Colossus

Colossus is a gigantic computing facility in the US state of Tennessee, more precisely in the city of Memphis. It is operated by xAI, Elon Musk’s AI company. Inside the hall stand hundreds of thousands of specialized chips, all working together on the same task. That task is training: a computer program reads huge amounts of text and learns from it which words typically follow one another. The result of such computing runs are the chat programs of the Grok series offered by xAI. The name Colossus comes from Greek and simply means giant.

Why xAI had to build its own facility

Modern AI systems are not created on ordinary computers. A single training run can take weeks and consume as much electricity as a small city. Anyone who wants to build such models therefore needs access to enormous amounts of computing power in one place. Without that access, it is simply impossible to keep up in development.

xAI started very late in 2023 compared to competitors such as OpenAI or Google. To close the gap, the company focused on building fast. According to its own statements, the first expansion stage of Colossus was built in about 122 days. For facilities of this size, two years is more typical. This speed was widely discussed in the industry and is one reason why the name became so well known.

Colossus also stands as a symbol of a broader trend. Data centers have become the decisive competitive factor in the AI industry. When investors assess the value of AI companies, they therefore also look at how many chips a company owns.

What’s inside the facility

The core components are graphics processing units, or GPUs for short. These chips were originally developed for video games because they can perform very many simple calculations simultaneously. That is exactly what is needed for training AI as well. Colossus started with around 100,000 such chips from manufacturer Nvidia and has since been expanded multiple times. A second facility called Colossus 2 is under construction.

For the chips to work together on a single model, they must communicate with each other extremely fast. To achieve this, they are connected via special network cables that transfer significantly more data than a normal internet cable. You can imagine it like a team working on a single, long calculation. If the connection is too slow, the fastest members are constantly waiting for the others.

The biggest practical problem is energy. Colossus requires several hundred megawatts of power, roughly the scale of a medium-sized power plant. Because the public power grid could not supply that immediately, xAI additionally installed gas turbines on the site. This drew protests from local residents and environmental groups, since the turbines emit pollutants. Another issue is cooling: the chips generate so much heat that air is no longer sufficient, so water is routed directly through the racks.

Colossus in the news

You cannot use Colossus directly. However, you encounter the facility indirectly as soon as you chat with Grok, xAI’s chat program. The image and video features of these models were also created on these machines. What is computed in Memphis therefore ends up in the app on your phone.

In business news, the name usually appears in two contexts. First, in funding rounds: xAI regularly raises billions of dollars, and a large portion of that flows directly into chips and electricity. Second, in connection with Nvidia, whose revenue depends heavily on such large-scale buyers. Colossus is often compared to similar projects, such as OpenAI’s Stargate initiative.

A common misconception is that Colossus is a single supercomputer like in old movies. In fact, it consists of a great many individual servers in racks working together as a network. The technical term for this is a cluster. Failing components are swapped out during ongoing operation, without the facility coming to a stop.