Kubernetes was born in a virtualized world. Every cluster assumed something underneath it with an API (EC2, vSphere, OpenStack, a hypervisor of some kind) and that assumption turned into a declarative contract. AI race broke it. Accelerators want the whole node, so the hypervisor goes away and hands back everything it was hiding: BMCs, firmware, PXE, vendor quirks. The mighty hardware, and bare metal does have an API. It’s just not the one you wanted. Then the bill arrives. Racks, power, cooling, depreciation from day one. At ten servers you can hand-roll it. At a thousand, such as production scale, the bottleneck stops being hardware and becomes operations, and the money goes straight to /dev/null. From my experience working with Mistral AI to build their bare-metal Kubernetes compute layer: the production-grade architecture to close the gap between racks and declarative infrastructure, and how to design so utilisation scales instead of headcount. No silver bullets, just the architecture and the reasoning behind it.