Pioneering the Gigawatt Era of AI Infrastructure

NVIDIA is at the vanguard of a transformative shift in artificial intelligence infrastructure, driving the development of gigawatt-scale AI factories. This ambitious vision is being realized through a collaborative effort involving over 50 industry partners, focusing on next-generation architectures and advanced power systems. The initiative centers on the NVIDIA Vera Rubin NVL144 open architecture rack servers and the NVIDIA Kyber rack-scale generation, both designed to meet the escalating demands of sophisticated AI workloads. A cornerstone of this evolution is the widespread adoption of 800-volt direct current (VDC) data center infrastructure, a strategic departure from conventional alternating current (AC) systems.

The Advantages of 800 VDC Infrastructure

The transition to 800 VDC power systems represents a significant leap forward in data center design and efficiency. This elevated voltage offers a cascade of benefits, including enhanced scalability to accommodate exponentially growing AI demands, improved energy efficiency that reduces operational costs and environmental impact, and decreased material usage, particularly in terms of copper. Furthermore, it provides a higher capacity for performance, crucial for the intensive computations required by modern AI models. These advantages are not theoretical; the electric vehicle and solar industries have already demonstrated the efficacy of 800 VDC infrastructure, validating its potential for large-scale applications.

Vera Rubin NVL144: A Scalable Architecture for AI Factories

The NVIDIA Vera Rubin NVL144 stands as a testament to NVIDIA's commitment to advancing accelerated computing. This architecture is engineered to deliver a substantial increase in AI performance, specifically catering to the complex requirements of advanced reasoning engines and AI agents. Its design is rooted in the flexible MGX rack architecture, which is being actively supported by a broad ecosystem of over 50 system and component partners. NVIDIA's decision to contribute the upgraded rack and compute tray innovations as an open standard to the Open Compute Project (OCP) consortium underscores its dedication to fostering industry-wide collaboration and standardization. The modular nature of the Vera Rubin NVL144, with its standards for compute trays and racks, allows partners to seamlessly integrate and scale their solutions, accelerating the deployment of gigawatt-scale AI infrastructure. Key features include an energy-efficient 45°C liquid cooling system, a novel liquid-cooled busbar for enhanced performance, and a 20x increase in energy storage capacity to ensure stable power delivery.

NVIDIA Kyber: The Next Generation of Rack Servers

Complementing the Vera Rubin NVL144, the NVIDIA Kyber rack server generation is poised to become a foundational element of future hyperscale AI data centers. This successor to the NVIDIA Oberon architecture is designed to house an exceptionally dense platform, capable of connecting up to 576 NVIDIA Rubin Ultra GPUs by 2027. The Kyber generation integrates innovations in 800 VDC power delivery, advanced liquid cooling, and refined mechanical design. A significant design innovation is the vertical orientation of compute blades, akin to books on a shelf, enabling up to 18 compute blades per chassis. This approach, combined with purpose-built NVIDIA NVLink switch blades integrated via a cable-free midplane, facilitates seamless scale-up networking and boosts rack GPU density. The transition to 800 VDC for in-rack components, from the previous 54 VDC, is critical for managing the high power demands. This architectural shift not only enhances performance and efficiency but also offers substantial cost savings by reducing the extensive copper cabling traditionally required for single racks, potentially saving tons of copper and millions of dollars.

Expanding the NVLink Fusion Ecosystem

Beyond hardware advancements, NVIDIA's NVLink Fusion ecosystem is experiencing significant growth. This initiative enables companies to integrate their semi-custom silicon into NVIDIA's highly optimized data center architecture, thereby reducing complexity and accelerating time-to-market for AI solutions. The ecosystem is expanding to include major players like Intel and Samsung Foundry, alongside existing custom silicon designers, CPU, and IP partners. This collaborative approach is vital for rapidly scaling AI factories to meet the demanding workloads of both model training and agentic AI inference.

An Open Ecosystem for Scalable AI Factories

The collective effort involving over 20 NVIDIA partners is crucial for delivering rack servers built on open standards, paving the way for the future of gigawatt AI factories. This open ecosystem approach, championed by initiatives like the OCP, ensures interoperability and accelerates innovation across the industry. By standardizing key architectural components and power systems, NVIDIA and its partners are collectively building the robust, efficient, and scalable infrastructure required to power the next generation of artificial intelligence.

Tag: nvidia vera rubin

NVIDIA and Partners Forge Path to Gigawatt AI Factories for Next-Generation Computing