SpaceX is Trying to Build a Data Center in Space | The Concept and the Barriers Ahead
The fact that SpaceX is seriously trying to build a "space data center" has been a topic of discussion in parts of the industry for some time. Its outlines were visible through FCC (U.S. Federal Communications Commission) filing documents and sporadically emerging technical information.
In May 2026, it was officially organized for the first time in the form of an IPO prospectus. Business structure, financial figures, risk perception—with these gathered in one document, it finally became possible to discuss the overall picture. This article deciphers that prospectus and related materials to explain what SpaceX's Orbital Data Center System (hereinafter "Orbital DC") concept is and what makes it difficult.
Terrestrial data centers are approaching their limits
First, from the background.
The power required for AI training is surging. There is a forecast that the power demand for global AI data centers will reach 347 GW by 2030. Power plants to supply that much electricity, water for cooling, land, and power grids. Securing these is rapidly becoming difficult on the ground.
SpaceX's concept starts from here. In space, power is infinite. Land and water are unnecessary. Could we build compute infrastructure in orbit that is freed from terrestrial constraints?
The future market size indicated by the prospectus is $28.5 trillion, of which 93% is tied to the AI segment. SpaceX has applied to the FCC for authorization to launch up to 1 million satellites with computing capabilities. Since the current Starlink constellation is about 10,000 satellites, that is 100 times the scale.
What does it mean to run AI in orbit?
The basic concept of an Orbital DC is to deploy a large number of satellites equipped with computing chips into orbit to form a distributed AI cluster.
The computing chips to be mounted are not limited to GPUs. While GPUs are often cited as the representative example, the prospectus assumes AI-optimized computing semiconductors in general. Not only general-purpose GPUs like Nvidia's H100, but also TPUs developed by Google and future custom ASICs (Application-Specific Integrated Circuits) are candidates. In fact, there is high demand for custom chips optimized for the harsh environment of space, and there are many undecided parts regarding who will handle their design and manufacturing.
The satellites are scheduled to be placed in Sun-Synchronous Orbit (SSO). The prospectus states "SSO," but considering the content of the application documents, the target appears to be a Dawn-Dusk orbit (an orbit that flies near the boundary line where constant sunlight is obtained between sunrise and sunset). As will be described later, the reason for choosing this orbit lies in power.
Communication between satellites will be conducted via optical links (lasers), and SpaceX is aiming for an inter-satellite communication speed of 1 Tbps with the next-generation Starlink. Data exchange with the ground will utilize the Starlink ground station network. The round-trip latency for sending inference requests to orbit and returning processing results will realistically be around 20 to 50 ms.
Why a Dawn-Dusk orbit? The story of power
The power source for an Orbital DC is sunlight. "Which orbit to choose" becomes the core of power design.
Normal low-orbit satellites spend about 30 minutes of each orbit (about 90 minutes) in the Earth's shadow (eclipse). While there is no sunlight, it must be supplemented by batteries. Since computing chips consume large amounts of power continuously, the battery weight would be enormous in an orbit that frequently enters an eclipse.
If you choose a Dawn-Dusk orbit, the satellites continue to fly near the boundary line of sunrise and sunset, and the time spent in the Earth's shadow is significantly reduced. Depending on altitude and season, it is possible to maintain almost constant sunlight throughout the year. This is considered the reason why SpaceX chose this orbit.
However, this orbit is also an orbit favored by SAR satellites (Synthetic Aperture Radar satellites) for Earth observation. SAR is a satellite that can observe the Earth's surface regardless of day or night, and because it performs radar transmissions that consume large amounts of power continuously for a certain period, it is concentrated in this orbit where sunlight is less likely to be interrupted. ICEYE, Capella, Umbra, Synspective, and Japan's ALOS-4 also use this orbit. This is the background to the orbital congestion problem described later.
The difficulty of cooling chips in space
Terrestrial data centers are cooled by air or water. They spin fans and circulate chilled water. Neither can be used in space.
Outer space is a vacuum. Since there is no air, convective cooling is impossible. Also, heat conduction does not work unless materials are in contact with each other, but there is no material in contact with the outside of the spacecraft. There is only one remaining means: radiative cooling.
Radiative cooling is a phenomenon where an object sheds heat by emitting electromagnetic waves (primarily infrared). Since outer space is a cryogenic "heat sink" at approximately 3K (-270°C), it is theoretically possible to discard large amounts of heat. However, because the amount of heat dissipation is proportional to the area of the radiator panel and its surface temperature, a large-area panel is required.
To provide a reference value, the International Space Station (ISS) dissipates a maximum of 70kW of heat using 422 square meters of radiator panels. That is 166W per square meter. If SpaceX aims for a computing payload of 40 to 100kW per satellite, it is calculated that hundreds of square meters of radiator panels would be required for each satellite alone.
Furthermore, there is a fundamental contradiction. Solar panels must be kept pointed toward the sun to generate power. On the other hand, radiator panels must be pointed toward the cold outer space away from the sun, because they would absorb heat instead if they received sunlight. In other words, panels that need to face the sun and panels that need to face away from the sun must coexist on the same satellite. To resolve this challenge where deployment directions are fundamentally opposed, it is necessary to design large radiator panels combined with spectral selective coating (a surface treatment that reflects sunlight while strongly emitting infrared rays) and orbital deployment and attitude control.
How SpaceX solves this problem is one of the most difficult technical challenges.
Radiation issues: Where is the difficulty?
Another major challenge when using AI computing chips in space is radiation.
First, to clarify, the low Earth orbit (LEO) altitude of about 500 to 600 km is also an area where NewSpace (emerging space industry) companies operate Earth observation satellites equipped with COTS (commercial off-the-shelf) components. While it is not completely free of radiation, it has been proven that COTS products can be used with ingenuity in design and testing if the mission duration is about 1 to 3 years.
The problem is that the situation is different for AI computing chips.
First, the operating density and cumulative exposure of the chips are completely different. Unlike sensors and communication equipment, AI computing chips are designed on the premise of constant full-load operation, and radiation exposure accumulates during long-term high-load operation.
Second, process miniaturization increases vulnerability. The fine processes adopted by state-of-the-art computing chips have extremely small circuits, making them prone to bit flips (a phenomenon where memory 0/1 is inverted) caused by radiation. Note that Google has radiation-tested its own TPU (Trillium generation) with a 67MeV proton beam irradiation device and published the results in a paper ("Towards a future space-based, highly scalable AI infrastructure system design", November 2025). While sensitivity to the High Bandwidth Memory (HBM) subsystem was confirmed, relatively good results were obtained, with no serious failures occurring up to a dose equivalent to three times the assumed exposure for 5 years. On the other hand, this test is only an evaluation under specific generations and conditions, and reliability under long-term operation remains a research topic.
There are countermeasures. These include bit flip detection and correction using ECC (Error Correction Code) memory, passive shielding using thin layers of water or aluminum, and fault tolerance using TMR (Triple Modular Redundancy). However, all of these involve trade-offs in cost, weight, and power consumption.
Launch costs: The world changed by Falcon 9, but still not enough
One of the reasons SpaceX was able to deploy Starlink in large numbers is that it dramatically reduced the cost of its own Falcon 9 rocket. By realizing the recovery and reuse of the first-stage rocket, the price of approximately $3,600 per kg has disrupted the conventional commercial launch market.
However, this is still not enough for orbital data centers to be profitable. To continue deploying and updating high-performance satellites equipped with computing chips, radiator panels, and solar arrays on a scale of 1 million units, it is said that launch costs must be reduced to $200/kg or less. That is about 1/18th of the current cost.
Starship is what will realize this goal. Both the first and second stages are fully reusable, and it can put over 100 tons into low Earth orbit in a single flight. Theoretically, it is possible to reach $200/kg.
But the reality is harsh. Of the five launches conducted in 2025, the first three ended in explosions, and commercial operation certification has not yet been obtained. The entire plan for orbital data centers is predicated on the success of Starship, and this is the first critical path.
The problem of running out of orbit
Reading the application SpaceX submitted to the FCC reveals another major problem.
As mentioned above, the Dawn-Dusk orbit targeted by orbital data centers is already crowded. SAR satellite constellations such as ICEYE, Capella, Umbra, and Synspective are rapidly deploying, and commercial and government satellites from all over the world are gathering in this orbital plane at an altitude of 500 to 800 km.
Radio waves are also a problem. SpaceX has applied to the FCC for the use of the Ka-band, but the application method is an unusual one with the condition of "non-interference, non-protected basis." They are applying outside the normal Processing Round (a process where operators using the same frequency band wait in line for review) and are even seeking an exemption from the milestone obligation for the deployment of 1 million units.
ITU international coordination is also incomplete. Government satellites from Japan, Europe, and Canada are concentrated in the Dawn-Dusk SSO, and the other countries hold the initiative in interference coordination with these agencies. If negotiations break down, deployment could be halted for years.
The application to the FCC is essentially a 'pre-reservation of orbital slots and spectrum,' and there is no obligation at this stage to prove the technical feasibility of the satellites. This point must be taken with a grain of salt.
Categorizing the four risks
Let's organize the points discussed so far as risks.
Technical Risks: If Starship is not completed, the premise of launch costs will collapse. The physical barrier of thermal management cannot be eliminated, even if it can be mitigated through design ingenuity. While a path to solving the radiation tolerance issue for computing chips has emerged, establishing long-term operational reliability is still ahead. Furthermore, there is the 'obsolescence trap.' AI chip performance doubles roughly every two years, but the satellite design, manufacturing, and launch cycle is five to six years. On the ground, equipment can be replaced, but that is not the case in space.
Regulatory and Orbital Risks: The risk that FCC reviews and ITU international coordination will be prolonged. The risk that the Dawn-Dusk SSO interference issue will develop into an international political problem.
Economic Risks: If terrestrial power issues are solved first by nuclear (including SMRs) or geothermal energy, the motivation to go to space will diminish. SpaceX's AI segment (post-xAI merger) recorded an operating loss of $6.35 billion for the full year 2025, and a scenario where funds raised after the IPO are absorbed by the AI business, pushing investment in orbital DCs to the back burner, is also realistic.
Human Risks: While the prospectus admits that 'if Musk leaves, it will have a significant impact on the company's direction,' it does not even provide key person insurance. The fact that his attention is divided by his concurrent roles at Tesla, X, and xAI is also disclosed as a risk.
Even so, SpaceX is serious
After listing all these risks, you might think, 'It's just a pipe dream after all.'But my assessment is different.
SpaceX's terrestrial AI data center, 'Colossus' (Memphis, Tennessee), is already operating with over 220,000 Nvidia GPUs. Anthropic has signed a contract to utilize the full capacity for approximately $1.25 billion per month. One of the uses for the IPO proceeds (up to $2 billion) is explicitly stated as 'expansion of AI compute infrastructure.'
Money is moving, and there are customers. The timeline of small-scale orbital demonstration in 2028 and reaching the break-even point in the first half of the 2030s might be too optimistic, butthere is no doubt that SpaceX is moving toward that goal.
And as this concept advances, technologies that SpaceX does not possess in-house will become necessary. I will leave the discussion of those technologies for the next article.
Mamochichi @ A Corner of the Space Industry, June 2026
The analysis in this article is based on the SpaceX prospectus (submitted to EDINET, May 2026), FCC application documents (January 2026), and various public materials. Descriptions regarding the Dawn-Dusk orbit include considerations from the application documents. These are personal views and not the views of the organization I belong to.