How much capital would a space-based data center require to become cost-competitive with its terrestrial counterparts? Deutsche Bank has conducted a detailed financial analysis.
On July 14th, analysts including Edison Yu from Deutsche Bank released the fourth report in their "Space Data Center Series," presenting a model to assess the economic viability of SpaceX's proposed Orbital Data Center (ODC).
Edison's calculations indicate that the current cost of a space data center is six times that of a ground-based facility. However, this premium could shrink to just 1.2 times following the deployment of AI1 satellites by 2029, with the potential to fall below ground costs by 2032.
Current Cost Disparity: A Six-Fold Difference
Referencing analysis from Epoch AI, Deutsche Bank outlines a baseline. Deploying 1 GW of AI computing power on Earth would require an initial capital expenditure of approximately $38 billion, with annual operational costs (covering power, maintenance, labor, etc.) around $900 million, leading to a total five-year cost of roughly $42.5 billion. Within this, computing hardware (like GPUs) accounts for about $21 billion, while non-computing infrastructure (facilities, cooling, power) makes up the remaining $21.5 billion.
NVIDIA CEO Jensen Huang recently noted at GTC Taipei 2026 that a new 1 GW "AI factory" could cost close to $100 billion, with roughly half tied to computing resources.
Based on the bank's model, deploying a data center of equivalent scale in space using current rocket and satellite designs results in non-computing costs that are approximately six times higher than those on the ground.
The primary drivers of this gap are two-fold: launch expenses and the satellite hardware itself.
Using a generic satellite design in 2027, with an assumed launch cost of about $1,429 per kilogram and non-computing satellite hardware costs around $50,000 per kilowatt, launching a constellation of 100 satellites would push the bill for these two components alone to $115 billion. In contrast, the equivalent ground-based infrastructure would cost only about $20 billion.
The Path to Cost Parity Within Five Years
The key variable for closing this gap is the Starship launch system.
Edison Yu's projections for Starship's cost trajectory are as follows: approximately $4,933 per kilogram in the early phase without booster reuse; dropping to $398 per kilogram with partial reuse; reaching $170 per kilogram with full reuse; and ultimately targeting $32 per kilogram with a fully reusable system achieving rapid turnaround.
By 2029, when the AI1 satellites are expected to be deployed, launch costs are forecast to fall to $398 per kilogram. Concurrently, non-computing satellite hardware costs are projected to drop to about $13,000 per kilowatt. This would bring the total non-computing deployment cost to around $27 billion, very close to the $23 billion estimated for ground-based infrastructure—reducing the cost multiplier from 6x to just 1.2x.
By the AI2 satellite phase in 2032, with launch costs potentially falling further to $170 per kilogram, the total non-computing cost could be approximately $15 billion, undercutting the projected $25 billion ground cost and resulting in a 0.6x cost factor. In the subsequent AI3 phase, with a target launch cost of $43 per kilogram, total costs could fall to about $9 billion, less than one-third of the ground-based estimate.
This cost curve is contingent upon SpaceX successfully advancing Starship's launch frequency and reusability as planned. This is why the report identifies SpaceX's "extreme vertical integration" as the most critical execution variable.
AI1 Satellite Computing Capability: Not Yet at Target
SpaceX plans to begin prototype deployment as early as the end of next year. FCC filings indicate the Starmind constellation, a massive low-Earth orbit satellite network for AI being developed by SpaceX, could ultimately scale to 1 million satellites.
Each AI1 satellite is designed for a power draw of 120-150 kW, with a stable operating power consumption around 120 kW—roughly equivalent to the power usage of a single NVIDIA GB300 NVL72 server rack.
SpaceX's target computing power density is 100 kW per ton. However, the initial AI1 satellites are expected to achieve only about 70 kW per ton. This means each Starship launch (assuming an 85-ton payload capacity) could deliver approximately 6 MW of computing power to orbit.
The model assumes density will improve to 85-90 kW per ton in the 2032 AI2 phase, with the 100 kW/ton target being reached in the AI3 phase and beyond.
The satellites are designed to be compatible with various chips, including NVIDIA GPUs, Google TPUs, Amazon Trainium processors, and Tesla's AI chips, which emphasize energy efficiency.
Thermal Management: A Fundamental Engineering Challenge in the Vacuum of Space
Cooling terrestrial data centers is relatively straightforward, using fans, air conditioning, and liquid cooling with air or water as mediums.
This is not possible in space. In a vacuum, heat can only be dissipated via thermal radiation, governed by the Stefan-Boltzmann law, where efficiency depends on temperature and surface area.
Existing satellites almost exclusively use passive radiators, which rely on material properties and geometry to conduct heat, consuming no power but limited by the satellite's physical size. Active radiators, which use pumps to circulate coolant, can handle higher thermal loads but consume power and carry a risk of mechanical failure—currently used only on structures like the International Space Station, China's Tiangong station, and crewed spacecraft such as Dragon.
The AI1 satellites will employ a dual-sided, actively deployable liquid-cooled radiator system. Each side has a dissipation capacity of 700 W/m², for a total of 1,400 W/m² across a combined surface area of 110 square meters. This design is necessary to manage the 120-150 kW high-power payload, which a purely passive system could not handle.
SpaceX is poised to become the first company to mass-produce such an active radiator design. Projections indicate that with mass production, radiator costs could fall from $8,000 per square meter in 2027 to $1,000 per square meter by the AI3 phase.
Solar Power: In-House Manufacturing with a 100 GW Ambition
Where will the satellites get their power? From solar energy.
An AI1 satellite requires approximately 600 square meters of solar panels. Initially, these will use silicon-based cells with an efficiency around 19%. In the longer term, Heterojunction (HJT) cells, with efficiencies up to 27%, offer benefits like bifacial light absorption and strong radiation resistance. Perovskite thin-film cells have theoretical efficiencies rivaling multi-junction cells and can be printed, offering ultra-lightweight properties.
SpaceX has begun construction on a solar cell factory in Bastrop, Texas. The facility has a planned capacity of 10 GW (5 GW per floor) across an area of about 1.1 million square feet, co-located with the existing Starlink production site. Construction started in late March 2026, equipment installation is underway, and the target is to achieve mass production by the end of 2027.
Elon Musk's broader ambition is to establish 100 GW of domestic solar cell manufacturing capacity in the United States within three years.
Communication Architecture: Optical Lasers, Avoiding Radio Spectrum
The AI1 satellites will not carry complex phased array antennas. Inter-satellite communication will rely entirely on Optical Inter-Satellite Links (OISL). Data will be routed within the Starmind constellation, then connected to the Starlink laser mesh, and finally transmitted down to Earth via ground stations.
The advantage is that the Starmind constellation itself will occupy almost no radio frequency spectrum.
The trade-off is that all data must eventually pass through Starlink ground gateways. The orbital data center will significantly alter data flow patterns. Starlink's existing gateway authorizations were designed for a consumer broadband model dominated by "downlink" traffic, which is ill-suited for the "heavy uplink" demands of AI inference workloads.
To address this, SpaceX is expanding gateway backhaul capabilities into higher frequency bands, including the E-band (already in use), V-band, W-band, and the D-band (proposed), which is primarily targeted for AI services. These higher bands have historically been considered unsuitable for consumer terminals due to challenges like rain fade and oxygen absorption. However, for high-capacity gateway stations equipped with large antenna arrays, these issues can be mitigated through site diversity and optical routing.
This year, the FCC also updated its satellite interference protection standards, replacing 1990s-era Equivalent Power Flux Density (EPFD) limits with performance-based rules. This change allows operators to deploy higher-power satellites and more co-frequency satellites in key Ku/Ka bands, theoretically increasing capacity by up to seven times for a given number of satellites.
Comments