Network Load Balancer and failure conditions across multiple AZs

To ensure I'm understanding, let me rephrase your example to address what I think you are asking.

You have a 3 AZ NLB with 100 targets per AZ in the same Target Group. 300 total targets.

In AZ-a you remove 99 of the targets, leaving 1 target. In the other AZs you still have 100 targets.

Assuming all targets are healthy and clients do not favor any of the IPs/AZs... Your 1 target remaining in AZ-a will get traffic based on the cross zone routing setting.

If cross zone load balancing is enabled, target 1 will get the same amount of traffic as each of the other 200 targets, or 1/200th of the traffic.

If cross zone load balancing is disabled, target 1 will get the same traffic as each of the other AZs, or 1/3rd of the traffic.

The reason for this is that NLB does load balancing at 2 layers, just like ALB and CLB. The first layer is DNS level load balancing, where all IPs in DNS are distributed to clients, and clients should be connecting to all IPs evenly. In ALB/CLB this can be a lot more total IPs (currently up to 100 total), but in NLB this is only 1 per AZ, or 3 in this case. The second level of load balancing is at the load balancer itself, and this is where the key aspect of what I think is your question comes into play. This is where each NLB IP/AZ considers all available healthy targets equally. If this includes all AZs, then you will get even amongst all targets in all AZs. If this is only the local AZ (local to where the request arrived), then you will get even amonst the targets in that AZ, which is only 1 in this case.