A benchmark that replays real agent sessions gives back less of the generational win

The interesting arithmetic is in neither post. On the fixed 8K/1K scenario, GB300 NVL72 was credited with up to 40x more tokens per megawatt than H200 on DeepSeek-R1-0528 ^[14]. On AgentX, running DeepSeek V4 Pro, the same rack is credited with up to 15x Hopper ^[6]. Different models and different baseline names, so this is not a clean subtraction, but the direction holds: the replayed workload returns roughly 37 percent of the static workload's generational advantage ^[19].

That gap is the content of the benchmark change. A fixed 8K/1K request has a known cache footprint and no dead time. A replayed Claude Code session accumulates context across turns, varies input and output length from request to request, and halts while a tool runs ^[2]. AgentX holds those reasoning and tool-call intervals at their recorded durations, which is what reproduces KV cache capacity pressure: the cache stays resident and provisioned while nothing is decoding ^[3]. Contexts that reach hundreds of thousands of input tokens do the rest ^[9]. Memory capacity and cache reuse set the ceiling, not peak decode rate.

Which makes one disclosure load-bearing. NVIDIA states that its Vera Rubin figures do not yet reflect Vera CPU performance for tool calling ^[7]. Tool calling is the behaviour AgentX was built to preserve, and the 30x is quoted at a single operating point of 160 tokens per second per user on a curve the benchmark deliberately sweeps across four user-experience targets ^[4][11][13].

There is also a denominator to watch. NVIDIA's DSX MaxLPS manages power at GPU, rack and workload level to fit up to 40 percent more GPUs inside the same megawatt budget ^[17]. Tokens per megawatt is therefore partly a power-management policy result. Anyone comparing that metric across vendors is comparing provisioning discipline alongside silicon, and the two are not separable from the published figure.

The cost claim moves faster than the power claim. Vera Rubin NVL72 is quoted at up to 35x lower cost per million tokens than GB300 while delivering up to 30x the throughput per megawatt, so cost per token improves about 17 percent more than energy per token does ^[16][21]. By NVIDIA's own framing, throughput per megawatt sets AI factory revenue and cost per million tokens sets the margin on it ^[8]. The extra 17 percent has to come from rack capital cost, utilisation, or the price of an accelerator hour, and neither post breaks it out.

Stack the two "up to" numbers and you get 450x over Hopper on this workload ^[20]. Neither post asserts that, and no single measurement supports it, which is roughly the hazard of a per-megawatt metric assembled from two hops with different baselines. Both posts are NVIDIA's, and both note the Vera Rubin results are pending SemiAnalysis review ^[5]. The suite is open source and every system receives the same recorded traffic ^[12], so the review is checkable in a way the old static number never really needed to be.