Nvidia’s Kyber rack-scale architecture, designed to include 2027 Rubin Ultra chips, will be delayed by more than 12 months until 2028, according to research firm Semianalysis.
The development comes just three months after the company’s CEO Jensen Huang publicly demonstrated the system on stage at the annual GTC event.
Kyber incorporates 144 of Nvidia’s most advanced chips in a single server enclosure, and the GPUs within Kyber are arranged vertically instead of horizontally.
According to SemiAnalysis, this problem occurs due to the complexity associated with a specific component of the system known as the “PCB midplane.”
“The Kyber NVL144 rack architecture has been postponed to 2028 as the PCB midplane continues to be challenging from a manufacturability standpoint,” the company said in a Monday post. The NVL576, a larger system that links eight racks via optical connections, may also be delayed or limited to low-volume production.
Nvidia’s fallback option, a design called NVL72x2 that bolted two current-generation racks together to match Kyber’s capabilities, has also been discontinued. According to a semi-analysis, cloud providers and hyperscalers strongly objected to the design due to its awkward layout and heavy operational burden, forcing NVIDIA to cancel the design.
This means, according to the semi-analysis, that NVIDIA does not have a “proven solution to scale up Rubin Ultra’s global size.”
According to SemiAnalysis, this shortage could give AMD and Google a rare opportunity to beat Nvidia in the high-end AI infrastructure market. That’s because the chips made by both companies outperform Nvidia in major AI labs. The company has not yet responded to CNBC’s inquiries regarding this matter.
This issue is specific to future Rubin Ultra processors and does not affect Nvidia’s existing processors. Rubin processors are currently in full production and will begin shipping later this fall to eight cloud customers, including Amazon Web Services, Microsoft Azure, and Google Cloud.
