Conference Presentation, Fireside Chat, Product Demonstration
Running Secure Agents on Multi-Tenant Sovereign AI Infrastructure | Protopia AI x HPE | RAISE 2026
- Sovereign AI providers, neoclouds, and enterprise IT departments are driving demand for solutions that enable sensitive agentic workflows on shared infrastructure without compromising trust or economics.
- Regulated industries such as FSI, healthcare, and the public sector present high-ROI use cases where data sensitivity is paramount, creating urgency as workflows shift toward multi-step agentic processes.
- Current friction arises from the need to balance the capture of sensitive agentic tokens with infrastructure cost efficiency, as isolating hardware to ensure data separation leads to over-provisioning and poor ROI.
- Risks of data exposure in shared environments include logs, observability pipelines, temporary storage, and internal network traffic, particularly when raw data is decrypted upon ingress despite transit encryption.
- Over-reliance on physical separation by dedicating GPUs for specific tenants results in utilization issues where isolated cells cannot meet token output volumes, potentially causing operators to face supply problems rather than demand deficits.
- To address these challenges, Protopia and HPE are deploying "Stained Glass Transform" technology, which converts raw data into a stochastic, randomized representation within the data owner's trust zone before leaving, eliminating the need for extensive hardware carve-outs.
- This transformation utilizes a neural network trained for specific target models like Nemotron Omni, operating as a one-way process without keys, achieving accuracy close to the baseline without introducing significant latency.
- The solution reduces exposure risk in logging and network traffic, enabling interactivity that traditional homomorphic encryption often lacks, though some customers may still require dedicated infrastructure for specific cases.
- Integration strategies include HPE services helping operators train transforms and deploy tools like Protopia's SafeCraw subagent, which sits within harnesses such as OpenShell to identify and transform sensitive requests prior to factory ingestion.
- Operators anticipating the shift to agentic workloads can expect revenue uplifts from serving sensitive information and potentially achieving hundreds of times larger factory throughput compared to baselines with lower GPU utilization.