xAI Plans to Push Colossus 2 Past 1.2 Million Nvidia GPUs Before December

by Geth Author

xAI Wants to Push Colossus 2 Past 1.2 Million Nvidia GPUs Before December

On September 25, 2026, Elon Musk outlined a phased expansion plan for Colossus 2, xAI's Memphis-area AI supercomputer, that would more than double the facility's current GPU count before the end of the year. Bloomberg reported the details from a company presentation.

Where things stand now

Colossus 2 currently operates with 110,000 Nvidia GB200 chips and 440,000 GB300s—approximately 550,000 GPUs total. At that scale it's already one of the largest AI training clusters in existence. The GB300 is part of Nvidia's Blackwell Ultra architecture, designed for the dense matrix operations that large language model training demands.

The expansion roadmap

Musk described three phases:

  • This week (late September): 220,000 additional GB300s coming online
  • November: Another 220,000 GB300s
  • Late December: A final 220,000 — "if we get lucky," Musk said

If all three tranches land on schedule, the total GPU count at Colossus 2 would reach approximately 1.21 million by year-end. That matches and slightly exceeds xAI's stated goal—earlier in 2026, the company announced plans to equip the facility with 1 million GPUs before the end of the year. The new roadmap, if it executes, would beat that target.

Why xAI is scaling this fast

Colossus 2 is what trains Grok. Grok 4.7, released four days earlier on September 21, is xAI's current flagship model—a 2.1-trillion-parameter system that xAI says required 4–8 weeks of training at Colossus-scale compute. To continue pushing model capability on that cadence, the infrastructure has to keep growing.

The broader AI compute race has intensified throughout 2026. OpenAI, Google DeepMind, and Anthropic are all building or operating clusters at similar scales, and Nvidia's GB300 supply remains constrained. Acquiring hundreds of thousands of chips and bringing them online within weeks—rather than months—represents a significant logistical and financial commitment. xAI's ability to execute on this roadmap will be a meaningful indicator of how effectively it can compete at the frontier.

The power problem

Colossus 2 is already reported to draw close to 2 gigawatts at peak load. Adding 660,000 more GPUs raises serious questions about power capacity at the Memphis site. xAI has been in discussions with the Tennessee Valley Authority and multiple regional providers about grid expansion. No specific commitments were disclosed alongside the GPU roadmap announcement.

The energy constraint is real and likely the biggest risk to the timeline. A delayed December tranche would put xAI at roughly 990,000 GPUs—still a historic scale for a single facility, but short of the symbolic million-GPU mark the company has been targeting.

Whether or not the December chips arrive on schedule, the September-to-November expansion alone would represent one of the largest single-site compute buildouts in 2026. In an industry where compute is competitive advantage, Colossus 2's trajectory matters beyond just xAI.