Musk Says Colossus 2 Will Add 660,000 Nvidia GB300 GPUs by Year-End, Taking xAI's Fleet to 1.44 Million Chips

Elon Musk has laid out the most aggressive compute timeline in the industry: Colossus 2, xAI's Memphis supercomputer, will add 660,000 Nvidia GB300 GPUs by the end of December. In a series of posts on X on September 25, Musk said the site currently operates 110,000 GB200 and 440,000 GB300 accelerators — about 550,000 chips in total — and that the new hardware would more than double that count within three months.

The expansion is scheduled in three tranches of roughly 220,000 GB300 GPUs each: one by the end of this week, another by the end of November, and the last by the end of December. If every tranche lands on time, Colossus 2 would hold about 1.21 million chips on its own, and xAI's combined fleet — including the roughly 230,000 accelerators already running at the original Colossus 1 site — would reach approximately 1.44 million.

The hardware bet is firmly on Nvidia's latest Blackwell generation. Musk described Colossus 2 as a GB300-first buildout: the GB300 is the more advanced evolution of the GB200, pairing Nvidia's Grace CPU family with the Blackwell Ultra GPU variant. Colossus 1, by contrast, was built primarily on the earlier Hopper generation.

Feeding that many chips requires power on the scale of a small city. Musk's operation is building a dedicated 1.2-gigawatt power plant to bring the new capacity fully online — roughly the output of a large conventional generating station. Grid interconnects in the United States can take years to approve, which is why the largest AI operators increasingly generate their own electricity behind the meter, a choice that carries its own questions about emissions and local resource use.

Musk did not sugarcoat the execution risk. "Activating a huge volume of compute quickly is very difficult," he wrote, while arguing that xAI — about three years old, by his count, versus six for Anthropic and ten for OpenAI — is scaling faster than its rivals. He claimed that if the company's growth rate holds, it will be "number one in about six months," and repeated his assertion that xAI can reach GPT-6-level AI within two to three months.

Skeptics have reason to wait for the milestones rather than the announcements. Chip deliveries slip, transformers and switchgear are back-ordered, and power projects face permitting delays, so the 1.44 million figure is best read as a ceiling rather than a promise. Rivals are not standing still either: Nvidia has said it will roughly double its own chip output and still expects to fall short of demand, while OpenAI, Meta and Google are all expanding frontier clusters in parallel.

What the race means in practice is that compute has become the currency of AI leadership — and for developers and businesses, the benefits arrive indirectly, through better or cheaper models rented over an API. If Colossus 2's tranches do come online by December, the immediate test is whether that scale shows up as capability gains at xAI's next release, not just as a bigger number on a slide.