Meet the Rubin Platform
Following the blockbuster Blackwell platform, NVIDIA has announced its next-generation AI architecture, codenamed Rubin, slated for a 2026 release. The platform isn’t just a single chip but an entire ecosystem of co-designed hardware. It features the powerful
Rubin GPU for computation, a new ARM-based central processing unit (CPU) named Vera for orchestrating tasks, and a suite of advanced networking components. These include the sixth-generation NVLink Switch, which physically connects GPUs, and the ConnectX-9 SuperNIC for broader network communication. The top-tier configuration, known as Rubin Ultra, is planned for 2027 and aims to push the boundaries of AI supercomputing even further.
The AI Scaling Challenge
In the world of artificial intelligence, “scaling” refers to the ability to make thousands of GPUs work together as one massive, coherent computer. As AI models grow to trillions of parameters, a single GPU is no longer sufficient. The primary challenge has shifted from the processing power of an individual chip to the network that connects them. If the links between GPUs are too slow, these expensive processors spend more time waiting for data than performing calculations, creating a massive bottleneck that wastes both time and energy. NVIDIA’s proprietary NVLink technology has been its key advantage, offering a high-bandwidth alternative to the standard PCIe connections used in most computers.
The Missing Piece: An Optical Solution
The headline claim revolves around the physical limits of the current interconnect technology. NVLink, like most internal computer connections, uses electrical signals sent over copper traces. As data rates double with each generation—with NVLink 6 reaching 3.6 TB/s per GPU—copper is hitting a wall. Signals degrade quickly over distance, limiting reliable connections to under a meter and generating significant heat. The industry-wide consensus is that the next major leap requires a move to optical interconnects, which use light to transmit data over fiber. Recent reports from July 2026 indicate that NVIDIA has had to delay several of its most ambitious rack-level designs due to manufacturing and design issues, suggesting a proven, mass-producible optical solution is not yet ready.
Why the Original Rubin Ultra Plan Was Scaled Back
NVIDIA's original plan for the 2027 Rubin Ultra GPU was incredibly ambitious, featuring a design with four compute dies (chiplets) on a single package. However, this created severe manufacturing problems. The large package size led to physical warping of the underlying substrate, causing connection failures and extremely low production yields. As a result, NVIDIA reportedly scrapped the quad-die design in mid-2026, reverting to a more manageable dual-die configuration similar to the standard Rubin GPU. This change effectively halves the potential performance of a single Rubin Ultra package, highlighting the immense difficulty in scaling hardware at the bleeding edge of chip manufacturing.
The High-Stakes Path Forward
Without a mature optical interconnect solution ready for the Rubin generation, NVIDIA is constrained in how far it can scale its systems. The announced NVLink 6 is a powerful iteration, but it remains a copper-based technology. This means larger AI factories, especially those connecting hundreds or thousands of GPUs across multiple racks, will continue to face a communications bottleneck that an optical solution would solve. The recent delays and the downscaling of the Rubin Ultra design suggest that the company's most advanced scale-up plans may not be fully realized until the subsequent 'Feynman' architecture, expected in 2028 or later. This engineering hurdle creates a potential window of opportunity for competitors like AMD and Google, who are also aggressively pursuing their own scaling technologies.
