Nvidia Debuts An AI Chip For Space-Based Data Centers

Orbiting data centers may sound like a moonshot, but Nvidia isn’t waiting for the concept to potentially take off. The company announced that it’s working on a chip designed to survive the rigors of space.

At Nvidia’s GTC event, the company revealed the Vera Rubin Space Module, which can run AI models, but from orbit. The module contains a GPU using Nvidia’s latest “Rubin” architecture, and promises to deliver an up to a 25-times performance leap from the H100 GPU, which arrived back in 2022.

But the chip seems engineered for specialized workloads on satellites or space stations, rather than one day processing your ChatGPT prompt from orbit. Nvidia noted the module can process data streams from “space-based instruments in real time,” creating a way to unlock “on-orbit analytics, autonomous scientific discovery, and rapid insight generation.”

The company didn’t announce a specific launch date for the chip. Nor did it mention SpaceX, the major player that’s been talking up orbital data centers. That company’s CEO, Elon Musk, is so bullish on the idea that he expects space-based data centers to eventually beat terrestrial data centers in both costs and efficiency in the near future. In addition, SpaceX has filed a regulatory request to operate up to 1 million satellites to support the orbiting data center project.

In contrast, Nvidia’s CEO Jensen Huang has been more cautious about space-based data centers and recently indicated that the concept needs more time to mature. “Well, the economics are poor today, but it is going to improve over time,” Huang said in an earnings call last month. Huang then mentioned that GPUs in space could excel at certain tasks, such as high-resolution satellite imaging; rather than relying on servers on Earth to process the imagery, a GPU on board a satellite could do so much faster.

It’s possible the Vera Rubin Space Module is all about laying the groundwork for a bigger business, but the current scope appears to be limited. At GTC, Huang pointed out that the company needs to overcome the challenge of cooling AI chips in orbit, where there is no air.

In the meantime, Nvidia noted it’s working on the Vera Rubin Space Module with partners including space-based solar-power developer Aetherflux, space-station maker Axiom Space, satellite-imagery provider Planet Labs, and a startup called StarCloud, which has been developing orbiting data centers.

Recommended by Our Editors

Last November, Starcloud launched an Nvidia enterprise GPU, the H100, into space using a test satellite. Starcloud was then able to successfully connect, train, and run AI models over the GPU. The company has since filed a request to launch up to 88,000 satellites.

Nvidia announced the Vera Rubin Space Module days after the company also began recruiting for an “Orbital Datacenter System Architect.”