NVIDIA has introduced its Vera Rubin NVL72 GPU, a system that the company says delivers 30 times the efficiency per megawatt compared to its Blackwell architecture. The chip also sets a new performance benchmark for agentic AI, according to the announcement.
The efficiency number that matters
The headline figure is the 30x efficiency gain per megawatt. That metric speaks directly to the power crunch facing data centers that run large AI models. As models get bigger, electricity consumption has become the limiting factor for scaling, so a GPU that does the same work with a fraction of the power changes the economics.
NVIDIA didn't disclose how it measured the figure, but the company uses a standard that compares workload throughput against electrical input. The Vera Rubin NVL72 is designed for dense racks, packing 72 GPUs into a single unit, which cuts interconnect overhead and reduces the energy wasted on moving data between chips.
What agentic AI demands
The term 'agentic AI' describes systems that don't just generate text or images, but take actions on their own—scheduling, negotiating, coding, running multi-step tasks with minimal human supervision. Those workloads are far more demanding than simple inference. They require rapid back-and-forth between reasoning and tool use, which makes latency and power efficiency critical.
NVIDIA's new chip targets that specific workload. The company says the Vera Rubin NVL72 sets a new benchmark for agentic AI performance, which suggests the architecture is tuned to handle the long inference chains and high-volume token processing those agents rely on. That's a shift from Blackwell's focus on training large models.
What's left to know
The announcement didn't include a release date, pricing, or any mention of which systems will ship with the Vera Rubin NVL72 first. NVIDIA also hasn't said whether existing Blackwell customers can upgrade or if they'll need to wait for a full platform refresh.
What's clear is that the efficiency advantage, if it holds up in real-world benchmarks, could pressure competitors to match the per-watt performance. For now, the only number on the table is the 30x claim, and the next step is seeing how that translates into actual deployments once the hardware reaches data centers.
