SpaceXAI targets expansion of Colossus 2 supercomputer to over one million processors

Elon Musk disclosed ambitious expansion plans for SpaceXAI's Colossus 2 cluster, potentially reaching approximately 1.21 million NVIDIA processors by the final months of 2026 provided that scheduled milestones are achieved. The facility now houses 110,000 GB200 chips alongside 440,000 GB300 units. Beginning next week, the operation intends to add 220,000 additional GB300s monthly through the year's end, contingent on infrastructure readiness. Musk emphasized that fiber-optic cable availability and network connectivity represent the most significant constraints on speed of expansion, with each increment of 110,000 chips reflecting the maximum density that can connect through a single central switch. Supporting such growth demands substantial investment in power delivery, cooling systems, and networking infrastructure.

1. Colossus 2 presently operates with 110,000 GB200 chips and 440,000 GB300s, with 220,000 additional GB300s slated to arrive next week
2. An extra 220,000 GB300 units are scheduled for November, followed by another 220,000 in late December assuming timelines are maintained
3. Full expansion would result in a total of 1.21 million chips, comprising 1.1 million GB300s
4. Fiber-optic cable capacity linked to central switches represents the primary networking bottleneck limiting expansion velocity
5. Each GB300 NVL72 rack incorporates 72 Blackwell Ultra GPUs, consuming 142 kilowatts of power and delivering 130 terabytes per second of NVLink bandwidth