Nvidia, Foxconn team up on Vera Rubin NVL144 AI factory platform

Nvidia is unveiling the specs of its Vera Rubin NVL144 MGX open architecture rack servers at the OCP Global Summit, with support from over 50 partners.

These servers are part of Nvidia’s plan to enable “gigawatt AI factories,” with ecosystem support for Nvidia Kyber, a rack system built to connect 576 Rubin Ultra GPUs and use 800-volt direct current (VDC) infrastructure.

Industry partners including Foxconn are building, while others like CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure, and Together AI are designing for 800 VDC data centers.

Vertiv has unveiled a reference architecture supporting Nvidia’s design, and HPE is announcing product support for Nvidia’s architecture.

Nvidia said moving to 800 VDC infrastructure can improve data center scalability and energy efficiency compared to traditional systems.

The Vera Rubin NVL144 features liquid cooling and modular design, while the Kyber platform aims to increase GPU density and reduce copper usage in racks.

🔗 Source: Nvidia

🧠 Food for thought

Implications, context, and why it matters.

800 VDC claims face code hurdles

Claims for 800 volts direct current (VDC) include 150% more power through the same copper, removing 200 kg copper busbars, and millions in savings. International Electrotechnical Commission (IEC), Underwriters Laboratories (UL), and National Electrical Code (NEC) approvals still decide timing.
Data centers allow people inside, so arc-flash limits, touch-safe gear, and maintenance rules are tighter yet missing. NVIDIA’s MGX supports Peripheral Component Interconnect Express (PCIe) and OCP 1, but 800 VDC readiness remains unclear.
Foxconn’s 40-megawatt Kaohsiung-1 plus CoreWeave and Oracle Cloud Infrastructure are designing around 800 VDC, yet gains beyond copper cuts are vague. Without OCP to global electrical code mapping, teams cannot judge if 2027 is realistic.

NVIDIA’s 45°C liquid cooling sparks bids

Vera Rubin NVL144 uses full liquid cooling at 45°C. That spurs demand for direct-to-chip (D2C) retrofits and builds as air systems stall at 30 to 100 kW racks common in AI rollouts 2.
Mechanical, electrical, and plumbing (MEP) engineering firms plus liquid cooling original equipment manufacturers (OEMs) with D2C cold plates can pursue bids. Data center contractors can do the same with CoreWeave, Lambda, Nebius, or others building 800 VDC sites.
D2C removes 70% to 80% of heat at the chip 2. Vendors not named here should seek requests for proposals (RFPs) and pilots, including modular container providers plus certification specialists. Liquid cooling has a 2 to 4 year return on investment (ROI) 2, so procurement can start well before 2027.