The question: where is the data processed?
In the cloud, the AI runs on someone else's servers – the provider's, or in your own cloud account. The data leaves your building for that. On-premise ("on site") means: the model runs on your own hardware, the data stays physically with you. That's the core – not performance, but the place of processing, and with it control.
It's not an either/or
Between the extremes there are gradations: public cloud (shared, provider-managed infrastructure), private cloud (dedicated but still remote servers), on-premise (your own hardware in-house), and hybrid combinations. The choice is a slider between convenience and control, not a switch.
What that means in practice
Data sovereignty
On-premise gives maximum control – the data never leaves your own network (relevant under strict requirements, see modules 7 and 20). Cloud means trusting the provider, safeguarded via contracts (DPA) and server location.
Cost
Cloud means pay-as-you-go, no hardware of your own. On-premise means a high upfront investment in hardware plus ongoing maintenance – it pays off more with high, steady utilization.
Effort & scaling
In the cloud, the provider handles operations and maintenance, and capacity can be scaled up and down flexibly. On-premise you have to run yourself, and capacity is limited by the hardware you have.
Why this matters for you as a decision-maker
For most companies, the cloud is the pragmatic default: no hardware, flexibly scalable, low operational effort. On-premise or private cloud becomes relevant with strict privacy/regulatory requirements, with very high steady utilization – or when you want to run an open model yourself anyway (module 26). Here too a hybrid is often the answer: most in the cloud, the sensitive part in-house.