The same GPU is not always the same GPU

Accelerator requests usually arrive with the two fields that appear on benchmark charts: model and memory. Those decide performance. They do not decide whether the thing will run in your rack.

Form factor does. A module built for a vendor baseboard and a PCIe card carrying the same chip need different chassis, different power delivery and different airflow. They are quoted through different channels on different lead times, and one is not a cheaper version of the other.

Cooling is the constraint people find last. Passively cooled cards assume server airflow that a workstation does not produce. Actively cooled cards assume intake clearance that a dense 1U does not have. Both look fine on the bench and then run hot in the rack.

Power is next. A card inside the slot budget on paper can still exceed what a particular supply and connector population will deliver, especially with several installed in one machine.

Interconnect is the one that is simply yes or no. If the workload depends on a high-bandwidth link between cards, anything without it is a different requirement, not a discount.

So name the machine alongside the card. We can check it fits before quoting, which is the point at which a mismatch costs nothing to fix.