太空数据中心？

15 December 2025DateIndustry Update IndustrialsSpace Technology Data centers in space? Edison Yu Following SpaceX confirming interest in an IPO next year, reportedly seeking a $1.5trillion valuation, and Elon Musk expressing strong support for deploying datacenters in space, we start taking a closer look at the concept. In short, there areclearly technical challenges to making this a viable endeavor but these seem to beengineering constraints as opposed to physics. Moreover, we are encouraged bythe fact that Google, OpenAI, and Blue Origin are all seemingly exploring ways todo this. Google’s Project Suncatcher is aiming to launch prototype satellites in 2027 Research Analyst Benjamin Black, CFAResearch Analyst Roshan Ranjit, CFAResearch Analyst Research Associate+1-212-250-2266 Satellite basics For background, a satellite is typically made up of two main parts:busandpayload.The bus is the main structural framework and support systems of the satellite. It actsas the "vehicle" that enables the satellite to operate in space, maintain its orbit, andsurvive the harsh environment of vacuum, radiation, and extreme temperatures.Key components include the physical frame, solar panels, thermal managementsystem,propulsion,and optical terminals.The payload is the specializedequipment or instruments that carry out the satellite's primary mission. For a data Why put these in space? With AI driving up demand for compute, Earth-based data centers appear to beencountering certain structural bottlenecks, mainly energy, cooling, and latency. Energy– at the right orbit such as in dawn-dusk SSO (sun-synchronousorbit), solar panels can harness the sun’s free power 24/7 and importantly,garner 40% higher intensity than on the Earth’s surface due to a lack of anatmosphere filtering/scattering sunlight. As such, operators can generate irradiance)and avoid dealing with expensive/complex power grids nCooling– represents a large burden, accounting for an estimated 40% ofenergy consumption, having to use large amounts of water/piping. In fact,Nvidia’s Jensen Huang recently commented for a 2-ton GPU rack, 1.95 tonis cooling mass. In space, cooling will require attaching a passive radiatoron the dark side of the satellite (i.e., part not facing the sun) which can dump nLatency- optical laser links traveling through vacuum are faster than fiberoptic cables on the ground, potentially by >40%. This is due to refractiveindex of glass and indirect path of travel for cables. In relation, satellites areincreasingly generating data in space (e.g., imagery, weather data, climatemonitoring). Currently, the data is downlinked to Earth for processing,which is slow and bandwidth-heavy. By having the data centers close to the Separately, another intriguing reason is sovereignty or even security. We havewitnessed many local communities protesting data centers being built in theirlocales and there could be geopolitical concerns in some regions. Terrestrial data What are the challenges? As we alluded to, while the rationale make a lot of sense in theory, there are severalchallenges to overcome both in terms of cost and engineering. commercial launch has a sticker price of ~$70m and let us assume grossmargin on that is ~40%, the cost is still likely around $30m, implying a $/kgcost of around $1.5k. According to Google’s Project Suncatcher white paper, launch costs would need to break below $200/kg to be viable whichin our view requires SpaceX Starship to be operational and launching on a compared to cooling on the ground. While space is cold, it is a vacuum orperfect insulator, meaning heat can only be removed via radiation (slow), asopposed to convection (fast, like a fan blowing air). GPUs are extremelypower-dense, generating heat in small, concentrated spots. To properlycool a large AI cluster, a DCS would require massive passive radiator panels.Therefore, we think some type of breakthrough is required in the radiator high-energy protons constantly bombard the satellite. Interestingly, theGoogle white paper discovered that based on simulations, TPU logic coreshandled radiation quite well but HBM began showing errors at much lowerdoses. There are some simple solutions for this such as wrapping theservers in heavy lead or aluminum. However, this would naturally add mass satellites may need to be upgraded with higher “space-grade” hardware toensure longevity, driving up costs. While there have been proposals fororbital transfer vehicles (OTVs) to provide maintenance, our view is the costof building one capable of carrying advanced maneuvers is too expensive(e.g., need robotic arm essentially to swap out components). Starlink bold future vision We forecast Starlink breaking 9 million subscribers exiting 2025E, representing adoubling YoY, once again demonstrating strong momentum and this excludes anyD2D users (via MNOs). Looking ahead, Musk indicated Starlink will develop amodified version of the V3 satellite for data centers. Recal