ASUS announced on November 17, 2026 that it is ramping volume Vera Rubin NVL72 shipments in the fourth quarter, shipping a fully liquid-cooled rack that pairs 72 Rubin GPUs with 36 Vera CPUs and lists nine NVLink 6 switches, 144 ConnectX-9 SuperNICs and four 110-kilowatt power shelves for a peak rack power demand of 227 kilowatts. NVIDIA first detailed the Vera Rubin POD at CES 2026 in its NVIDIA Vera Rubin POD technical blog, spanning 40 racks, 1,152 Rubin GPUs and 60 exaflops with 10 PB/s aggregate scale-up bandwidth across the racks. For infrastructure buyers, the shift from Blackwell to Rubin racks is "not a socket-for-socket refresh"; it changes how power is provisioned, how heat is rejected and how the NVLink 6 fabric is wired inside each rack.
Rack power delivery moves to 227 kW and full liquid cooling
The ASUS XA VR721-E3 AI POD uses four 110-kilowatt power shelves to support a peak rack power demand of 227 kilowatts. That is a substantial increase even though NVIDIA states that all MGX racks share identical power, cooling and mechanical envelopes for easy transitions. Buyers who have already built to the MGX power and cooling envelope can reuse the physical infrastructure, but upstream electrical service, busway and breaker coordination must be sized for the higher load. ASUS is working with Schneider Electric, AVEVA and IBM on digital-twin facility co-design, and its Control Center Data Center Edition monitors infrastructure health, power and cooling in real time. For facility planning, see power and cooling considerations for AI deployments and check the Open Compute Project (OCP) specifications for rack envelopes.
The Vera Rubin NVL72 reference rack is fully liquid-cooled, removing the 72 Rubin GPUs and 36 Vera CPUs from air-cooled heat rejection. ASUS also shows an HGX Rubin NVL8 in 3U hybrid-cooled and 2U liquid-cooled form factors for smaller AI nodes. At 227 kW peak rack power, direct-to-chip liquid cooling becomes the default for dense NVL72 deployments; operators must plan for coolant distribution units, facility water loops and leak detection. ASUS’s digital-twin work extends the thermal model beyond the rack to the building level, with the Omniverse DSX Blueprint allowing racks to be simulated before deployment.
NVLink 6 fabric rewires the rack interior
The scale-up fabric is the most consequential change inside the rack. Vera Rubin NVL72 uses an NVLink copper spine to pair 72 Rubin GPUs with 36 Vera CPUs, and the ASUS rack configuration lists nine NVLink 6 switches alongside 144 ConnectX-9 SuperNICs and 18 BlueField-4 DPUs. The NVLink 6 chiplet carries 36 custom 400G SerDes links, and NVLink-C2C chiplet bandwidth to the Vera CPU doubles to 1.8TB/s, according to SemiAnalysis’s Vera Rubin analysis. Across a Vera Rubin POD, NVIDIA reports 10 PB/s aggregate scale-up bandwidth across 40 racks. This means the interconnect topology must be planned as a rack-level subsystem: the copper spine and nine NVLink 6 switches are fixed elements, while scale-out uses ConnectX-9 SuperNICs. MGX NVL racks use NVLink; MGX ETL racks use Spectrum-X Ethernet or Groq 3 LPU links. A related BlueField-4 STX rack combines Vera CPU and ConnectX-9, scaling out via Spectrum-X Ethernet, and the Spectrum-6 SPX networking rack uses a 102.4 Tb/s switch with 200 Gb/s co-packaged optics. For guidance on matching fabric choices to workload requirements, see choosing the right interconnect for AI workloads and the broader AI data centres coverage hub.
Deployment checklist and open risks
Several infrastructure-relevant details remain open. Initial Rubin HBM4 shipments may fall below NVIDIA’s 22TB/s target, landing near 20TB/s, and memory suppliers are reportedly struggling with HBM4 pin speed requirements. Micron is said to be behind Samsung and SK Hynix and effectively out of Rubin HBM4, which could constrain 288GB-per-GPU capacity if qualification slips. SemiAnalysis estimates a $500B Rubin buildout with large downstream market dislocations, but that is an estimate, not a committed roadmap. ASUS’s six-layer validation framework, detailed August 24, 2026, blocks failing systems before shipment, but whether it can be reproduced consistently across customer sites is not yet proven.
- Plan for 227 kW peak rack power with four 110-kilowatt power shelves in the ASUS XA VR721-E3 design; reuse existing MGX power and cooling envelopes, but verify upstream capacity.
- Treat liquid cooling as mandatory for NVL72, while HGX Rubin NVL8 offers hybrid-cooled and 2U liquid-cooled alternatives for smaller pods.
- Budget for the NVLink 6 copper spine, nine NVLink 6 switches and 144 ConnectX-9 SuperNICs per NVL72 rack; scale-out options split between Spectrum-X Ethernet and Groq 3 LPU links depending on MGX rack type.
- Validate HBM4 supply before committing to capacity: Rubin GPUs target 288GB HBM4 at 22TB/s, but early shipments may arrive near 20TB/s.
- Use digital-twin facility co-design along the lines of ASUS with Schneider Electric, AVEVA and IBM to pre-simulate power and cooling before racks arrive.