Derik Hammer explains how cloud data warehouses differ from their on-prem cousins:
Given the data processing needs of a data warehouse, they tend to be implemented on massively parallel processing (MPP) systems. The MPP architecture replies upon a shared nothing concept for distributing data across various slices. Compute nodes are layered on top of the storage and processes queries for data residing in its local slice. The control node is responsible for taking a query and dividing it up into smaller queries to be run in parallel on the compute nodes.
Read the whole thing.