NVIDIA's Kyber Rack Midplane Showcased at GTC 2025, Featuring Advanced Interconnects

NVIDIA's next-generation Kyber rack midplane, a critical component for its future AI infrastructure, was recently showcased at the NVIDIA GTC 2025 event. The innovative design, which replaces traditional cabling with a central printed circuit board, aims to significantly boost GPU density and network performance in large-scale AI factories. This development is a key step in preparing for the upcoming NVIDIA Rubin Ultra GPUs and the NVL576 generation.

The Kyber rack architecture is designed to house a high-density platform of 576 NVIDIA Rubin Ultra GPUs by 2027, succeeding the Oberon rack design. A central feature is its cable-free midplane, which integrates NVLink switch blades at the back, enabling seamless scale-up networking and faster assembly. This design is expected to eliminate miles of copper cabling, enhancing reliability and reducing power loss.

"if people want to see what Kyber rack midplane looks like. This is from NVDA GTC DC three weeks ago," stated social media user TheValueist, highlighting the physical presence of the technology.

The Kyber system also incorporates innovations in 800 VDC power delivery and advanced liquid cooling, moving away from traditional 415 or 480 VAC systems. This transition to 800 VDC is crucial for handling the increasing power demands of future AI workloads, allowing for 150% more power transmission and potentially eliminating heavy copper busbars. Industry partners, including Amphenol, are actively involved in the ecosystem supporting these next-generation data center architectures.

Amphenol, a global leader in interconnect products, plays a vital role in providing high-performance connectors and cable assemblies essential for such advanced systems. While specific details of Amphenol's direct contribution to the Kyber midplane were not detailed in public announcements, the company is a known supplier for high-speed interconnects in NVIDIA's ecosystem. This collaboration underscores the complex partnership required to build the next era of AI computing infrastructure.

NVIDIA plans to contribute the upgraded rack and compute tray innovations as an open standard to the Open Compute Project (OCP) consortium. This move aims to foster broader industry adoption and accelerate the development of gigawatt-scale AI factories. The Kyber rack is anticipated to be a foundational element for hyperscale AI data centers, delivering superior performance and efficiency for generative AI workloads in the coming years.