Elon Musk's SpaceXAI Plans to Deploy 1.44 Million Nvidia AI Chips by Late 2026

SpaceXAI Accelerates Nvidia AI Chips Deployment to 1.44 Million GPUs

Elon Musk revealed an aggressive timeline to expand SpaceXAI's Colossus 2 data center with 660,000 additional Nvidia GB300 GPUs by the end of December 2026. The billionaire announced on X that 220,000 Nvidia GB300 GPUs will become operational next week, followed by another 220,000 units in November1

1

. A final batch of 220,000 GB300 chips should arrive by late December "if we get lucky," Musk added2

2

. The Colossus 2 data center currently operates 110,000 GB200 chips and 440,000 GB300 chips3

3

. This expansion would bring the Memphis facility to approximately 1.21 million processors by year-end, with about 1.1 million being GB300s4

4

.

World's Largest AI Supercomputer Reaches Million-GPU Milestone

Combining both Colossus sites positions SpaceXAI as operating the world's largest AI supercomputer. Colossus 1 currently runs 150,000 H100, 50,000 H200, and 30,000 GB200 GPUs1

1

. When added to Colossus 2's projected capacity, the full Memphis site would reach approximately 1.44 million GPUs by December 2026, comfortably exceeding SpaceXAI's original one-million-GPU target4

4

. The 110,000-chip increments reflect "the number of fiber optic cables that can be plugged into a central switch," highlighting networking constraints in scaling clusters3

3

. This achievement stands out given that SpaceXAI is only three years old, while Anthropic and OpenAI are six and ten years old respectively1

1

.

AI Arms Race Intensifies as SpaceXAI Targets OpenAI and Anthropic

Musk's expansion directly supports SpaceXAI's ambition to match Anthropic's Fable class and OpenAI's GPT-6 models within two to three months. "Our AI efforts are only 3 years old, vs 6 and 10 years old for Anthropic and OpenAI. If our second derivative remains strong, SpaceX will reach pole position in about 6 months," Musk stated2

2

. The Colossus 2 data center focuses exclusively on Nvidia's Blackwell GB300 GPUs, which feature the same Grace CPU as the GB200 but include a newer GPU variant called Blackwell Ultra2

2

. This strategic choice ensures no bottlenecks during training of the Grok AI model. Meanwhile, Colossus 1's combination of Hopper and Blackwell GPUs proved inefficient for training Grok, leading Musk to rent it out to Anthropic for inference instead1

1

.

Power Supply Challenges Drive 1.2-Gigawatt Plant Construction

Power infrastructure has emerged as one of the biggest bottlenecks in AI compute infrastructure buildouts. SpaceXAI addressed this by building its own 1.2-gigawatt power plant to support operations1

1

. The approach has sparked controversy, with a lawsuit filed against the company for running unpermitted gas turbines. SpaceXAI promised to remove the temporary 59 natural-gas turbines over one year as the permanent power plant comes online3

3

. Musk stated that SpaceXAI would end the year with over 2 gigawatts of computing power, with targets reaching closer to 10 gigawatts than 5 by 20274

4

. Jensen Huang estimated that one gigawatt of compute costs roughly $50 billion to $60 billion, with Nvidia chips and components making up about $35 billion of that total4

4

.

Commercial AI Compute Revenue Offsets Infrastructure Costs

SpaceXAI has transformed Colossus from an internal training facility into a commercial AI compute business. Reflection AI signed a deal to pay SpaceXAI $150 million per month starting July 1 for access to GB300 capacity at Colossus 23

3

. Google separately signed a multiyear agreement paying roughly $920 million per month for compute capacity3

3

. Anthropic gained access to more than 300 megawatts at Colossus 1, with SpaceXAI agreeing to lease its entire first data center to the AI competitor3

3

. The initial buildout cost approximately $18 billion for 550,000 GPUs, with newer GB300 chips potentially costing even more given the generation's higher price point4

4

.

AI Infrastructure Expansion Plans Extend Beyond Earth

Musk's ambitions for AI infrastructure expansion extend well beyond the current Memphis facilities. He stated that SpaceXAI will grow its data center capacity sevenfold by 2027, aiming for 50 million H100-equivalent GPUs by 20301

1

. These supercomputer sites won't be limited to ground-based operations. SpaceX plans to launch an orbital data center system called Starmind AI1 with a million satellites, though Jensen Huang called it a "dream" for now1

1

. SpaceX and Nvidia are jointly developing the compute payload for this first satellite, built around Nvidia's Rubin GPUs and Vera CPUs4

4

. During SpaceX's first earnings call as a public company in early August, Musk told analysts the company would build its AI infrastructure exclusively on Nvidia, calling the Vera Rubin the best architecture and the best AI computer4

4

. Nvidia holds a significant equity stake in SpaceX, while Musk himself is the company's largest individual shareholder4

4

.