Nvidia builds compute for orbit
Nvidia unveiled hardware designed for AI data centers in space, extending its edge-compute push beyond terrestrial datacenters and opening new architectures for low-latency global autonomy and resilient comms. That move changes the frontier for simulation, AV fleets, and defense platforms that rely on distributed, high‑performance inference and training. (fool.com)
NVIDIA unveiled the Space‑1 Vera Rubin Module during its GTC keynote on March 16, 2026, framing it as a purpose‑built computing platform for orbital data centers and autonomous space operations. (investor.nvidia.com) NVIDIA said the Rubin GPU on that module delivers up to 25× the AI inferencing compute of an H100 for space workloads and the company used a 50 petaflops NVFP4 performance figure in its orbital compute pitch. (nvidianews.nvidia.com) Technical details released during the announcement describe the Rubin graphics component as a 3 nm design with roughly 336 billion transistors, 88 compute cores, and a memory subsystem based on LPDDR5X. (siliconangle.com) The product rollout pairs space‑specific edge platforms (NVIDIA IGX Thor and NVIDIA Jetson Orin) for on‑orbit inference with NVIDIA’s RTX PRO 6000 Blackwell Server Edition for high‑throughput ground processing, with IGX Thor and Jetson Orin listed as available today while the Space‑1 Vera Rubin Module is slated for later availability. (nvidianews.nvidia.com) NVIDIA named six commercial partners already testing or integrating the stack—Aetherflux, Axiom Space, Kepler Communications, Planet Labs PBC, Sophia Space, and Starcloud—signaling coordinated launch and integration efforts across small‑sat and commercial space firms. (nvidianews.nvidia.com) Starcloud previously launched Starcloud‑1 on Nov. 2, 2025 carrying an NVIDIA H100, and later reported training and running inference on a NanoGPT/Gemma‑family model in orbit—an operational precedent cited by vendors and NVIDIA during the Space‑1 briefing. (datacenterdynamics.com) NVIDIA emphasized use cases such as real‑time geospatial intelligence and autonomous space operations, claiming RTX PRO Blackwell can analyze imagery archives up to 100× faster than legacy CPU batch systems, and noting orbital designs must address radiation‑tolerant thermal rejection and solar power constraints. (nvidianews.nvidia.com)
