What Is the Kyber Platform?
To understand the gravity of the alleged delay, one must first understand what Kyber represents. It is not just another graphics card; it's NVIDIA's blueprint for the next generation of 'AI factories'. Unveiled by CEO Jensen Huang, the Kyber architecture
is the successor to the company's current Oberon rack design. It is engineered to house an unprecedented density of GPUs, such as the upcoming Rubin and Rubin Ultra chips, in massive server racks. Configurations like the Kyber NVL144 are designed to link 144 GPUs into a single, cohesive supercomputer. This move towards extreme densification is crucial for training and running the increasingly complex AI models that power everything from chatbots to scientific discovery. By creating these highly integrated systems, NVIDIA aims to solve the immense communication and power challenges that arise when trying to make thousands of chips work as one.
The Billion-Dollar Bottleneck
According to the SemiAnalysis report that surfaced on July 6, the entire Kyber program has hit a formidable wall: a single, incredibly complex component known as the midplane PCB, or what NVIDIA calls the 'orthogonal backplane'. This is no ordinary circuit board. It is a monstrous, 78-layer piece of hardware, built from a hybrid of advanced materials, designed to eliminate the need for a jungle of traditional copper cables. In a system with 144 GPUs, using cables would add significant weight and, more importantly, cause signal degradation that would cripple performance. The midplane allows compute and switch trays to connect directly, but manufacturing it at scale has proven to be an immense challenge. The report claims these difficulties are the direct cause of a production delay of over 12 months, pushing the expected launch of the Kyber NVL144 system from 2027 to 2028.
Cancellations and Performance Cuts
The ripple effects of this manufacturing hurdle are significant. The report alleges that NVIDIA's backup plan, a transitional architecture known as NVL72x2, was rejected by major cloud service providers and subsequently canceled, leaving a potential gap in its product roadmap. Furthermore, SemiAnalysis claims that NVIDIA has also canceled the more powerful four-chip version of its upcoming Rubin Ultra GPU. This would mean that only a smaller, two-chip version will be available, effectively halving the potential peak performance per rack, even if the Kyber system eventually ships. These compounding issues suggest that the challenge is not just one of timelines, but also of capability, potentially forcing NVIDIA to recalibrate its performance targets for the next generation of AI infrastructure. The delay also puts pressure on NVIDIA's vast supply chain, from memory and ODM partners to the specialised PCB manufacturers at the heart of the issue.
A Window for Competitors?
For years, NVIDIA has maintained an iron grip on the AI hardware market, but this reported stumble could provide a crucial opening for its rivals. Competitors like AMD, with its forthcoming MI500X accelerator, and Google, with its custom TPUv8i 'Broadfly' chips, are aggressively pursuing the same hyperscale customers. A significant delay and potential performance reduction in NVIDIA's top-tier scaling solution could make these alternatives more attractive to cloud giants desperate for AI computing power. The race to build out massive AI training clusters is relentless, and any perceived gap in a market leader's portfolio is an opportunity for others to gain ground. While NVIDIA's market position remains formidable, the Kyber delay, if it holds true, represents one of the most significant strategic challenges the company has faced in its recent era of dominance. As of early July, NVIDIA had not issued a public response to this specific report from SemiAnalysis.
