Rhino’s big idea: let the data stay home

In the fall of 2020, twenty institutions faced a question that was urgent, clinical and inconvenient: could an AI model predict how much oxygen a patient with COVID-19 would need? Each institution possessed part of the answer. None possessed everybody else’s patients. The EXAM study tried a different arrangement. Keep the records in place. Let the model learn across them. Collaboration would require a shared method, without requiring a shared warehouse.

Twenty hospitals, one awkward question

Published in Nature Medicine in September 2021, EXAM supplied the research foundation for Rhino’s commercial idea. Co-founder Ittai Dayan had led AI development and deployment at Mass General Brigham; fellow co-founder Yuval Baror brought experience building production AI systems. Their company, initially known as Rhino Health, launched publicly in 2021. The experiment offered a persuasive starting point: institutional boundaries need not define the boundaries of a model’s education.

There is a familiar trap here. A model learns the habits of one hospital and appears splendidly competent. Then another hospital arrives with different scanners, coding practices or patients. Competence turns out to have a postcode. More varied data can help researchers test that problem, but acquiring it is a project in its own right. Privacy, ownership and security are substantial obligations. The hospital’s reluctance is part of the design brief.

The algorithm was only the beginning

Rhino’s own account describes an instructive revision. It set out to build an enterprise federated-learning solution, then found collaborators needed more than training. Records needed preparation. Researchers needed statistics. Models needed validation and deployment. A consortium needed to work again next Tuesday. The company widened its scope to federated computing: running approved workloads across separate data environments and bringing back permitted results.

Rhino FCP supplies a web interface, Python SDK and APIs to coordinate that work across cloud and on-premises systems. Local clients execute workloads; an orchestration layer organizes the collaboration. Permissions, project isolation and audit trails govern access. Privacy techniques can constrain outputs. Keeping raw data local does not mean no information moves: model updates, statistical aggregates and approved results are the point of the exchange.

The quieter product is data harmonization. At Sheba Medical Center’s ARC Innovation, Rhino’s Harmonization Copilot addresses clinical information that arrives in incompatible forms. Its tools map data toward standards such as FHIR and OMOP. Current offerings, including Data Activation and the Data Harmonization Engine, emphasize human review and maintaining pipelines as source systems change. Two hospitals can agree to collaborate and still disagree on what a column means. Software has to attend to both problems.

A drugmaker opens the door, carefully

Lilly TuneLab makes the arrangement tangible. Announced in September 2025, it lets biotech partners use Lilly’s drug-discovery models on their own proprietary data. Rhino provides backend infrastructure, deploying NVIDIA FLARE within a multi-cloud architecture. Partners can run inference and contribute to model improvement through federated fine-tuning. The valuable exchange involves access and learning, while the underlying datasets retain their separate owners.

“AI only moves the needle when it meets real data, real constraints and real partners.”Ittai Dayan, on Lilly TuneLab, September 2025

Rhino also serves the FAITE Consortium, whose founding members include AbbVie, Amgen, AstraZeneca, Johnson & Johnson and UCB. Its task involves predictive models for biologics using federated and active learning. These are sophisticated organizations with existing technical teams. Rhino’s place in the market is the infrastructure between them: deployment, controls and continuing operations that make a shared program possible.

A promising trial is still a trial

Banking presents the same institutional puzzle with different consequences. A fraudulent payment can cross several banks; each bank sees only its own slice. Google Cloud’s December 2024 description of Swift’s collaboration assigns Rhino the core federated-learning platform and Capgemini the implementation and integration. The model visits participating institutions, learns locally and contributes updates to an aggregated model.

Swift reported that result after experiments involving thirteen financial institutions. It planned a second phase using real transaction data. The distinction matters. An artificial transaction set can demonstrate that collaboration is technically useful. It cannot, by itself, establish how much money a bank will save in production. Rhino’s opportunity is credible without borrowing certainty from an experiment that has further work to do.

Buying the machinery of cooperation

This is enterprise software sold through subscriptions and negotiated contracts, with implementation support and forward-deployed engineers. Rhino’s $15 million Series A, announced in May 2025 and led by AlleyCorp, brought reported funding above $30 million. Buyers also have a separate bill to consider: local or cloud compute, integration and ongoing operations. AWS Marketplace lists a $999,999 annual “Units” dimension without defining its scope. Treat that as a contract listing, not a typical project budget.

The alternatives depend on the job. A capable team can assemble a federation around open-source frameworks; Flower and NVIDIA FLARE are also part of Rhino’s ecosystem. Centralizing data can be practical when everyone has permission to do so. Rhino’s proposition becomes more relevant when institutions must retain control and the collaboration needs repeated, governed execution. Distributed infrastructure still demands compatible definitions, functioning local environments and people willing to cooperate.

The lesson worth copying is modest: pick a shared question, agree on definitions, specify which outputs may travel, then test on another institution’s data. Rhino’s ELISE case study describes Emory researchers validating a breast-cancer algorithm against 500 locally held cases in Israel. That is a concrete way to discover whether an algorithm’s apparent competence survives a change of address.

In June 2026, Rhino added an MCP interface for conversational access to platform tools. In August it announced Microsoft Marketplace availability alongside AWS and Google Cloud. Its October 19-20 Boston summit is due to examine the journey from pilots to production. The next question is less theatrical than inventing an algorithm: can separate institutions make the arrangement work, safely and repeatedly, on an ordinary working day?