Kimi K2.5/K2.6/K2.7-Code 1T — GB300 NVL72 vs H200
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and H200 (NVIDIA Hopper) on Kimi K2.5/K2.6/K2.7-Code 1T. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
GB300 NVL72: 7910 tok/s/GPU, $0.09 per million tokens at 60 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T. H200 is unmeasured here.
At 104 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T, GB300 NVL72 delivers 576 tok/s/GPU at $1.28 per million tokens; H200 hasn't been benchmarked at this target.
GB300 NVL72 hits 194 tok/s/GPU for $3.76 per million tokens at 148 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T. No H200 data at this operating point. (Numbers reflect the default 1k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:7910.3H200:— | GB300 NVL72:576.5H200:— | GB300 NVL72:194.0H200:— |
| Cost ($/M tok) | GB300 NVL72:$0.091H200:— | GB300 NVL72:$1.285H200:— | GB300 NVL72:$3.759H200:— |
| tok/s/MW | GB300 NVL72:3731285H200:— | GB300 NVL72:271928H200:— | GB300 NVL72:91527H200:— |
| Concurrency | GB300 NVL72:~3377H200:— | GB300 NVL72:~119H200:— | GB300 NVL72:~28H200:— |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.