SpaceX's Colossus 2 AI data center will receive 220,000 NVIDIA GB300 GPUs next week, with two additional tranches of the same quantity arriving later this year, totaling approximately 660,000 new GPUs.

Nearly two years after announcing plans to expand the Colossus supercomputer to over a million GPUs, Elon Musk has signaled that the end goal is finally within reach. On September 25, 2026, he announced via X that 220,000 Nvidia GB300 GPUs will be operational by the following week, with another 220,000 coming online in November and a third batch of 220,000 units by late December "if we get lucky."

Musk detailed the composition of SpaceXAI's GPU infrastructure across two Colossus clusters. Colossus 1 houses 150,000 H100, 50,000 H200, and 30,000 GB200 GPUs, while Colossus 2 contains 110,000 GB200 and 440,000 GB300 GPUs. The three planned tranches of 220,000 GB300 units would bring the total GB300 count to 1.1 million, with overall GPU capacity reaching 1.44 million units across both facilities.

This milestone represents a significant achievement for a company only three years old, particularly given that rivals Anthropic and OpenAI are six and ten years old, respectively. Notably, Colossus 1—which combines Hopper and Blackwell GPUs—proved inefficient for training Grok; Musk has rented it to Anthropic for inference work instead. Colossus 2, by contrast, operates exclusively on Blackwell GPUs, eliminating architectural bottlenecks.

While GPU availability remains challenging given Musk's well-documented appeals to Nvidia CEO Jensen Huang, power infrastructure presents the more acute constraint. SpaceXAI solved this by deploying its own generation capacity, though the approach has sparked controversy with neighboring communities. Unpermitted gas turbines triggered a lawsuit, leading SpaceXAI to commit to their removal—though only over a one-year transition as its own 1.2-GW power plant comes online.

SpaceXAI is not alone in pursuing million-GPU scales. Broadcom disclosed in 2024 that three unnamed hyperscale customers are targeting this threshold by 2027. For Musk, however, the one-million-GPU milestone marks merely an interim step. He has stated that SpaceXAI will expand data center capacity sevenfold by 2027 and aims for 50 million H100-equivalent GPUs by 2030. SpaceX is also exploring orbital expansion, planning an Orbital Data Center System with a million satellites, though Jensen Huang has characterized the concept as a "dream" for the present.