NVIDIA Vera Rubin general standard consolidates Taiwan and global supply chains, with an optimistic outlook on performance

NVIDIA Senior Director of AI Infrastructure Dion Harris reveals details about the development of the next-generation AI system Vera Rubin and its global supply chain deployment, including computing power upgrades, implementation of universal standards, energy optimization, and major investments in U.S. infrastructure. Vera Rubin will succeed the current Blackwell architecture and become the new universal benchmark for large-scale AI infrastructure, deeply integrating hardware design with the global supply chain. The market is optimistic that this will drive strong performance for NVIDIA.

Vera Rubin Universal Architecture Standard Connects Global Supply Chains

The current Blackwell architecture contains approximately 1.2 million components, while the next-generation Vera Rubin architecture will increase this number to 1.3 million. NVIDIA adopts a universal standard architecture, with over 80 suppliers and 350 factories worldwide implementing this new standard. The supply chain spans more than 20 countries, including Taiwan, Japan, South Korea, the U.S., China, Mexico, Israel, Vietnam, and Thailand. The core silicon chips are produced by TSMC, with Foxconn responsible for chassis assembly, and Taiwanese suppliers playing a key role.

HBM4 Solves Computing Energy Consumption

Vera Rubin addresses the biggest bottleneck threatening AI development: energy consumption. The new Vera CPU’s performance per watt is about twice that of the previous Grace CPU. Unlike Blackwell, which welds memory onto the circuit board, Vera Rubin introduces a removable, replaceable SoC Memory Module (SoCAMM) design, increasing system maintenance flexibility. For high-load storage needs, the system uses high-bandwidth HBM4 memory supplied by SK Hynix and Samsung, solving data access latency issues. NVIDIA’s proprietary NVLink switch increases data transfer rates to 3.6TB per second, paired with a copper wire network extending two miles behind the chassis, significantly improving computational efficiency per unit of power.

Liquid Cooling System Replaces Fans, Saves Water

Vera Rubin is NVIDIA’s first system to adopt a 100% liquid cooling system. During initial deployment of Blackwell, some customers reported overheating issues, which the technical team attributed mainly to human installation errors, such as improper sealing of liquid cooling valves, rather than design flaws. Most systems are now stable. Vera Rubin’s compute trays completely replace fans, hoses, and cables, using cold plates and a closed-loop water system to cool processors directly. Although liquid cooling appears water-intensive, the closed-loop system reduces water consumption compared to traditional evaporative cooling, offering water-saving benefits.

NVIDIA Invests $500 Billion in U.S. AI Infrastructure

Despite NVIDIA’s current market dominance, it is prepared to invest up to $500 billion in AI infrastructure in the U.S., utilizing robotics for automated assembly to mitigate supply chain disruptions, geopolitical risks, and tariffs.

Google and Other Customers Develop In-House Chips but Maintain Collaboration with NVIDIA

NVIDIA’s competitors, such as AMD, plan to launch their first rack-scale system, Helios, to gain market share. Major customers like Microsoft, Google, Amazon, and Meta are actively developing in-house chips like Trainium or TPU but continue to collaborate with NVIDIA. The development roadmap extends beyond the Rubin architecture; the R&D team has demonstrated a prototype of the next-generation Kyber architecture, which significantly increases the number of GPUs in a rack from 72 to 288. Kyber reduces internal cabling dramatically, quadrupling computational density while increasing weight by only about 50%. This design aims to lower system failure rates and enable ultra-low latency communication through fewer connection points and better integration. Vera Rubin Ultra and Kyber architectures are respectively expected to launch in 2026 and 2027. The AI hardware market is shifting from single-chip performance competition to system integration and global supply chain resilience, with NVIDIA already ahead of the curve, leading other competitors.

This article, “NVIDIA Vera Rubin Universal Standard Consolidates Taiwan and Global Supply Chains, Market Outlook Bright,” first appeared on Chain News ABMedia.