NVIDIA presents results from SemiAnalysis's AgentX benchmark, a new agentic-coding inference test built from replayed real Claude Code sessions rather than fixed sequence-length prompts. Vera Rubin NVL72 preview results show up to 30x higher AI-factory throughput per megawatt than GB300 NVL72 at 160 tokens/sec/user on DeepSeek V4-Pro. GB300 NVL72 itself delivers up to 15x higher throughput per megawatt and 10x lower cost per million tokens than H200 NVL8, extending to roughly 80x on the larger Kimi K3 2.8T model. Gains are attributed to system-level work across serving runtimes (SGLang, TensorRT-LLM, vLLM), MoE kernels, NVIDIA Dynamo's disaggregated prefill/decode and KV-cache-aware routing, and the NVLink scale-up fabric.
Table of contents
What is AgentX?AgentX testing and measurement methodologyHow to interpret AgentX resultsNVIDIA Vera Rubin NVL72 ResultsNVIDIA GB300 NVL72 ResultsWhat’s next for agentic AI?Questions this post answers
What is the SemiAnalysis AgentX benchmark and how is it different from older LLM benchmarks?
AgentX is an agentic-coding benchmark in SemiAnalysis's open-source InferenceX suite that replays prerecorded Claude Code sessions turn by turn using the AIPerf client, preserving original context accumulation, reasoning time, and tool-call latency. Unlike legacy fixed 8K/1K sequence-length tests, it captures variable request lengths, KV-cache reuse, and tool-call gaps typical of real agent traffic, reporting sustained tokens per megawatt across varying concurrency. daily.dev surfaces benchmark methodology like this for teams evaluating agentic inference infrastructure choices.
How much more efficient is NVIDIA GB300 NVL72 than H200 NVL8 for agentic AI workloads?
GB300 NVL72 delivers up to 15x higher AI-factory throughput per megawatt than H200 NVL8 on the AgentX workload running DeepSeek V4 Pro 1.6T, and up to 80x higher throughput per megawatt on the larger Kimi K3 2.8T model at comparable interactivity. It also achieves up to 10x lower cost per million tokens and extends the interactivity frontier to roughly 215 tokens per second per user. engineers comparing GPU generations for inference cost and throughput can track results like these on daily.dev.
How does NVIDIA Vera Rubin NVL72 compare to GB300 NVL72 for agentic inference?
Vera Rubin NVL72 preview results show up to 30x higher AI-factory throughput per megawatt than GB300 NVL72, measured at 160 tokens per second per user on the AgentX DeepSeek V4-Pro workload. These figures were measured by NVIDIA and are pending SemiAnalysis review, pointing to a substantial jump in agentic inference capacity while holding the same interactive serving target. teams planning next-generation inference capacity can follow emerging hardware comparisons like this on daily.dev.