --image nixos-25-11-nvidia
$ machine0 ssh train

Provisioning takes minutes. Once you are in, the GPU is visible to the usual tools and to containers:

$ nvidia-smi
$ docker run --gpus all <image> nvidia-smi

## The NVIDIA image

nixos-25-11-nvidia is the image to reach for on NVIDIA hardware. It includes:

  • NVIDIA drivers and CUDA
  • GPU container support, so docker run --gpus all works without extra setup
  • The SSH user nix (not the default user of the Ubuntu images)

It is available on gpu-4000ada-1, gpu-6000ada-1 and gpu-h100-1. Because it is NixOS, the whole machine configuration is code: a flake you can version, rebuild and roll back. If you want a reproducible research environment, this is the shortest path to one.

Other GPU sizes use the standard images (ubuntu-24-04-loaded, nixos-25-11-loaded and the minimal variants), where you install the drivers and toolkit you need.

## When a size is out of stock: standing orders

GPU capacity comes and goes. Rather than polling, place a standing order:

$ machine0 order my-gpu --size gpu-h100-1 --region tor1 --order-expiry "2 days"

The order sits unbilled until capacity appears. When it does, the VM is created automatically and you get an email. Expiry defaults to 7 days and can be set up to 30. If the order expires before capacity shows up, nothing is charged.

This is the practical answer to "the H100 I want is sold out": queue for it, carry on, and get an email when it is yours.

## Keep the environment, drop the GPU

A GPU environment takes time to set up: drivers, libraries, datasets, checkpoints. Two features let you keep that work without paying for the GPU.

Suspend. machine0 suspend train snapshots the disk, releases the GPU and the instance, and bills only image storage at $0.078/GB/month. machine0 start train brings it back on a GPU. The disk is intact; running processes are not, so checkpoint your training loop to disk before suspending. The VM gets a new IP on resume.

Move the snapshot to a CPU size. A 1x GPU snapshot has a 500 GB or 720 GB disk, which fits the 5xl (720 GB disk, $2.970/hr) and 6xl (900 GB disk, $3.714/hr) CPU sizes. Start the snapshot there to inspect results, serve a small model on CPU, or do data preparation, then move it back to a GPU size for the next run. The same works in reverse.

To turn a configured VM into a reusable image:

$ machine0 images new train train-env
$ machine0 new train-2 --image train-env --size gpu-h100-1

Images are versioned. Note that a snapshot captures everything on disk, including any secrets you left there, so rotate credentials on VMs you snapshot.

## Cost examples

JobSizeTimeCost
Fine-tune on one H100gpu-h100-13 hoursAbout $14.55
One hour on 8x H200gpu-h200-81 hour$39.34
A week of experiments, 4 hours a day on one H100, suspended in betweengpu-h100-128 hoursAbout $135.83 plus storage
Keeping a 720 GB GPU snapshot while suspendedstorage1 monthAbout $56

The storage figure is 720 GB x $0.078.

Keep datasets on a disk. Persistent disks attach to GPU VMs, at creation with --attach disk:<name> path:</abs/path> or later with machine0 disks attach. Create the disk in the same datacenter as the VM (machine0 disks new models --size 500 --region tor1). Disks are billed separately at about $0.1667/GB/month and survive both suspend and VM destruction.

## Suspending between runs

The habit that keeps GPU bills small is simple: suspend when the run finishes.

$ machine0 suspend train    # after the job writes its checkpoint
$ machine0 start train      # next session

An agent can do this for you. The remote MCP server exposes vm_suspend and vm_start, and the TypeScript SDK has machines.suspend and machines.start, so a training script or an orchestrating agent can release the GPU the moment it is done.

## Access without a commitment

GPU VMs on machine0 are on demand. There is no reservation, no minimum term and no upfront payment beyond the wallet top-up (minimum $5, refundable if unused). Create a machine for an hour, suspend it for a month, destroy it when the project ends, and pay for exactly that.

## FAQ

How is GPU time billed?

Per minute, from the moment the VM is running, including while it is stopped. Suspend releases the GPU and bills only image storage until you resume.

Which regions have GPUs?

tor1 (Toronto), nyc2 (New York), atl1 (Atlanta), fra1 (Frankfurt), lon1 (London), ams3 (Amsterdam) and sgp1 (Singapore). GPU sizes are not offered in nyc3 or sfo3, and a GPU VM needs one of these datacenters named explicitly.

Which image should I use for CUDA work?

nixos-25-11-nvidia. It includes NVIDIA drivers, CUDA and GPU container support, the SSH user is nix, and docker run --gpus all works. It is available on gpu-4000ada-1, gpu-6000ada-1 and gpu-h100-1.

What if the GPU size I want is sold out?

Place a standing order with machine0 order. It is created automatically when capacity appears, you get an email, and nothing is billed until the machine runs. Orders expire after 7 days by default, up to 30.

Can I keep my GPU environment without paying for the GPU?

Yes. Suspend the VM, or start its snapshot on a CPU size with enough disk: a 1x GPU snapshot fits 5xl or 6xl. Move it back to a GPU size when you need the hardware.

Is there a minimum commitment?

No. GPU VMs are on demand and billed per minute. Create one for an hour and destroy it.

## Sources

## Related

  • Pricing: Per-minute CPU and GPU prices, what gets billed, wallet and safeguards.
  • Cloud computers for AI agents: Persistent VMs vs sandboxes, with a comparison of machine0, E2B, Daytona and Fly.io Sprites.
  • Remote MCP server: Let Claude Code, Claude Desktop or claude.ai create, start, stop and SSH into your VMs.
  • SDKs: The TypeScript SDK, the CLI with --json output, and the remote MCP server.