A pricing comparison from SemiAnalysis puts AMD’s MI355X ahead of Nvidia’s GB200 NVL72 on cost per token when both serve Moonshot’s 2.8T-parameter Kimi K3. The gap holds across the whole interactivity range, and it widens the harder the system is pushed.
The headline number sits at the top end. At roughly 165 tokens per second per user, the AMD part lands at $0.15 per million tokens against $0.22 for Nvidia’s rack, a 46 percent advantage. Priced at the looser 71 tokens per second, MI355X reads $0.05 against $0.07, a 35 percent edge. The middle setting splits the difference at $0.14 versus $0.16.
The input column matters as much as the output. SemiAnalysis carries GB200 NVL72 at $1.86 per chip per hour and MI355X at $1.50, both drawn from its July 2026 pricing surveys. The benchmark holds FP4 precision with 8K input and 1K output context.
Why the win comes with a catch
Concurrency is where the tradeoff appears. At the loosest setting, Nvidia’s rack serves about 38 users per chip while the AMD part handles 44. Push to peak interactivity and Nvidia drops to roughly 8 against AMD’s 4. The per-token victory therefore rides on serving half as many people at once.
That shape suits batch coding and document workloads more than chat products with unpredictable traffic.
A methodology built by one vendor’s model
The figures come from SemiAnalysis’s own total cost of ownership model, not a neutral third party, and the firm states the assumption plainly. Chip prices and utilization rates drive nearly everything published.
For buyers, the takeaway is less about which accelerator wins than about matching silicon to traffic patterns.
