Reference
Incident log
Every incident and maintenance window, newest first.
Live state is on the status page. This page is the record. An incident is any time an Anvil GPU could not run a job for a reason on our side, or the API or console failed for more than a minute.
Entries
| Date | Where | Kind | Length | What happened |
|---|---|---|---|---|
| 2026-09-17 | F1 node B | Incident | 13 min | gpu11 dropped off the PCIe bus. It left scheduling within 40 s. One Strike job ended with reason hardware and was not billed. Node B placed no new jobs for 13 min while the GPU was reseated and tested. |
| 2026-09-08 | F1 | Maintenance | 2 h 0 min | NVLink switch firmware update, announced 2026-09-01. Ember jobs got their 30 s notice. No Strike or Anvil job was running past the start of the window. |
| 2026-08-30 | API | Incident | 4 min | Job submission returned 500 after a config push and was rolled back. Running jobs and the heat feed were not affected. |
| 2026-07-28 | F1 | Commissioned | · | Listed after 96 h of burn in at 97.4 % mean load on all 16 GPUs. |
How entries are written
Start time in UTC, what was affected, length to the minute, what happened to running jobs, and whether those minutes were billed. Written within one business day of the end. If the cause is still unknown, the entry says so and is updated.
Following it
GET /incidents returns the log as JSON, and an Atom feed is at /status/feed.xml. Anvil teams get an email for every incident on their node.
Forge F1 listed 2026-07-28.