Network and operational overhead

Add bandwidth, PUE-already-in-power, and the people/parts load that does not show up on the GPU nameplate.

Bandwidth

Assume 2 KB egress per generated token (tokenizer + JSON envelope). At 160 tokens/s:

160 tok/s × 2 KB × 3600 s  =  1.15 GB / compute-hour
1.15 GB × $0.09 / GB       =  $0.104 / compute-hour

Cite AWS egress. On-net or settlement-free peering can zero this line; do not invent a lower number without a bill.

8–15% of hardware + power as annual opex. On the mid 24/7 A100 pair:

($1.103 + $0.115) × 0.10  =  $0.122 / compute-hour

Use 10% unless you have payroll that says otherwise.