xAI to Add 660,000 GB300 Chips by December, Taking Colossus 2 to 1.1 Million Nvidia GPUs
Elon Musk posted a detailed chip ramp schedule for xAI’s Colossus 2 supercomputer in Memphis, Tennessee, planning three separate waves of Nvidia GB300 additions that would lift its total GPU count to roughly 1.21 million by late December.
Current Inventory and the Three Waves
Colossus 2 currently runs 110,000 GB200 chips and 440,000 GB300 chips. Musk’s expansion plan adds GB300 units only:
| Wave | Chips | Target Date |
|---|---|---|
| Wave 1 | +220,000 GB300 | Late September 2026 |
| Wave 2 | +220,000 GB300 | November 2026 |
| Wave 3 | +220,000 GB300 | December 2026 (“if we get lucky”) |
After Wave 3, Colossus 2 would carry 110,000 GB200 and 1,100,000 GB300 chips — up from the current combined count of 550,000.
From Grok Training to Shared Compute Rack
Colossus 2 was built primarily to train Grok, xAI’s flagship AI model. The facility has since shifted to a hybrid model: Google has contracted for roughly 110,000 GPUs from xAI’s Memphis campus, while Anthropic separately leases the original Colossus 1 facility. Between the two agreements, xAI’s Memphis campus effectively serves as shared frontier compute infrastructure for its direct rivals.
Musk has described Colossus as the “world’s largest AI supercomputer.” The three-wave expansion, if completed on schedule, takes GB300 deployment alone past the 1-million mark — a scale no single facility has publicly declared before.
What the Announcement Did Not Address
Musk’s post on X outlined the timeline and chip counts but gave no specifics on power capacity at the Memphis site. Large-scale GB300 deployments are power-intensive: each GB300 NVL72 rack draws 132–142 kW, meaning 660,000 additional chips across roughly 9,200 racks would require more than 1.2 GW of additional capacity.
The facility’s power situation has been contentious. The Memphis campus has operated a fleet of natural gas turbines that became the subject of Clean Air Act scrutiny and a mid-2026 state regulatory order. Musk’s chip ramp announcement made no reference to how the additional 1.2+ GW requirement will be met.