AMD MI355X Up to 46% Cheaper Than GB200 NVL72 on Kimi K3 2.8T

At the highest interactivity level in SemiAnalysis's InferenceX head-to-head on Kimi K3 2.8T, AMD's MI355X lands at $0.15 per million tokens against $0.22 on Nvidia's GB200 NVL72. The page calls the AMD part "46% cheaper," per the InferenceX comparison.

The gap narrows at other operating points. At 71 tok/s/user, the page reports "MI355X edges GB200 NVL72" with the AMD part at $0.05 per million tokens against $0.07, "a 35% cost-per-token gap." At 118 tok/s/user, MI355X comes in at $0.14 against $0.16 on GB200 NVL72, a 20% advantage.

The input column: GB200 NVL72 at $1.86 per chip per hour, MI355X at $1.50, both quoted as "owning hyperscaler" rates from SemiAnalysis's July 2026 pricing surveys and AI Cloud TCO model. The benchmark holds precision at FP4 and context at 8K in, 1K out.

Concurrency is where the trade-off shows. At the loosest setting, 71 tok/s/user, GB200 NVL72 is listed as serving around 38 users against MI355X's 44. Pushed to 165 tok/s/user the Nvidia part lands near 8 users and the AMD part near 4, so MI355X's per-token win at peak interactivity rides on roughly half the users per chip.

This is our 326th chip story in 90 days and the first SemiAnalysis InferenceX comparison on Kimi K3 2.8T we have tracked. The methodology is one vendor's model.