Elon Musk in profile in a dark jacket, hands clasped, listening intently, with other attendees blurred behind him

Elon Musk at an event on internet connectivity in the Amazon, in Porto Feliz, Brazil, in May 2022. Image: Cleverson Oliveira / Brazilian Ministry of Communications / Wikimedia Commons, CC BY 2.0, cropped

Elon Musk is going all in on Nvidia. SpaceXAI’s Colossus 2 data center already runs about 550,000 of Nvidia’s top AI chips, and Musk says he wants to add up to 660,000 more by the end of December, more than doubling it, as Bloomberg reported. He laid out the timeline in a reply on X early on Friday:

The numbers

Colossus 2, just outside Memphis in Southaven, Mississippi, is SpaceXAI’s second giant training cluster. According to Musk’s post, it has:

  • Today: 110,000 GB200 and 440,000 GB300 chips, about 550,000 in total.
  • Next week: another 220,000 GB300s fully operational.
  • November: a further 220,000.
  • Late December: 220,000 more, “if we get lucky.”

If every wave lands on time, Colossus 2 alone would reach roughly 1.2 million chips. Add the original Colossus, with its 150,000 H100s, 50,000 H200s and 30,000 GB200s, and SpaceXAI’s Memphis operation would be running about 1.44 million Nvidia GPUs by the end of the year. That’s ahead of the company’s own earlier goal of 1 million GPUs in Memphis in 2026.

These figures come from Musk alone, and SpaceXAI hasn’t published anything to back them up. Musk has a long record of optimistic timelines, and his own “if we get lucky” suggests the December batch is far from certain.

Why 110,000?

The oddly specific batch sizes have a practical explanation, according to Musk. In a follow-up reported by TeslaNorth, he said:

The strange multiple of 110k is due to the number of fiber optic cables that can be plugged into a central switch.

Elon Musk, on X

In other words, the cluster grows in blocks sized by its networking, which is what lets hundreds of thousands of chips work together on one training run.

What it’s for

All that compute trains and runs Grok. SpaceXAI launched Grok 4.7 on September 21, and it powers the Grok Bot agent in Teslas. Musk has argued that sheer scale will put SpaceXAI in “AI pole position,” TeslaNorth reports, even though Grok 4.7 landed as a close second rather than a clear leader.

It’s also a huge order for Nvidia, whose flagship GB300 is among the most expensive chips it sells, and chip stocks were rising on Friday morning. On top of that, Musk wants to put compute in space: SpaceX’s Starmind plan, which we covered in our orbital data centres explainer, aims to run AI chips on satellites.

Why it matters

The AI race is increasingly a race for chips and power, and Musk is signalling he intends to win it on raw scale. If he hits even most of this timeline, SpaceXAI will run one of the biggest AI clusters on Earth by Christmas; the question is whether all that hardware can turn Grok into a leader rather than a close second.

Sources: Elon Musk on X, Bloomberg via Yahoo Finance, TeslaNorth, TipRanks.

Related