Batch prediction jobs submitted to locations/global sit in JOB_STATE_QUEUED until they are killed at the 72h mark with “The job is queued beyond the maximum allowed time (72h)”. Roughly 4,000 jobs over four days on a paid project, none of which ever reached JOB_STATE_RUNNING. Synchronous generateContent works normally on the same project and the same model throughout.
The same one-line input submitted to us-central1 starts in under two minutes. Submitted to global it never starts at all. Only the location differs, and a GA model (gemini-2.5-flash) hangs on global too, so it is not model-specific starvation.
This is already filed and assigned as Google Issue Tracker — open since 2026-01-16 with no update since January. Full detail, the controlled experiment and a quota hypothesis are in comment #4 there. If you are seeing stuck global batch jobs, a +1 on that issue would help move it.
Posting here mainly to find out who else is affected. Two specific questions for anyone who has hit this. First, has anyone found a model that does dispatch for batch on locations/global? Second, has anyone found a region that accepts gemini-3-flash-preview for batch? We tested ten and it 404s or is rejected in all of them, which leaves global as the only endpoint serving it for batch — and global is the one that never dispatches.