Shortly after the AI boom began, Nvidia’s goal was to deliver new generations of GPUs much more quickly. Fueled by tens of billions of dollars in quarterly revenue, R&D has raced from Hopper to Blackwell to Rubin. However, the ambitious rack-scale Kyber system based on Rubin Ultra chips appears to be unfeasible by 2027. A delay of over a year gives trailing competitors AMD and Google a rare opportunity to compete at the top with their AI products.
The news comes from SemiAnalysis, which states that a particular PCB is very difficult to manufacture. This “midplane” consists of 78 layers and is highly prone to defects. The larger NVL576, which, as the name suggests, is equipped with 576 GPUs, is also experiencing manufacturing issues. In this setup, eight racks are connected via optical links.
Kyber is no ordinary product. It is a server cabinet that packs 144 of Nvidia’s most powerful chips into a single unit, allowing them to work together as one giant computer. A contingency plan to link two existing racks has been scrapped. Cloud customers reportedly found that proposed design cumbersome and expensive to maintain.
An end to the breakneck pace?
From 2016 to 2022, the codenames Pascal, Volta, Ampere, and Hopper succeeded one another as flagship GPUs for Nvidia’s enterprise market. The vast majority of attention was still focused on the consumer market, particularly for gaming. When the AI hype, fueled by ChatGPT, rapidly demanded greater scale, it was initially the Ampere- and Hopper-powered A100s and H100s that led the way. Since then, Blackwell, Blackwell Ultra, Rubin, Rubin Ultra, and subsequently Feynman have formed the roadmap. But this roadmap now appears unfeasible.
The delay is one problem in a series of setbacks that raise questions about Nvidia’s annual release pace. In hindsight, it highlights the wisdom of the previous production pace, although in the current competitive landscape, full of potential rivals and customers with deep pockets, it’s very tempting to deliver faster than ever. Nvidia continues to perform strongly financially. The current Rubin systems are in full production and will be shipped this fall to eight cloud partners, including the big three hyperscalers: AWS, Microsoft Azure, and Google Cloud.
Roadmap under pressure
That Nvidia itself is investing heavily to maintain this pace was already evident when the company entered the bond market for the first time in five years to raise at least $20 billion. Capital expenditures in the AI sector are set to grow to approximately $660 billion this year. According to NVIDIA’s own developer blog, Kyber forms the basis for even larger systems such as the NVL576 and even the NVL1152. The end of this growth is therefore nowhere in sight; the only question remaining is whether the pace of that growth will remain as it is now.