How enterprises can securely run sensitive agentic AI workloads on shared AI factory infrastructure to reduce costs, improve GPU use, and unleash stronger ROI.

Every customer wants to invest heavily in AI factories, expecting them to become multitenant platforms for innovation, productivity, and business transformation. However, many organizations discover an unexpected challenge once they move beyond experimentation and pilots.

The most valuable AI workloads often require access to sensitive data such as customer records, financial transactions, source code, healthcare information, and proprietary intellectual property.

As a result, these workloads are frequently isolated onto dedicated infrastructure, reducing utilization and limiting the return on AI investments. They end up creating compute silos that are almost always highly underutilized.

When sensitive workloads cannot safely run on multitenant infrastructure, organizations create:

• Dedicated clusters
• Coarse-grained separate environments
• Additional governance layers
• Duplicated operational processes

This creates what many IT organizations experience as an isolation tax with consequences such as:

• Lower GPU utilization
• Increased infrastructure costs
• Slower AI adoption
• Fragmented operating models

Instead of becoming a shared business platform, the AI factory becomes a collection of disconnected environments with complex ownership and limited ROI visibility.

As enterprises move deeper into AI adoption, AI factory operators face a complex security mandate: protect sensitive data through complex operational surfaces, secure shared infrastructure, safeguard model IP, and enforce governance without constraining the high-value workloads the factory was built to run.

A reference design for private enterprise agentic workflows on zero trust AI factories

Solutions combining NVIDIA OpenShell, NVIDIA Nemotron, Protopia AI Stained Glass Transform (SGT) with HPE AI factory solutions demonstrate how sensitive agentic workflows can safely operate within multitenant AI factory environments. The result is simple: organizations can run their highest-value agentic workloads on the same infrastructure they already own.

Together, these components enable secure, scalable agentic AI on shared infrastructure.

• NVIDIA OpenShell is an open source runtime for building and running autonomous, self-evolving agents more safely. OpenShell sits between your agent and your infrastructure to govern how the agent executes, what it can see and do, and where inference goes. It enables claws to run in isolated sandboxes, with fine-grained control over privacy and security.
• NVIDIA Nemotron is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.
• Protopia AI SafeClaw integrates into the agent loop as a subagent, ensuring sensitive data is stochastically transformed locally before leaving the enterprise boundary, without disrupting orchestration frameworks.
• Protopia AI Stained Glass Transform protects the inference path by removing plaintext exposure across logs, memory, and system layers, enabling sensitive workloads to run securely on shared infrastructure.

AI factory solutions within the joint portfolio NVIDIA AI Computing by HPE bring these elements together into a governed, multitenant platform—combining infrastructure, security, and operating models to increase utilization and mitigate the isolation tax.

How it looks in practice

One example of a sensitive agentic AI workload illustrates how these components come together in practice.

A fashion retailer operating across multiple locations struggles with siloed data across CRM, inventory, and sales systems, leading to delayed decision-making and limited visibility into store operations.

Using NVIDIA OpenShell and NVIDIA Nemotron models, the company deploys AI assistant agents enabling natural language queries and real-time, role-based insights across stores.

SafeClaw and Stained Glass Transform ensure sensitive data is protected throughout the workflow, allowing the solution to run securely on a multitenant AI factory—improving utilization while enabling faster, more consistent operations at scale.

HPE Services help ensure robustness and consistency across the solution components integration, aligning access control with data, agents, and user interfaces applying the shared gateway implementation.

Figure 1. Example of end-to-end customer integration

Many customers underestimate the complexity of operationalizing these capabilities. Success requires more than deploying technology. Organizations must address:

• AI governance
• Security architecture
• Sovereignty requirements
• Data classification
• Model lifecycle management
• Agent orchestration
• Operational processes
• Change management

Figure 2. Boundaries representation

The challenge shifts from can technology work? to how do we safely operationalize it at enterprise scale?

HPE Services helps customers:

Assess—Identify workloads, use cases that can benefit from secure agentic AI

Design—Create architectures aligned with security, governance, and sovereignty requirements

Implement—Integrate AI platforms, models, data pipelines, and enterprise systems

Operationalize—Establish AI factory operating models, governance frameworks, optimization, and lifecycle management

Scale—Move from isolated pilots to enterprise-wide adoption

The combination of product innovation and services expertise accelerates time to value while reducing risk.

Agentic AI is the new application model for the enterprise. As organizations deploy thousands of AI agents operating across business processes, the ability to securely run sensitive workloads on multitenant infrastructure is becoming a strategic requirement rather than a technical preference.

The organizations that succeed will not simply adopt new AI technologies. They will build operating models that combine:

• Secure AI platforms
• Governance frameworks
• AI factory architectures
• Enterprise-scale operational practices

Transforming AI from isolated experiments into measurable business outcomes.

The future of AI factories will not be determined solely by model performance or infrastructure scale. It will be determined by an organization's ability to securely operationalize its most valuable AI workloads. The architectures for secure agentic AI are in production today, and useful agentic AI is already generating the token demand AI factories were built to serve. Enterprises that act now mitigate the isolation tax, restore the utilization their investment was justified on, and capture the full economic potential of their AI investments; for those that wait, that potential sits stranded on capacity they have already paid for. Technology provides the foundation, but it is the combination of product innovation and services expertise that ultimately turns AI capacity into business value.

CTA: For more information, visit: HPE.com/ai-services

By Authors:

Raffaele Tarantino, WW AI & Data GTM Strategy Lead, HPE AI Services

Raffaele Tarantino is the WW AI & data GTM strategy lead within Advisory and Professional Services at HPE. He is responsible for the messaging and sales enablement of the services portfolio. Raffaele has 10 years of experience in HPE roles ranging from private cloud consultant to compute specialist and AI architect as a member of the worldwide practice. Raffaele designed the Machine Learning Development Services and contributed to the HPE AI Services – Generative AI Implementation launch.
Linkedin account (Raffaele Tarantino)