Measured on real agentic coding trajectories, Nvidia's rack multiplies the throughput per megawatt by thirty compared to the GB300 NVL72, and divides the cost of the token by 45.
Nvidia points to energy efficiency metrics for its Vera Rubin NVL72 rack, obtained from AgentX, SemiAnalysis' loadset built from real-world agentic coding sessions—including context growth, tool calls, and subagents—rather than just isolated queries.
On these loads, served at a high interactive rate and with a model of the caliber of DeepSeek V4 Pro, the manufacturer claims up to thirty times the throughput per megawatt of the GB300 NVL72 and a cost per token divided by 45. The GB300 itself provided about fifteen times the efficiency of the Hopper architecture: the gap widens from one generation to the next.
The argument is based on a sensitive point: an agentic task consumes about fifteen times more tokens than a traditional conversation, since it combines research, tool calls and synthesis. In data centers limited not by the purchase budget but by the power connection, the number of tokens per megawatt becomes the real measure of capacity.
Read the full article on the original source
Go to the source →