SpaceX’s Starmind Plan Could Bring 5 Gigawatts of AI Computing to Orbit by 2028
SpaceX may be preparing for one of its most ambitious projects yet: building AI data centers in space. According to analyst Jeff Pu of GF Securities, Elon Musk’s company could reach as much as 5 gigawatts of space-based AI computing capacity by 2028, powered by a new generation of satellites designed specifically for high-performance computing.
The plan, often referred to as Starmind, appears to build on the same large-scale satellite deployment strategy that made Starlink a major force in global internet connectivity. Instead of simply transmitting broadband service from orbit, these next-generation spacecraft would carry powerful AI computing hardware, potentially turning low Earth orbit into a massive distributed data center.
At the heart of the project is SpaceX’s AI1 satellite design. SpaceX has indicated that these satellites are being built to support a 250-kilowatt compute payload per spacecraft. That payload is expected to include NVIDIA’s Vera Rubin AI platform, a next-generation system designed for intensive artificial intelligence workloads.
Elon Musk has previously described the Starmind satellite as essentially an optimized Vera Rubin NVL72 computer adapted for space. That detail is important because traditional AI data centers on Earth face growing challenges around electricity demand, cooling, land use, and infrastructure. SpaceX’s idea is to move part of that computing burden into orbit, where satellites can draw power from solar arrays and radiate heat into the cold environment of space.
Jeff Pu believes the future version of these AI satellites could be even more powerful than the initial AI1 design. His estimate suggests that SpaceX could deploy around 10,000 space-optimized NVIDIA Vera Rubin NVL72 GPU racks by 2028. He refers to these specialized space computing racks as “Kessel” racks.
According to Pu, each Rubin GPU in these racks could have a thermal design power of around 4,000 watts. That would be a major increase compared with typical expectations for Earth-based AI GPUs, highlighting how aggressively SpaceX and NVIDIA may be pushing performance for orbital computing.
The rack design would also need to be very different from conventional data center hardware. Pu describes the Kessel rack as a compact version of the NVL72 system. Instead of using a tall 48U rack format often found in terrestrial data centers, SpaceX and NVIDIA are believed to be arranging the hardware horizontally in a much smaller 9U configuration.
That design shift is not just about saving space. A satellite payload must survive the extreme vibration, acceleration, and mechanical stress of a rocket launch. A lower, more compact rack layout could help improve stability and make it easier to integrate high-performance AI hardware into a spacecraft.
Pu estimates that SpaceX’s next-generation AI satellites could support between 350 kilowatts and 400 kilowatts of peak compute power, compared with the 250-kilowatt level associated with the AI1 satellite. If accurate, this would represent a significant step forward for space-based artificial intelligence infrastructure.
The analyst also outlined some of the hardware features that may be required for such systems, including a roughly 60-layer PCB midplane, advanced high-speed connectors, PTFE materials, hybrid M9 technology, and single-loop liquid cooling connected to external radiators. He estimated that the midplane alone could carry an average selling price of more than $20,000.
Cooling remains one of the most important challenges for AI in space. On Earth, AI data centers require enormous cooling systems to remove heat from GPUs and networking equipment. In orbit, there is no air to carry heat away, so satellites must rely on radiators that emit thermal energy directly into space.
Musk has said that aluminum radiators in space could radiate around 1,300 watts per square meter due to the extremely low surrounding temperatures. That could make orbital computing attractive in theory, though the engineering demands would be substantial.
If SpaceX can make the concept work, Starmind could open an entirely new category of infrastructure: AI data centers in orbit. The company has already proven that it can mass-produce, launch, and operate thousands of satellites through Starlink. Applying that model to artificial intelligence computing could create a new market for space-based cloud services, machine learning workloads, and high-performance data processing.
Musk has suggested that the opportunity for space AI computing could target a market worth tens of trillions of dollars, potentially far larger than current estimates. Whether that forecast proves realistic remains to be seen, but the direction is clear: SpaceX is no longer thinking only about communications satellites. It is exploring the possibility of turning space into the next frontier for artificial intelligence infrastructure.
If the projections hold, by 2028 SpaceX could be operating a massive orbital AI network with gigawatt-scale computing capacity, powered by advanced NVIDIA hardware and supported by purpose-built satellite engineering. That would mark a major evolution in both the AI industry and the commercial space economy.
