Chips & Compute

ASUS Ramps Vera Rubin NVL72 Shipments as Rack Race Shifts to Execution

ASUS is ramping Q4 shipments of NVIDIA Vera Rubin NVL72 racks and betting on digital twins, six-layer validation and automated commissioning to stand out as rivals build the same hardware.

ASUS Increases Vera Rubin NVL72 Output for Rack-Scale AI Deployments
ASUS Increases Vera Rubin NVL72 Output for Rack-Scale AI DeploymentsTho-Ge / pixabay
By Sophie Lindqvist5 min read

Why it matters

  • ASUS announced on November 17, 2026 that it is ramping volume shipments of NVIDIA Vera Rubin NVL72 systems in Q4.
  • The XA VR721-E3 AI POD holds 72 NVIDIA Rubin GPUs, 36 Vera CPUs and peaks at 227 kilowatts per rack.
  • ASUS validates racks through a six-layer framework detailed on August 24, 2026, blocking shipment on any failed threshold.
  • The AI Factory workflow combines NVIDIA's Omniverse DSX Blueprint, introduced October 28, 2025, with partners Schneider Electric, AVEVA and IBM.
  • At SC26 in Chicago, ASUS also shows HGX Rubin NVL8-based systems and the Vera-CPU XA P2N-E2 for agent orchestration.

ASUS said on November 17, 2026 that it is increasing volume shipments of NVIDIA Vera Rubin NVL72 systems during the fourth quarter, pointing to robust market demand for rack-scale AI infrastructure. The company is using SC26 in Chicago to present an AI Factory reference architecture that spans design simulation, factory validation, deployment automation and ongoing management.

The timing tells you what has changed in this market. Multiple system builders now offer Vera Rubin NVL72 platforms, so the hardware alone no longer separates vendors. ASUS is positioning manufacturing consistency, pre-deployment testing and facility planning as its differentiators, with the stated goal of making large-scale deliveries more predictable.

The stakes are straightforward: these racks are expensive, power-hungry and hard to retrofit once installed. If a vendor can't deliver consistent, pre-validated systems, customers pay for facility rework and idle GPUs.

What is ASUS actually shipping?

At the center of the SC26 display sits the XA VR721-E3 AI POD, a fully liquid-cooled rack packing 72 NVIDIA Rubin GPUs and 36 Vera CPUs. ASUS designed it for frontier-model training, complex reasoning and high-throughput inference — the three workloads driving current datacenter buildouts.

The spec sheet explains why delivery is risky:

  • Nine NVLink 6 switches
  • 144 ConnectX-9 SuperNICs
  • 18 BlueField-4 DPUs
  • Four 110-kilowatt power shelves
  • Peak rack power demand of 227 kilowatts

A rack pulling 227 kilowatts leaves little margin for assembly errors or facility surprises. That density makes the production process part of the system's delivery risk, not a separate afterthought.

How does ASUS validate a rack before it ships?

ASUS runs a layered approach that starts before volume manufacturing. Its Predictive Virtual Lab assesses potential signal-integrity problems during design, catching faults while they're still cheap to fix.

On the factory floor, automation at the L10 node and L11 full-rack stages handles optical inspection, protective-cover removal, module-positioning assistance and precision fastening. ASUS says this set of steps reduces assembly variation across units — the consistency problem that plagues high-density hardware.

Testing then continues through a six-layer validation framework the company detailed on August 24, 2026. It covers:

  • Hardware functionality
  • Firmware and remote-management checks
  • Repeated power cycling and stress testing
  • NVIDIA diagnostic tools
  • Cross-GPU and network performance

The framework has teeth, at least on paper. Any test item that misses its threshold stops the system from advancing while teams isolate the cause, remediate it and repeat validation. Powered aging assessments also monitor thermal behavior, power stability and system health before final inspection and shipment.

The takeaway: ASUS is treating validation as a gate, not a checkbox, which matters when a single faulty connection can sideline a multi-million-dollar rack.

Why simulate power and cooling before buildout?

ASUS's AI Factory workflow starts before a rack reaches the customer at all. It combines NVIDIA's DSX reference design with the Omniverse DSX Blueprint, which NVIDIA introduced on October 28, 2025. Working with Schneider Electric, AVEVA and IBM, ASUS brings IT infrastructure, power, cooling and facility models into one simulation environment.

The blueprint uses digital twins to co-design buildings, power, cooling and AI infrastructure. That lets engineers examine constraints and design trade-offs before anything gets physically deployed.

In a May 13, 2026 technical account, ASUS described how rack systems become validated simulation assets carrying power, thermal and connection-point metadata. Planners can then evaluate water temperature, flow and pressure, cooling-distribution-unit placement, piping and rack heat load alongside facility power constraints.

The purpose is simple: expose cooling or facility problems before equipment arrives. Once a 227-kilowatt rack lands on site, changes can require expensive rework. Simulation moves that discovery earlier in the timeline, when it costs software iterations rather than construction.

What happens after installation?

ASUS Professional Services covers site-readiness assessment, power and cooling planning, systems integration and on-site installation. The ASUS Infrastructure Deployment Center then automates operating-system and software installation, cluster configuration and network setup.

Its zero-touch workflow lets servers register, configure and join a cluster when powered on. Version-controlled scripts make deployments repeatable, so a second site gets the same setup as the first.

Once systems run, ASUS Control Center Data Center Edition provides centralized infrastructure-health monitoring, power and cooling visibility, firmware and software updates, and security controls. The through-line is automation at every stage — from factory fastening to cluster join — aimed at removing human variation from a process where variation is costly.

What else is on show at SC26?

Alongside the NVL72 rack, ASUS is displaying hybrid-cooled 3U and fully liquid-cooled 2U systems based on NVIDIA HGX Rubin NVL8. It is also showing the Vera-CPU-based XA P2N-E2, aimed at multi-step agent execution and orchestration — a nod to the shift toward agentic AI workloads that need sustained, coordinated compute rather than single-shot training runs.

ASUS servers are available worldwide, though individual products and configurations vary by region.

Can ASUS repeat this at scale?

The open question for the production expansion is whether ASUS can reproduce this chain of simulation, validation and automated commissioning consistently across customer sites. A reference architecture demonstrated at a trade show is one thing; repeating it across dozens of facilities with different power, cooling and network conditions is another.

The company's practical pitch is that the workflow limits costly facility changes, integration delays and GPU idle time as these densely integrated racks move into service. Whether that pitch holds depends on execution — which, by ASUS's own framing, is now the whole game.

Source: unite.ai

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Nexalus and RENDA Launch Nova8, a Liquid-Cooled AI Server Built for Retrofit
  2. NVIDIA Locks In 3.5x More Substrate Capacity From Kinsus By 2028
  3. Dell expands AMD-based AI infrastructure with new servers and rack systems
  4. AMD Ships 256-Core Zen 6 EPYC 'Venice' Chips as AI Agents Drive CPU Demand
  5. Nvidia launches open-weight Nemotron models for enterprise agents and AI PCs

« Previous articleNext article »