NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt

AgentX replays production-style Claude Code sessions with reasoning, tools, growing context and recorded timing. NVIDIA reports up to 30-times higher throughput per megawatt for Vera than GB300 at 160 tokens per second per user, while GB300 is reported at up to 15-times H200 throughput on DeepSeek V4 Pro.

The benchmark combines tokens per megawatt with interactivity, latency and time to first token; the stack includes SGLang, TensorRT-LLM, vLLM, DeepGEMM, Dynamo and NVLink.

This is NVIDIA’s own developer post, and it says Vera results await SemiAnalysis review. The methodology is useful, but the comparisons remain vendor-reported.

Signal Desk's read: AgentX is more valuable than its chart. Stateful requests create cache pressure and uneven shapes, so realistic replay is closer to purchasing reality. Yet this is a stack score, not a chip score.

Operators should reject throughput numbers that arrive with unacceptable TTFT or completion latency, especially across many turns.

The post distinguishes generation-only interactivity from end-to-end interactivity. It says fixed 8K/1K tests are now maintenance-mode scenarios.