ASUS has introduced a new deskside AI supercomputer designed to bring data-center-class computing power directly to enterprise offices and research labs, eliminating the need for cloud-based AI infrastructure for many organizations. The ExpertCenter Pro ET900N G3, built on NVIDIA's DGX Station GB300 architecture, delivers up to 20 PFLOPS (petaflops) of AI performance and includes 748GB of coherent unified memory capable of supporting AI models with up to 1 trillion parameters.

Why Are Companies Moving AI Workloads Off the Cloud?

The shift toward on-premises AI infrastructure reflects growing concerns about data privacy, latency, and operational costs. Organizations using cloud-based AI services must send sensitive data to remote servers, introducing governance challenges and potential security risks. The ET900N G3 addresses these concerns by enabling enterprises to train, fine-tune, and run large AI models locally while maintaining complete control over where data is processed and stored.

ASUS positions the platform as an alternative to traditional cloud deployments, highlighting several practical advantages. Local AI infrastructure reduces latency for real-time AI operations, provides predictable operational costs without cloud subscription fees, and eliminates dependence on external service providers. For organizations handling proprietary data or operating in regulated industries, these benefits can justify the upfront investment in local hardware.

What Makes the ET900N G3 Different From Traditional Workstations?

The most significant innovation in the ET900N G3 lies in its memory architecture rather than raw computing power. ASUS combines 496GB of LPDDR5X system memory with 252GB of HBM3e graphics memory to create a unified 748GB memory pool connected through NVIDIA NVLink-C2C technology, which provides a 900GB per second bidirectional bridge between processors. This coherent memory design eliminates bottlenecks that typically occur when CPU and GPU memory operate separately, allowing developers to work with substantially larger AI models on a single local system.

The system achieves impressive performance metrics under demanding workloads. During testing using vLLM and the Qwen open-source AI model, the ET900N G3 achieved approximately 864 tokens per second output throughput and approximately 1,600 tokens per second combined input and output processing. For context, tokens are small units of text that AI models process; higher token throughput means faster responses and more efficient handling of large documents.

How to Evaluate On-Premises AI Infrastructure for Your Organization

• Data Governance Requirements: Assess whether your organization handles sensitive data that cannot be processed in cloud environments due to regulatory requirements or competitive concerns. Local AI infrastructure provides complete control over data location and access.
• Latency Sensitivity: Determine if your AI applications require near-instantaneous responses. On-premises systems eliminate network delays associated with cloud services, which can be critical for real-time decision-making applications.
• Total Cost of Ownership: Compare upfront hardware investment against ongoing cloud subscription costs. Organizations with consistent, high-volume AI workloads may achieve lower long-term costs through local infrastructure despite higher initial capital expenditure.
• Model Size and Complexity: Evaluate whether your organization needs to run AI models with hundreds of billions or trillions of parameters. The ET900N G3's 748GB unified memory supports substantially larger models than typical enterprise workstations.
• Development Team Expertise: Consider whether your team has the technical capability to manage on-premises AI infrastructure, including hardware maintenance, software updates, and system optimization.

The ET900N G3 supports the full NVIDIA AI software stack, CUDA-X libraries, and NVIDIA NemoClaw workflows, providing developers with familiar tools and frameworks. The system also supports NVIDIA Multi-Instance GPU technology, which allows the Blackwell GPU to be partitioned into as many as seven isolated instances that multiple developers or workloads can share simultaneously. This capability enables organizations to maximize hardware utilization across different teams and projects.

ASUS designed the platform with future expansion in mind. The company plans to support Windows-based AI development environments, which could broaden adoption across enterprise teams accustomed to Windows workflows. Support for autonomous AI agent development and interconnected AI workflows suggests the platform is positioned for emerging use cases beyond traditional model training and inference.

The launch of the ET900N G3 reflects a broader industry trend toward decentralizing AI infrastructure. Futurum Group's market analysis projects the edge semiconductor market reaching $339.6 billion by 2030, reflecting continued investment in AI-capable systems deployed closer to end users and workloads rather than concentrated in distant data centers. This shift suggests that organizations increasingly view local AI infrastructure as a strategic necessity rather than a niche alternative.

For enterprises evaluating whether to invest in local AI infrastructure, the key question centers on whether the advantages in data control, latency, and operational predictability justify the capital investment and ongoing maintenance responsibilities. The ET900N G3 demonstrates that vendors are packaging enterprise-grade AI computing into systems designed for direct deployment in office and research environments, making sophisticated AI infrastructure accessible to organizations that previously relied exclusively on cloud services.