Startup Starcloud Makes History with 'Large-Scale Model Training in Space Orbit'
2025-12-11 / Read about 0 minute
Author:小编   

Last month, orbital data center startup Starcloud made a groundbreaking move by launching a satellite. This satellite was outfitted with NVIDIA H100 chips, marking humanity's first-ever endeavor to train large language models in the vast expanse of space.

At present, the satellite is actively running applications that are based on Google's open - source model, Gemma. It has successfully trained the LLM NanoGPT using the complete collection of Shakespeare's works.

Starcloud has ambitious plans. It aims to construct a 5 - gigawatt orbital data center. The primary goal is to slash the energy costs of orbital data centers, making them 10 times more cost - effective compared to their ground - based counterparts.

Through the in - orbit operation of Starcloud - 1, the company will conduct a thorough assessment to verify the feasibility of building a space - based data center. The rationale behind sending NVIDIA chips into space is rooted in the limitations of ground - based power capacity. In contrast, space - based data centers have the potential to harness solar power more efficiently.

Currently, the 'space computing power satellite' has already commenced processing commercial services. This enables real - time intelligence processing, a significant leap in data processing capabilities.

Looking ahead, the satellite that Starcloud plans to launch in October 2026 will be a powerhouse. It will be equipped with multiple NVIDIA H100 chips and the next - generation Blackwell platform. Meanwhile, the satellite scheduled for launch next year will be fitted with cloud platform modules.

Starcloud - 1 also serves an important purpose. It will verify the operational risks associated with orbital computing power satellites. These satellites typically have a lifespan of approximately 5 years.

Despite the myriad of uncertainties, numerous tech companies are still fervently promoting the development of space - based data centers. Google, for instance, has plans to launch its self - developed TPU into space for operational use. On the other hand, Lonestar Data Holdings is concentrating its efforts on establishing a commercial data center on the Moon.