Your Docker image,
resumed on the cheapest box.

A drop-in wrapper for Hugging Face Trainer and PyTorch Lightning jobs on interruptible GPUs. No refactor, no new framework — when a spot instance goes away, your run keeps going on the next-cheapest one.

Zero refactor

Point it at the image
you already run.

ComputePulse wraps the training image you run on Vast.ai or RunPod. At launch it supports two frameworks — where it can hook checkpointing and resume without touching your code.

Docker

The wrapper sits around the image you already built — nothing to rebuild.

Hugging Face Trainer

Checkpointing and resume are hooked automatically, no script changes.

PyTorch Lightning

Supported at launch alongside Hugging Face Trainer — more later.

Built for Vast.ai and RunPod spot instances first.
Easy to point at

Watch your own run resume.

The researcher stops thinking about which machine is safe and just starts the run on ComputePulse. When a spot instance is reclaimed, this is what shows up in the log.

Sample log
$ computepulse run --image ghcr.io/you/finetune-7b:latest
> watching checkpoint dir ./ckpt
> training on vast.ai, node 41c2 (spot)
> interruption detected — node 41c2 reclaimed
> relaunching on runpod, cheapest available node
Resumed from 14:05 UTC checkpoint, 3 min lost
{ event: "run_resumed", node: "runpod-a91f" }
How it works

Two moves, no hands
on the keyboard.

Step 1

Wrap the image

Point ComputePulse at the Docker image you already run on Vast.ai or RunPod. No refactor, no new framework — just the Hugging Face Trainer or PyTorch Lightning job you already have.

Step 2

It resumes for you

When the spot instance is reclaimed, ComputePulse relaunches your job on the next-cheapest provider from the last checkpoint — no manual restart.

Pricing

Priced at what it saves,
not what it manages.

A monthly subscription, tiered by the GPU spend ComputePulse manages. Each tier is priced near 10% of that tier's typical savings, billed through NanoCorp's hosted Stripe checkout.

Solo
$49/month

Up to $3k/mo managed GPU spend

For one researcher running interruptible GPUs alone.

Lab
$199/month

Up to $15k/mo managed GPU spend

For a small group running several fine-tunes in parallel.

Team
$599/month

Up to $50k/mo managed GPU spend

For a team with mission-critical runs across providers.

Savings calculator

What the cheap machines save you.

Published on-demand prices against the interruptible marketplace, per GPU and hours of training per day. The gap is what you keep — if a lost run doesn't take it back.

24 h320 h

On-demand

$1,776/mo

Interruptible, resumed

$1,104/mo

You keep

$672/mo

Published prices, checked October 2026 — on-demand from RunPod's Secure Cloud list, interruptible from Vast.ai marketplace medians. Both move by the hour; the real saving depends on the machine your run lands on. The plan below costs a small share of this number.

Protect a first run

Stop thinking about
which machine is safe.

Start the run on ComputePulse. It keeps going, wherever the price drops next.