Tianfeng Guo Mingqi: NVIDIA AI Server Power Supply Planning Roadmap

NVIDIA is elevating its power supply strategy to a new strategic height, extending its technological moat from chip computing power to the entire power architecture of data centers through the next-generation Kyber platform, aiming to define the standards for future AI factories.

According to the latest report by Tianfeng International Securities analyst Guo Mingqiang, NVIDIA's next-generation AI server power strategy "Kyber" is advancing on two fronts, with mass production targets set for the end of 2026, earlier than the market's general expectation of 2027. This development indicates that NVIDIA is accelerating the validation of its 800 VDC (high-voltage direct current) architecture and high-density cabinet design.

Guo Mingqiang pointed out that the reference design scope of the Kyber project has significantly expanded, no longer limited to the GPU and cabinet level, but now includes the entire data center's power supply and infrastructure planning, including the application of 800 VDC/HVDC distribution and solid-state transformers (SST). This move signifies that, starting from the Kyber generation, the importance of power architecture within NVIDIA has been elevated to a strategic position equal to that of semiconductors.

Dual Progress: Kyber Project's Early Mass Production and Risk Control

According to Guo Mingqiang's industry survey, NVIDIA's AI server power planning is advancing on two fronts.

The Kyber project plans to achieve mass production by the end of 2026, a timeline that is ahead of market consensus, helping NVIDIA to complete the mass production validation of the 800 VDC architecture and high-density cabinet design ahead of schedule.

The VR200 NVL288 (which includes only two canisters), due to its relatively low design and mass production difficulty, has a significantly higher probability of successful mass production compared to more complex models.

Guo Mingqiang believes that if the goal is to ship by the end of 2026, the controllable risks of the VR200 NVL288 will benefit NVIDIA in entering the 800 VDC generation earlier.

Strategic Upgrade: Reference Design Extending from Cabinets to Data Centers

The key difference between the Kyber project and the existing Oberon cabinets lies not only in its high-density blade design's different appearance but also in the upgrade of its strategic connotation.

NVIDIA has incorporated the power distribution network (Power Distribution Network; PDN) of data centers into the reference design scope.

This means that NVIDIA is attempting to establish a complete ecological standard from chips, cabinets to the entire data center power infrastructure.

Guo Mingqiang analyzes that while competitors remain focused on chip computing power and software, NVIDIA is forming a new, insurmountable competitive barrier by defining the underlying architecture of the "AI factory."

Technical Trade-offs: The Efficiency and Feasibility Debate of the 800VDC Conversion Path

Guo Mingqiang stated that the Kyber platform will adopt 800 VDC and configure dedicated power cabinets.

In the core power conversion process, NVIDIA faces a critical technical path choice: how to efficiently and safely reduce 800 VDC to a 12 V load point (Point-of-Load; POL).

Currently, there are two solutions. The first is "single-stage conversion," which can achieve the highest conversion efficiency but poses significant challenges to the design of power management chips and system safety.

The second is "Intermediate Bus Architecture (IBA)," which, although sacrifices some efficiency, has more mature technology and is more favorable for achieving large-scale production.

Ming-Chi Kuo predicts that in the next one to two years, NVIDIA will need to balance efficiency advantages with mass production feasibility. However, from a long-term trend perspective, the single-stage conversion solution with the highest efficiency remains the ultimate direction for technological convergence.