2026-11-04 · Fonte：Nvidia Blog
NVIDIA says Vera Rubin NVL72 achieves up to 30x more work per watt, as OpenRouter data shows agentic AI consumes 15x more tokens than simple chat.
NVIDIA is positioning its Vera Rubin NVL72 platform as an answer to the surging computational appetite of agentic AI, claiming in a company blog post that the system delivers up to 30 times more work per watt. The company frames the figure as a new efficiency standard for AI agents, tying hardware economics directly to the workload type the announcement puts at center stage.
The claim is grounded in usage data from OpenRouter, which shows that agentic AI workloads consume 15 times more tokens than a simple chat request. The gap reflects how agents operate: rather than answering a single prompt and stopping, they chain together many model calls to complete a task, multiplying token consumption at each step.
NVIDIA illustrates the dynamic with an investment-research scenario. An agent asked to research a company for an investment decision queries financial databases, searches news and filings, invokes a sub-agent to run peer comparisons and model valuations, and then synthesizes the results. The example shows why agentic token consumption compounds: a single user request sets off database queries, document searches and delegated sub-tasks, each generating additional inference work that a one-shot chat exchange never triggers.
The per-watt framing is the crux of NVIDIA's argument. As agents multiply the model calls behind a single task, efficiency per watt becomes the yardstick for how much usable work a system can extract from constrained power. By claiming up to 30 times more work per watt, NVIDIA is presenting Vera Rubin NVL72 as calibrated to exactly this token-heavy, multi-step pattern of demand. In that framing, the relevant benchmark for AI infrastructure is not peak performance but output per unit of energy.
Tags: NVIDIA · Vera Rubin NVL72 · OpenRouter · AI Agents · Energy Efficiency
萬安算交所 AIXX · AIXX · Exchange verificado · Ir ao mercado