GLM5.2 runs 2x cheaper on AMD MI355X than Nvidia Blackwell

Wafer.ai demonstrated GLM5.2 inference at 2626 tokens/second per node on AMD's MI355X GPU, achieving significantly lower costs than Blackwell-based alternatives. For indie teams running inference at scale, this challenges the assumption that Nvidia's latest silicon is the only viable path—AMD hardware now offers a concrete cost-per-token advantage.