Nvidia launches space computing to bring AI into orbit

With Earth-bound data centers running out of power and land, Jensen Huang is looking toward low Earth orbit.

At GTC, NVIDIA officially unveiled its Space Computing initiative, bringing accelerated computing architectures to orbital data centers, geospatial satellites, and autonomous spacecraft.

From orbit to ground stations

Headlining the announcement is the NVIDIA Space-1 Vera Rubin Module. Designed specifically for orbital data centers, the Rubin-based module promises up to 25x more AI inference performance than a terrestrial H100 GPU.

It packs a tightly coupled CPU-GPU architecture with high-bandwidth interconnects to run large language models and foundation models directly in orbit. The Vera Rubin module is slated for release at a later date.

For missions launching sooner, NVIDIA made its IGX Thor and Jetson Orin edge computing platforms available today. Both target size, weight, and power (SWaP) constraints onboard smaller satellites.

On the ground, NVIDIA rolled out the RTX PRO 6000 Blackwell Server Edition GPU, which it claims processes massive satellite imagery archives up to 100x faster than legacy CPU batch systems.

Heat sinks do not work in a vacuum

A solid roster of commercial space players has already signed up, including Planet Labs, Axiom Space, Kepler Communications, Aetherflux, Sophia Space, and Starcloud.

Planet Labs, for instance, uses the platform alongside NVIDIA CorrDiff AI models to process raw satellite pixels into real-time geospatial intelligence before beaming it down. Starcloud plans to build purpose-built orbital data centers to run both training and inference in space.

Yet, the laws of physics remain unforgiving. As Jensen Huang acknowledged, space offers neither air conduction nor fluid convection. Cooling high-power AI silicon entirely through thermal radiation is a massive engineering hurdle that teams are still racing to solve.

Why it matters

We think this launch addresses the single biggest bottleneck in modern earth observation: the downlink choke point.

Modern constellations capture petabytes of imagery, but satellites can only dump data when flying directly over ground stations. That pipeline is hopelessly narrow. By running models like CorrDiff onboard, satellites can discard empty cloud layers and send down only actionable alerts—like tracking wildfires, oil spills, or military movements—in near real time.

For space startups and defense contractors, deploying Jetson Orin today is a no-brainer for immediate edge AI experiments. However, betting on multi-kilowatt orbital data centers powered by Vera Rubin remains a high-risk long game.

Strategically, NVIDIA is exporting its CUDA lock-in to the aerospace sector, challenging slow, custom radiation-hardened ASICs with modular, commercial off-the-shelf platforms. If NVIDIA solves the thermal radiation puzzle, it will turn orbit into the ultimate cloud extension; if not, orbital data centers will remain a very expensive sci-fi dream.