Colossus 1

Colossus 1 is the supercomputer that trains Grok, the AI behind xAI's chatbot. It is one huge machine, housed in a roughly 785,000-square-foot former Electrolux appliance plant in Memphis, Tennessee, and it came online in September 2024 with 100,000 of NVIDIA's flagship H100 graphics chips. A graphics processing unit (GPU) is the specialized chip that does the heavy parallel math behind modern AI, and the first phase was built, start to finish, in 122 days. Elon called it the most powerful AI training system in the world.

What is Colossus 1?

Colossus 1 is an AI training cluster, which means tens of thousands of GPUs wired together so they work as one enormous computer. It has one job, and that is to train the Grok family of large language models, the AI systems that read and write text inside xAI's assistant. It has since grown into the anchor of a bigger fleet that includes a second site, Colossus 2.

Why Colossus 1 Exists

xAI joined an AI race that was already crowded with well-funded rivals, and large language models mostly get better by training on more data, across more chips, for longer. To close that gap quickly, xAI needed a very large cluster, and it needed one fast. Building a data center from scratch can take years, but taking over a shuttered factory that already had heavy industrial power and a huge open floor let the team get the first 100,000 GPUs running within months.

How it Works

Colossus 1 works less like one machine and more like a tightly woven web of GPUs that all have to think as one brain. Each server moves data at 3.6 terabits per second over NVIDIA Spectrum-X Ethernet, a standard networking approach xAI chose over more exotic options, and the full setup reaches a combined memory bandwidth of 194 petabytes per second, with storage past one exabyte, which is a billion gigabytes. Feeding all of that power is the hard part, because training loads swing hard as the chips spike and drop together. So xAI paired the site with Tesla Megapack batteries to smooth out the draw.

The local grid had to grow to keep up. Power approved for the site from the Tennessee Valley Authority, delivered through Memphis Light, Gas and Water, started at 150 megawatts in November 2024 and doubled to 300 megawatts once a second approval cleared in February 2026.

The mix of chips changed as fast as the building filled up. When Elon laid out the jump to 200,000 GPUs, he was specific about it: the original H100s plus a batch of newer H200s, with later expansions adding GB200 Blackwell-class hardware on top.

The Economics of Colossus 1

Each new generation of chip costs more and does more work per dollar, and a cluster this size starts losing value the moment it turns on. So in May 2026 xAI opened Colossus 1, by then more than 220,000 mixed H100, H200, and GB200 GPUs, to Anthropic, the maker of Claude, adding capacity for Claude Pro and Claude Max users. The older cluster became steady income while the newest training work moved to the Blackwell-based Colossus 2, and the same chips that train Grok now also serve a direct competitor. In June 2026 Google agreed to pay SpaceX about $920 million a month, from October 2026 to June 2029, for the use of about 110,000 NVIDIA GPUs, and by the end of June SpaceX had signed $14.1 billion of cloud-services agreements. The full build runs into tens of billions of dollars.

Current Status of Colossus 1

Colossus 1 opened at 100,000 GPUs and roughly doubled within months. Elon Musk said in late September 2026 that it held about 230,000 GPUs (150,000 H100s, 50,000 H200s, and 30,000 GB200s) and Colossus 2 about 550,000, about 780,000 in all, with a stated plan to push toward 1 million. Colossus 1 draws 300 megawatts of approved grid power, and SpaceX's quarterly results put SpaceX's total AI computing power at 1.7 gigawatts on September 30, 2026.

What Comes Next

Colossus 2 is where Grok 5 is being trained, on NVIDIA's Blackwell GB200 and GB300 chips. On January 8, 2026, xAI announced a third data center in Southaven, Mississippi, a more than $20 billion investment named MACROHARDRR after MACROHARD, which Elon said would push Colossus training compute to about 2 gigawatts. The Anthropic deal also mentioned interest in multi-gigawatt orbital compute, meaning data centers in space.

The Bottom Line

Colossus 1 turned an empty appliance plant into xAI's first training cluster. By 2026 Anthropic was renting it, and the newest Grok training had moved to Colossus 2.