AMD undercuts Nvidia in Kimi inference cost comparison

AMD’s MI355X delivers lower modeled inference costs than Nvidia’s GB200 NVL72 when running Kimi K3 2.8T, according to SemiAnalysis’s InferenceX comparison. At the highest tested interactivity level, 165 tok/s/user, costs reach $0.15 per million tokens for AMD versus $0.22 for Nvidia. SemiAnalysis describes that result as “46% cheaper.” The benchmark holds precision at FP4 and context at 8K in, 1K out.

The comparison also favors AMD at less demanding operating points. At 71 tok/s/user, MI355X costs $0.05 per million tokens against $0.07 for GB200 NVL72, with a reported 35% cost-per-token gap. At 118 tok/s/user, the respective costs are $0.14 and $0.16, described as a 20% advantage. These estimates use “owning hyperscaler” rates of $1.86 per chip per hour for Nvidia and $1.50 for AMD, drawn from SemiAnalysis’s July 2026 pricing surveys and AI Cloud TCO model.

Concurrency complicates the cost advantage. At 71 tok/s/user, Nvidia serves around 38 users per chip compared with AMD’s 44. At 165 tok/s/user, Nvidia supports near 8 users while AMD supports near 4. AMD’s lower per-token cost at peak interactivity therefore comes with roughly half the users per chip. The results depend on SemiAnalysis’s modeling assumptions holding at scale.