# AI21 moves GPU scheduling from manual requests to Kueue automation

> AI21 Labs replaced Slack‑style GPU negotiations with Kueue, cutting manual interventions and job starvation on a 10,000‑GPU GKE cluster.

Oossa · 2026-10-04 · https://oossa.com/en/ai21-moves-gpu-scheduling-from-manual-requests-to-kueue-automation

AI21 Labs switched its shared GPU pool of about 10,000 GPUs on Google Kubernetes Engine from a manual #gpu-resources channel to Kueue, an open‑source job queueing system. Kueue automatically admits, prioritizes and preempts jobs, reducing the need for people to coordinate requests. The change lowered manual intervention rates and the time critical jobs spent without resources. AI21 worked with the Kueue team, prompting new features like Admission Fair Sharing.

## The facts

- Around 10,000 GPUs are managed in a single GKE cluster.
- AI21 saw an 83% reduction in time‑to‑start for high‑priority workloads.

## Why it matters

For AI researchers at AI21, the automation means less time spent negotiating GPU access and faster starts for important training jobs.

## Sources & references

1. [From manual negotiation to automated scheduling: How AI21 manages its GPU fleet with Kueue Google Cloud recently published our story on AI Hypercomputer, including an 83% reduction in time-to-start fo](https://www.ai21.com/blog/how-ai21-manages-its-gpu-fleet-with-kueue/) – AI21 Labs

Last updated: 2026-10-04