ASUS is pitching a joined-up AI infrastructure offer, from NVIDIA-based data-centre servers to edge systems and tools for planning power and cooling. The practical point is that enterprise AI factories need more than GPUs: they also need the CPUs, networking and facilities to keep agent workloads running.
Watch Desk analysis
What happened
At NVIDIA’s GTC event in Berlin on 21 and 22 October, ASUS announced an enterprise offer combining AI servers and infrastructure, workflow software, and deployment services. The sponsored Register article describes two systems built around NVIDIA’s eight-GPU HGX Rubin NVL8 board: a 3U hybrid-cooled server and a 2U fully liquid-cooled model. ASUS says the hybrid design is an easier transition into AI factory hardware, while the liquid-cooled system is aimed at compute density and power efficiency.
For more flexible builds, ASUS also showed servers based on NVIDIA’s modular MGX blueprint, including an eight-GPU liquid-cooled system and a 2U server using NVIDIA Vera processors for agentic work. Beyond the data centre, the reported line-up includes developer systems and a rugged edge computer for industrial use.
Why it matters
The announcement reflects a shift in how companies are planning AI infrastructure: not just around model training, but around inference and agent workloads that repeatedly plan tasks, run code and call tools. ASUS executive Paul Ju argues that those loops put more demands on CPUs as well as GPUs. That is his account of the bottlenecks, not an independently measured comparison of these systems.
The offer also reaches beyond the server rack. ASUS says its digital-twin work with NVIDIA’s DSX platform can model equipment layouts and cooling requirements before construction, then help operators test proposed changes before applying them to live infrastructure. For buyers, that could make planning and ongoing operations part of the same package as the hardware.
Our read
The useful idea here is the breadth of the pitch: data-centre servers, edge hardware and facility planning are being presented as one AI infrastructure problem. But this is a sponsored account, and ASUS’s performance and serviceability figures are company claims, not independent test results. Buyers should treat the menu as a starting point for comparing systems, not as a verdict on which configuration wins.
What to watch
- When the announced Rubin-based systems will be available, and at what cost.
- Whether ASUS substantiates its claimed maintenance-time savings.
- How the digital-twin tools perform against real facility plans and operating data.
- Whether enterprise customers adopt a full-stack offer or keep mixing suppliers.
Discussion spark: Would you buy an integrated AI infrastructure package from one vendor, or keep servers, software and facility planning with separate suppliers?
Sources and evidence
- Building AI factories at enterprise scale (5 November 2026, 00:01 UTC)
- GSLM2: Multi-Stream Speech Representation Learning and Generative Spoken Language Modeling | Research – AI at Meta (5 November 2026, 00:01 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.