Artificial intelligence (AI) data centers require large land areas, high power, produce noise pollution and require a lot of water. The solution to these issues is to build data centers in space. Next November, Starcloud, a startup company under NVIDIA Inception, will launch the first satellite equipped with data center-class GPUs as the first step towards building a data center in space.
The Starcloud AI satellite, which is the size of a mini-fridge and weighs 60 kilograms, will be equipped with NVIDIA H100 GPUs. This gives it 100 times the processing power of other systems in space today. Starcloud founder Philip Johnston predicts that within 10 years, all AI data centers will be built in space. In the video shared, Starcloud’s AI data center has a power of 5GW and it requires solar panels covering an area of about 16 square kilometers.
His reasoning is that space offers unlimited power because it only requires solar power. No water will be wasted for cooling systems because the heat generated will be released directly into the freezing cold of space. He added that the lifetime carbon emission reduction of a space data center is 10 times lower than that of a data center on the ground.
The idea of building a data center in space is not a new idea, with China having already started construction of its first space supercomputer in May this year. Under the Star Compute program, 2,800 AI satellites will form the Three-Body Computing Constellation. When fully operational, it will be a network of space supercomputers with the capacity to process 1,000 Maps of Operations per Second (POP).
The Starcloud team is now focusing on launching their first AI satellite, which will then run Google’s Gemma AI model to prove that LLM can be processed in space. If successful, Starcloud will be the third to successfully do so, after Booz Allen and HPE on the International Space Station (ISS) in 2024 and China on the Taingong Space Station in August.
