Your stack, running, in 1.8 seconds.
A shell in under two seconds.
The default template is PyTorch; no flag needed. The environment is already warm.
Three doors to the same box.
JupyterLab, VS Code in the browser, or SSH. Same instance, same files, your pick per session.
An API when you want one.
Expose ports for Gradio, FastAPI, or any web service. Demo a model without moving it.
What is a GPU template?
A template is a managed container with the deep-learning stack already standing: the framework, CUDA and cuDNN, the notebook server, the editor. You bring code and data. It boots in 1.8 seconds because there is nothing left to assemble, and it stays yours to shape: install anything with pip or apt, and your storage persists across stop and start.
The trade against a VM is deliberate. A VM hands you the machine and the responsibility for it; a template hands you the morning back. Same GPUs, same rates, same per-minute billing, and the comparison below draws the line item by item.
Pick a stack, not a Dockerfile.
Every template ships with its framework, CUDA and cuDNN configured. pip and apt take it from there.
PyTorch
Tensor computation, autograd and GPU acceleration in a Pythonic API.
TensorFlow
End-to-end machine learning platform from Google.
FastAI
High-level PyTorch wrapper for rapid training, vision to NLP.
Axolotl
Fine-tune open LLMs like Llama and Mistral from a YAML config.
ComfyUI
Node-based workflow UI for Stable Diffusion pipelines.
Automatic1111
The Stable Diffusion web UI, with its extension ecosystem.
Fooocus
Text-to-image with minimal prompt engineering.
Kohya
LoRA, DreamBooth and QLoRA fine-tuning for Stable Diffusion.
Ollama
Run open LLMs with a simple CLI and API; bring your own UI.
One card, or eight of them.
Pick the GPU, then the count. Rates below are for the whole instance, on demand, billed by the minute.
| GPU | Total VRAM | vCPU | RAM | On demand |
|---|---|---|---|---|
| NVIDIA H200 SXM | 141 GB | 28 | 300 GB | $3.99/hr |
| NVIDIA H100 SXM | 80 GB | 16 | 200 GB | $2.69/hr |
| NVIDIA RTX Pro 6000 Blackwell | 96 GB | 28 | 160 GB | $1.89/hr |
| NVIDIA A100 80GB | 80 GB | 16 | 112 GB | $1.49/hr |
| NVIDIA A100 40GB | 40 GB | 16 | 112 GB | $0.89/hr |
| NVIDIA A30 | 24 GB | 16 | 112 GB | $0.41/hr |
| NVIDIA L4 | 24 GB | 32 | 124 GB | $0.44/hr |
vCPU and system RAM scale with the number of GPUs in the instance.
Instance storage is $0.00014/GB/hour. A paused template is charged storage only.
Work the way you already work.
One instance, three doors. Pick per session; the files are the same behind all of them.
JupyterLab
The notebook server is running before you ask. Open the URL, attach to the kernel, go.
VS Code in the browser
The full editor against the instance filesystem. Extensions, terminal, debugger. Or connect from your own IDE.
SSH
A real shell for people who live in one. Keys, tmux, rsync, everything you expect.
From notebook to URL, no re-deploy.
The model you just fine-tuned can answer requests from the same instance it trained on.
Expose what is running
Serve with FastAPI, Gradio, Streamlit or ComfyUI. Expose the port at launch with --http-ports, or one-click expose an API or notebook URL from the dashboard.
Demo without shipping
Hand a teammate the URL instead of a checkpoint. Pause the instance when the demo ends and the charge stops with it.
Graduate when it is real
When traffic deserves autoscaling and scale-to-zero, the same model moves to Serverless. A template endpoint is one box; that is the point, and the limit.
Templates or VMs?
Same GPUs underneath, at the same hourly rate. The difference is how much of the stack you want to own, and how fast you want to be working.
| Template (managed container)This page | On-Demand VM | |
|---|---|---|
| Time to first shell | 1.8 seconds | About 90 seconds |
| Pre-built stacks | PyTorch, ComfyUI and more, ready | You build the image |
| How you reach it | JupyterLab, VS Code in the browser, or SSH | SSH |
| Custom packages | pip and apt, freely | Anything the kernel allows |
| Access | Root inside the container | Root on the machine |
| Kernel and drivers | Maintained by us | Yours to choose and patch |
| Docker and Kubernetes | Not available inside | Your own daemon or cluster |
| Storage across stop and start | Persists | Persists |
| Billing | Per minute | Per minute |
| Pick this when | you want to start training right away | a container is not a machine |
From template to production.
What every template sits on.
The same platform under every instance, whichever GPU you picked and however many of them.
Your GPUs are never shared. On VMs you control the kernel, drivers and OS. Your own isolated VPC with IP ranges you choose. One filesystem, attached to as many of your own instances as you need.
Admins manage members, allocate credits, and see team usage.
Choose where your data lives. India and Europe today, more regions coming soon.
What people run on them.
Workloads researchers actually run, and the card that handles each one.
| Workload | Runs well on | From |
|---|---|---|
| Fine-tuning open models | NVIDIA A100-80GB | $1.49/hr |
| Image generation with ComfyUI | NVIDIA RTX Pro 6000 | $1.89/hr |
| Long-running experiment sweeps | NVIDIA L4 | $0.44/hr |
| Notebook prototyping | NVIDIA L4 | $0.44/hr |
Every template, from the terminal.
Create, connect, copy files and tear down without opening the dashboard. Your agents can drive it too.
A pre-built container with a primary framework, CUDA and cuDNN configured, and the notebook server and editor already running. You bring code and data; pip and apt handle the rest.
Yes, to a running environment. The stack is pre-built and warm, so launch is allocation, not assembly. A VM, which boots a whole machine, takes about 90 seconds; that is the trade the comparison above draws.
Both. SSH, JupyterLab and VS Code in the browser all reach the same instance and the same files.
Yes, with pip or apt, and they live on your persistent storage. What you cannot do inside a container is swap the kernel or run your own Docker daemon; that is what On-Demand VMs are for.
Yes. Expose ports at launch or one-click expose an API or notebook URL, and requests reach it while the instance is up. For autoscaling endpoints that sleep at zero, use Serverless.
GPU charges stop the moment you pause. You pay storage only, at $0.00014/GB/hour, and your environment and data are kept.
Yes. Storage follows the instance, not the card, so you can train on an A100, pause, and resume on an L4 for inference.
Not every GPU is offered at every count. NVIDIA H200 SXM, NVIDIA H100 SXM, NVIDIA RTX Pro 6000 Blackwell, NVIDIA L4 run at 8 per template. The rest are capped at 4, and the configurations table shows exactly which.
A $10 minimum to add credit. Billing is per minute with no minimum rental period. There is no free trial.