Elon Musk is pushing the artificial intelligence race from Earth’s surface to low Earth orbit. On Monday, SpaceX (SPCX) CEO Musk announced on social platform X that the company’s first Starmind AI satellites, equipped with NVIDIA (NVDA) chips, are scheduled for their initial launch in the fourth quarter of 2027 and will reach “significant scale” in 2028. This timeline represents an acceleration from the company’s previously disclosed 2028 plan in its prospectus.
On the day of the announcement, NVIDIA officially confirmed that SpaceXAI will deploy its next-generation Vera CPU to accelerate agentic AI workloads and will send an optimized version of the Vera Rubin NVL72 rack-level system into orbit as the computing core for the first-generation Starmind satellites. This move marks the official transition of the world’s first space-based AI data center project from concept to engineering implementation. As the first CPU designed specifically for AI agents, Vera features 88 NVIDIA-designed Olympus cores, equipped with LPDDR5X memory offering bandwidth up to 1.2 TB/s, and can complete specific tasks up to 1.8 times faster than traditional x86 CPUs.
Both NVIDIA executives and the President of SpaceXAI have pointed out that this architecture can effectively handle large volumes of coordination scheduling and data processing, allowing GPUs to focus on core tasks, thereby enabling higher-performance AI agents and greater energy efficiency output.
Starmind is SpaceX’s planned orbital AI data center constellation project. The company submitted an application to the U.S. Federal Communications Commission in January 2026, proposing to launch up to 1 million computing satellites. Each Starmind satellite is essentially an optimized Vera Rubin NVL72 rack-level system, achieving high-speed inter-satellite data transmission via laser links to form a distributed space-based AI supercomputer. Musk posted on X that this satellite-form data center is “simpler, lower-cost, higher-density, and lighter-weight than traditional racks,” while requiring no large land occupation and facing relatively less cooling pressure.
Musk also set more ambitious computing power expansion targets: planning to have over 2 gigawatts of AI computing capacity by the end of 2026, and “closer to 10 gigawatts than 5 gigawatts” by the end of 2027, with cumulative computing power across multiple projects potentially reaching 20 GW by the end of next year. He acknowledged that some projects may not be fully completed on schedule but projected that approximately 15 GW of online computing capacity could be achieved by the end of 2027, at which point SpaceX will secure a “very large proportion” of NVIDIA’s GPU supply.
This collaboration represents a comprehensive upgrade in the relationship between SpaceX and NVIDIA. Previously, Musk had announced during an earnings call that the company had decided to exclusively adopt NVIDIA technology for AI computing infrastructure, stating that its Vera Rubin architecture is “the best AI computer available today.” The two companies have announced joint development of the computing payload for the Starmind AI1 satellite. NVIDIA’s previously disclosed 13F filing showed that it holds approximately $21 billion in SpaceX shares, indicating that this deep integration has now extended from the capital level to the technical level.
The announcement of this collaboration comes on the eve of NVIDIA’s upcoming quarterly earnings release. The narrative of space-based AI data centers adds a new dimension to NVIDIA’s AI infrastructure story. Market reactions to the prospects of space data centers have been mixed, with supporters seeing advantages in regulatory avoidance, while critics firmly believe that the operational costs of ground-based data centers will remain far lower than space-based solutions for years to come.
