# NomadGPU

> NomadGPU runs GPU jobs (fine-tuning, batch processing, inference) on low-cost marketplace GPUs from Vast.ai, on the
> client's own account. It picks reliable hosts, saves checkpoints to object storage every few minutes and moves the
> job to a new machine when a host fails, with hard spending limits. The client pays the GPU platform directly;
> NomadGPU charges only a service fee.

## What it does

- Picks a host by reliability, network, disk and PCIe speed and the full hourly price including traffic.
- An agent in the container uploads checkpoints and reports a heartbeat; a supervisor outside the GPU machine detects
  a lost host, a GPU that stopped responding or a job with no progress, rents a replacement and resumes from the last
  checkpoint.
- Spending limits per job, day, week and month: a warning at 80%, at 100% the machine is released and progress is kept.
- Notifications in the account, by email and in Telegram.

## Measured numbers

- QLoRA fine-tune of Qwen2.5-7B-Instruct, 300 steps, one RTX 4090 on Vast.ai: about $0.19 from the Vast.ai invoice,
  versus $1.35–1.80 for the same job on one A100 at AWS on-demand rates (86–89% cheaper on this job).
- Live supervisor numbers (uptime, machine failures handled, median time to resume) on the public status page:
  https://status.nomadgpu.com/s/status — based on the operator's own test jobs, so the sample is small.

## Pricing

- Self-serve: 8% of GPU spend, at least $5 in months with GPU use, at most $19 per month; no fee in months without jobs.
- Starter: $39 per month, for GPU bills up to about $1,500 per month.
- Managed: from $300 per month, for GPU bills from about $1,500 per month.
- First 14 days free on Self-serve and Starter. GPU time is always billed by the GPU platform to the client's account.

## Limits (stated honestly)

- The GPU host can see the data while a job runs; not for personal, medical or confidential data. A data-center-only
  mode lowers this risk but does not remove it.
- One GPU machine per job; no multi-node training.
- Start-up is slower than a big cloud: each new machine downloads the image and the model.
- No SOC 2 or ISO 27001 certification.

## Links

- Home: https://nomadgpu.com/
- Security: https://nomadgpu.com/security.html
- Terms of Service: https://nomadgpu.com/terms.html
- Privacy Policy: https://nomadgpu.com/privacy.html
- Status: https://status.nomadgpu.com/s/status
- Sign up: https://app.nomadgpu.com/portal/signup
- Contact: ops@nomadgpu.com