# GPU sandboxes for agents

Replicas agents can now spin up their own cloud GPUs to run training, inference, rendering, and CUDA work, with zero setup on your end.

- Author: Connor Loi
- Published: 2026-10-19
- Category: Announcements
- Canonical: https://replicas.dev/blog/gpu-sandboxes

Replicas agents can now spin up their own GPUs. Ask an agent to fine-tune a model, run inference, render a scene, or test CUDA code, and it starts a cloud GPU, runs the job, copies the results back, and shuts the GPU down. You don't need a cloud account, a driver install, or a separate machine.

<video autoplay controls loop muted playsinline preload="metadata" poster="/blog/gpu-sandboxes-hero.png" aria-label="An agent starting an H100 sandbox to fine-tune a model, then an L4 for rendering" style="width: 100%; border: 1px solid var(--accent);">
  <source src="https://d1y8o87cun9pyo.cloudfront.net/2026-10-19/gpu-sandboxes-launch.mp4" type="video/mp4" />
</video>

## Why agents needed GPUs

Workspaces don't have GPUs. That was fine for most engineering work, but ML work kept ending the same way: the agent wrote the training script, then a person had to copy it to a GPU box, run it, and paste the results back. The agent couldn't check its own work, so it couldn't finish the task.

GPU sandboxes, powered by [Modal](https://modal.com), fix that. The agent decides when a task needs a GPU, picks one, and verifies the result itself.

## Ask in the chat

Describe the task. For example:

- **"Fine-tune our support model on last month's tickets."** The agent starts an H100, trains, and pulls the checkpoints back into the workspace.
- **"Benchmark the new embedding model against the current one."** The agent runs both on the same GPU and reports the numbers.
- **"Render the product scene at 4K."** The agent starts an L4, renders, and copies the frames back.

Under the hood, the agent uses the `replicas gpu` CLI, which is installed and signed in in every workspace:

```bash
replicas gpu run --gpu H100 --pull outputs -- python train.py
```

That one command starts an H100, copies the code in, streams the output, copies `outputs/` back, and stops the GPU. For longer jobs the agent can keep a sandbox open, run work in the background, and follow the logs.

## Any GPU

T4, L4, A10G, L40S, A100, H100, H200, and B200, with up to 8 GPUs per sandbox. Each workspace can only see its own GPU sandboxes, provider credentials never enter the workspace, and sandboxes stop when the job goes idle or the workspace goes to sleep. GPU time is billed at cost and counts toward included usage.

Learn more in [GPU sandboxes](https://docs.replicas.dev/features/gpu-sandboxes).