How a federated learning project works, and how the costs are structured
A Mesh project has five stages: framing, agreeing on a common feature vocabulary, a pilot with at least three actors, measurement, then extension. Billing rests on a one-time setup and an annual licence per member, the structure is public, the amount is discussed case by case. One point to note: forming a consortium usually takes 6 to 12 months. That is our sales cycle, not an observed delivery lead time.
The stages of a project
We describe here a proposed method, not a sequence timed across dozens of projects. The order of the stages is stable; their duration depends on the sector, the number of actors and the data maturity of each.
- Framing. We define the use case (a detection, a forecast), the candidate actors, and what "better" will mean: the success metric, decided before starting.
- Agreeing on a common feature vocabulary. Horizontal federated learning requires each actor to describe its data in the same variable space: same columns, same units, same conventions. This is not the same sector or the same database schema: it is a shared measurement reference, built together. It is often the most structuring stage.
- A pilot with at least three actors. The value of a collective model only emerges from three participants onward. Below that, there is no collective to demonstrate. Each one trains locally, on its own data, which never leaves.
- Measurement. On the participants' real data, we compare the collective's model to the best model an actor gets alone. The gain is observed at this stage; it is never promised upfront.
- Extension. If the pilot convinces, we widen it: more actors, more use cases, continuous training. Each member can then use the collective's engine and, if it receives it, keep training it privately on its own unshared data.
The pilot, the heart of the proof
The pilot is where assumptions are replaced by figures. Two things are checked: that the common vocabulary holds on real data, and that the collective's model brings a measurable gain over the best standalone model. That gain is read on a metric decided at framing, often the AUC for a detection problem, and it is measured on each participant's data, never on a demo set.
Our public scenarios illustrate the mechanism: in a maintenance demonstration, the network's model catches faults a lone actor missed. That is a demonstration, not a customer result: the only measurement that commits is your pilot's. To size the value side, see the ROI calculator.
How the costs are structured
The structure of billing is public; the amount is discussed case by case, depending on scope.
- A one-time setup. It covers the upfront engineering: the use case, the common vocabulary, installing the nodes and exchanging the encryption keys at each member.
- An annual licence per member. This is the recurring core: right to use the resulting model, coordination, maintenance. The metric is the member, because value only emerges from three actors onward.
We publish no amount on this page: a figure out of context misleads. The price is compared to the estimated value for your organization (that is what the ROI calculator is for), and it is set in conversation. Let's talk.
How long? The sales cycle, not the delivery lead time
The only duration we state is the bootstrap: bringing together at least three actors ready to co-train a model usually takes 6 to 12 months. That is the time to convince, align leaderships and frame the common vocabulary: a sales cycle, not a "delivery lead time". We show no deployment duration as if it had been observed across many projects: that would be a promise nothing supports today.
FAQ: Process and costs
How many actors are needed to start a project?
At least three. Below three participants, there is no collective model to train or demonstrate: the value of federated learning comes from pooling learning across several actors.
How is billing structured?
A one-time setup, then an annual licence per member, from three actors onward. The structure is public; the amount is discussed case by case and compared against the value estimated for your organization.
How long does a project take?
The actual duration is the bootstrap: bringing together at least three actors usually takes 6 to 12 months. That is a sales cycle, not a delivery lead time. We do not present any deployment duration as an observed fact.
Do actors need an identical data schema?
They need a common feature vocabulary: same variables, same units, same measurement conventions. It is neither the same sector nor the same database: it is a shared measurement reference, built together at framing.
Let's frame your use case
We look together at the scope, the common vocabulary, and how a pilot would be measured on your data.