Private Inference for Multi-Tenant AI Factories

Eliminate GPU carveouts. Run sensitive workloads on shared infrastructure without exposing tenant data to the model or the hosting environment.

stained Glass for AI Factories

Make every tenant's most sensitive data available for AI

Private inference is what separates an AI Factory running at full utilization from one where sensitive workloads never reach production or tie up valuable GPUs.

Enterprise AI

Unlock sensitive data tiers on shared AI Factory compute. HR, legal, finance, clinical, and code workloads that currently sit in dedicated carveouts or never reach production move to shared capacity. The factory runs at full utilization.

Sovereign AI

Sensitive government, defense, and citizen data cannot leave the jurisdiction or touch foreign-controlled infrastructure. Stained Glass enables efficient, cost-effective sovereign AI factories that never take custody of raw data.

Train SGT for Your Model

Tenants train model and SGT together on shared AI Factory infrastructure. They leave with both, ready for private inference anywhere. Don’t let your customers leave your AI Factory with just a model.

AI-as-a-Service

MSPs, telcos, and SIs serve sensitive workloads without taking custody of end-customer plaintext. End-customer data policy objections go away. Your platform becomes the reason they succeeded.

AI Factory ROI requires full use of data and compute

Dedicated infrastructure guarantees isolation. It also guarantees underutilization, inflated cost, and half your AI Factory sitting idle.

20–40%

GPU utilization when relying on hardware isolation for sensitive workloads. Compute costs are consistent regardless of usage.

90%

GPU utilization on shared infrastructure with Stained Glass. Same factory footprint. Dramatically higher token output.

10X savings

Using Stained Glass to comingle sensitive workloads on shared AI Factory infrastructure vs. siloed compute.

~15ms

Added latency from Stained Glass Transform. Model inference runs at full speed. No performance tradeoff.

Private by design, not by isolation

Protopia Stained Glass sits at the tenant boundary, transforming sensitive inputs into protected representations before they reach the serving environment. Every tenant keeps plaintext under their own control. The factory runs on shared capacity, at scale.

Plaintext stays in the tenant boundary

The AI Factory operator never takes custody of plaintext prompts or context. No log, no memory buffer, no temporary file contains raw data.
No model retraining or infrastructure changes

SGT is a post-training step using less than 1% of training resources. Deploys as a drop-in privacy layer via an OpenAI-compatible API. Works with Llama, Mistral, and other open-weight LLMs.
Accuracy parity with plaintext

Validated across HellaSwag, MMLU, TruthfulQA, and ARC benchmarks. Delta under 5% in production workloads. Full fidelity for text, code, tabular data, and video.

Benefits

Protopia Stained Glass unlocks AI Factory value

From inference to agentic workflows, Stained Glass Transform secures sensitive inputs at every stage of AI Factory deployment. Explore how Protopia enables safe, scalable LLM adoption on shared infrastructure.

Activate Restricted Data

Your highest-value data is often what AI can’t touch because sending it to a shared serving environment exposes it in plaintext. HR records, legal files, clinical notes, code repositories. Stained Glass transforms it at the source so it reaches the model as a protected representation. The data becomes usable. The exposure risk goes away.

Better Economics at Scale

Every hardware carveout reserves capacity for one tenant that could serve many. When plaintext never enters the serving environment, sensitive workloads move to shared capacity, token volume goes up, and cost per outcome comes down. The same AI Factory footprint does more. GPU utilization moves from 20–40% to 90%.

Serve Regulated Markets

Regulated industries, sovereign programs, and multi-tenant operators have all been told dedicated infrastructure is the only path for sensitive workloads. Stained Glass gives operators a concrete alternative and opens demand from customers and use cases that previously required their own stack to serve. Finance, healthcare, defense, and public sector.

Faster than FHE. Lighter than enclaves. Smarter than masking.

Approach	Eliminates plaintext	Low Compute Usage	Full accuracy
Protopia Stained Glass
Homomorphic Encryption (FHE)			~
Confidential Computing (TEE)		~
Token Masking / Redaction	~

CUSTOMER STORIES

See private inference in action

Test real-world multi-tenant inference use cases on your existing infrastructure — no model retraining or refactoring needed.

MGT AI Experience Center for SLED

Together, MGT, Protopia AI, and HPE are delivering a practical path for SLED agencies to deploy governed AI with the performance to scale and the privacy and control to protect sensitive data.

Trustworthy AI Factory Architecture with HPE

Securely orchestrate data between agents, apps, and endpoints on HPE Private Cloud AI. Validated across five selling patterns for MSPs, telcos, and enterprise operators.

Secure Inference for NVIDIA NIMs

Deploy Stained Glass as a drop-in privacy layer on NVIDIA NIM inference microservices. Works with Llama, Mistral, and other open-weight models on your existing AI Factory.