---
title: "Some nodes lost access to a GPU"
description: "Resolved · Sep 30, 2026"
base_url: "https://status.autoresearch.sfcompute.com"
canonical: "https://status.autoresearch.sfcompute.com/events/report/3126"
---

# Some nodes lost access to a GPU

[Status](/.md) › [Events](/events.md) › Some nodes lost access to a GPU

Sep 30, 2026 · 1 day ago · 8 hours

## Updates

### + Resolved — Sep 30, 10:45 PM

The affected hosts have been repaired or replaced, and GPU availability is back to normal.

### x Investigating — Sep 30, 2:50 PM

A GPU on a small number of hosts has become unavailable, so a node or job running there may see CUDA errors or fewer GPUs than expected. We are repairing or replacing the affected hosts. If your workload is on one, checkpoint what you can and recreate the node - new nodes land on healthy hosts.

---

_Powered by [openstatus.dev](https://openstatus.dev)_