Data centre networking is the set of high-capacity switches and routers that join the thousands of servers in a data centre to each other and to the outside world. Every request that reaches an app, and every piece of data that one server sends to another, crosses this network.
Most of the traffic in a data centre never leaves the building. Consider a single search on a shopping site. The web server asks a search server, the search server asks a database server, and the database server reads from a storage system. Each of these messages travels between servers inside the building. So a data centre network is designed above all for traffic between servers, with very high bandwidth and very low delay. The page on bandwidth and latency explains these two measures.
The main products here are data centre switches and routers. Switches join the servers in each rack, and join the racks to each other. Routers connect the building to telecom operators and to other data centres.
A buyer can get a switch in three ways. Most buyers buy a branded switch, with the maker's hardware and software sold as one package. Some of the largest buyers use plain hardware and write their own software. Others buy white-box networking: plain switch hardware with a network operating system from another company, such as the open-source SONiC. This separation of software from hardware is called hardware-software disaggregation.
AI has created a network of its own. An AI cluster joins thousands of GPUs, which exchange results after every step of training. AI cluster networking does this with InfiniBand, a specialist technology, or with Ethernet at much higher speeds.
The network is built with redundancy. In most company data centres, every server connects to two switches, and every switch has several paths to the rest of the network. So one failed switch or cable does not cut any server off. The largest cloud companies often connect each server to only one switch, and move its work to servers in other racks if that switch fails.