A Visual Breakdown of NVIDIA’s Vera Rubin AI Cabinet

From Super‑Server to AI Factory

NVIDIA’s Vera Rubin NVL72 cabinet marks a shift from simple hardware stacking to a co‑designed system that integrates GPUs, CPUs, networking, liquid cooling, power, and management software, defining a new standard for AI infrastructure.

Core Specifications

The NVL72 system houses 72 Rubin GPUs and 36 Vera CPUs, linked by a sixth‑generation NVLink switch delivering up to 260 TB/s of intra‑cabinet bandwidth. When treated as a single AI supercomputer, it can train MoE models with only a quarter of the GPUs required by the previous Blackwell generation and reduces per‑token inference cost by tenfold.

Six‑Chip Collaborative Design

NVIDIA Vera CPU : 88 custom Olympus cores, space‑threading technology, and 1.8 TB/s NVLink‑C2C links directly to GPUs, acting as a data‑movement engine.

NVIDIA Rubin GPU : Third‑generation Transformer engine, up to 50 PFLOPS inference at NVFP4 precision, paired with HBM4 memory for trillion‑parameter MoE models.

NVLink 6 Switch : Provides up to 3.6 TB/s per GPU, serving as a high‑speed “overpass” for GPU communication.

ConnectX‑9 SuperNIC : 800 Gb/s Ethernet per port, doubling bandwidth over the previous generation and enabling horizontal scaling.

BlueField‑4 DPU : Offloads data‑center tasks and offers hardware‑level security isolation.

Spectrum‑6 Ethernet Switch : Silicon‑photonic integration connects multiple NVL72 cabinets, forming a large‑scale AI cluster “neural network”.

100% Liquid Cooling

With power approaching 200 kW, traditional air cooling reaches its limit. Rubin implements the world’s first fully liquid‑cooled AI platform, eliminating all fans. The coolant enters at 45 °C, allowing passive dry‑cooler heat rejection in many climates, dramatically lowering PUE and water consumption. Micro‑channel and large‑cold‑plate designs further boost heat‑exchange efficiency.

Industry Impact

BOM Cost Surge : The VR200 cabinet’s material cost rises 95% over the previous GB300, reaching roughly $7.8 million per unit.

Memory Becomes the Biggest Winner : Memory’s cost share jumps from 5‑10% to 25‑30% (a 435% increase), reducing GPU’s BOM share from 65% to 51%.

PCB and Materials Upgrade : PCB content grows 233% due to added layers, CCL upgrade to M8/M9, and new mid‑boards.

Liquid Cooling and Power : Full‑liquid cooling becomes standard (+12% value), and power architecture shifts toward 800 V HVDC to meet rising power demands.

Future Outlook

Vera Rubin platforms are slated for mass production in the second half of 2026, ushering in an era where AI compute is delivered as a “cabinet‑as‑computer” system. The ensuing AI dividend will radiate from individual chips to storage, packaging, materials, and thermal solutions across the supply chain.

Notes: All cost ratios are illustrative estimates from industry supply‑chain calculations, not official NVIDIA disclosures. Image sources are credited to “图财社”.