Every organization building with AI eventually runs into the same limit. The data that would make its models and agents most useful sits with partners, in other business units, or across borders. Moving it means months of approvals, new security exposure, and copies that someone has to govern.
Federated computing solves this by sending the code to wherever the data lives. Each organization runs approved computations inside its own environment and shares only the results, so sensitive records never leave the site.
The approach grew out of federated learning, and it now covers everything from simple patient counts to AI agents working across sites. In this post, we explain how federated computing works, how it relates to federated learning and edge computing, why regulation and the rise of AI agents have made it more pressing, and how it compares to other ways of working with sensitive data.
What Is Federated Computing?
Federated computing is a way to analyze data, train models, and run AI across multiple organizations without moving the underlying data. Each participant runs approved code inside its own environment, and only aggregated results go back to the team or system coordinating the work. The approach is also known as compute-to-data, because the code travels and the data stays where it is.
That covers a wide range of work. A federated query can count eligible patients across ten hospitals, while a transformation converts lab values to standard units at every site before anyone compares them. Federated training fits a single machine learning model to cohorts held by different institutions, and a federated deployment can even run one partner’s software on another partner’s data, behind that partner’s firewall.
Throughout, data custodians keep the final say. Their existing security measures stay in place, they approve which code runs against their data, and they can revoke a collaborator’s access at any time. For another layer of protection, teams can apply privacy-enhancing technologies such as differential privacy and k-anonymization, which limit what any single result can reveal about an individual.
How Does Federated Computing Relate to Federated Learning and Edge Computing?
Federated computing overlaps with several terms that people often use interchangeably. Each one describes a different piece of the same approach, and knowing where they differ makes it easier to scope a project or compare platforms.
- Federated learning (FL): Federated learning trains one machine learning model across data held at different sites. Each site trains locally and shares only model updates, such as weights or gradients, which are combined into a global model. FL was the starting point for the field and remains one of its most common workloads.
- Federated statistics: Federated statistics runs analyses such as descriptive statistics, regressions, and survival analyses at each site, then combines the results. Many research and real-world evidence questions call for statistics rather than a trained model, so this workload often comes first.
- Edge computing: Edge computing runs applications close to where data is generated. The term usually comes up in latency-sensitive settings like connected devices, and it also describes how federated computing works, since every computation runs at the edge inside each participant’s environment.
- Distributed computing: Distributed computing splits a large problem across many machines to get more processing power. Federated computing distributes work too, though its goal is keeping each organization’s data under that organization’s control.
In practice, most federated projects combine several of these. A single study might harmonize data at each site, run federated statistics to confirm cohort sizes, and then train a model with federated learning, all without the data leaving any site.
Why Is Federated Computing So Important Now?
Federated computing first took hold in academic research consortia. Since then, tighter regulation, rising breach costs, and the spread of AI agents have made it relevant to any enterprise that works with sensitive data.
Regulators are moving in the same direction. In Europe, the European Health Data Space regulation entered into force in March 2025, and its rules for secondary use of health data apply from March 2029 for most data categories and March 2031 for the rest, including genetic data, according to the European Commission. Approved users will work with that data inside secure processing environments, and the regulation states that only non-personal data should be downloaded from them.
The US is tightening its own safeguards. HHS has proposed updates to the HIPAA Security Rule that would require encryption of electronic protected health information at rest and in transit, along with multifactor authentication and network segmentation. Each extra copy of patient data an organization creates adds one more system those safeguards have to cover.
These rules respond to a real cost. Healthcare remained the most expensive industry for data breaches in 2025, at an average of $7.42 million per incident, according to IBM’s Cost of a Data Breach Report. The 2024 Change Healthcare attack affected about 192.7 million people, according to HHS. The more data an organization pools in one place, the more a single breach can expose. Federated computing reduces that exposure, because records stay with the institutions that collected them and no central copy exists for an attacker to reach.
Adoption has kept pace. Published federated learning research has increased more than 13-fold since 2020, and healthcare and life sciences is its fastest-growing application, according to Rhino’s analysis of the OpenAlex research database.
AI agents raise the stakes again. Gartner predicted that 40% of enterprise applications would include task-specific AI agents by the end of 2026, up from less than 5% in 2025. Security controls are lagging behind. In the same IBM report, 13% of organizations reported breaches of AI models or applications, and 97% of those lacked AI access controls. Agents need data to be useful, so how they reach sensitive data safely has become a central question in enterprise AI planning.
How Can AI Agents Work With Sensitive Data That Can't Move?
In regulated industries, the most valuable data is also the data least likely to be handed to an autonomous system. Security teams won’t approve agent access until every action is scoped and reviewable.
Federated computing gives agents a governed way in. The agent sends queries to each site, those queries run inside that site’s environment, and the agent gets back only counts and aggregate results. The underlying records stay put, and each data owner decides what the agent can run.
That changes how quickly teams get answers. A trial feasibility question that once took weeks of data requests can now draw on everything an organization already holds across functions, regions, and licensed sources, with every query running under its data owner’s policies.
How Does Federated Computing Compare to Other Secure Data Sharing Approaches?
Federated computing is one of several ways to work with sensitive data, and each fits some situations better than others. The three you’re most likely to weigh it against are centralization, data clean rooms, and trusted execution environments.
- Centralizing data: Pooling data in one warehouse or lake remains the most common pattern, and it works well within a single organization’s governance. Across organizations, it raises re-identification risk, adds egress and duplicate storage costs, and creates a copy that starts going stale the moment it transfers. Data transfer agreements also slow legal review.
- Data clean rooms: Clean rooms are familiar and quick to set up, which makes them a common choice for joint analysis. In many of them, each party’s data moves into someone else’s environment, and supported workloads lean toward queries and reporting more than model development.
- Trusted execution environments (TEEs): TEEs isolate computation in hardware, so nothing outside the enclave can see the data while it is in use. When the TEE sits in a cloud or partner environment, the data still has to travel there, and the security guarantee depends on the chip maker’s implementation. TEEs and federated computing also work well together. Projects on the Rhino FCP can require that every code run executes inside a TEE, with confidential computing supported on AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure.
Federated computing has its own setup costs. Each site installs software inside its own environment, and data has to be harmonized to a common model before results line up across sites.
The Rhino FCP shortens both steps. Deployment typically takes two to four weeks, mostly for security review and hardware provisioning, and the software installs in under 10 minutes. The Data Harmonization Engine then proposes schema mappings for data teams to review and approve, which cuts harmonization from months to days. For a closer look at these approaches and others, including de-identification, synthetic data, and multi-party computation, check out our plain-English guide to secure data sharing.
How Does Rhino’s Platform Deliver Federated Computing?
Rhino’s approach started with federated learning. Co-founder and CEO Dr. Ittai Dayan led the EXAM Study, published in Nature Medicine in 2021, which trained a COVID-19 outcomes model on data from 20 institutions worldwide without pooling patient records. The resulting model improved AUC by 16% on average and generalizability by 38% compared with models trained at a single site. Bringing that work into production showed that consortia needed more than model training. They also had to query, clean, harmonize, and analyze data across sites, so Rhino built a complete federated computing platform to support all of it.
The Rhino Federated Computing Platform (Rhino FCP) runs inside each participant’s cloud or on-premises environment, so raw data stays behind that participant’s firewall. Three solutions built on the Rhino FCP address the problems customers bring to us most often.
- Rhino Data Activation readies an organization's own data for AI and enterprise-wide use. The Data Harmonization Engine maps local schemas and vocabularies to a common data model, so a health system can put data from every department and site to work without centralizing it.
- Rhino Federated Intelligence lets organizations build and run federated networks with their partners. The consortium owns its network on the Rhino FCP, and each member keeps control of its data while contributing to shared analyses and models.
- Rhino Private Agentic AI lets teams put AI to work on data they can't move or can't expose. Agents connect through Rhino's MCP integration and run their queries inside each site's environment, under governance that each data owner controls.
Across all three solutions, the Rhino FCP adds the controls and tooling that make federated work practical at enterprise scale.
- Layered security: The Rhino Client needs no inbound firewall ports and communicates only outbound over TLS. Data stays encrypted at rest, in transit, and in use, with options for confidential computing, encrypted model weights and code, and differential privacy.
- Fine-grained governance: Project- and site-level permission policies, field-level permissions, and hidden collaborators let each site decide exactly what others can see and run, and a full audit log records every significant action.
- Rhino MCP Server: Any MCP-compatible AI assistant or agent can run federated analytics, harmonize data, and monitor jobs through natural language with the Rhino MCP. Results come back aggregated, and each user sees only the projects they’re already authorized to access.
- Interactive Containers: Teams can run any Linux-based application, such as Jupyter Notebooks or an annotation tool, in an interactive session at the data custodian’s site. That gives organizations an alternative to standing up a dedicated trusted research environment (TRE).
- Autoscaling and VM Pools: The Rhino FCP provisions compute sized to each workload and spins it down when the job finishes. For GPU-heavy training and inference, VM Pools reserve capacity in advance, so jobs start quickly and queue instead of failing when a pool is full.
- Workflows and Trusted Integrations: Workflows chain data imports, code runs, and conditional steps into one repeatable pipeline, and Trusted Integrations connect Rhino Clients to external tools such as AI models and data repositories.
The Rhino FCP already runs in 14 of the top 20 hospitals on Newsweek’s Best Smart Hospitals list. In one deployment, a global pharmaceutical company uses it to give more than 140 biotech partners access to 30+ drug discovery models through federated learning. Those partners have run more than 3 million inference jobs so far, while each one’s proprietary data stays in its own environment.
Put the Data You Can't Move to Work
Your most valuable data shouldn’t sit idle just because it can’t leave its source. Federated computing lets you activate sensitive data for AI, analytics, and agents where it lives, on terms each data owner sets. Organizations working under the strictest privacy and security requirements already use it to move AI from pilot to production.
Want to see how it would work with your data? Request a demo and our team will walk you through it.
Frequently Asked Questions
Is Federated Learning the Same as Federated Computing?
No. Federated learning is one type of federated computing, focused on training machine learning models across distributed data. Federated computing covers that and every other workload that runs where the data lives, from queries and statistics to data transformation, custom code, and AI agents.
Does Raw Data Ever Leave the Site?
No. All code runs inside each site’s own environment, and only the results that site approves, such as aggregate statistics or model weights, ever leave it. Sites can also apply differential privacy or minimum cell counts to limit what those results reveal.
Can AI Agents Run on Federated Data?
Yes. An agent can query every site in a federated network and get back aggregate results while the underlying records stay in place. With Rhino Private Agentic AI, each data owner decides what an agent is allowed to run.
Does Federated Computing Help With HIPAA and EHDS Compliance?
Yes. Federated computing makes compliance easier because regulated data moves and gets copied far less often, which shrinks the number of systems in scope. Each organization’s own policies and controls still determine compliance. The Rhino FCP meets HIPAA, GDPR, ISO 27001, and SOC 2 Type II requirements, with details in the Trust Center.
