At GTC 2026, Jensen Huang officially expanded the NVIDIA ecosystem beyond the biosphere. The unveiling of **Vera Rubin Space-1** marks the birth of the "Orbital AI Factory," a dedicated hardware stack designed to process the trillions of data points generated by satellite constellations in real-time.
The Architecture: Radiation-Hardened Agentic Compute
Processing AI in space is a thermal and radiological nightmare. Traditional **H100s or B200s** would fail within minutes under the bombardment of high-energy protons in Low Earth Orbit (LEO). The **Vera Rubin Space-1** solves this through a "Triple-Redundant Logic" architecture at the gate level.
Unlike terrestrial chips that prioritize density, Space-1 utilizes a **sub-14nm FinFET** process optimized for latch-up immunity. It features a dedicated **Radiation-Aware Scheduler** that dynamically migrates "Inference Kernels" away from silicon clusters experiencing transient bit-flips, ensuring 99.999% uptime in the vacuum of space.
Thermal Management in a Vacuum
Without air for convection, cooling a high-TDP AI chip is impossible using traditional fans. The **Space-1 module** utilizes a proprietary **Phase-Change Cold Plate** integrated directly into the satellite's structure. Heat is moved via heat pipes to large-scale deployable radiators, tapping into the 2.7K ambient temperature of deep space.
This allows the Space-1 to maintain a consistent **300W TDP** per node, delivering **4.5 Petaflops of Agentic FP8 throughput**—roughly 100x the capacity of current "hardened" space processors like the RAD750.
Key Specs: NVIDIA Vera Rubin Space-1
- - **Core Count:** 12,288 Radiation-Hardened Tensor Cores.
- - **Memory:** 128GB HBM4E with Error-Correcting Code (ECC) Level 3.
- - **Connectivity:** Integrated 800G Laser-Link Interconnect for inter-satellite swarms.
- - **Security:** Hardware-isolated OpenShell runtime for autonomous spacecraft agents.
Starcloud Partnership: The First November Launch
NVIDIA is partnering with **Starcloud** to deploy the first cluster of 12 satellites equipped with Space-1 modules in November 2026. These satellites will act as a "Distributed Orbital Cloud," allowing smaller, cheaper cubesats to "rent" inference time via laser links.
This enables "Agentic Spacecraft" that can independently identify forest fires, track illegal fishing, or adjust their own orbits to avoid space debris without waiting for instructions from ground control. By moving the intelligence to the data source, NVIDIA is effectively eliminating the bandwidth bottleneck that has limited space exploration for decades.