Musk says xAI's Memphis supercomputer could double to 1.2m chips

Elon Musk says the supercomputer that trains Grok could more than double in size by the end of the year. In a reply on X at 03:44 UTC on 25 September 2026, he put Colossus 2 at 110,000 Nvidia GB200 and 440,000 GB300 chips, then listed three more batches of 220,000 GB300s: one “fully operational next week”, one in November and, “if we get lucky”, one “by late December”.

Added up, the three batches take Colossus 2 from 550,000 chips to 1,210,000, 2.2 times its size today. Musk was answering a post that put xAI’s two clusters at about 670,000 GPUs between them; his own figures give 780,000. Bloomberg reported the plan the same day.

Colossus 2 would reach 1.21 million chips by late December

If all three batches arrive on Musk’s schedule, Colossus 2 will hold 1,210,000 Nvidia chips by late December 2026, against 550,000 on 25 September. He gave no date for the November batch beyond the month, and hung the third batch on luck.

The plan is the same shape as the one Musk trailed in July 2025, when he wrote that “the first batch of 550k GB200s & GB300s” would start going online at Colossus 2 “in a few weeks”. Fourteen months later, 550,000 is the number running.

A GB300 rack holds 72 Nvidia GPUs

The GB300 is Nvidia’s Grace Blackwell Ultra platform, which Nvidia sells as the GB300 NVL72: a liquid-cooled rack that joins 72 Blackwell Ultra GPUs and 36 Grace CPUs into one machine, according to Nvidia’s product page. Nvidia says the Blackwell Ultra GPU carries 1.5 times the dense FP4 compute of the earlier Blackwell GPU, the chip inside the GB200 that makes up the rest of Colossus 2.

xAI counts these chips as GPUs. Its own account of Grok 4.5 says the model “was trained across tens of thousands of NVIDIA GB300 GPUs”. On that reading, the 440,000 GB300s already running would fill about 6,100 of Nvidia’s 72-GPU racks.

Colossus 2 has been training Grok since January

Musk declared Colossus 2 operational on 17 January 2026, calling it the “First Gigawatt training cluster in the world” and promising “Upgrades to 1.5GW in April”. Colossus 2 is the cluster xAI trains its Grok models on.

The money came earlier that month. xAI announced on 6 January that it had raised $20 billion in its Series E, above the $15 billion it had targeted, with Nvidia and Cisco Investments among the investors backing “the largest GPU clusters in the world”. It said it had ended 2025 with “over one million H100 GPU equivalents” across Colossus I and II, and that Grok 5 was “currently in training”.

xAI now publishes under the SpaceXAI name.

Its newest model, [Grok 4.7](/ai/llms/grok-4-7/), launched on 21 September 2026 and "uses a new, larger base model compared to Grok 4.6", according to its release.

Colossus 1 took 122 days to build, and Anthropic now has access to it

Colossus 1, xAI’s first supercomputer in Memphis, came online with 100,000 Nvidia H100 chips in September 2024 after a 122-day build, and SpaceXAI agreed in May 2026 to give Anthropic access to it. “From start to finish, it was done in 122 days,” Musk wrote at the time, and xAI’s Colossus page says it then “doubled it in 92 days to 200k GPUs”. It now holds 150,000 H100, 50,000 H200 and 30,000 GB200 chips, 230,000 in all, Musk wrote on 25 September.

Colossus 1 no longer serves xAI alone. On 6 May 2026 SpaceXAI signed an agreement “to provide access to Colossus 1” to Anthropic, which “plans to use this additional compute to directly improve capacity for Claude Pro and Claude Max subscribers”. Anthropic also “expressed interest in partnering to develop multiple gigawatts of orbital AI compute capacity”, SpaceXAI said.

How Colossus 2 compares with OpenAI’s and Meta’s sites

Musk’s figure of one gigawatt, rising to 1.5 gigawatts, puts Colossus 2 in the same class as the biggest AI sites its rivals have announced, though those are measured across whole campuses or several sites. The figures below are each company’s own, as stated on the date shown.

Meta says its Louisiana site will deliver “5 gigawatts of compute capacity to house Hyperion, our largest multi-gigawatt AI cluster”. OpenAI, Oracle and SoftBank put Stargate at “nearly 7 gigawatts of planned capacity” across five new sites, the Abilene flagship in Texas and projects with CoreWeave in September 2025, on the way to a $500 billion, 10-gigawatt commitment. Anthropic’s deal with Microsoft and Nvidia covers $30 billion of Azure compute and “additional compute capacity up to one gigawatt”.

The next step for Colossus 2 is the nearest one: 220,000 GB300s that Musk says will be fully operational next week, which alone would take the cluster to 770,000 chips.