Nvidia Announces Next-Gen 'Vera Rubin' AI Platform
Nvidia CEO Jensen Huang announced the company will begin shipping its next-generation "Vera Rubin" AI platform in the second half of 2026. The move follows the rapid adoption of its current Blackwell architecture, which is fueling a global shift toward "AI Factories." Nvidia leadership expects enterprise AI spending to accelerate, cementing its role as the primary architect of the agentic AI era.
The "Vera Rubin" platform marks an acceleration in Nvidia's release schedule, shifting to a one-year cadence instead of its prior two-year cycle. This rapid iteration, with Rubin following the recently shipped Blackwell architecture, is designed to meet the surging demand of the AI industry and solidify Nvidia's market dominance by rendering competitor roadmaps obsolete more quickly. The architectural leap from Blackwell to Rubin is tailored for the shift from generative AI to "agentic AI." While Blackwell excels at generating content, Rubin is optimized for reasoning and continuous, multi-step task execution. This is reflected in a design that prioritizes inference speed using new low-precision math formats (NVFP4), even sacrificing some performance in traditional high-precision scientific computing. The Vera Rubin platform is a full-stack system, not just a single chip. The flagship NVL72 configuration integrates 72 Rubin GPUs and 36 ARM-based Vera CPUs into a single liquid-cooled rack. It debuts HBM4 memory for faster data access and NVLink 6, which provides 3.6 TB/s of GPU-to-GPU bandwidth, enabling the massive data movement required for trillion-parameter models. This integrated, rack-scale design is the foundation of the "AI Factory" concept—a purpose-built system for turning raw data into intelligence, measured in token throughput. Nvidia is moving beyond selling individual components to providing the entire blueprint for these factories, including networking (ConnectX-9), data processing units (BlueField-4), and even digital twin software (Omniverse) to design the data centers themselves. The market for agentic AI is projected to expand from around $5-$7 billion in 2024-2025 to nearly $200 billion by 2034. Nvidia aims to capture this growth by significantly lowering the cost of AI decision-making, claiming the Rubin platform can offer up to 10x lower inference cost per token compared to Blackwell. This performance leap comes with immense power demands, with a single Rubin rack projected to consume over 250 kW, necessitating direct liquid cooling. To secure production, Nvidia has reportedly booked more than half of TSMC's advanced 3nm chip packaging capacity for 2026, creating a significant supply chain moat against competitors.
