Nvidia revealed new details about the Vera Rubin platform during a technical workshop last week at its headquarters in Santa Clara, California. Executives briefed a small group of journalists on the system's increased power and efficiency. The key takeaway is that Nvidia, long known for GPUs, is increasingly positioning itself as a CPU supplier to power AI agents.
Integrating CPUs and GPUs for AI Agents
GPUs remain the primary hardware for training and running AI models, but the shift toward more complex agentic systems has boosted demand for CPUs, which handle data flow orchestration, networking, and other software tasks. This drives Nvidia's ambition to sell complete AI systems rather than just chips. Vera Rubin succeeds the Grace Blackwell hybrid superchip and is the linchpin of Nvidia's near-term AI future. It offers one CPU for every two GPUs: in the Vera Rubin NVL72 system, there are 36 Vera CPUs for 72 Rubin GPUs. Nvidia also sells the Vera CPU standalone, reportedly ready for Chinese customers as early as August.
Sponsored Protocol
Decupled Performance and Tripled Memory Bandwidth
Nvidia claims the Vera Rubin NVL72 system will process ten times more tokens per watt than Grace Blackwell. The Vera CPU is faster at agentic AI tasks compared to rival AMD and Intel CPUs, though benchmarks used slightly older competitor generations. Localized memory subsystems offer nearly three times the memory bandwidth of Blackwell, appealing amid high-bandwidth memory shortages. The system is 100% liquid-cooled, reducing energy for cooling compared to air cooling.
Plug-and-Play Installation and Cable-Free Design
The NVL72 racks are touted as much more plug-and-play: Nvidia calls it "cable-free compute" and "hot-swappable", cutting installation time from hours to minutes. During a brief tour of an Nvidia data center lab in Silicon Valley, executives revealed that OpenAI already has a Vera Rubin rack in use. CEO Jensen Huang did not attend the workshop; he was in Japan announcing robotics AI partnerships. The briefings were led by Ian Buck, vice president of accelerated computing, and Andrew Bell, senior vice president of hardware engineering.
Sponsored Protocol
Market Challenges and AMD Competition
Nvidia is sensitive to delays after previous-generation Blackwell chips overheated in custom racks, forcing design changes. The Vera Rubin marketing push comes just ahead of AMD's annual conference, where it revealed details about its competing Helios AI rack. AMD has grown its data center CPU market share with x86 architecture, while Nvidia uses ARM. Executives Buck and Hannah Coutand emphasized that Vera Rubin abandons the chiplet architecture for a monolithic design, avoiding a "heavy tax" on memory bandwidth and data movement.
Sponsored Protocol
As AI infrastructure expands, new security threats emerge, such as the worm discovered by CrowdStrike hiding in AI development pipelines that steals tokens and destroys data. Meanwhile, deepfake fraud losses have reached $3.7 billion, with social media as the top vector. These incidents highlight the need for robust AI solutions like those Nvidia aims to provide with Vera Rubin.
For further reading, see the original article on WIRED or the Wikipedia entry for Nvidia.
Source: https://www.wired.com/story/nvidia-wants-to-own-every-chip-inside-an-ai-data-center
