SpaceXAI to add 660,000 GPUs, nearing 1.44 million total
Elon Musk announced on X that SpaceXAI will bring 220,000 Nvidia GB300 GPUs online by next week, with another 220,000 arriving in November and a final batch of 220,000 by late December, contingent on luck. These additions will join the 110,000 GB200 and 440,000 GB300 units already operating at Colossus 2, as well as the mixed Hopper and Blackwell fleet at Colossus 1. This expansion brings the…
Key points
- SpaceXAI will add 660,000 Nvidia GB300 GPUs by late December 2026.
- Total operational GPUs will reach 1,440,000, including 1.1 million GB300s.
- Musk aims for 50 million H100-equivalent GPUs by 2030 and orbital data centers.
The rapid build-out highlights SpaceXAI’s aggressive strategy despite being only three years old, compared to OpenAI’s ten and Anthropic’s six. To address power constraints, the company built its own 1.2-GW power plant, though this has led to legal disputes over unpermitted gas turbines. Musk stated that Colossus 1 is inefficient for training Grok and has been rented to Anthropic for inference, while Colossus 2 uses solely Blackwell GPUs to avoid bottlenecks.
Looking ahead, Musk aims to grow data center capacity sevenfold by 2027 and reach 50 million H100-equivalent GPUs by 2030. He also plans to launch an orbital data center system with a million satellites, a concept Nvidia’s Jensen Huang has described as a dream for now. Broadcom previously noted that three hyperscale customers are targeting similar scale by 2027.
Elon Musk's SpaceXAI to add another 660,000 AI GPUs this year, nearing a total of 1.44 million in operation
Tom's Hardware · 25 September 2026
Loading the full article…
This text was published by Tom's Hardware and written by Jowi Morales. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- SemiAnalysis reveals China’s 24GW datacenter capacity, surpassing EMEA · 1 src
- Ricursive Intelligence aims to cut chip design cycle from years to weeks · 1 src
- Meta CEO Mark Zuckerberg announced the Muse Charm AI device at Connect event · 21 src
- Microsoft removes Copilot+ PC label from new Surface devices · 3 src
- Meta launches Ray-Ban Meta Audio glasses and Ray-Ban Meta Gen 3 with expanded styles and features · 27 src
Comments
via GitHub Discussions