At AI Tech 2026 in Seoul, ASUS set out an AI factory platform spanning rack-scale Vera Rubin systems, AI-native and object storage, deployment tooling and a governance layer. It puts ASUS alongside Supermicro and GIGABYTE as a full-stack AI infrastructure supplier rather than a server vendor.
What was announced
- Announced at AI Tech 2026 in Seoul, covering compute, networking, storage, deployment software, infrastructure management and AI governance.
- The rack-scale AI POD XA VR721-E3 is based on NVIDIA Vera Rubin NVL72; ASUS claims 10 times the performance per watt of the previous generation.
- XA NR1I-E12LR and XA NR1I-E12L are based on NVIDIA HGX Rubin NVL8 for training, inference and post-training.
- The 2U XA P2N-E2 uses NVIDIA MGX with two NVIDIA Vera CPUs and up to two dual-slot GPUs, aimed at agentic workloads.
- The ESC8000-E12P supports NVIDIA RTX PRO 6000 and RTX PRO 4500 Blackwell Server Edition GPUs for enterprise inference and vision AI.
- Storage additions: UF920-E3-RS24 built on NVIDIA STX, OJ340A-RS60 object storage and the VS320D-RS26N.
- Conventional server refreshes include RS700-E12-RS4U and RS720-E12-RS12U on Intel Xeon 6, and RS720A-E14B-R32U and RS500A-E14B-R12U on AMD EPYC 9006.
- ASUS is using the NVIDIA DSX Sim Blueprint with Schneider Electric, AVEVA and IBM to build digital twins of proposed AI factories before installation.
- Edge additions include the PE3000N using the NVIDIA Jetson Thor T5000 module and the fanless RUC-2000 series on Intel Core Ultra Series 3 rated up to 180 AI TOPS.
The ETON view
Every large server vendor is now selling the same story: not a server, a factory. The commercial substance for a buyer is narrow but real. Digital twin planning before installation is genuinely useful if you are committing to a 72-GPU rack, because the failure mode on these projects is almost never the compute, it is power density, cooling loop design and the hall being unable to take the rack you ordered. If a vendor will model that with you, take them up on it.
The part to be careful about is the software layer above the hardware. Quota and billing, MLOps portals and governance tooling are how a hardware vendor turns a one-off purchase into a platform relationship, and they are rarely portable. If you already run Kubernetes and Slurm, buy the hardware and keep your own stack. The claim worth pressing on is the 10 times performance per watt for Vera Rubin NVL72 against the previous generation: that is an ASUS claim with no published test behind it, so treat it as a starting point for a proof of concept rather than a planning number.
The quieter and more useful line in the same announcement is the refresh of ordinary rack servers on Xeon 6 and EPYC 9006. That is what most estates actually need this year. We source ASUS, GIGABYTE, Dell and HPE without a vendor quota to hit, so if the requirement is 20 general purpose 1U and 2U nodes rather than a rack-scale AI pod, we will say so and quote on price, lead time and support terms.
Related infrastructure
Category: AI & GPU · Vendor: ASUS, NVIDIA, AMD, Intel · Technology: Vera Rubin NVL72, HGX Rubin NVL8, NVIDIA MGX, NVIDIA STX storage, EPYC 9006, Xeon 6 · Last verified: 06 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
