Within 5 years it will approach, and within 10 years surpass ground-based data centers! This investment bank has built a cost model for space data centers.

Within 5 years it will approach, and within 10 years surpass ground-based data centers! This investment bank has built a cost model for space data centers.

Here is the English translation: ---

How much money does a space data center need to burn before it can break even with ground operations? Deutsche Bank did the math.

According to WindChase Trading Desk, on July 14, Deutsche Bank analysts Edison Yu and others released the fourth report in their "Space Data Center Series," modeling the economic feasibility of SpaceX's Orbital Data Center (ODC).

Edison estimates that currently, the cost of a space data center is six times that of one on the ground, but by 2029, after the deployment of the AI1 satellite, it could be reduced to 1.2 times, and by 2032, it could fall below ground-level costs.

How expensive is it now? Six times more

Deutsche Bank cites analysis from Epoch AI, assuming the upfront capital expenditures for deploying 1 GW of AI computing power on the ground is $3.8 billion, with annual operating costs (electricity, maintenance, labor, etc.) of about $0.9 billion, bringing the total 5-year cost to about $4.25 billion. Of this, computing hardware (GPUs, etc.) accounts for $2.1 billion, and non-computing aspects (infrastructure, cooling, power, etc.) are about $2.15 billion.

Nvidia CEO Jensen Huang recently mentioned at GTC Taipei 2026 that the cost of a new 1 GW "AI factory" might approach $10 billion, about half of which is related to computing power.

According to the bank’s estimate, using current rocket and satellite designs to deploy a space data center of equivalent scale, the non-computing costs are about 6 times higher than on the ground.

Where's the gap? Two parts: launch costs and the satellites themselves.

In 2027, using a generic satellite plan, launch costs are about $1,429 per kilogram, and the cost for satellite non-computing hardware is about $50,000 per kilowatt. Shooting 100 satellites up there, just these two items can stack the bill up to $11.5 billion, while ground operations of equivalent scale need only about $2 billion.

Why can it shrink to 1.2 times within five years?

The crucial variable is Starship.

Edison Yu's projection for Starship launch cost is as follows: Early stage with no rocket reuse, about $4,933 per kilogram; partial reuse drops to $398; full reuse reaches $170; ultimately achieving "full reuse + rapid turnaround," the goal is $32 per kilogram.

By 2029, when the AI1 satellite is officially deployed, launch costs are expected to fall to $398 per kilogram, satellite non-computing costs to about $13,000 per kilowatt, and total non-computing deployment costs to about $2.7 billion—almost no difference from the ground’s $2.3 billion. The cost multiple narrows from six times to 1.2 times.

By 2032 in the AI2 satellite phase, launch costs further drop to $170 per kilogram, with total non-computing cost about $1.5 billion, lower than the ground’s $2.5 billion, a cost ratio of 0.6. Going further to the AI3 phase, launch costs aim for $43 per kilogram, total costs about $0.9 billion—less than a third of ground costs.

This curve is predicated on Starship’s launch frequency and reusability advancing as planned. That’s why the report considers SpaceX’s "extreme vertical integration" the most crucial execution factor.

The current status of AI1 satellite computing power: not yet up to standard

SpaceX plans to start prototype deployment as early as the end of next year. FCC filings show the Starmind constellation could reach up to 1 million low-earth-orbit satellites. Starmind is SpaceX’s massive space-AI satellite constellation under development.

Each AI1 satellite is designed for power output of about 120-150 kilowatts, with actual stable operation at about 120 kW—roughly equivalent to the power consumption of a Nvidia GB300 NVL72 rack.

SpaceX’s target compute density is 100 kW/ton, but early AI1 can only reach about 70 kW, meaning each Starship launch (assuming 85 tons payload) can send up about 6 MW of computing power.

According to the model assumptions, in the AI2 phase (2032), density increases to 85-90 kW/ton, with AI3 finally achieving the 100 kW/ton target.

Satellites can support multiple chips—Nvidia GPU, Google TPU, Amazon Trainium, and Tesla AI chips (focused on energy efficiency).

Cooling: No air in space, this is a hard engineering constraint

Cooling in ground data centers is simple—fans, AC, water cooling, using air or water as the medium.

Not in space. In vacuum, heat can only be radiated away, following the Stefan-Boltzmann law; cooling efficiency depends on temperature and radiator surface area.

Current satellites almost exclusively use passive radiators, relying on material properties and geometric structure to conduct heat, consuming no electricity, but cooling capacity is limited by satellite size. Active radiators require pumps to drive fluid circulation, consume power and risk faults, but can handle higher heat loads—currently only used on the ISS, Tiangong, and crewed spacecraft (Dragon).

AI1 satellites use double-sided actively expandable liquid-cooling radiators, single-sided cooling capacity is 700 W/m², double-sided totals 1,400 W/m², with a total coverage area of 110 square meters. This is to handle the high load of 120-150 kW—pure passive design simply can’t cope.

SpaceX will be the world’s first company to mass-produce such active radiator designs. Estimates show that as mass production advances, radiator costs can drop from $8,000/m² in 2027 to $1,000/m² in the AI3 phase.

Solar Cells: Self-built factory, target 100 GW

Where does the satellite electricity come from? Solar power.

AI1 satellites need about 600 square meters of solar panels, initially using silicon-based cells with about 19% efficiency. In the long term, heterojunction (HJT) cells can reach 27% efficiency, are dual-sided for light absorption, and have strong radiation resistance; perovskite thin-film cells offer theoretical efficiency comparable to multi-junction cells, and can be printed and are ultra-light.

SpaceX has started building a solar cell factory in Bastrop, Texas, with planned capacity of 10 GW (5 GW each on two floors), plant area about 1.1 million square feet, co-located with the existing Starlink production base. Construction started at the end of March 2026, equipment installation underway, targeting mass production by the end of 2027.

Musk’s larger goal: to establish a U.S. domestic solar cell production capacity of 100 GW within three years.

Communications Architecture: Optical lasers, not using radio spectrum

AI1 satellites are not equipped with complex phased-array antennas; all satellite-to-satellite communications rely on Optical Inter-Satellite Links (OISL), data is routed within the Starmind constellation, then connected to the Starlink optical network, and finally relayed through ground stations.

The benefit: The Starmind constellation itself uses almost no radio frequency spectrum.

But the price is: all data must eventually pass through Starlink ground gateways, and orbital data centers will significantly change the direction of data flow—Starlink’s original gateway authorization was designed for "downlink focused" consumer broadband, not suited for "large uplink" AI inference scenarios.

For this reason, SpaceX is expanding higher frequency gateway relay capabilities, including E-band (already in use), V-band, W-band, and D-band (proposed), mainly for AI business use. These high-frequency bands are traditionally unsuitable for consumer terminals due to rain fade and oxygen absorption, but for high-capacity gateways equipped with large antenna arrays, the problems can be overcome with site diversity and optical routing.

This year, the FCC also updated satellite interference protection standards, replacing the 1990s Equivalent Power Flux Density (EPFD) limits with performance-based standards, allowing operators to deploy higher power, more co-frequency satellites in key Ku/Ka bands, theoretically increasing the capacity for the same number of satellites by around seven times.

 

 

 

~~~~~~~~~~~~~~~~~~~~~~~~

The above exciting content comes from WindChase Trading Desk.

For more detailed interpretation, including real-time analysis and frontline research, please join 【WindChase Trading Desk Annual Membership】

Risk Warning and DisclaimerThe market has risks and investment needs caution. This article does not constitute personal investment advice, nor does it take into account the special investment goals, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. If you invest accordingly, you assume full responsibility.