The narrative of space data centers is becoming popular, and Musk's vision for space computing power is not completely detached from economic logic. However, according to SemiAnalysis's calculations, the key to determining when it becomes a reality is not superficial narratives such as "space has free energy", but multiple reality constraints such as chip supply, launch costs, cooling systems, and lifespan and maintenance reliability.
Musk has frequently talked about orbital computing power this year. In February of this year, Musk predicted on the Dwarksh Podcast that in five years, the annual AI computing power in space may exceed the cumulative total on Earth, and mentioned the scale of "hundreds of gigawatts/year" of space AI. SpaceX also stated in its S-1 filing on May 20th that its long-term goal is to launch 100 gigawatts of computing power into space annually, and believes that space computing power will significantly expand the scale of AI computing and improve token economy.
SemiAnalysis stated in its in-depth report released on June 3rd that deploying AI data centers into orbit today using existing technology still incurs significantly higher costs than on the ground. Taking a 30.5kW B300 cluster in 2026 as an example, the total capital cost of the space deployment project is 4.1 million US dollars, with a ground deployment cost of 1.4 million US dollars; Converted to monthly total cost of ownership, space costs approximately $100900 per month and ground costs approximately $27700 per month.
According to estimates, in the baseline scenario, the cost of computing leveling for space and ground data centers may not reach parity until around 2040. By the early 2030s, the cost of space data centers may still be about 30% higher than on the ground, but this is enough to open the window for the first large-scale deployment. In its' Musk scenario ', if the expansion of ground data centers is limited and chip production capacity continues to expand, the time for space computing power to approach parity may be advanced to the early 2030s.
Sending computing power to the sky now, the account still cannot be calculated
Using the B300 cluster from 2026 as a reference, the difference is very clear.
A B300 cluster with 30.5kW and 16 GPUs has a total project capital expenditure of approximately $4.1 million for deployment in space and $1.4 million for ground deployment. Converted into monthly total cost of ownership, space costs $100925 per month and ground costs $27724 per month.
Let's switch to the more common approach of cloud services:
Space deployment TCO: $8.64/hour/GPU
Ground deployment TCO: 2.37 USD/hour/GPU
Space deployment LCOC: $10.91/hour/GPU
Ground deployment LCOC: 2.49 USD/hour/GPU
LCOC is closer to the true cost of computing power than TCO because it takes into account availability, redundancy, and fault absorption. Ground clusters only require an additional cost increase of about 5%; Due to radiation effects, space clusters cannot be repaired on-site, resulting in a cost increase of approximately 26%.
The cost difference is mainly not in the GPU itself. The capital expenditure for IT equipment is almost the same on both sides: about 981000 US dollars in space and about 986000 US dollars on the ground. The gap lies in the essence of the 'data center'.
The capital expenditure for the data center deployed in space is about 3.1 million US dollars, while on the ground it is only 382000 US dollars. The launch cost for one project is 1.6 million US dollars. More troublesome is the lifespan: space data center facilities are depreciated for 5 years, while ground facilities are depreciated for 15 years. The result is that the capital cost of the space data center is reduced to $6.29 per GPU hour, while on the ground it is only $0.36, a difference of about 17 times.
That's also why 'free solar energy' cannot directly become 'cheap computing power'. Electricity costs are important in ground data centers, but they are not the only variable in the overall TCO. For space, launch, heat dissipation, structure, power supply, lifespan, and reliability are the top priorities.
'Free solar energy' and 'free heat dissipation' are both said too easily
The four most common reasons for optimism in space data centers need to be rewritten.
Firstly, there is no 24-hour sunlight in low orbit. The International Space Station and most star chains are in low Earth orbit, orbiting the Earth about 15 times a day, with an average of only about 60% of the time being exposed to the sun. The theoretical value of solar irradiance is 1361W/square meter, but low orbit data centers may only capture about 800W/square meter on average over 24 hours. When entering the shaded area, a battery is also required to power 100% of the IT load.
A more suitable orbit for data centers is the sun synchronous orbit, especially the orbit near the terminator. It can face the sun most of the time, but there may still be a maximum of 35 minutes of eclipses per day. A decrease in battery demand does not mean disappearance.
Secondly, space cooling does not mean free heat dissipation. Ground data centers can rely on air and water systems to transport tropical areas, while space has almost no medium and cannot rely on convection, only radiation for heat dissipation. The radiator system of the International Space Station can only dissipate 70kW of heat, with an area of 325 square meters and a cost of 340 million to 500 million US dollars. This system is technologically advanced and costly, but it illustrates a fact: one of the biggest structural constraints on orbital computing power is heat dissipation.
Thirdly, the fast speed of light in a vacuum does not necessarily mean low user latency. Low Earth orbit satellites orbit the Earth about 15 times a day, and typically only take 5 to 7 minutes to pass through a certain ground station window. If you miss this window, the data will have to be redirected through inter satellite links or routed to other gateways. If a satellite serving American users is over the Indian Ocean, a multi hop inter satellite link may result in a one-way delay of 30 to 80 milliseconds. Optical ground links are also affected by atmospheric interference, requiring more ground stations to be distributed globally.
Fourthly, space is not without 'capacity constraints'. The dawn dusk sun synchronous orbit is just a narrow subset of the low orbit, not an infinite parking lot. The estimated overall carrying capacity of low Earth orbit ranges from 100000 to over 1 million satellites, but sun synchronous orbit requires specific altitude and inclination relationships, with common concentration areas ranging from 600 to 800 kilometers. The truly suitable track for continuous lighting in the morning and evening is narrower. As for the L1 point of the Sun Earth Lagrange, it is indeed possible to see the Sun for a long time, but the round-trip path from Earth to L1 is about 3 million kilometers, and the propagation of light takes about 10 seconds, making the delay meaningless.
The power on the ground will be tight, but not so tight that it can only go up to the sky
The space data center needs to become a 'must-have', provided that all available supply layers on the ground are exhausted, rather than the ground power being tight.
This framework divides the newly added ground supply into four layers:
Grid connected power supply
Transforming Bitcoin mining sites and existing electrified land
Power generation behind the meter, i.e. self-contained power supply
Industrial capacity and manpower expansion
The first layer is grid connected power supply, which is the cheapest on paper, with infrastructure costs of approximately $12 million to $15 million per MW. However, the real problem is queuing. The grid connection cycle of PJM in Northern Virginia is actually close to 7 years. The reliability margin of the power grid in the ISO region of the United States has decreased from 70.2GW in 2021 to 18.3GW in 2025, further narrowing to 15.9GW in 2026, turning negative in 2027, and a total gap of about 40GW by 2030. This may sound bad, but the ground is not just the power grid.
The second layer is the renovation of existing power assets. The conversion of encrypted mining sites is the most typical. Projects such as Core Scientific, IREN, Cipher Mining, Applied Digital, TeraWulf, etc. will have a total contracted renovation capacity of approximately 2GW by the end of 2026 and approximately 5GW by the end of 2027. Overall, the supplied land and renovation sites can contribute 8 to 10GW of supply in the near future, with encrypted mining sites accumulating approximately 8GW by 2028. The cost is approximately 10 to 15 million US dollars/MW, which is similar to or even lower than the grid connection plan.
The third layer is behind the table power generation. This used to be like a last resort, but now it has become a realistic option. The reason is straightforward: the annual revenue of the AI cloud contract is approximately $12 million to $13 million per MW of critical IT load, with a 200MW capacity launched 6 months ahead of schedule, and a net present value of potentially $400 million to $500 million. As long as the demand for computing power is strong enough, self built power generation and additional capital expenditure are also reasonable.
The comprehensive cost of generating electricity behind the grid is approximately $110 to $170/MWh, and the electricity prices in major markets in the United States may already reach the level of $150/MWh. By 2028, behind the table power generation may contribute half of the newly added AI data center power capacity, while by 2025 this proportion will be less than 7%. The confirmed key IT capacity is expected to reach approximately 26GW by the end of 2030, and undisclosed projects may have even higher capacity.
The fourth layer is the harder industrial bottleneck: transformers, oriented silicon steel, copper, gas turbines, construction manpower, and cooling equipment. Large power transformers have long lead times, and copper prices have risen by nearly 20% in the past year. Modularization and digitization can reduce on-site labor by more than 50%, but when computing power construction reaches the level of hundreds of gigawatts, skilled working hours will still become a tangible constraint. After entering this layer, the cost will exceed $20 million/MW, and the amount of excess depends on how much new capacity the industry needs to extract in a short period of time.
So, the ground supply is not infinite. But it is also not a single-layer system that is about to reach its peak. To win in space, we must wait until the ground reaches the fourth layer and the cost significantly increases before we have a chance.
The chip is the first to truly hinder AI expansion, and Terafab is the key variable
Space data centers cannot solve the upstream problem: without chips, there can be no clusters.
The current constraints have shifted from data center capacity to semiconductor production, especially TSMC N3 advanced process, HBM, and DRAM production capacity. AI related demand is expected to consume nearly 60% of TSMC N3's output by 2026 and about 86% by 2027, almost squeezing the demand space for smartphones and CPUs.
The memory is also tight. HBM consumes approximately three times the wafer capacity of regular DRAM per bit. The proportion of AI related demand in the total DRAM wafer production capacity is expected to increase from 12% in 2023 to approximately 70% in 2027.
This is more difficult to expand rapidly than electricity. There are multiple technical routes for power projects, including the diversion of pressure from electrified land, gas turbines, and post metering power generation; Advanced wafer fabs need to first build clean rooms, install equipment, and then conduct process validation. Capital is not the only constraint, time and technological accumulation are equally constraining. The more realistic window of relief is more like 2032 to 2034, rather than 2027 to 2029.
Musk is clearly aware of chip constraints. SemiAnalysis points out that this is precisely the background of the Terafab Initiative.
When Musk released Terafab in March 2026, he described it as a "1 terawatt computing power factory per year". Tesla, SpaceX, and xAI will jointly build in Austin with a budget of $20 billion to $25 billion, with an initial goal of 100000 wafers per month and ultimately moving towards 1 million wafers per month, which is approximately 70% of TSMC's current global output. The project scope includes logic, storage, masking, advanced packaging, and testing, with approximately 80% of the computing power allocated for space and 20% for ground use.
SemiAnalysis believes that even if Terafab only achieves partial goals, it will still be a meaningful success. But the numbers themselves are extremely strict, according to its Foundry Model, the global 300mm foundry capacity will exceed 4 million wafers per month by 2025. If Terafab reaches 1 million pieces per month, it will be equivalent to 24% of the global OEM production capacity or 68% of TSMC's production capacity.
The bigger challenge lies in process IP and storage. SemiAnalysis believes that Tesla does not manufacture IP, and the design, interconnection, lithography, etching formula, and yield engineering of GAA transistors are all in the hands of existing manufacturers. If Terafab ultimately reaches mass production, a more realistic path may be to operate an integrated wafer fab based on authorized nodes, rather than developing advanced processes from scratch.
Storage is even more difficult. HBM, LPDDR, and NAND correspond to different processes, with IP concentrated in the hands of manufacturers such as Samsung, SK Hynix, and Micron. SemiAnalysis believes that long-term supply contracts or joint investments with existing DRAM manufacturers are the more realistic path.
When can it ascend to heaven: the baseline is around 2040, and the radical scenario is advanced to the early 2030s
SemiAnalysis's baseline scenario assumes that key engineering issues such as radiation impact and GPU reliability will be fully alleviated by around 2040, and major cost items such as emissions, heat sinks, and solar energy will be scaled down to reduce costs. At the same time, AI demand and chip production capacity will significantly increase.
In this scenario, the cost gap between space and ground data centers will gradually narrow from over four times by 2026 and reach parity around 2040. Afterwards, the cost of leveling computing power in space may be lower than on the ground.
But this does not mean that commercial deployment of space data centers will not occur until 2040. SemiAnalysis states that by the early 2030s, space data centers may only be about 30% more expensive than on the ground, which could open the window for the first large-scale space data centers.
Another more radical "Elon Musk scenario" assumes that the increase in ground data center capacity will peak in 2028 and remain low for decades, while chip production expansion continues to advance. In this scenario, space becomes the only alternative path for large-scale AI data center deployment, and the potential market for space data centers could reach the level of adding hundreds of gigawatts annually, approaching cost parity by the early 2030s.
In other words, the commercialization schedule of space computing power depends on the relative speed of two directions: how quickly space systems can reduce costs and how severely ground data centers are constrained.
Investors should focus on five verification points
The conclusion of SemiAnalysis has implications for the market that space AI data centers should not be simply understood as launch capability stories, nor should they be understood solely as power arbitrage stories. It is a system engineering that spans across semiconductor, power, aerospace manufacturing, and data center economics.
Firstly, can the advanced logic and HBM production capacity be broken through. If chips continue to be a bottleneck, both space and ground will be limited.
Secondly, can the launch cost be significantly reduced. SemiAnalysis mentioned that SpaceX envisions the future launch cost of Starship to decrease from around $1400 to $1800/kg for Falcon 9 to around $250/kg, which is a necessary condition for the cost curve of space data centers.
Thirdly, can radiators, solar arrays, and battery systems be scaled down to reduce costs. Heat dissipation is not an ancillary issue, but a core engineering constraint for orbital calculations.
Fourthly, how to solve reliability and maintenance issues. Approximately 3% to 6% of GPUs in ground clusters experience failures that require manual intervention each year. Space deployment requires addressing this issue through robots, higher reliability, over configuration, or a combination of solutions.
Fifth, is the ground data center really stuck for a long time. In the SemiAnalysis baseline scenario, even if space and ground reach cost parity, ground capacity is still sufficient, and going to space is more of a preference and optimization; Space will only become necessary when regulations, permits, power grids, and industrial capacity continue to suppress ground expansion.
Therefore, Musk's "space computing dream" is not without a path, but its key lies not in the slogan, but in the cost curve. According to the SemiAnalysis model, the real turning point is not today, nor can it be achieved solely through rocket reuse; It requires simultaneous progress in chip, emission, heat dissipation, solar energy, and track maintenance. The baseline answer is cost parity around 2040, while the radical answer is approaching parity starting in the early 2030s.
