GPUs on Kubernetes are bring-your-own-everything, starting with the bill

The New Stack explains why GPU accounting on Kubernetes is far harder than CPU and memory: NVIDIA's DCGM reports at the physical device level, time-slicing and MPS make every sharing pod look identical, and teams end up billing on memory because compute attribution is unreliable.

• DCGM reports per physical device, not per pod, by default
• Three pods sharing one GPU can look like 300% utilization
• GPUs are requested as whole numbers with no scheduler overcommit
• Accurate compute attribution between teams is not possible today