An administrator has configured a VM with passthrough GPU to allow direct access to GPU...

Comprehensive and Detailed Explanation From Exact Extract of Nutanix End User Computing documents:

According to Nutanix AHV Administration Guide regarding GPU Pass-Through:

Hardware Binding: GPU Pass-Through (DirectPath I/O) binds a virtual machine to a specific physical PCIe device on the host. Unlike vGPU (which abstracts the hardware), passthrough creates a 1:1 dependency.

HA Behavior: Nutanix AHV supports High Availability (HA) for GPU-enabled VMs in the sense that if a host fails, the system attempts to restart the VM on another host (Cold Migration). Option C ("Not HA-capable") is incorrect because the system attempts the restart; it does not simply abandon the VM by design.

Failure Cause: The restart fails (the "behavior" described) because the destination host must have the exact same GPU hardware resource available (free and matching the profile/addressing). If the other nodes in the cluster do not have a free GPU of the same type, or if they are already fully utilized, the VM cannot power on, resulting in the "Insufficient GPU resources" error.