Elon Musk's SpaceXAI is adding 660,000 Nvidia GB300 GPUs across the final months of 2026, pushing its Colossus supercomputer cluster toward 1.44 million total accelerators and marking the largest single-site GPU deployment disclosed to date.
The rollout schedule
Musk said on X that 220,000 GB300 units will be operational by early October, another 220,000 in November, and a final 220,000 by late December "if we get lucky." Those batches sit on top of 440,000 GB300s and 110,000 GB200s already running at Colossus 2, plus 150,000 H100s, 50,000 H200s and 30,000 GB200s at the original Colossus 1 site. The combined GB300 count reaches 1.1 million.
Power and the neighborhood
The build-out required SpaceXAI to supply its own electricity. The company ran unpermitted gas turbines, drawing a lawsuit from local residents; it has agreed to remove them over a year as a dedicated 1.2-gigawatt power plant comes online. Power, not chip allocation, has become the binding constraint for data centers at this scale.
The competitive context
SpaceXAI is three years old. Anthropic and OpenAI are six and ten. Broadcom disclosed in 2024 that three unnamed hyperscalers are targeting million-GPU clusters by 2027. Musk has said SpaceXAI will grow capacity sevenfold by 2027 and aims for 50 million H100-equivalent GPUs by 2030, including an orbital constellation of one million satellites, a concept Nvidia's Jensen Huang has called a "dream" for now.
Colossus 1 finds a tenant
The first Colossus site, a mix of Hopper and Blackwell silicon, proved inefficient for training Grok. SpaceXAI rented it to Anthropic for inference workloads instead. Colossus 2 runs purely on Blackwell, eliminating the interconnect bottlenecks that plagued the heterogeneous design.