NVIDIA Vera Rubin NVL72 Claims Up to 30x More Work Per Watt, Setting a New Efficiency Bar for AI Agents

NVIDIA published a blog post claiming its Vera Rubin NVL72 platform sets a new efficiency standard for AI agent workloads, with up to 30x more work per watt.

The post cites OpenRouter data showing that agentic AI workloads consume 15x more tokens than a simple chat request, positioning agents as a major driver of inference demand.

Using investment research as an example, the post walks through an agent's workflow: querying financial databases, searching news and regulatory filings, invoking a sub-agent for peer comparisons and valuation modeling, then synthesizing everything into a report.

Such multi-step reasoning, tool calling, and sub-agent collaboration multiply token consumption per task, making cost per token a decisive factor in whether agent applications can scale.

In NVIDIA's framing, the economics of AI factories are defined by delivered output — work per watt, token costs, and utilization — rather than by raw accelerator counts.

Vera Rubin NVL72 is NVIDIA's next-generation rack-scale platform for AI factories, and this efficiency push targets agent workloads, one of the fastest-growing inference scenarios, as the company seeks to defend its position in inference computing.

The question now is whether the 30x figure holds up in third-party benchmarks and real customer deployments, and what it means for cloud purchasing decisions and token pricing.