Nvidia Unveils Liquid-Cooled Vera Rubin Architecture for Next-Gen AI Factories

Nvidia made a significant splash at the 2025 OCP Global Summit in San Jose, unveiling its vision for "gigawatt AI factories" based on the new Vera Rubin architecture. These data centers are designed to support the next generation of massive AI models with unprecedented efficiency and scalability.

Key Technical Changes

• Vera Rubin NVL144 Architecture: Nvidia introduced the Vera Rubin NVL144, an open architecture rack server that features a 100% liquid-cooled design. This is crucial for managing the high thermal loads generated by advanced AI workloads.
• Central PCB Midplane: The architecture includes a central printed circuit board (PCB) midplane, which facilitates faster assembly and modular expansion bays. These allow for flexible scaling of networking and inference capabilities as needed.

Why It Matters

For practitioners, this new architecture means more efficient data centers that can handle the growing demands of AI models without overheating or breaking the bank on cooling costs. The modular design also offers flexibility in expanding infrastructure, which is essential for keeping up with rapid advancements in AI technology.

Implementation Details

• Liquid Cooling: Liquid cooling is a game-changer for high-performance computing (HPC) and AI workloads. It can significantly reduce energy consumption and improve thermal management compared to traditional air-cooled systems.
• Modular Expansion Bays: These bays allow for the addition of networking and inference modules, making it easier to scale up or down based on current needs. This is particularly useful in dynamic environments where resource requirements can fluctuate.

Open Standard

Nvidia has donated the Vera Rubin NVL144 architecture to the Open Compute Project (OCP) as an open standard. This move encourages widespread adoption and innovation, allowing any company to implement this design in their data centers. It also aligns with the growing trend of open-source hardware in the tech industry.

Ecosystem Support

• Kyber Server Rack Design: Nvidia's ecosystem partners are ramping up support for the Kyber server rack design, which will eventually support 576 Rubin Ultra GPUs when they become available in 2027. This will enable companies to build highly scalable and powerful AI infrastructure.
• Meta Platforms Inc. and Oracle Corp.: Both Meta and Oracle have announced plans to standardize their data centers on Nvidia's Spectrum-X Ethernet networking switches, further solidifying the company's position in the AI hardware market.

Benchmarks and Performance

While specific benchmarks for the Vera Rubin NVL144 are not yet available, the combination of liquid cooling, modular design, and support for next-generation GPUs suggests significant performance improvements. The ability to connect 576 GPUs in a single rack will likely set new standards for AI inference and training workloads.

Conclusion

Nvidia's introduction of the Vera Rubin NVL144 architecture marks a significant step forward in the development of efficient, scalable AI data centers. By donating this design to the OCP and garnering support from major players like Meta and Oracle, Nvidia is positioning itself at the forefront of the AI hardware revolution.