NVIDIA has introduced its Vera Rubin NVL72 platform, claiming a 30-fold improvement in throughput per watt for agentic AI workloads. The company says the efficiency gain also lowers token costs, a metric that directly affects how much businesses pay to run AI agents.
Why energy efficiency is now the bottleneck
Agentic AI differs from a simple chatbot. These systems plan, call tools, and execute multi-step tasks, so they demand far more compute per request. Each step means more tokens processed and more energy burned. Throughput per watt measures how much work a chip does for every unit of electricity, so a 30x jump means the same task could be done with a fraction of the power.
That matters because data-center power costs have become a wall for AI expansion. Companies running large-scale agentic systems can quickly see their electricity bills outpace hardware costs. A 30x efficiency claim, if it holds up, would push that wall back significantly.
What lower token costs mean for enterprises
Token costs are the price per unit of input and output text that models process. Agentic workloads are token-hungry—a single agent run can consume thousands of tokens. Cutting that cost per token is what makes the difference between an AI experiment and a production system.
NVIDIA's claim points directly to that math. Lower tokens cost per unit of compute means a company could run the same agentic workflows for a smaller monthly bill, or push more work through the same hardware budget. The exact numbers aren't public, but the ratio is clear.
The missing details
The announcement didn't include a release date, pricing, or benchmark results from independent tests. NVIDIA's claims are based on its own internal measurements. So the 30x number is a headline, not yet a proven spec.
Customers will be waiting for the hardware to ship and for third-party reviews to see if the efficiency holds up outside a test environment. Until then, the NVL72 remains a promise—one with a lot of financial weight for anyone building agentic AI at scale.
