AI architecture: AI architecture: cloud, private cloud, on-premise, hybrid – and the decoupled application layer.
How an AI architecture can be structured: applications, shared data core, control, execution, model layer, compute nodes and operating model.
In short: A layered AI architecture separates application, control, model layer and infrastructure. Compute nodes, known as nodes, run the models. The separation can make it easier to switch models and infrastructure.
Layers instead of a monolith.
Binding directly to one model can make a later switch harder. Separate customer and project data can lead to duplication. A layered architecture separates applications, shared data core, control, execution, model layer, compute nodes and operating model.
The architecture is designed to largely decouple models and infrastructure from the application layer and to scale horizontally with additional nodes – without absolute guarantees, but with clear interfaces.
Four models, one architecture.
| Operating model | Models run | Data | Typical for |
|---|---|---|---|
| Cloud | with the provider | with the provider | non-critical tasks, getting started |
| Private cloud (CH) | dedicated in a Swiss data centre | there | Control without your own hardware |
| On-premise | on your own premises | on your own premises | regulation, trade secrets |
| Hybrid | per data class | per data class | most companies |
Questions to ask of any architecture
- Can I switch the model without changing the application?
- Can I grow by adding nodes?
- Do applications share customers, projects and users?
- Does the data class decide where a task runs?
- Are there approval points and monitoring?
What this article is based on.
- Own public architecture description of the Swiss AI Platform and own project experience at Swiss AI Labs; no external study
Products, Labs and case studies.
Frequently asked questions: AI architecture
Is this only for large companies?
No. Even a single product can use separate layers. Costs depend on the product and operating model.
Do you show your internal architecture?
Publicly, we show the principle and layers; internal topology and rules remain internal.
Have a specific process assessed.
The AI potential assessment examines the benefits, prerequisites and risks of the process.