Nvidia's next major AI system, the Kyber NVL144 server rack, has been delayed by more than a year to 2028, according to semiconductor research firm SemiAnalysis. CEO Jensen Huang presented the system at Nvidia's GTC conference in March as the platform for the company's 2027 Rubin Ultra chips.

SemiAnalysis traced the delay to the midplane, a large circuit board that connects the chips inside the cabinet, which it said remains "challenging from a manufacturability standpoint." The firm also reported that Nvidia has cancelled its backup design, leaving the company without a proven way to scale up Rubin Ultra. Nvidia did not respond to requests for comment.

Tech stocks in Asia and circuit board suppliers fell after the report, while AMD rose in early trading and Nvidia stock held just under the $200 mark. The delay puts fresh attention on whether Nvidia can keep its promise of a new AI platform every year, and on which of its rivals benefit if it cannot.

NVIDIA Ising: 7 Things to Know About Its Open-Source AI Models for Quantum Computing
At its Quantum Day event, the company unveiled the world’s first open-source family of AI models built to automate calibration and error correction.
TechloyDamilare Odedina

What went wrong with Kyber

Kyber is a server cabinet that holds 144 of Nvidia's most advanced chips in a single system. The chips stand upright in trays rather than lying flat, a layout that packs in more computing power and moves data between the chips faster.

Subscribe for free to continue reading this article

Subscribe Subscribe

Already Have an Account? Log In