According to the headline of the NVIDIA Blog post, the NVIDIA Vera Rubin NVL72 sets a new benchmark for AI task efficiency: up to 30imes more work per watt.
The post also cites OpenRouter data, stating that such workloads consume 15imes more tokens than a simple chat request. An investment decision analysis by a company is mentioned as an example of this workload.
The practical implication of the claim lies in evaluating computational efficiency not just by hardware power, but by the volume of work performed. However, the available package contains only a metadata synopsis, so the calculation methodology and comparable conditions remain unknown.
editorial commentary
Why it matters
If the metric is confirmed in independent tests, efficiency per unit of energy will become a critical criterion for multi-step AI tasks. The next observable signal would be the publication of the comparison methodology or results from third-party measurements. Significant uncertainty remains as currently only a synopsis from a single source is available.