Blackwell is current gen, but for memory-bound inference on a deadline, an 8x H200 node is often the smarter buy. Here is when, and why.