Deutsche Bank models the cost of SpaceX's orbital data centers, forecasting cost parity by 2032.

Deutsche Bank has released a series of research reports on space-based data centers, analyzing the economic feasibility of SpaceX’s orbital data centers. The report finds that non-computing costs for current space-based data centers are six times higher than those on Earth, primarily due to high launch costs and expensive satellite hardware. As Starship launch costs decline (from $4,933 per kg to $32 per kg) and satellite technology advances, the cost multiplier is projected to fall to 1.2 by 2029 and potentially drop below ground-based costs by 2032. The AI1 satellite employs dual-sided active liquid cooling, and SpaceX has already begun construction of a 10 GW solar panel factory, targeting 100 GW of production capacity within three years.

Article author and source: Zhui Feng Trading Desk

How much money does a space data center need to spend to break even with a ground-based one? Deutsche Bank ran the numbers.

According to Zhui Feng Trading Platform, on July 14, Deutsche Bank analyst Edison Yu and others released the fourth report in their "Space Data Center" series, conducting a modeling analysis of the economic feasibility of SpaceX's Orbital Data Center (ODC).

Edison estimates that the current cost of space-based data centers is six times that of ground-based ones, but will be reduced to 1.2 times after the deployment of the AI-1 satellite in 2029, and could fall below ground-based costs by 2032.

Deutsche Bank, citing analysis from Epoch AI, assumes that the upfront capital expenditure for deploying 1 GW of AI computing capacity on the ground is $38 billion, with annual operating costs (including electricity, maintenance, labor, etc.) of approximately $900 million, resulting in a total cost of about $42.5 billion over five years. Of this, AI computing hardware (such as GPUs) accounts for approximately $21 billion, while non-computing components (infrastructure, cooling, power, etc.) amount to about $21.5 billion.

NVIDIA CEO Jensen Huang also mentioned at GTC Taipei 2026 that the cost of a new 1 GW "AI factory" could approach $100 billion, with approximately half of that related to computing power.

According to this calculation, deploying a space-based data center of equivalent scale using existing rockets and satellite designs results in non-computing costs approximately six times higher than on the ground.

Where is the difference? Two areas: launch costs and the satellite itself.

In 2027, using a general satellite solution, the launch cost is approximately $1,429 per kilogram, and the non-computing hardware cost of the satellite is about $50,000 per kilowatt. Launching 100 satellites would result in just these two costs totaling $115 billion, whereas an equivalent ground-based system would cost only about $20 billion.

The key variable is Starship.

Edison Yu assumes the following cost trajectory for Starship launches: approximately $4,933 per kilogram without spacecraft reuse;降至$398 per kilogram with partial reuse;降至$170 per kilogram with full reuse; and ultimately targeting $32 per kilogram with "full reuse + rapid turnaround."

By 2029, when the AI1 satellite is officially deployed, launch costs are projected to fall to $398 per kilogram, non-computing satellite costs to approximately $13,000 per kilowatt, and total non-computing deployment costs to around $27 billion—nearly matching the $23 billion on Earth, reducing the cost multiplier from 6x to 1.2x.

By the AI2 satellite phase in 2032, launch costs will further decrease to $170 per kilogram, with non-computing total costs reaching approximately $15 billion, below the $25 billion on Earth, reducing the cost multiplier to 0.6x. In the subsequent AI3 phase, the target launch cost is $43 per kilogram, with total costs around $9 billion—less than one-third of Earth-based costs.

The validity of this curve assumes that the Starship launch frequency and reusability progress as planned. This is why the report identifies SpaceX’s “extreme vertical integration” as the most critical execution variable.

SpaceX plans to begin prototype deployment as early as the end of next year; FCC filings indicate that the Starmind constellation could ultimately consist of up to one million low-Earth-orbit satellites. Starmind is a massive space-based AI satellite constellation currently under development by SpaceX.

Each AI1 satellite is designed with a power capacity of approximately 120–150 kW, with actual stable operational power consumption around 120 kW—roughly equivalent to the power consumption of an NVIDIA GB300 NVL72 rack.

SpaceX's target computational power density is 100 kW per ton, but AI1 can initially achieve only about 70 kW, meaning each Starship launch (assuming an 85-ton payload) can deliver approximately 6 MW of computational power.

According to the model assumptions, the power density will increase to 85–90 kW/ton in the AI2 stage by 2032, reaching the target of 100 kW/ton only after AI3.

The satellite is compatible with multiple chips—including NVIDIA GPUs, Google TPUs, Amazon Trainium, and Tesla AI chips (optimized for energy efficiency).

Cooling in ground-based data centers is straightforward—using fans, air conditioning, or water cooling, with air or water as the cooling medium.

Space won't work. In a vacuum, heat can only be dissipated through radiation, following the Stefan-Boltzmann law, where cooling efficiency depends on temperature and surface area.

Existing satellites almost entirely rely on passive radiators that dissipate heat through material properties and geometric design, consuming no power but limited by satellite volume. Active radiators require pumps to circulate liquid, consuming power and introducing failure risks, but can handle higher thermal loads—currently used only on space stations (ISS, Tiangong) and crewed spacecraft (Crew Dragon).

The AI1 satellite uses a double-sided active deployable liquid-cooled radiator with a single-side heat dissipation capacity of 700 W/m², totaling 1400 W/m² on both sides, covering a total area of 110 m². This is necessary to handle the high-power payload of 120–150 kW, which passive cooling alone cannot manage.

SpaceX will become the first company in the world to mass-produce this type of active cooling design. Estimates show that as mass production progresses, the cost of the heat sink can decrease from $8,000 per square meter in 2027 to $1,000 per square meter in the AI3 phase.

Where does the satellite get its power? From solar energy.

The AI1 satellite requires approximately 600 square meters of solar panels. Initially, silicon-based cells with an efficiency of about 19% will be used. In the long term, heterojunction (HJT) cells can achieve up to 27% efficiency, offer bifacial light absorption, and exhibit strong radiation resistance; perovskite thin-film cells have theoretical efficiencies comparable to multi-junction cells and are printable and ultra-lightweight.

SpaceX has begun construction of a solar cell factory in Bastrop, Texas, with a planned capacity of 10 GW (5 GW per floor) and a facility area of approximately 1.1 million square feet, co-located with the existing Starlink production site. Construction commenced at the end of March 2026, equipment installation has already started, and the goal is to achieve mass production by the end of 2027.

Elon Musk's broader goal is to establish 100 GW of domestic solar cell production capacity in the United States within three years.

The AI1 satellite does not carry complex phased-array antennas; inter-satellite communication relies entirely on optical inter-satellite links (OISL). Data is routed within the Starmind constellation, then connected to the Starlink laser mesh, and finally transmitted back via ground stations.

The benefit is that the Starmind constellation itself uses almost no radio spectrum.

However, the cost is that all data must ultimately pass through Starlink ground gateways, and orbital data centers would significantly alter the direction of data flow—Starlink’s original gateway authorization was designed for downstream-heavy consumer broadband services and is not suited to AI inference scenarios requiring massive upstream traffic.

To this end, 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 at AI services. Historically, these high-frequency bands have been unsuitable for consumer-grade terminals due to issues such as rain fade and oxygen absorption; however, these challenges can be mitigated through site diversity and optical routing when deployed at high-capacity gateway stations equipped with large antenna arrays.

This year, the U.S. Federal Communications Commission (FCC) also updated its satellite interference protection standards, replacing the 1990s-era equivalent power flux density (EPFD) limits with performance-based standards, enabling operators to deploy higher-power, more co-frequency satellites in critical Ku/Ka bands—potentially increasing capacity by approximately sevenfold for the same number of satellites.