Recently, a report about Beijing accelerating the layout of space data centers has attracted widespread attention from the industry. According to reports, a team from Beijing Xingchen Future Space Technology Research Institute is exploring the deployment of large-scale AI computing infrastructure in space orbit, attempting to provide new solutions for artificial intelligence computing using the special environment of space, and has released a plan for the solution.
At first glance, from ground data centers to space data centers, this once unattainable idea is quietly becoming a reality. Upon closer examination, in the face of the many unknowns in the space environment, is this a revolutionary strategic layout or a unrealistic technological fantasy? Still awaiting verification.
01
Building data centers in space
What are the advantages?
The exploration of the universe by humans is less than one percent, but the idea of building data centers in space has a long history and meets industry expectations. A senior industry insider told reporters that initially the industry discussed space computing power, with a focus on the fact that space does not require large refrigeration systems.
Especially in recent years, with the rapid development of artificial intelligence, data centers represented by large intelligent computing centers are facing difficulties such as land resource scarcity, power grid pressure, and massive consumption of water resources. The International Energy Agency report shows that by 2030, the electricity demand for global data centers is expected to more than double, and artificial intelligence will be the main driving force behind this surge in electricity consumption.
So, is space suitable for building data centers as a new way out of the computing power dilemma on the ground? After researching, the author found that there are many theoretical advantages to building data centers in space.
Firstly, as a huge energy source, the sun can release 3.86 × 10 ㎡⁶ watts of energy per second, and the loss of solar energy in space is smaller. Deploying data centers in a synchronous orbit during dawn and dusk, solar panels can work 24 hours a day without interruption, and the efficiency can reach 8 times that of the ground.
Secondly, data centers deployed in space can adopt closed-loop cooling systems that dissipate heat directly into outer space through infrared radiation, without the need for water resources.
Finally, space has spatial advantages that are not limited by limited land resources.
02
Large scale AI computing power in space
Is it trustworthy?
Even though deploying data centers in space has obvious advantages, the public's focus is still on whether the technology can be achieved. In response to this, Beijing Xingchen Future Space Technology Research Institute also gave a positive answer and announced a specific construction plan, which is to deploy computing power star clusters in the morning and evening orbits 700-800 kilometers above the ground, and build and operate a centralized large-scale data center system with a power exceeding GW (gigawatts).
In terms of composition, the system specifically includes spatial computing power, relay transmission, and ground control subsystems. Among them, space computing power, as the operational part placed in space, is planned to deploy multiple space data centers, each with a power of about 1GW, capable of accommodating server clusters at the level of one million cards, to carry out space-based data relay transmission and computing services.
Overall, AI computing power in space has become an artificial intelligence industry disguised as aerospace. Moreover, in order to demonstrate that this exploration is not a 'castle in the air', the team also provided a specific timeline for the next decade.
Specifically divided into three stages, steadily advancing through the path of "verification first, improvement later, and scale later". From 2025 to 2027, breakthroughs will be made in key technologies such as energy and heat dissipation for space data centers, and experimental satellites will be iteratively developed. The first phase of the computing power constellation will be constructed, with a planned total power of 200kW and a computing power scale of 1000POPS, achieving the goal of "day to day calculation" application; From 2028 to 2030, breakthroughs will be made in key technologies such as in orbit assembly and construction of space data centers, reducing construction and operation costs, constructing a second phase computing constellation, and achieving the goal of "Earth Day Computing" application; From 2031 to 2035, achieve large-scale mass production and network launch of satellites, establish a large-scale space data center through in orbit docking, and support future 'space-based computing'.
A series of plans demonstrate the team's ambition to explore new tracks and expand this field into a large-scale comprehensive industry that runs through new materials and integrated circuits, satellite communication, and aerospace. While bringing new directions to industry chain enterprises, it also greatly increases the uncertainty of industry integration.
03
Not an isolated case
Space computing power has already become a new track
Although space data centers have significant theoretical advantages and clear paths are currently planned, there is still a gap between ideal and reality in moving from "drawings" to "orbit".
It is worth mentioning that space computing power has been favored by the industry in recent years. Looking at the world, technology giants, startups, and traditional aerospace companies have been competing to enter space data centers, initiate technology verification and industrial practices, and actively seize this emerging market.
SpaceX founder Musk stated that the large-scale deployment of solar powered AI satellites is the only path to achieving an annual deployment of 1 terawatt of AI computing power. SpaceX's Starlink V3 satellite and Starship rocket will create conditions for the deployment of AI satellites. Google has launched the "Sun Catcher Program" and plans to launch two prototype satellites equipped with TPU chips in early 2027 to verify laser communication and radiation resistance capabilities. The goal is to achieve 615MW of computing power by 2029. Amazon, on the other hand, is using the "Kuiper Plan" to attempt to replicate the cloud computing experience of ground-based AWS into space, with the goal of deploying 3.2GW of orbital computing power by 2030.
Startup Starcloud has also partnered with SpaceX to launch a satellite carrying Nvidia H100 GPU and Google Gemini large model in November this year, completing the first data center level GPU in orbit test.
On the Chinese side, in addition to the AI computing power sky plan proposed this time, Beijing Zhongke Tianhua Technology Co., Ltd. has taken the lead in exploring the field of "supercomputing sky". As early as 2022, the Sky Computing team led by its founder had launched the Aurora 1000 spaceborne intelligent machine equipped with domestically produced high-performance AI chips into space. It has been running stably in orbit for more than 1000 days, successfully verifying the possibility of a high-performance AI computing system going into space.
In May of this year, China Aerospace Science and Technology Corporation and Zhijiang Laboratory launched 12 computing satellites worldwide, with plans to build a computing power network of 100 satellites by 2027.
Frequent industrial actions not only verify feasibility, but also indirectly reflect the value of this emerging track.
04
Practice landing on high mountains, long road ahead
Is it really feasible?
In the long run, the industry still needs to face common issues such as severe technological challenges in extreme environments, high research and development costs for satellites, and how to realize commercial value.
As Wang Kun, Director of Infrastructure Research at Omdia Data Center, pointed out in an interview, in terms of implementation, actual deployment still faces multiple obstacles. Firstly, there is the difficulty and cost of launch. Currently, the cost per kilogram of payload for satellites is still relatively high; Next is power supply, as solar panels combined with energy storage are difficult to meet the power requirements of modern data centers. Not only that, cooling is also a problem. In space, heat dissipation can only be achieved through thermal radiation, and it becomes even more difficult to dissipate heat when exposed to direct sunlight.
The most important thing is radiation protection. At present, the cost of aerospace grade chips is very high, and chips on the ground may encounter problems in space due to various radiations. Compared to others, data uploading and returning are relatively easy to solve problems, "said Wang Kun.
Apart from technical validation and cost assessment, the implementation of scenarios has not been clearly defined. It can be foreseen that in addition to supporting AI training, space data centers will also have the potential to open up new scenarios such as real-time Earth observation, deep space exploration, and real-time disaster warning.
In this regard, the author believes that the implementation of scenarios should also avoid "technology for the sake of technology". The early exploration of the industry should focus on demand scenarios that cannot be replaced by the ground, and rationally explore and develop based on scenario priorities to avoid resource waste. Especially in the current situation where the supply and demand of intelligent computing centers have not been well matched, exploring technological directions from the demand side is more suitable for industrial development.
Overall, the current space computing industry is still in a stage where technology is feasible but not economical, and scenarios exist but are not clear. In consideration of long-term development, it is similar to the competitive logic of satellite Internet, and its essence is the competition of "orbit resources+technology ecology". If you want to take the initiative in this competition, you need to avoid blind investment and take the rational path of "heaven and earth coordination, scene focus".
Therefore, returning to its roots, the "sky high" of AI computing power is not a question of "whether or not to do it", but a question of "how to do it rationally".
