NVIDIA Vera Rubin Architecture Revealed

TL;DR

The Full Story

NVIDIA's Vera Rubin platform represents the most aggressive hardware co-design effort in GPU history. Rather than treating the GPU as a discrete component slotted into a standard server, Vera Rubin treats the entire rack as a single compute unit.

Cableless Compute Tray

The most striking departure from previous architectures is the elimination of internal cabling within compute trays. GPU-to-GPU interconnect, GPU-to-NIC, and power delivery all happen through custom PCB backplanes and blind-mate connectors. This reduces signal integrity issues, simplifies manufacturing, and enables significantly higher rack density.

Each compute tray integrates:

NVLink 6 and Spectrum-6

NVLink 6 delivers double the bandwidth of NVLink 5 (used in Blackwell), enabling efficient scaling across 576+ GPU clusters without the bandwidth cliff that plagues current NVL72 configurations. Spectrum-6 switches provide 102.4 Tb/s of aggregate switching capacity — enough to wire an entire AI training cluster with a single tier of switches.

Power Rack Architecture

Perhaps the most underappreciated innovation is the centralized Power Rack. By moving AC-to-DC conversion outside the compute trays, NVIDIA achieves:

So What? — Market Impact

Vera Rubin isn't competing against individual GPUs — it's competing against the entire server OEM ecosystem. NVIDIA is effectively designing the datacenter rack, reducing Dell, Supermicro, and HPE to assembly partners rather than design partners.

For cloud providers evaluating 2027 AI infrastructure purchases, Vera Rubin's co-designed approach promises 2-3x better performance-per-watt compared to today's GB200 NVL72. The question is whether hyperscalers will accept NVIDIA's increasing control over the hardware stack.

AMD's "Helios" rack-scale offering and Intel's Gaudi 4 will need to match this level of system integration to remain competitive.

Sources