The world's first 1TB memory AI GPU, NVIDIA's Jensen Huang showcases the Rubin Ultra tray

IT Home reports on March 18 that technology media Tom’s Hardware published a blog post yesterday (March 17), saying that at the 2026 GPU Technology Conference (GTC), Nvidia will showcase its next-generation data center GPU tray, Rubin Ultra, with plans to launch it in 2027.

Rubin Ultra single-tray is equipped with 4 compute chiplets and up to 1TB of HBM4E memory. This breakthrough specification makes it the industry’s first AI accelerator with TB-level memory, providing a solid hardware foundation for handling extremely large-scale AI models.

IT Home cites the blog post, saying that based on on-site images, this four-chiplet packaged Rubin Ultra uses a brand-new packaging technology. The heatsink obscures the internal details, and it is currently unclear whether the chip has completed tape-out. However, its compact package size suggests it may use a stacked design.

Also worth noting is that this tray almost completely eliminates cable connections, which is expected to significantly simplify the server assembly process. However, the media also interprets that Nvidia may in the future directly sell complete trays, thereby weakening the role of partners to nothing more than rack assemblers, who would no longer need to manufacture motherboards and server trays.

To support Rubin Ultra’s powerful compute performance, Nvidia will roll out a new rack-scale design called Kyber. Unlike traditional horizontal trays, Kyber adopts a vertical tray layout and comes by default with a liquid cooling system. This new rack can integrate 144 GPU enclosures within a single rack.

Compared with the current Oberon NVL72 system based on 72 Rubin GPUs, the Kyber NVL144 system with Rubin Ultra not only doubles the number of GPU modules within a single enclosure, but also doubles the total number of enclosures, ultimately delivering at least a fourfold performance improvement.

In terms of network interconnect, the Kyber rack will upgrade the NVLink switch from the sixth generation to the seventh generation. The new-generation switch increases the number of supported GPUs further while maintaining the ultra-high bandwidth of 3600 GB/s. At the same time, Nvidia also plans to introduce a brand-new CX9-1600G Ethernet processor to comprehensively accelerate the efficiency of data center horizontal expansion communications.