800G

Data is measured in bits, the smallest unit of information. A gigabit is a billion such units. 800G is the short form for a connection that transports 800 gigabits per second. For comparison: a good home connection might manage one gigabit per second. An 800G line is therefore roughly eight hundred times faster. Such cables aren’t found in homes, but in large data centers — the halls full of computers where AI services run.

Why AI data centers depend on the connection

A large AI model isn’t trained on a single computer, but on thousands of graphics chips simultaneously. Each of these chips works on part of the task. After every computing step, they must exchange their intermediate results. Only then can they proceed. If the connection is too slow, expensive chips sit idle waiting for data.

This is exactly why network technology has suddenly become a topic on the stock market. A modern AI chip costs tens of thousands of euros. If it spends a third of its time waiting for data, a third of that investment yields nothing. Operators therefore speak of utilization as the decisive metric. Faster connections are cheaper than even more chips.

A common misconception is that 800G has something to do with users' internet access. That is not the case. These connections are usually only a few meters to a few kilometers long and remain within a single data center or between neighboring sites.

Light instead of electricity in the cable

Such volumes of data can barely be transported over copper cables across longer distances anymore. Instead, fiber optics are used: thin strands of glass through which light is sent. A small component at the end of the cable, the so-called transceiver, converts electrical signals into flashes of light and back again. These plug-in modules are the actual product that is talked about in the news.

The 800 gigabits don’t arise from a single signal. Typically there are eight parallel channels, each carrying 100 gigabits per second. You can picture them as eight lanes of a highway. The lanes run either over separate fibers or over a single fiber using light of different colors. At the other end, the data streams are reassembled.

The bottleneck here is less about speed than about heat. Each of these modules consumes power and gives off waste heat, and dozens of them sit in a single rack. That’s why the industry is working on methods to bring the light conversion closer to the chip. Less distance for the electrical signal means less power loss.

800G in quarterly figures and product announcements

You don’t encounter 800G directly in everyday life — you use it without noticing. Every request to a chatbot eventually runs over such connections. The term becomes visible mainly in business news. Manufacturers of network technology and transceivers regularly state in their quarterly reports how many 800G modules they have shipped.

The term is also a kind of timestamp. A few years ago, 400G was considered cutting-edge technology; today 800G is the standard for new AI installations, and 1.6T at 1600 gigabits per second is already on manufacturers' roadmaps. Anyone reading an announcement can roughly gauge from this figure how up-to-date an installation is. That’s why these abbreviations frequently appear in announcements about new data centers.