Driven by the AI boom, numerous global tech giants such as OpenAI, Google, Microsoft, Meta, and Amazon are aggressively building new AI data centers. However, they are also facing energy supply dilemmas, even needing to build their own power plants to ensure energy supply. This has led major spacefaring nations like China and the United States, along with leading tech companies, to set their sights on space, which boasts virtually limitless solar energy, hoping to build data centers in space.
Space Data Centers Are About to Enter the Actual Deployment Phase
A space data center is a computing facility that builds square-kilometer-scale large spacecraft through in-orbit rendezvous and docking. Adopting a modular design, it is primarily deployed in the 700-800 km dawn-dusk orbit or even on the lunar surface. It aims to utilize the inexhaustible solar energy in space as a power source and leverage the low-temperature space environment to solve the critical cooling issue of data centers.
Since last year, the concept of space data centers has been gaining momentum. In particular, over the past few months, as Tesla and SpaceX CEO Elon Musk has publicly discussed the space data center plan multiple times, the concept has continued to heat up.
In May 2025, China's space computing satellite constellation, the "Three-Body Computing Constellation," was successfully launched and entered networking. Since then, 12 computing satellites have been in orbit for over half a year. The "Three-Body Computing Constellation" will also join hands with multiple satellite enterprises in 2026 to complete a constellation layout of over 50 computing satellites, further promoting the application and development of AI in space.
In October 2025, AI cloud service company Crusoe announced that it would partner with space data center startup Starcloud to send NVIDIA's AI chips into orbit, creating the world's first "space AI data center."
Cully Cavness, Co-founder, President, and COO of Crusoe, stated: "We believe space will ultimately be critical to the future of computing because it offers a new solution to a key scaling bottleneck for AI infrastructure—access to abundant, stable, and clean energy. Since its inception, Crusoe has focused on combining computing infrastructure with novel energy resources. By partnering with Starcloud, we will extend this energy-first philosophy from Earth to the next frontier: outer space."
It is reported that the first batch of H100 GPUs is scheduled to launch with satellites in November 2025, ushering in the true era of "space AI." However, no further information is available at this time.
In early November 2025, Elon Musk publicly stated on the "X" platform that SpaceX plans to deploy data centers in space by expanding its future Starlink V3 satellites, which feature high-speed laser links. Responding to discussions about the feasibility of building an autonomous space data infrastructure, Musk wrote, "SpaceX will do this," indicating that the idea of using upgraded Starlink satellites as the foundation for orbital data centers is part of the company's long-term roadmap.
In December 2025, Musk confirmed that SpaceX's space AI data center plan is one of the key rationales for SpaceX's initial public offering (IPO) in 2026. Musk explained that deploying data centers in orbit, or building satellites capable of carrying computing resources like AI chips, would require "a lot of cash," and the IPO would fund these ambitions.
Musk even looked further into the future, stating that to reach an annual capacity of 100 terawatts, it would be necessary to manufacture satellites on the Moon and launch them using mass drivers. Because the Moon's escape velocity is only 2,500 meters per second and it has no atmosphere, mass drivers could work perfectly.
On January 22, 2025, at the World Economic Forum (WEF) Annual Meeting in Davos, Switzerland, Musk once again pointed out that as the production of AI chips grows exponentially, the power supply has failed to keep up, which will affect the efficiency of AI data centers and hinder the training and deployment of AI models. However, deploying data centers in space can effectively solve the energy supply problem through solar power.
"Building solar-powered AI data centers in space is a no-brainer. You just point the solar panels at the Sun and the radiators away from the Sun, and then the only problem left is heat dissipation," Musk said. "In the next three years, the cheapest place to build AI infrastructure will be in space."
In mid-January 2026, the National Aeronautics and Space Administration (NASA) announced that it had successfully completed critical tests for the Power and Propulsion Element (PPE), a key module of the lunar Gateway space station, successfully verifying the normal operation of the element. The supporting solar arrays have also been completed, capable of outputting up to 60 kW of power, laying the energy foundation for subsequent long-term deep space missions in lunar orbit. This news has also added new heat to the recently hot topic of "space data centers."
It can be said that space data centers have currently moved from technical verification and are about to enter the phase of commercial constellation deployment.
What Are the Advantages of Space Data Centers?
1. Nearly Unlimited Physical Space
Compared to building data centers on the ground, which requires purchasing large tracts of land and constructing facilities, or renting large factory spaces—both requiring significant capital investment—space theoretically has no spatial limits. Ultra-large-scale facilities can be expanded in low Earth orbit or on the lunar surface without paying exorbitant land prices.
2. Highly Efficient Utilization of Solar Energy in Space
In space, due to the absence of an atmosphere and other obstructions, the solar radiation intensity is constant at 1360 W/m² (the solar constant), with a theoretical peak more than 4.5 times that of the ground. If solar panels are deployed in geosynchronous orbit, the annual power generation time ratio can reach over 99%, achieving nearly 24-hour continuous and efficient power generation (even for low Earth orbit satellites, which orbit the Earth every 90 minutes, the sunlight ratio exceeds 60%). Moreover, due to the vacuum environment, solar panels will not suffer from efficiency reduction caused by dust coverage. Combining solar radiation intensity and sunlight duration, the annual power generation per unit area of solar panels in space is about 8 to 12 times that of the ground (depending on the orbit type).
Currently, the typical power generation efficiency of the most advanced space solar cells can reach about 32%, meaning each square meter of photovoltaics in space can generate about 435 W of power. Assuming a total power conversion loss of about 13%, the usable power after conversion is 378 W/m².
In the entire process from solar power generation to powering GPUs/ASICs, a lot of heat is ultimately generated, which needs to be entirely dissipated by radiators. However, in space, heat can only be dissipated through radiation, and the efficiency of radiator panels cannot reach 100%. Additional power may be needed to drive extra cooling components. If connected to computing devices, there will be additional losses (power supply, voltage regulation, local temperature rise). The final actual effective usable power is estimated to be 250-330 W/m². Taking a median value, the electrical energy ultimately available for computing per square meter of photovoltaic panels in a space data center is about 300 W.
Assuming a small space data center requires 1 MW of usable computing power:
Required photovoltaic panel area: 𝐴=1,000,000/300=3333m²
This is equivalent to: a large 50m × 70m solar array, which is about 40% smaller than the solar panels of the International Space Station (ISS) (the ISS has about 2500 m²).
If a larger-scale data center, such as a 10 MW level, is to be deployed, it would require 33,333 m², which is still feasible and equivalent to the size of a football field. Current deployable solar structure technologies (such as ROSA) can completely achieve this.
3. Space Solar Cell Technology Already Meets the Requirements
In terms of space solar cell power generation technology, early satellites used silicon cells with an efficiency of only about 15%. Now, the mainstream is triple-junction gallium arsenide cells, which divide the solar spectrum into three bands and absorb them with different material layers. The maximum efficiency has exceeded 32%, becoming the standard choice for current high-power spacecraft.
Currently, space solar cells are also developing towards more junctions: four-junction, five-junction, and even six-junction cells are under research, aiming to further broaden the spectral absorption range, with theoretical efficiencies potentially breaking through 40%. Meanwhile, technologies like inverted metamorphic multijunction cells have improved material flexibility, laying the foundation for a new generation of lightweight panels.
In terms of solar cell structural design, compared to ground-based solar systems, space solar power is not affected by weather and does not require protective structures like glass. The design is developing towards flexible and lightweight arrays.
Currently, the structural design of space solar panels is built around two key factors: volume and reliability. Because the size of the panels affects launch costs, and long-term reliability requires resistance to the harsh space environment, including temperature variations, radiation, and micrometeoroid impacts. Since ground-standard designs produce solar panels that are rigid, bulky, heavy, complex to operate, and relatively more expensive to launch into space, with relatively limited power generation capacity, the solar arrays for the Gateway space station adopt a brand-new Roll-Out Solar Array (ROSA) design. Based on flexible solar panels and composite materials, it offers the advantages of compact design, affordability, and autonomous deployment.
Therefore, during launch, ROSA can be rolled up like a carpet for storage, significantly reducing volume and weight while maintaining a large surface area. Once in space, it can automatically deploy to form a large light-receiving structure, supporting high-power output requirements without increasing the launch burden. Additionally, ROSA is scalable and modular, with flexible DSS design to meet various mission needs. ROSA can be scaled down for small satellite applications or made extremely large for deep space missions.
△ Spacewalkers Shane Kimbrough and Thomas Pesquet appear tiny in front of the main solar panels of the International Space Station as they work to complete the installation of the roll-out solar arrays on the P-6 truss structure.
When deployed in space, traditional large solar arrays mostly rely on motor-driven deployment, which involves complex systems and requires additional power and manual control intervention. ROSA, on the other hand, utilizes the strain energy stored in composite materials, deploying automatically in a principle similar to a spring releasing energy. This reduces the number of mechanical components and improves overall deployment reliability, making it particularly suitable for long-duration, unattended deep space missions.
It is reported that although ROSA is smaller than traditional solar arrays, it offers excellent performance, with each panel capable of generating over 30 kW of power, depending on the size. ROSA uses high-efficiency solar cells, and future versions may even attempt to use concentrators to enhance performance. Additionally, the composite booms provide structural rigidity, capable of withstanding dynamic environments, multiple frequencies, and collisions with debris or micrometeoroids, providing reliability for long-term missions.
4. Potential to Significantly Reduce Current Cooling Components
The dark side of space approaches absolute zero (as low as -270°C). Therefore, deploying data centers in space eliminates the need for the massive energy-driven water/air cooling systems required by ground data centers, potentially reducing cooling costs.
5. Reducing Network Latency and Covering Global Edge Areas
A space data center is essentially a distributed supercomputer. Hundreds or thousands of satellites need to be interconnected at high speeds, just like servers in ground-based data rooms. Inter-Satellite Links (ISL), as the core physical layer technology for this architectural transformation, are no longer just an auxiliary means for satellite communication, but the backbone nerves for building an integrated space-ground information network.
Traditional satellite communication modes heavily rely on frequent interactions between satellites and ground stations (Gateways). Under this architecture, satellites are essentially signal repeaters in space, and their data transmission is limited by the geographical distribution of ground stations, visibility time windows, and the availability of spectrum resources. This "satellite-ground dependency" model struggles to meet global coverage needs, especially in oceans, polar regions, and geopolitically sensitive areas, where the deployment of ground stations often faces physical or political infeasibility.
The introduction of inter-satellite link technology has thoroughly broken this shackle. By establishing direct data transmission channels between satellites, ISL enables data to be multi-hop routed in space, transmitted from one end of the Earth to the other without ground transit. This not only greatly reduces reliance on ground infrastructure, achieving true global seamless coverage, but also leverages the advantage of light speed propagation in a vacuum to achieve lower end-to-end latency than ground-based fiber optic networks.
The physical layer implementation of inter-satellite links is mainly divided into two categories: Radio Frequency (RF) links and Optical (Laser/OISL) links. Although RF technology has mature applications in the aerospace field for half a century, with the exponential growth in bandwidth demand, the limitations of physical layer characteristics make optical communication (OISL) gradually become the inevitable choice for next-generation large-scale aerospace applications.
6. High Security and Business Continuity
Space facilities are difficult to physically destroy or maliciously infiltrate by humans, and are not affected by ground risks such as earthquakes, floods, or wars.
7. Is the Comprehensive Cost Only 5% of That on the Ground?
Although space data centers do not require cooling systems, sending the required core chips, solar panels, and many other essential components to space and assembling them into a sufficiently large scale is still very costly and time-consuming at present.
However, Musk believes that once SpaceX's Starship achieves full reusability, the marginal cost per launch could be only about $1 million (with a Starship payload of 100 tons). Calculated at a power density of 100 kW per ton, theoretically, launching 1 million tons of payload annually is expected to achieve 100 GW of space AI computing power per year.
As long as launch costs can be significantly reduced, factoring in the saved land and construction costs, cooling costs, energy costs (requiring only solar panels), and low maintenance costs, building data centers in space is obviously more cost-effective than on the ground.
According to Starcloud's calculations, the 10-year total cost of ownership (operating cost) for a 40-megawatt cluster is about $8.2 million, while the corresponding ground-based system is about $167 million.
Challenges Facing Space Data Centers:
1. Launch Cost Issues
Although rocket recovery is lowering costs, it still reaches thousands of dollars per kilogram. Even if SpaceX's Falcon 9 rocket reduces the low Earth orbit launch cost to about $1,500/kg (2023 data), the launch cost alone for a standard rack (about 500 kg) would be $750,000.
Referring to the historical data of the International Space Station (ISS): the space cost of deploying and maintaining a 1 kW power system is about $1 million to $2 million, while the construction cost for 1 kW in a ground data center is only $2,000 to $5,000 (a difference of over 500 times).
If Musk's Starship can truly reduce the launch cost per ton to $10,000, this problem will be easily solved.
However, the approximately $1 million cost per launch is just a "long-term aspirational goal" for Musk, and the prerequisites are: fully reusable launches (Super Heavy + Starship), extremely high reuse counts for the heat shield, mass production and maturity of engines with costs dropping by two orders of magnitude, and production scale reaching hundreds to thousands of Raptor engines per year. Obviously, these are long-term visions, not immediately achievable levels.
2. Radiation Resistance Issues
Electronic devices are easily interfered with or even destroyed by cosmic rays in the space environment. Therefore, for chips to work normally in space, they need to achieve radiation resistance, resulting in extremely high manufacturing costs. Currently, the price of space-grade radiation-hardened chips can be 10 to 100 times that of similar ground products.
Specifically, the selection of space AI chips needs to fully consider the impact of high-energy radiation. The technical routes are divided into two types:
One is to use special semiconductor processes (such as SOI - Silicon on Insulator) and hardened designs (such as triple modular redundancy circuits) to resist radiation from the physical bottom-level design. However, the disadvantages are also obvious: the process is relatively backward, with a large gap compared to current AI computing chips.
The other is to improve software and hardware fault tolerance based on existing chips, which may also be the mainstream route for space data centers. Directly using high-performance computing chips from the ground to solve radiation problems through software algorithms and system architecture. Technically, by increasing the number of chips through redundant design to compute simultaneously (TMR), once a chip's result is inconsistent, it is restarted.
Additionally, utilizing the inherent robustness of AI algorithms (neural networks are insensitive to a small number of weight errors) to tolerate bit flips. Meanwhile, memory scrubbing technology is used to regularly repair flipped bits. Furthermore, Google's Suncatcher project explicitly proposed using TPUs for on-orbit computing. Tests showed that its v6e TPU still had no hardware failures after experiencing a radiation dose equivalent to 5 years, although the memory (HBM) subsystem is relatively sensitive.
Of course, if the market demand for radiation-hardened chips is high enough, their costs are also expected to drop significantly.
3. Cooling Issues
On Earth, heat dissipation is mainly through conduction (contact with heat sinks) and convection (flow of air or liquid). In the near-vacuum of space, although the dark side temperature can reach absolute zero (as low as -270°C), without air, convection and conduction barely exist. The only way heat is lost is "thermal radiation" — that is, equipment radiates heat directly into the universe in the form of infrared rays. This process is relatively slow, and the efficiency depends on the object's temperature (proportional to the fourth power of the absolute temperature) and the surface emissivity (the darker and rougher, the higher the radiation efficiency). Therefore, its cooling design challenges are completely different from those on the ground.
Therefore, the heat generated by the chips inside the space data center must be collected through an extremely efficient internal cooling system (likely using pumped liquid metal or special coolant). Then, this heat needs to be transferred to external radiator panels. This process itself requires energy and complex piping. The area of the radiator panels must be large enough because the radiation cooling efficiency is relatively low. A high-power data center might require radiator panels as large as a football field.
However, there is also a radical cooling technology—the Liquid Droplet Radiator, which directly sprays hot liquid droplets into space. The droplets radiate heat during flight and are then collected by a collector after cooling. This method can achieve a huge surface-area-to-mass ratio and is a potential ultimate solution for GW-level cooling, but it faces risks of droplet loss and contamination.
Although these solutions will bring about cost increases, from a longer life cycle perspective, compared to the "air cooling" or "liquid cooling" solutions on the ground that require additional energy consumption to drive, the costs will still be lower.
4. Maintenance Issues
When a ground data center fails, it can be repaired immediately by ground engineers. However, repairing a space data center is obviously very difficult. Although most software issues can be solved through backup systems or remote networks, physical issues make it difficult to send personnel to space for repairs. Therefore, ultra-high reliability design and redundant hardware systems are required.
5. Backup Energy
Space data centers are powered by solar energy. If in low Earth orbit, they will be affected by orbital shadow periods. Therefore, they need to be equipped with backup energy storage systems to work continuously, which will also bring about cost increases.
6. Communication Limitations
In terms of communication, data centers have very high data throughput requirements. Although the introduction of inter-satellite link technology can solve communication between satellites, the data ultimately needs to serve the ground. The current space-to-ground communication bandwidth is limited (Starlink's single satellite bandwidth is about 20 Gbps, far lower than the Tbps level of submarine optical fibers).
The currently feasible solution is to use extremely narrow laser beams. Its advantages include extremely high bandwidth (reaching hundreds of Gb or even Tb per second), strong anti-interference capability, and no restriction by radio frequency spectrum regulations. However, this requires higher pointing accuracy (microradian level) and cannot penetrate clouds at all, necessitating the selection of clear-sky ground stations or the use of air/orbit relays. Therefore, it is necessary to combine laser inter-satellite links to form a space backbone network, routing data in orbit and only downloading it to the ground at optimal locations. On the ground, arrays of multiple distributed small antennas are used to replace single giant antennas to reduce costs and improve reliability.
Additionally, to reduce the data transmission bandwidth pressure between ground and space data centers, data can be sent to space in the form of physical media during satellite launches. When data needs to be downloaded to the ground, preprocessing and intelligent compression are required, downloading only the most valuable result data to fundamentally reduce the amount of data that needs to be transmitted.
7. Security Issues
The increase in space debris may trigger a chain reaction of orbital collisions.
Conclusion:
The concept of space data centers closely links two of humanity's most cutting-edge technological fields—AI and aerospace engineering—painting a grand blueprint of utilizing the universe's infinite energy and space to break through the bottlenecks of Earth's computing power and energy.
The advantages of space data centers are disruptive: nearly limitless solar energy, a naturally ultra-low temperature cooling environment, vast orbital space resources, and the potential to build a global seamless low-latency network. These advantages directly address the core pain points of ground data centers regarding energy, land, cooling, and network coverage.
However, astronomical launch costs, stringent radiation protection requirements, complex space cooling engineering, as well as the difficulties of on-orbit maintenance and space-to-ground communication bandwidth limitations, are all problems that need to be continuously overcome.
From current progress, space data centers have begun to gradually move from technical verification to the process of actual deployment. However, they still face many challenges in the deployment of massive computing power and have not yet formed an economic closed loop. Their current core value may lie in serving demands with "space-native" attributes, such as specific scenarios like on-orbit data processing and global edge computing coverage.
In the view of Zhizhixun, due to existing technical and cost limitations, the deployment of space data centers in the short term may still be limited to a relatively small scale (below 1 MW). Considering the reduction of energy consumption, cooling burden, and cost pressure, deploying ASIC chips that are more efficient for specific computing tasks in space data centers is far more suitable than deploying general-purpose GPUs.
As for whether space data centers can truly become the mainstream infrastructure for AI computing power in the future, it is not only a hardcore technological marathon revolving around cost, reliability, and efficiency but also an ultimate test of humanity's large-scale aerospace engineering operational capabilities. Although the ideal has already stepped into reality, the road ahead is destined to be filled with both stars and thorns.
Editor: Zhizhixun - Langkejian
Partial sources: NASA, Starcloud, weforum, Jinyuan Securities