Elon Musk laid out the next act of his computing buildup in a few spare sentences on X this week: the Colossus 2 data center outside Memphis, already the largest artificial-intelligence computer on Earth, would nearly double its Nvidia chips before the end of the year.
The facility now runs 110,000 Nvidia GB200 chips alongside 440,000 of the newer GB300 chips, according to Musk’s post. Three more waves would follow. A first batch of 220,000 GB300s was due online within a week, a second batch of 220,000 in November, and, “If we get lucky, yet another 220k GB300 by late December,” Musk wrote.
Add them together and the cluster would absorb as many as 660,000 additional GB300 processors by year-end, roughly doubling its current count. Operated by Musk’s artificial-intelligence company xAI, the site would then hold well over a million Nvidia accelerators in total, a single facility with more computing power than most countries possess.
The scale is difficult to place against anything that came before. When xAI built the first Colossus in 2024, it stood up around 100,000 of Nvidia’s H100 chips in a matter of months, a pace the industry read as a signal of how quickly frontier computing was moving. Colossus 2 has already moved far beyond it, drawing on the order of a gigawatt of electricity. Analysts at Epoch AI put the capital cost so far at roughly $35 billion, with the full buildout approaching $58 billion.
The GB300 sits in the middle of that plan. It belongs to Nvidia’s Blackwell line, a generation that produces more computing per watt than the H100 chips that defined the previous rush. For Musk, the chips are the raw material for Grok, xAI’s model, and for the argument he has pressed all year: that whoever assembles the most compute first will lead in artificial intelligence.
The chips are also the raw material for Grok, the model xAI sells to consumers and businesses. Grok has trailed OpenAI’s ChatGPT and Google’s Gemini in market share but not in ambition, and Musk has cast it as the vehicle through which xAI will turn its computing lead into revenue. Every wave of GB300s announced in the post is, in that telling, more Grok, faster.
Musk has turned Memphis into a statement of intent. The site sits near Southaven, Mississippi, in a region he has chosen for speed, and he has promoted the pace of construction almost as much as the hardware inside it. The first Colossus took its name from the code-breaking computer built at Bletchley Park during the Second World War, and Musk has leaned on that lineage to cast the project as a national-scale effort rather than a private one.
The stakes are financial as well as technical. xAI has raised tens of billions of dollars across successive funding rounds and has been in merger talks with X, the social network Musk owns. A combination would fold the model maker, the data centers and the platform into one company whose backers are wagering that a lead in raw computing power converts into models customers will pay for.
The buildup also matters to Nvidia. The GB300 is the successor to the B200, and xAI’s orders, hundreds of thousands of units at a time, rank among the largest single-customer commitments in the chipmaker’s history. When Musk writes of adding 660,000 chips, he is describing demand Nvidia must manufacture and deliver on a timeline that would have looked absurd two years ago.
The race around xAI has only intensified that pressure. OpenAI has locked in tens of billions of dollars of capacity through Microsoft and Oracle, and Google has been expanding its own tensor-processing-unit fleets for the same purpose. Each camp now measures progress in how many accelerators it can put to work in a given quarter, and Musk’s post was a direct entry in that competition.
Power is the constraint nobody can design around. Colossus 2 draws its electricity from the Tennessee Valley Authority grid, and xAI has installed gas turbines on site to bridge the gap between what the grid can deliver and what the cluster demands. The arrangement has drawn scrutiny from local utilities and regulators, and it explains why Musk’s expansion plan is phased rather than immediate.
There are reasons for caution. Musk himself noted that the December batch depended on schedule alignment, and the industry’s recent record is full of data-center projects that slipped as power, grid connections and cooling grew harder to manage at this scale. Memphis has already tested those limits.
Yet the direction is unambiguous. Musk has spent two years telling the market that the winner in AI will be the company that moves fastest to assemble the most compute, and each post about Memphis advances that argument. The latest one pushes the number higher than it has ever been, and leaves little doubt about where he expects the lead to come from.
