Running More Secure AI Agents Locally with ASUS Ascent GX10 and NVIDIA NemoClaw

Why enterprises are rethinking cloud-only AI agents

Autonomous AI agents are moving beyond cloud-only deployments. As organizations begin to operationalize agentic AI, limitations around cost, data access, and control are becoming more visible. Reference stacks like NVIDIA NemoClaw, combined with purpose-built AI hardware like the ASUS Ascent GX10, introduce a path toward persistent, safer, self-hosted AI agents that can run on local infrastructure.

For many teams, cloud-based agents are effective during early experimentation. But at scale, token-based pricing introduces unpredictable operating costs, especially for always-on systems. At the same time, sensitive data often cannot be exposed to external services due to internal policies or regulatory requirements. These constraints limit how useful a cloud-dependent agent can be in real business environments.

Local deployment changes that equation. By running agentic AI on infrastructure you control, organizations can enable deeper system access, improve data security, and establish more predictable cost structures.

From experimentation to production: infrastructure requirements for agentic AI

Many developers first explore agentic AI using compact systems such as mini PCs or repurposed laptops. These environments are sufficient for lightweight tasks or cloud-assisted inference. However, production use cases introduce new requirements.

Agentic workflows rely on large language models to reason, plan, and execute actions across systems. As these workflows become more complex, they demand:

• High memory capacity for larger models and longer context windows
• Consistent compute performance for low-latency inference
• Reliable storage for logs, datasets, and task history
• Tailored efficiency for 24/7 operation

These requirements often exceed what entry-level systems can provide. Organizations moving toward always-on automation typically need dedicated infrastructure that can support sustained workloads without competing for resources.

A purpose-built approach: ASUS Ascent GX10

The ASUS Ascent GX10 is designed specifically for AI development and deployment at the edge. While its compact form factor is similar to a mini PC, its architecture is built for a very different purpose.

Rather than functioning as a general-use system, the GX10 operates as a dedicated AI node. It is typically managed over the network and deployed as part of a controlled environment. This approach aligns with enterprise best practices for isolating AI workloads and maintaining clear boundaries around data access.

At the core of the system is the NVIDIA GB10 Grace Blackwell Superchip, supported by the NVIDIA AI software stack. The platform delivers up to 1 petaflop of AI performance and includes 128GB of unified system memory. This combination enables local execution of advanced models that would otherwise require cloud resources or significantly larger infrastructure.

The GX10 can also scale to meet increasing demands. With NVIDIA ConnectX-7, two systems can be linked to significantly scale memory and support larger models or more complex multi-agent workflows.

Why memory architecture matters for agentic AI

For many AI workloads, memory capacity is a primary constraint. Traditional systems split resources between system RAM and GPU VRAM, which can create bottlenecks when running large models or managing extended context.

The GX10 uses a unified memory architecture with 128GB of LPDDR5x. This allows models such as Llama 3.1 70B or Nemotron-3 120B to run entirely in local memory while still supporting large context windows.

In practical terms, this enables:

• More advanced reasoning capabilities within agents
• Longer and more coherent multi-step interactions
• The ability to process larger internal datasets
• Reduced latency during inference and task execution

For enterprise use cases, these capabilities translate into more reliable automation and better performance across complex workflows.

Reducing latency in autonomous workflows

Agentic systems continuously alternate between reasoning and execution. Delays in this loop can reduce effectiveness, particularly in time-sensitive or multi-step processes.

The compute architecture of the GX10 is designed to minimize these delays. By accelerating inference and reducing memory bottlenecks, the system enables faster transitions between decision-making and action.

This is especially relevant for use cases such as:

• IT automation and system management
• Internal knowledge assistants
• Workflow orchestration across business applications

Faster iteration within the agent loop improves responsiveness and allows agents to handle more complex tasks with fewer interruptions.

Secure execution with NemoClaw

NVIDIA NemoClaw introduces an optimized ​reference​ stack for deploying AI ​agents ​with stronger security and control. A key component is NVIDIA OpenShell, which provides a sandboxed environment for executing commands and managing files.

This architecture allows AI agents to interact with local systems while maintaining guardrails around access and behavior. Tasks such as organizing files, generating reports, or interacting with internal tools can be performed without exposing the system to unmanaged risk.

For enterprise environments, this level of control is essential. It enables teams to adopt autonomous workflows while aligning with existing security policies and governance requirements.

The ASUS Ascent GX10 is designed to support these deployments as an agent-ready platform, providing the hardware foundation needed to run NemoClaw effectively.

Deployment efficiency and total cost considerations

High-performance AI infrastructure has traditionally required complex, large-scale systems. Multi-GPU workstations and data center resources can deliver strong performance, but they also introduce higher costs, increased power requirements, and greater deployment complexity.

The GX10 offers an alternative approach. By integrating compute, memory, and networking into a compact system, it reduces the overhead associated with building and maintaining custom infrastructure.

From a cost perspective, local deployment can also reduce reliance on ongoing cloud consumption. For always-on agents, this shift can lead to more predictable long-term operating expenses.

Flexible configurations for different workloads

The ASUS Ascent GX10 is available in multiple storage configurations, including 1TB, 2TB, and 4TB options. All configurations include the same core compute and memory architecture, allowing organizations to select storage based on their specific workload requirements. Storage can also be expanded over time, providing flexibility as data needs grow.

Building a foundation for agentic AI

Agentic AI is moving quickly from experimentation to real-world application. As organizations evaluate how to deploy these systems, infrastructure decisions are becoming increasingly important.

Running AI agents locally provides greater control over data, improves consistency in performance, and enables deeper integration with internal systems. With ​reference ​platforms like NemoClaw and purpose-built hardware such as the ASUS Ascent GX10, organizations can begin deploying secure, always-on AI agents that align with enterprise requirements.

For teams looking to move beyond proof-of-concept and into production, this approach offers a practical path forward. The ASUS Ascent GX10 is your turnkey solution for OpenClaw and more, and it’s available today. Follow the links below to purchase one of these compact AI supercomputers today.

ASUS Ascent GX10
Storage	Price (USD)	US	CA
1TB PCIe 4.0 SSD	ASUS price starting at $3,499	ASUS Newegg Amazon CDW Insight	ASUS Amazon CDW Insight
2TB PCIe 4.0 SSD	ASUS price starting at $3,999	ASUS Newegg Amazon CDW Insight
4TB PCIe 5.0 SSD	ASUS price starting at $4,699	ASUS Newegg Amazon CDW Insight