A GPU machine you rent by the hour.
Full shell over SSH, your files on a volume, one GPU or several. For dev boxes, training, fine-tuning and serving your own models.
Browse every GPURuncrate Compute
A GPU and a shell, in about four minutes.
Your agent launches any GPU from L40S to B200 on a ready image, then runs the job over SSH, from $0.22/hr. Stop it and the bill stops.
Launch 8 or more GPUs yourself, or ask us for a custom quote.
Terminal
The command: runcrate instances create --name demo --gpu A6000
Output
exit 00.00 s
Median launch to running4 min 14 s
GPU models on offer21
Cheapest GPU-hour$0.22/hr
What you get
Compute that behaves like a tool. Not a project.
--gpu H100:2
Pick the GPU by name, from L40S to B200. One for a dev box, several on one instance for training.
gpu, then count
--template ubuntu-train
Boot a crate, a ready image: Devbox, Finetune with PyTorch and CUDA, or Inference with vLLM or SGLang.
Devbox · Finetune · Inference
runcrate ssh
A real shell from your terminal or the browser. Each session is signed with a short-lived certificate.
no keys to manage
runcrate cp
Copy files and whole folders to the instance and back.
cp -r ./data my-gpu:/workspace/
volumes
Datasets and checkpoints on storage that outlives the instance. Attach it to the next one on a bigger GPU.
mounts at /workspace
instances delete
Billed for the time it runs: by the second on most GPUs; a few bill in whole minutes, rounded up, including about 30 seconds of teardown after you delete. No commitment.
the meter stops
The same balance and API key pay for instances, storage and model calls, and your team shares workspaces.
Two products, one platform
Instances and crates. The machine, and the image it boots.
A ready image your instance boots.
Devbox with VS Code and Jupyter, Finetune with PyTorch and CUDA in place, Inference with vLLM or SGLang one command away. Pick one when you launch and skip the setup.
Start from a crateExamples
Ask your agent. Or read the code it writes.
Connected by the line, Claude Code or Codex calls the Runcrate tools for you: launch, copy files, run, delete.
Connect your agent first
Add Runcrate's MCP server to yourself: run curl -fsSL"https://runcrate.ai/mcp.md"and follow it.
Add to Cursor
Install in VS CodeWorks inClaude CodeCodexOpenCodeClineRoo CodeKilo CodeCursor
Pricing
The lineup, at today's prices. From the lowest-priced configuration of each.
GPU
From, per hour
H200
141GB of memory
$5.13/hr
H100
80GB of memory
$2.85/hr
A100 80GB
80GB of memory
$1.73/hr
L40S
48GB of memory
$1.26/hr
RTX 6000 Ada
48GB of memory
$0.96/hr
RTX 5090
32GB of memory
$0.72/hr
RTX 4090
24GB of memory
$0.58/hr
Instances bill from your dollar balance while they run: by the second on most GPUs; a few bill in whole minutes, rounded up, including about 30 seconds of teardown after you delete.
Deleting an instance stops its meter. Volumes bill separately, for the space they hold.
The listed rate is the rate. No commitment and no quota request.
Options
Change only what you need. Everything else has a default.
- Pick the GPU
--gpu - Default: required
- By name, with a count after a colon:
H100:2is two H100s on one instance. - Boot a crate
--template - Default: none
ubuntu-devbox,ubuntu-trainorubuntu-inference.runcrate templateslists what each has installed.- Choose where it runs
--region - Default: auto
- Pass a preferred region, or leave it out and the instance goes where your GPU is free.
- Take an exact offer
--type-id - Default: none
- An id from
runcrate instances types, in place of--gpu: one configuration at one price. - Attach a volume at boot
storage_id - Default: none
- Your files are at
/workspacewhen the instance starts. From the SDK and the API. - Run a script on boot
launch_script - Default: none
- Install, download, start training: it runs as the instance comes up. From the SDK and the API.
Compared with a typical cloud
Built for AI work, not for every workload at once.
| Compared on | Typical cloud | Runcrate Compute |
|---|---|---|
| Billing | Per hour, often with a minimum | By the second on most GPUs |
| Discounted rate | One to three year commitments | The listed rate, no commitment |
| Prices | List price plus a private discount | Public hourly rates |
| Setup | Machine image, drivers, access policy | Pick a GPU and a crate |
| Big jobs | Quota requests | Reserve a cluster, every offer side by side |
Talk to an engineer
Need a custom setup?
Self-serve covers most jobs, and you can reserve a multi-node cluster online. For a custom setup, tell us the workload and we come back with a rate card within 24 hours.
Prefer Slack? Create a Slack Connect channelFAQ
The questions every team asks.
An instance is a GPU machine you rent by the hour, with a full shell. A crate is the ready image it boots: Devbox for general work, Finetune with PyTorch and CUDA for training, Inference with vLLM or SGLang one command away for serving. Pick one when you launch.
Over 449 real launches between September 19, 2026 and October 18, 2026, the median time from launch to a running instance with an address was 4 min 14 s. Nine in ten were running within 6 min 50 s.
You pay while the instance runs, from your dollar balance: by the second on most GPUs; a few bill in whole minutes, rounded up, including about 30 seconds of teardown after you delete. Volumes are billed separately for the space they hold.
Volumes survive the instance. Detach one, attach it to another instance, resize it as your data grows. Your files stay put when the GPU is gone.
Yes. Pass a preferred region when you launch (runcrate instances create --region), or leave it out and Runcrate places the instance where your GPU is available.
Launch 8 or more GPUs yourself, by the hour, or prepay a machine for 1 to 12 months at a discount for paying ahead, billed by Runcrate. For a multi-node cluster, reserve one online: we verify your company and your need, then take your exact spec to GPU owners that fit. For a custom quote, use the form on this page.