AI That Never Exposes Patient Data
Run frontier AI models on hardware you control, keeping every byte of patient data inside your institution. HIPAA-compliant by architecture, not just by policy.
Compliance by Architecture, Not by Contract
On-premises AI keeps PHI inside your network, eliminating the compliance gap that cloud services create. No BAA needed — your data simply never leaves.
Research Data Stays Private
When inference runs on hardware you control, data never traverses an external network or becomes subject to another organization's policies.
No Data Leaves Your Network
All inference runs on hardware you control. Patient data and research prompts never traverse external networks or reside on third-party servers.
No Model Training on Your Data
Clinical prompts never reach model providers. The architecture physically prevents your data from being incorporated into training datasets.
Institutional Review Board Compliance
Research data stays within your institution's approved computing environment throughout the entire research lifecycle, satisfying IRB data governance requirements.
Multi-Site Research Without Data Sharing
Share model weights across institutions, not patient data. Each site runs the same models on its own data without transferring PHI, enabling collaboration with full data sovereignty.
Clinical and Research Risks of Cloud AI
Cloud AI convenience obscures risks that can jeopardize research programs, institutional reputations, and patient trust.
Inadvertent PHI Disclosure
Even de-identified records can contain re-identification signals. Once pasted into a cloud prompt, that data exists on third-party servers outside institutional control — an untracked PHI disclosure violating HIPAA's minimum necessary standard.
Training Data Leakage
Cloud providers may use your prompts to train their models unless you opt out — and opt-out mechanisms vary and can change without notice. Research hypotheses and preliminary findings could surface in other users' outputs.
Audit Trail Gaps
Cloud AI services may not integrate with institutional logging systems. When a researcher queries patient data, that interaction may not satisfy HIPAA audit requirements — and you cannot verify what the cloud provider retains.
Cross-Border Data Concerns
US-hosted AI services expose Canadian data to the CLOUD Act, letting US agencies compel access regardless of physical location. This conflicts with PIPEDA, provincial health privacy laws, and institutional data residency requirements.
Top 10 Ways Faraday Helps Medical Researchers
Faraday removes the constraints cloud AI imposes on data handling, query volume, and compliance. Here are ten ways it supports medical research:
- HIPAA-Compliant Literature Review — Search and synthesize literature with frontier AI while keeping your research directions completely private.
- Clinical Trial Data Analysis — Process clinical data on-premises, maintaining chain-of-custody for regulatory submissions without PHI ever leaving your network.
- Grant Proposal Drafting — Generate applications without exposing preliminary research directions to cloud providers.
- Medical Image Pre-Screening — Run vision models on imaging data that stays within your hospital network.
- Patient Record Summarization — Generate longitudinal summaries from EMR data without PHI leaving your infrastructure, at zero per-token cost.
- Drug Interaction Checking — Query real medication lists against AI without sending patient data to third-party APIs.
- IRB Document Preparation — Draft IRB submissions and consent forms locally without exposing proprietary research methodologies.
- Systematic Review Automation — Screen thousands of abstracts at zero marginal cost per query, removing the per-token cost barrier.
- De-identification Validation — Verify PHI removal before sharing datasets, ensuring compliance with HIPAA Safe Harbor and Expert Determination methods.
- Collaborative Model Fine-Tuning — Fine-tune models on institutional data with all weights under your control, retaining full ownership of derived models.
On-Premises vs. Cloud AI
For medical researchers, the choice between cloud and on-premises AI is a compliance, financial, and risk decision with direct implications for patient privacy.
| Consideration | Cloud AI for Medical Research | Faraday On-Premises |
|---|---|---|
| Pricing model | Per-token billing — high-volume analysis becomes prohibitively expensive | $299 / $599 / $999 a month includes hardware, models and support — zero per-token cost |
| Data location | PHI processed on third-party servers outside institutional control | PHI never leaves your network — architectural guarantee, not just a contract |
| Model training | BAA does not prevent model training on your data — opt-out mechanisms are unreliable | No model training on your data — the architecture physically prevents it |
| Audit trails | Audit trails may not satisfy institutional compliance requirements | Full audit trail on your infrastructure — integrates with institutional logging |
| Jurisdiction | Data subject to US CLOUD Act | Canadian data under Canadian law — no CLOUD Act exposure |
Frequently Asked Questions
Is on-premises AI HIPAA compliant for clinical research data?
On-premises AI removes the biggest HIPAA risk of cloud AI, which is sending PHI to an outside vendor. Faraday Machines systems are designed for HIPAA compliance, with end-to-end encryption and audit logging, and all inference runs on hardware inside your institution, so no BAA with an AI provider is needed. Overall compliance still depends on your institution's access controls, policies and IRB requirements.
Can researchers use AI on patient data without breaking IRB data governance rules?
Yes, if the AI runs inside the approved computing environment. With Faraday Machines, the model runs on hardware within your institution, so research data stays inside the environment your IRB approved for the whole research lifecycle. Nothing is sent to a model provider, so prompts and preliminary findings cannot become training data for anyone else's model. See the security and compliance page for details.
How does multi-site medical research work with on-premises AI?
Each participating site runs the same open-weight models on its own Faraday Machines hardware, against its own patient data. Sites share model weights, prompts and methods rather than PHI, so collaborators get consistent AI tooling without transferring records between institutions. Because nothing goes to a US-hosted AI service, this also avoids the CLOUD Act exposure that cloud AI creates for Canadian health data.
Which AI models can analyze medical images on-premises?
Faraday Machines ships three models with image input: Qwen3.8-27B, which natively understands images and video and runs on a single Starter unit; GLM-5.3-Flash, natively multimodal, on the Growth tier as a compressed 1–2-bit build or at 4-bit on Scale; and MiniMax M3, multimodal across text, images and video on the Scale tier. They are general-purpose models, so use them for pre-screening and summarization, not as validated diagnostic tools.
How much does private AI cost for a research lab?
A lab of 2–9 researchers fits Faraday Machines' Starter tier at $299 US a month; groups of 10–19 fit Growth at $599 and 20–29 fit Scale at $799, each on a 36-month plan including the hardware, updates, maintenance and support. There is no per-token charge, so screening thousands of abstracts for a systematic review costs nothing extra. Larger clusters can run bigger models such as GLM-5.3 and Kimi K3.
Secure AI for Medical Research
Run frontier AI models on hardware you control. No PHI leaves your network, no compliance gaps. Schedule a consultation.
Schedule Consultation