Zero Patient Data Leakage: Six Clinical AI Agents Run on a Single System

A system has emerged that enables health care organizations to run state-of-the-art clinical decision-making AI without sending patient data outside their facilities.

According to BioPharma Dive on the 8th (local time), a new reference architecture based on the Dell Pro Max (GB300) and NVIDIA Grace Blackwell Ultra GB300 superchip has been unveiled. The system runs multiagent clinical AI entirely on premises, processing patient data only within an organization’s own infrastructure.

Designed for research labs and enterprise AI teams, the system delivers 20,000 teraflops (TFLOPS) of computing performance at FP4. It also has enough memory capacity to host models with up to 1 trillion parameters.

According to a recently released NVIDIA playbook, the system runs six specialized AI agents on a single supercomputer. It consists of one orchestrator and five specialized agents responsible for patient data, tests and vital signs, medications, clinical analysis, and molecular visualization.

At its core, NVIDIA Nemotron 3 Super, a 120-billion-parameter mixture-of-experts (MoE) model, runs locally using containerized inference.

The architecture supports air-gapped deployment, keeping data on the device. Patient data is processed locally and does not pass through hosted large language models (LLMs) or external AI services. Only a small number of read-only outbound connections are allowed under an implicit-deny network policy.

The system is designed to complement, rather than replace, existing clinical workflows. For example, in identifying care gaps, it automates the process by which quality teams manually compare test results with benchmarks, identifying patients whose results fall outside target thresholds.

However, the basis for clinical decisions is stored in files, not model weights. If an organization changes a threshold from 9.0% to 8.5% to align with its standards, agents reflect the updated value starting with the next query. Because they read human-readable skill files at query time without retraining, they can immediately adapt when clinical guidelines change.

Dell AI Factory helps place workloads where appropriate, whether locally, in a data center, or in the cloud. The Dell Pro Max GB300 has enough GPU memory density to run frontier-class LLMs and protein structure prediction models simultaneously on a single deskside system.

In regulated industries, data is intellectual property (IP). Organizations in these fields often require training data and model weights to remain within their own firewall-protected infrastructure. On-premises infrastructure can also convert variable token costs into predictable capital investments.

Although the playbook is a proof of concept rather than a finished product, it signals that the computational barriers to running sophisticated AI locally have fallen below a level that most health systems and biopharmaceutical organizations can afford, BioPharma Dive reported.

The playbook has been validated by NVIDIA and is available for rapid deployment through Dell AI Factory. Health care and biopharmaceutical organizations can download it, deploy the full six-agent system in about 1 hour, and evaluate it for their own use cases. Because the architecture is consistent from deskside systems to data centers, organizations can scale what they have validated locally to AI Factory without retraining or switching platforms.