- Own the layer above bare metal: managed Kubernetes and GPU VMs as products with service levels, monitoring, adoption, and direction.
- Build to the criteria behind the top tier of ClusterMAX, measured on paying customers and fleet readiness.
What you'll do:
- Stand up managed Kubernetes on bare metal from scratch with hosted control planes and GPU/network stack.
- Build GPU VMs and turn one pool of capacity into reserved, on-demand, and spot products with quota and preemption.
- Build isolation between customers and make failures internal, not customer-facing.
- Make every operation work for agents and people through API, CLI, and agent tools.
Good fit if you have:
- Built and run production software that people depended on, owning it when it broke (2-10 years experience).
- Run production Kubernetes on bare metal including provisioning, upgrades, and node operations.
- Debugged GPU failures in production: drivers, Xid errors, NCCL hangs.
- Working knowledge of KVM/QEMU, KubeVirt, and GPU passthrough.
- Owned a product other developers paid for or depended on, including its service levels.
- Used coding agents effectively on infrastructure and know which changes a person must check.
Nice to have:
- Hosted control planes with Gardener or k0smotron.
- Bare-metal provisioning with Ironic, Metal3, and Redfish.
- Network fabrics like Netris or Hedgehog, and BlueField DPUs.
- GPU partitioning: MIG or NVLink partitions through Fabric Manager.
- Quotas and preemption with Kueue.
Perimeter Compute turns unused building power into GPU clusters for superintelligence inference. It is a small, highly leveraged team with experience building GPU infrastructure at Lambda and HPE, commercial buildings at Tishman Speyer and Trane, and routing inference at Subquadratic.