Our control plane has been public for a while, but I don't think I ever posted it here. We've been working on AI infra for a few years, and have built a highly optimized set of primitives for GPU inference and sandboxes. This includes all sorts of bells and whistles like embedded CAS, a custom image format, multiple container runtimes supported (gVisor/runc), and an opinionated python SDK. Totally self-hostable (of course, you'll need to find some GPUs) on any k8s cluster with a helm chart.