SpaceX plans to scale its AI compute capacity to roughly 10 gigawatts by the end of 2027, and it is building the whole stack on Nvidia hardware.
The plan was laid out during SpaceX’s Q2 2026 earnings call in early August, alongside its xAI unit. Capacity stood at approximately 1.4GW at the close of Q2 2026.
Reaching 10GW would mean growing to more than seven times that base in about 18 months.
The Nvidia bet and the orbital wildcard
The expansion will run primarily on Nvidia’s Vera Rubin architecture, specifically the NVL72 and Kyber configurations. Musk described the partnership as exclusive and called Vera Rubin the best architecture available in the industry.
The interim milestone is more modest. SpaceX projects it will exceed 2GW by the end of 2026.
The 2027 target carries some wiggle room of its own. According to the research findings, the company aims for something potentially closer to 10GW than 5GW by year-end 2027.
SpaceX is developing an orbital AI platform called Starmind, built on the same Vera Rubin systems as its terrestrial servers. Initial Starmind satellite launches are planned for 2027, with a substantial scale-up anticipated in 2028.
Who gets the compute, and what it could earn
SpaceX intends to keep the majority of the capacity in the family. Roughly 60% is earmarked for internal use, spanning xAI, the Grok chatbot, SpaceX itself, and Tesla.
The remaining 40% would go to outside customers. Google and Anthropic are named as notable anticipated engagements.
Musk predicted the 10GW buildout could generate between $300 billion and $500 billion in annual recurring revenue, assuming pricing of $30 to $50 per watt.
Capital expenditures were estimated at between $15.8 billion and $18.4 billion in Q2 2026 alone.
A track record of moving fast
The company’s confidence rests partly on its Memphis supercluster. That site set records by bringing approximately 100,000 Nvidia H100 GPUs online in 122 days.
As of October 2026, the targets remain in place despite ongoing supply chain challenges affecting components such as printed circuit boards and cooling systems.
What this means for Nvidia, rivals, and the AI arms race
For Nvidia, a customer committing this much future capacity to a single architecture is a significant endorsement. It also concentrates risk. If Nvidia hits production delays on Vera Rubin, SpaceX’s timeline absorbs the hit directly, with no alternate supplier baked into the plan.
Selling capacity to competitors of xAI means SpaceX will be both a rival and a landlord to firms like Anthropic and Google.
The things to watch are concrete. Look for whether SpaceX clears 2GW by the end of 2026, whether the first Starmind satellites launch on schedule in 2027, and whether quarterly capital spending keeps climbing from the Q2 range.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.