SemiAnalysis: AMD MI355X Out-Earns GB200 NVL72 on Kimi K3 Inference

SemiAnalysis is making a claim about the business of serving a model, not only about speed. In a post on X, it says that when serving Moonshot AI's Kimi K3, a 2.8-trillion-parameter model, AMD's MI355X earns a 53.6% profit margin against 44.3% for Nvidia's GB200 NVL72, and that AMD also beats Nvidia's B300 and B200 on the same measure. The figures come from AgentX, SemiAnalysis's open-source benchmark for agentic inference, the work of running a trained model to answer AI agents. SemiAnalysis's earlier write-up on AgentX describes the benchmark itself.

SemiAnalysis gives two reasons for the gap: AMD's lower total cost of ownership, meaning the full cost of buying and running the hardware, and better performance on this workload. A margin of this kind folds both into one number, because it depends on how much a system can serve and on what that output costs to produce. A system can win on margin by being cheaper as well as by being faster, and SemiAnalysis says AMD did both here.

The post frames the comparison per gigawatt of capacity, but its text does not give the assumptions that turn benchmark throughput into a margin, such as the price charged per token or the hardware and power costs used, and the text gives no margin figures for B300 or B200. SemiAnalysis also credits two engineers with optimizing the 2.8-trillion-parameter model, so the result reflects tuned software on one model and one workload. Read it as SemiAnalysis's modeled economics for serving Kimi K3, which favor MI355X under assumptions its text does not state, rather than as evidence that AMD hardware out-earns Nvidia across models.