NVIDIA's Vera Rubin platform provides comprehensive scalability designed for the demanding requirements of agentic AI and large-scale AI factories. It represents a significant architectural shift towards full-stack, tightly integrated supercomputing systems, moving beyond individual components to engineered entire systems for maximum efficiency and scale. The platform is built to operate at rack-scale and POD-scale deployments, where multiple server racks function cohesively as a single, powerful AI supercomputer. This integrated approach allows for seamless scaling of complex, multi-step autonomous AI workflows, from massive pre-training to real-time inference and post-training tasks. For instance, a Vera Rubin POD can encompass 40 racks, integrating 1,152 Rubin GPUs and delivering 60 exaflops of performance, showcasing its ability to handle immense computational loads.
The scalability of Vera Rubin is underpinned by a suite of purpose-built, interconnected components. The Vera CPU, designed specifically for agentic AI and reinforcement learning, can support over 22,500 concurrent CPU environments in a single rack, enabling extensive testing and validation of AI agents with increased efficiency and speed compared to traditional CPUs. The Rubin GPU, a core element, features a new Transformer Engine that boosts NVFP4 inference performance to up to 50 petaFLOPS. Critical for inter-component communication and data movement, the sixth-generation NVLink unifies up to 72 Rubin GPUs into a single performance domain, offering 3.6 terabytes per second (TB/s) of bandwidth per GPU and 260 TB/s of low-latency connectivity across the system. Complementing these, the ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and the newly integrated Groq 3 LPU (for low-latency, large-context inference) work in concert to manage and optimize data flow, offload tasks, and minimize communication bottlenecks, which are crucial for scaling AI workloads efficiently.
Vera Rubin’s scalability also translates into significant performance and economic advantages for AI operations. The platform offers up to 10 times more inference throughput per watt and a tenth of the cost per token compared to previous Blackwell systems, alongside 4 times better training performance. When combined with the Groq 3 LPX inference accelerator, this can further increase inference throughput by up to 35 times per megawatt. This focus on "tokens per watt" and "cost per token" is central to NVIDIA's vision of "AI factories," which are designed to produce intelligence outputs at industrial scale with maximum energy efficiency and cost-effectiveness. Major cloud providers, including Amazon Web Services, Google Cloud, and Microsoft Azure, are planning to integrate Vera Rubin into their offerings, demonstrating its suitability for scalable, shared AI infrastructure that can meet the growing demands of enterprises and developers for increasingly complex agentic workflows.