NVIDIA Trained AI in Space | Now It’s Building Data Centers Up There! (it's OVER for Competitors)

Author: Evolving AI - Bewertung: 1x - Views:5

NVIDIA just helped launch something no one has ever done before: real AI training in space. In this video, we break down how Nvidia-backed startup Starcloud used an H100 GPU on the Starcloud-1 satellite to train a NanoGPT transformer on Shakespeare entirely in orbit, then followed it up by running Google’s Gemma model for live LLM inference from space. From there, we dive into Starcloud’s much bigger plan: 5-gigawatt orbital data centers powered by massive solar and cooling panels, autonomous “tile” assembly built with Rendezvous Robotics, and future satellites carrying multiple H100s and Blackwell B200 chips deployed in bulk via SpaceX Starship’s “Pez dispenser” payload system. We also zoom out to the coming arms race for computing above Earth. You’ll see how laser links between thousands of satellites could beat fiber on latency and reliability, why Starcloud focuses on “utilities-first” economics while Nvidia supplies the chips, and how early use cases like wildfire detection and SAR imagery from Capella Space prove the concept. Then we compare the rest of the field: Jeff Bezos’ vision of gigawatt data centers in orbit, SpaceX’s own orbital compute ambitions, Eric Schmidt’s Relativity Space bet, Google’s Project Suncatcher TPUs in space, and even lunar data centers from Lonestar and Aetherflux. The real question is no longer which chip is fastest, but who controls the orbital infrastructure that feeds those chips. And if you want the real story behind the world’s fastest-moving AI breakthroughs, make sure to like and subscribe to Evolving AI for daily coverage.