Networking chips are the main chips inside network equipment: the switches and routers of a data centre, the routers of an office, and the radio units on a mobile tower. A chip is a small, flat piece of silicon that carries billions of transistors, microscopic electrical switches. A networking chip is built for one job. It receives data, reads where the data must go, and sends it out again, billions of times a second.
Data on a network travels in small pieces called packets. Each packet carries the address of the computer it is going to.
Take, for example, a switch in a data centre in Mumbai. A switch is a box that joins many servers to the network. Inside it is one switch chip with 64 ports, and each port carries 400 gigabits per second. That is 25.6 terabits per second in total. A packet arrives on port 12 from a server. The chip reads the packet's destination address, looks it up in a table stored on the chip, and sends the packet out of port 40, towards another server. This takes less than a microsecond, a millionth of a second. The chip does the same for billions of packets every second, on all 64 ports at once. A general-purpose processor could not do this fast enough. So the switch chip does it in fixed circuits, designed only for this task.
The switch also has a small CPU of its own. It runs the switch's operating system, which fills the chip's address tables and reports problems. In white-box networking, a data centre buys the switch hardware from one company and the operating system from another. This is possible because the same few networking chips sit inside switches from many makers.
Who buys networking chips? Not the data centre. The buyers are the companies that make switches, routers and tower equipment. Getting a chip chosen as the part inside another company's product is called a design win. Once a chip is chosen, it usually ships in every unit for as long as that product is sold. So a few large chip companies make most of the world's networking chips. They are fabless: they design the chips, and a foundry, mostly in Taiwan, makes them. Much of the design and testing is done in Hyderabad, Bengaluru and Pune.
AI has made these chips more important. In AI cluster networking, thousands of GPUs train one model and exchange results many times a second. If the network is slow, every GPU waits. So switch chips for AI have more ports, faster ports and shorter delays. On a mobile tower, a different family of networking chips processes the radio signals inside the radio units. Inside a server, the links between one chip and the next are handled by chip interconnect.