Colossus 2 Cluster

Colossus 2 is a massive computing facility belonging to xAI, the company owned by Elon Musk. It is located in the city of Memphis in the U.S. state of Tennessee. In such facilities, a very large number of computers are housed in halls and connected together to form a single giant machine. Experts call this kind of arrangement a cluster. Colossus 2 is the successor to the first facility, Colossus, which xAI built in the same city in 2024. Its task is to train the company’s AI models, that is, to let them learn from huge amounts of text.

Why a single hall determines the lead

Modern language models don’t get smarter because someone teaches them better rules. They get smarter mainly because more computing power and more data are put into them. This observation is known as the scaling law. Whoever owns the largest computing facility can therefore train a more powerful model in the same amount of time than the competition.

That is why Colossus 2 is also an economic signal. Facilities of this size cost several billion dollars, mainly because of the chips. With it, xAI bought its way into contention with competitors such as OpenAI, Google, and Anthropic, who operate similar facilities. For investors, the number of chips is one of the few tangible metrics in the AI race.

The second reason for the attention is power. Colossus 2 permanently requires an output in the range of several hundred megawatts, equivalent to the consumption of a medium-sized city. xAI therefore installed gas turbines on the site because the public grid could not supply power quickly enough. Residents and environmental groups in Memphis complained about the emissions, and this turned into a legal dispute. AI data centers have thus become an issue in energy and environmental policy.

What stands in the halls of Memphis

The actual computing work is done by graphics processing units, or GPUs for short. These are chips that were originally developed for video games and can carry out a very large number of simple calculations simultaneously. That is exactly what an AI model needs. Colossus 2 contains chips made by manufacturer Nvidia, and an expansion toward one million GPUs has been announced. Eight of them at a time sit in a flat enclosure, several enclosures stand in a cabinet, and a cabinet is called a rack.

The wiring is crucial. All chips work on the same model and must constantly exchange their intermediate results. For this, extremely fast networks are used, in which each chip is directly connected to many others. If this network is too slow, expensive chips just end up waiting for data. You can imagine it like a huge construction site: a thousand workers only help if materials and coordination flow fast enough.

Then there is the cooling. Each chip releases almost all of its energy as heat. Air cooling is no longer sufficient at this density, so cooling liquid runs directly past the chips. A common misconception is that such a facility is constantly computing answers for users. Training is the main purpose; answering individual requests, so-called inference, often runs on different machines.

Colossus 2 in headlines and in its own chat app

The name most often appears in business news. Reports about new funding rounds for xAI, about large orders placed with Nvidia, or about permits in Memphis almost always mention the facility. The share prices of chip manufacturers and energy suppliers also react to such news.

Indirectly, one encounters Colossus 2 every time one uses the chatbot Grok, which is built into the platform X. This model’s capabilities come from training carried out in these halls. When a new version of Grok is released, the computing facility is the reason it can do more than the previous one.

It is useful to distinguish this from a normal data center, such as that of a cloud provider. There, thousands of mutually independent programs run for many customers. A training cluster like Colossus 2, by contrast, behaves like a single, very large computer with a single task.