Jensen Huang Against the Gloom: 'Vera Rubin is Already in Production, Huge Quantities on the Way'

After an initial official statement, Jensen Huang dismisses rumors about delays in the development of the next-generation AI platform Vera Rubin. On the sidelines of an event in Japan, the CEO of NVIDIA labeled the circulated rumors from recent weeks as unfounded, emphasizing that the production of the new architecture has already begun and that the company is preparing for particularly high volumes.

The manager stated: "The news about delays of Vera Rubin is not true. Vera Rubin is already in production. Huge production volumes are on the way." This statement reinforces what NVIDIA had already communicated in previous months. Huang's recent intervention seems primarily aimed at reassuring investors and customers about the roadmap for the next generation of AI products, which will include Vera CPUs, Rubin GPUs, and Vera Rubin NVL72 rack-scale systems.

However, the CEO's comments do not directly address part of the recent rumors. Reports from SemiAnalysis suggest that the Kyber NVL144 rack-scale system based on Rubin Ultra has accumulated over a year of delay, possibly shifting from 2027 to 2028. According to this information, the main cause is the production complexity of the central PCB (midplane) needed to connect 144 Rubin Ultra GPUs through the new NVLink 7 fabric based on copper interconnections. The difficulty does not concern the chips or other electronic components, but rather the creation of the sophisticated printed circuit board infrastructure required to manage high-speed electrical connections between the different elements of the system. This complexity has made large-scale production of the project challenging.

When asked about the topic, NVIDIA chose not to delve into the details. A spokesperson for the company stated that "our roadmap remains unchanged," without confirming or denying the rumors of possible delays. The company reiterates that the planned products will continue to be part of its strategy, but it has not clarified whether the initially projected launch dates will remain unchanged.

Even the most ambitious NVL576 system could be subject to potential slowdowns. The project involves the integration of eight Oberon racks connected via co-packaged optics (CPO) technologies among NVSwitch switches. The same sources assert that the development is facing technological challenges, leading to potential delays or limited availability in the initial production batches.

The existence of the NVL576 project suggests that NVIDIA is developing NVSwitch connections based on CPO for the Rubin generation. In theory, this technology could represent an alternative to the complex copper architecture envisioned by Kyber NVL144, but it is currently unclear if it can offer the same network topology, bandwidth, and comparable latency, or if it has reached a sufficient level of maturity for large-scale deployment.

The rumors also come after other reports that NVIDIA has abandoned the Rubin Ultra variant equipped with four dies, opting instead for a two chiplet configuration that, according to rumors, would provide lower performance than the original project.