Wafer AI serves GLM5.2 on AMD MI355X at 2626 tok/s, 2x cheaper than Blackwell

Wafer AI achieved 2626 tok/s/node throughput (213 tok/s single stream) on GLM5.2 using AMD MI355X hardware, at over 2x lower cost than NVIDIA Blackwell. The company overcame AMD's lack of day-0 support by using AI agents for kernel and model optimization. Performance was measured on a 20k input / 1k output workload with 60% cache hit rate.