AI infrastructure
Your GPUs run jobs. We run the control plane.
For AI infrastructure and ML platform teams running training and inference on their own GPU servers. Your Kubernetes control plane runs on our infrastructure, so a failed GPU node never takes the API server with it.
The problem
1.
More machines to run
A highly available control plane needs at least three more machines to buy, rack, patch, and watch, next to hardware you bought to run jobs.
2.
Bursty load on etcd
Large training and batch jobs create pods, events, and watches in bursts. When etcd or the API server struggles, scheduling stalls for every job.
3.
Hardware fails at scale
GPUs, NICs, and drives fail. When control plane members share racks and power with your workers, a hardware fault can become a cluster outage.
What Kestrel runs for you
A dedicated, highly available Kubernetes control plane. Your workers stay on your nodes.
Dedicated API server and etcd
Your own API server and etcd cluster, sized for your cluster and shared with no one.
GPU nodes join like any node
Workers are standard Kubernetes nodes. Install the GPU drivers, device plugins, operators, and batch schedulers you already use.
Separate failure domains
Drain, reimage, or lose a GPU server, and the control plane keeps scheduling.
Regular etcd snapshots
Cluster state is backed up with regular etcd snapshots, so it outlives a bad day.
Questions
Things people ask us.
Do you run our GPUs or host our models?
No. Your GPU servers, models, and data stay in your environment. We run only the Kubernetes control plane.
Can we keep our GPU tooling?
Yes. Workers are ordinary Kubernetes nodes, so the GPU drivers, device plugins, operators, and batch schedulers you install run as they would on any cluster.
Does it work for both training and inference?
Yes. The control plane schedules pods. Whether those pods train models or serve them is up to you.