The Scaling Challenge in AI
In the world of artificial intelligence, speed is king, but scale is the kingdom. Training and running the massive AI models that power everything from chatbots to scientific discovery requires thousands of graphics processing units (GPUs) to work together
as a single, colossal computer. The key to making this happen isn't just the power of the individual chips, but the network that connects them—the interconnect. A faster, more efficient interconnect allows for bigger models and quicker results, which is why NVIDIA has invested heavily in its proprietary NVLink and NVSwitch technologies. These systems act as the central nervous system for its AI supercomputers, allowing hundreds or even thousands of GPUs to share data at blistering speeds, far surpassing standard networking technologies. This capability has been a cornerstone of NVIDIA's market dominance, enabling the creation of so-called "AI factories."
What is Kyber and What Went Wrong?
Kyber was NVIDIA's ambitious next step in this journey. Unveiled as a next-generation rack-scale architecture, Kyber was designed to be the backbone for future AI systems, connecting the upcoming Rubin Ultra GPUs. A key innovation was its PCB midplane, a complex circuit board intended to connect compute and switch components directly, eliminating a mass of internal cabling and enabling unprecedented density and performance. However, according to recent reports from semiconductor analysts, this critical midplane has run into significant manufacturing difficulties. The complexity of the 78-layer board, which requires incredibly fine traces, has reportedly proven too challenging for mass production, leading to a major delay. Reports now suggest the Kyber NVL144 rack system has been pushed back by over a year, to 2028.
No Easy Plan B
The trouble doesn't appear to stop with a simple delay. Faced with the Kyber midplane manufacturing issues, NVIDIA reportedly explored a transitional solution. This backup plan, dubbed NVL72x2, involved placing two of its current-generation racks back-to-back and connecting them with copper cabling. This would have offered a way to increase the scale of GPU clusters without relying on the troubled Kyber architecture. However, according to analyst reports, this design has now been canceled entirely. This leaves NVIDIA without a publicly announced, proven replacement to achieve the scale promised by Kyber in the near term. The company will continue to ship its existing and highly successful Blackwell and Rubin-based systems, but the path to the next level of massive scale-up now seems less clear.
An Opening for Competitors?
NVIDIA's interconnect technology has long been a key part of its competitive moat. While competitors like AMD and custom chip-makers at Google and Amazon are producing powerful AI accelerators, replicating NVIDIA's full-stack solution—from the chip to the software to the system-scale interconnect—is a massive challenge. This delay could provide an opening. A consortium of rivals including AMD, Intel, Google, and Microsoft has already formed to create an open standard alternative to NVLink called the Ultra Accelerator Link (UALink). While this effort is still in its early stages, a stumble in NVIDIA's own roadmap could give it and other competing technologies more time and opportunity to gain traction. With the pace of AI not slowing down, customers who need to build ever-larger clusters may be more willing to explore alternatives if NVIDIA's premier scale-up solution is on hold.
