SpaceX designed an orbital Vera Rubin. Radiation comes next.
The New Stack Steven J. Vaughan-Nichols
SpaceX and Nvidia want to run Vera Rubin AI chips in orbit, with a first launch eyed for Q4 2027. The hard part isn’t the launch — it’s keeping a rack full of delicate silicon alive in radiation and vacuum.
Based on reporting by The New Stack, Steven J. Vaughan-Nichols — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
SpaceX and Nvidia are betting that the Vera Rubin NVL72 can leave the data center and survive in low Earth orbit. SpaceX says the first launch is targeted for the fourth quarter of 2027, and Elon Musk says the plan includes “significant scale in 2028.” That’s not a side experiment. It’s the computing core for SpaceXAI’s first Starmind AI satellite.
On Earth, the NVL72 is a rack-scale AI system built around 72 Rubin GPUs, 36 Vera CPUs and networking gear such as ConnectX-9 SuperNICs. Musk has said SpaceX will build exclusively on Nvidia because he thinks the Vera Rubin architecture is the best one, and he wants to use it on the ground as well as in orbit. He also argued that the NVL72 is a cleaner design than a standard rack and would cost less and work better.
And orbit changes everything. No gravity. No technicians. No stable grid power. No big building to hide the cooling hardware in. SpaceX says the orbital version would use closed-loop liquid cooling inside the spacecraft, plus large deployable radiators that dump heat as infrared radiation. That may sound neat on a slide deck, but heat rejection in vacuum is a brutal engineering problem, not a slogan.
Radiation is the other problem waiting in the wings. NASA’s figures for low Earth orbit below 500 km put trapped-particle dose rates at 100 to 1,000 rad(Si) per year, which is enough to wear on electronics over time even if it doesn’t kill them outright. A real orbital Rubin would need redesigns across GPUs, CPUs, memory, networking, power, firmware and operations. The likelier near-term result is not a fully hardened space chip, but a radiation-tolerant system built from selected commercial Nvidia parts, with shielding, error correction and lots of fault handling doing the heavy lifting.
My take — AI-written commentary, not fact-checked reporting
Space people keep trying to turn “it works in the lab” into “it works in orbit,” and orbit keeps winning. The cool part is the ambition; the less glamorous part is that vacuum, heat and radiation do not care about conference slides. This is exactly why so much of AI hardware hype sounds heroic on land and looks much older once the power bill, the failures and the physics show up.
Read more about this at: The New Stack