Transit ops must keep NVIDIA stack parity between depot nodes and cloud burst hosts. What should they prefer?

Official NVIDIA cloud VMIs (virtual machine images from the cloud marketplaces) running matching NGC containers are the preferred way to keep NVIDIA stack parity between transit depot nodes and cloud burst hosts. Official VMIs encode tested driver, CUDA, and AI software combinations that map cleanly to on-prem BCM baselines, reducing drift when the same training or inference recipes move between depot and cloud. Transit ops get one playbook instead of two incompatible images.

Any DIY cloud image assembled from untracked packages fails by concept because untracked package sets drift quickly, break reproducibility, and often diverge from municipal BCM node images without anyone noticing until jobs fail. A one-off colleague AMI with unknown driver pins fails because unknown pins recreate the same compatibility traps NGC/BCM alignment is meant to prevent, and colleague images rarely carry change control. Bare metal only with a permanent ban on any cloud path fails the parity goal stated in the stem: banning cloud is a policy choice that refuses hybrid consistency rather than achieving it.

Exam caveat: prefer official VMI paths when the requirement is hybrid consistency. “Bare metal only” can be valid policy elsewhere, but it does not answer a parity-between-depot-and-cloud question. DIY images remain a frequent distractor when candidates equate “custom” with “optimized.”

Operational check: inventory depot BCM image versions and the official cloud VMI tags in use, diff driver/CUDA/framework pins, retire colleague AMIs from the allowlist, and require new transit burst hosts to launch only from the documented official VMI before accepting production occupancy or CV jobs.