AI consumes enormous power, solved with solar power from outside the Earth
Nvidia’s ‘H100’ goes to space… SpaceX, low launch cost is key
[실리콘 디코드] Google and NVIDIA compete for ‘space data center’… AI hegemony, on track
View enlarged image
NVIDIA and Google are quickly moving the stage of competition to outer space orbit to solve the problem of rapidly increasing computational demand in the AI era. DigiTimes, an IT media outlet, reported on the 9th (local time) that just two days after NVIDIA launched its flagship GPU (graphics processing unit) into orbit, Google announced the introduction of TPU (tensor processing unit) chips, sparking the race to build a gigawatt (GW) scale ‘space data center’.
On November 2nd, local time in the United States, NVIDIA made history by successfully sending its powerful H100 GPU into orbit for the first time ever, making AI infrastructure available in space. Just two days later, on the 4th, Google CEO Sundar Pichai fired back through social media by announcing that Google’s TPU chip would soon head to space as well.
SpaceX founder Elon Musk responded to CEO Pichai’s announcement, saying, “It’s definitely a good idea.” SpaceX is expected to play a key role in Google’s space data center plan. The current trend of rocket launch costs falling exponentially is making Google’s ‘orbital computing network’ concept more than just a fantasy, but a realistic goal with economic sustainability.
Google overcomes Earth’s resource limits with ‘solar AI’
Google unveiled ‘Project Suncatcher’ on November 4th. The plan is to build the world’s first space-based data center powered by solar power using a constellation of interconnected satellites. The key is to efficiently supply the enormous power required for AI calculations with solar power in low Earth orbit (LEO) and form an AI integration network through optical communication between satellites to perform sustainable AI calculations 24 hours a day. The main goal of this project is to solve the problem of enormous energy consumption in space, which is considered one of the biggest challenges in AI technology.
Google’s strategy is to deploy a satellite network equipped with a custom-designed TPU processor in space and leverage falling launch costs to make it cost-effective. To verify this idea, Google has already partnered with Planet, a satellite imaging company, and plans to launch two prototype satellites into orbit by early 2027.
CEO Pichai said, “‘Project Suncatcher’ seeks to build a scalable space-based machine learning system by utilizing the abundant solar energy outside Earth,” and added, “The goal is to overcome Earth’s resource limitations through continuous solar-based AI computation in orbit.”
He also emphasized that early trial results are promising. As a result of Google’s simulation of low-orbit radiation conditions with a particle accelerator, ‘Trillium TPU’ proved strong radiation resistance and passed all tests without damage.
Google also presented economic feasibility through a related research paper. The analysis is that if Space
In other words, if launch costs drop below a certain threshold, space, rather than Earth, could become a more efficient space for data center deployment. Of course, the current launch cost is more than 10 times higher than this standard. Google’s prediction relies entirely on Space
NVIDIA, 5GW orbital center ‘fighting back’… Space construction within 10 years
Meanwhile, the mission to send the H100 GPU into space ahead of Google was carried out by ‘Starcloud’, a startup company supported by NVIDIA’s ‘Inception’ program. Starcloud is preparing ‘Starcloud-1’, a satellite-type data center weighing approximately 60 kg equipped with H100 GPU, and announced that it chose NVIDIA chips because they provide unrivaled performance in all areas such as training, fine tuning, and inference, which are essential for implementing data center-level computing power outside the Earth’s atmosphere.
StarCloud has huge ambitions to build an orbital data center of 5 gigawatts (GW), approximately 4 km long, by deploying NVIDIA’s next-generation ‘Blackwell’ GPU into orbit. They plan to dramatically improve energy efficiency and cooling performance compared to terrestrial data centers by using cooling using the natural vacuum of space and infinite solar energy. “Within the next 10 years, all new data centers will be built in space,” said StarCloud CEO Philip Johnston.
The competition between Google and NVIDIA to solve rapidly increasing computing demand and power problems in the AI era is leading to the construction of space solar power-based data centers. The industry believes that large-scale space orbital data centers will become core infrastructure for the AI industry in the future.
Global Economics Reporter Park Jeong-han park@g-enews.com
