InferenceX Research
Head-to-head AgentX results for new accelerator systems: tokens per dollar of TCO, per megawatt and per chip at matched interactivity.
11 min read
At 200 tok/s/user on Qwen3.5-397B, TPU v7 FP8 delivers 1.85x the tokens per $1 TCO of GB300 NVL72 FP8 on external TCO and 2.17x on internal TCO. GB300 NVL72 FP4 on Dynamo TensorRT-LLM leads at every matched point on both TCO bases. On Kimi K3, GB300 NVL72 leads on external TCO and TPU v7 on internal TCO from ~53 to ~149 tok/s/user.
Read the results11 min read
On InferenceX AgentX at 100 tok/s/user (120 on GLM-5.2), Vera Rubin NVL72 leads GB300 NVL72 by 1.8x to 2.0x in total tokens per $1 TCO and in token throughput per MW on all four models with Vera Rubin results. The per-dollar ratio for each model peaks where every GB300 NVL72 config serves 4 agents or fewer, up to ~10.2x on MiniMax-M3.
Read the results