Summary of Key Points
In its IPO prospectus, SpaceX no longer refers to itself as a rocket company but as a “space AI company.” Rockets are used to deliver payloads into space, while Starlink is responsible for building a global network in space, with the ultimate goal of establishing a space data center. Musk’s compensation is tied to the achievement of a space data center with 100 terawatts of computing power (12.5 times the total current global power generation capacity). However, the company acknowledges potential technical challenges that may prevent this from becoming a reality. Meanwhile, ground-based data centers face limitations due to issues such as electricity and cooling, making space data centers an attractive solution. Despite these challenges, startups like Starcloud have already successfully sent GPUs into space and completed AI training, indicating initial progress in space computing. However, large-scale deployment of space computing requires a significant reduction in launch costs.
I. SpaceX’s Rebranding: From Rocket Company to “Space AI Empire”
The most surprising aspect of SpaceX’s IPO is its new positioning. It no longer sees itself as just a company that builds rockets and launches satellites but as a provider of “space AI infrastructure”:
- Rockets as Delivery Vehicles: They are used to transport the hardware for data centers (servers, chips) into orbit.
- Starlink as Space WiFi: Satellites form a global network that enables high-speed communication between space data centers and between space and Earth.
- Space Data Centers as the Core Business: These centers operate directly in orbit to provide AI computing power, addressing the shortcomings of ground-based data centers.
Musk’s compensation plan is even more ambitious: He will receive stock only if SpaceX succeeds in building a space data center with 100 terawatts of computing power. This is an enormous goal, as the total global power generation capacity currently amounts to about 8 terawatts. Success would mean huge profits for Musk; failure would result in him losing out on his compensation. However, the company’s legal department has clearly stated in the risk assessment that orbital AI computing is still in its early stages and the technology may not be profitable.
II. Why Are Ground-Based Data Centers a Limitation? Physical Barriers to AI Power Expansion
To understand the rationale behind space data centers, it’s important to consider the challenges faced by ground-based data centers:
- Insufficient Electricity: The computing power needed to train large models is growing rapidly; by 2030, global data center electricity consumption could approach Japan’s annual usage (about 1 trillion kWh), but grid infrastructure is struggling to keep up. In some areas of the United States, it can take 7–12 years to connect a data center to the grid.
- Cooling Challenges: AI chips generate significant heat, and large data centers require millions of liters of water for cooling, which has led to opposition from residents in drought-stricken regions.
- Land Approval Issues: Building data centers requires land and environmental approvals, and communities may oppose them due to concerns about noise and water consumption.
In the words of tech investors, “The expansion of AI computing power is hindered by electricity, land, water, and approval processes.” This is where SpaceX sees an opportunity in space, as these issues could be avoided in orbit.
III. The “Dream” and “Challenges” of Space Data Centers
Space data centers seem ideal:
- Unlimited Energy: Solar energy in geosynchronous orbit provides a constant supply of power (no need to connect to the grid or wait for electricity).
- Faster Communication: Laser communication between satellites is faster than fiber optics in a vacuum.
- No Ground-Based Constraints: They don’t require land, water, or community approval.
However, reality is more complex:
- Heat Dissipation: There’s no air or water in space, so heat must be dissipated through radiation, which is very inefficient. The cooling system of the International Space Station (the size of a basketball court) can only dissipate 70 kilowatts of heat, which is less effective than a single AI server rack.
- Radiation Damage: Cosmic rays in space can cause data bits to flip from 0 to 1 or 1 to 0, affecting the accuracy of AI calculations. Commercial GPUs are not resistant to radiation, and specialized radiation-resistant chips are still several generations behind.
- Maintenance Challenges: There are no “space maintenance workers” to repair damaged chips; redundant hardware must be used as a backup, but if that fails, the damaged components become space debris.
Calculations show that building a 1-gigawatt space data center would cost three times more than a ground-based one. Amazon’s cloud CEO stated, “We don’t even have enough rockets to launch millions of satellites, and the costs are still too high.”
IV. Startups Taking the Lead: Sending GPUs into Space for AI Training
While SpaceX was developing its plans, a startup called Starcloud has already made progress:
- Founded in 2024, it became a unicorn with a valuation of $1.1 billion in just 17 months.
- In November 2025, it used a Falcon 9 rocket to launch a satellite equipped with NVIDIA H100 GPUs and completed the first space AI training experiment, using Shakespeare’s works to train the nanoGPT model and running Google’s Gemma model.
- Its current valuation is $2.2 billion, and SpaceX is considering investing in it.
Starcloud’s approach is to start small: Its next satellite will carry the latest Blackwell chips, capable of handling commercial tasks for AWS and Google, and it has a larger cooling system. If this becomes profitable, it will prove the commercial viability of space computing. While SpaceX hasn’t yet launched any computing satellites, Starcloud has already taken the lead in this field.
V. The Present and Future of Space Computing: Options Rather than Guaranteed Success
What can space computing do today? It can process data collected by satellites—e.g., meteorological and surveillance satellites that generate large amounts of images. These images can be analyzed directly in orbit using AI to identify useful information (such as forest fires or unusual ships), significantly improving efficiency. This is a real-world application that drove NVIDIA’s development of space-specific AI components.
However, to replace ground-based data centers, launch costs need to drop to one-tenth of their current level (from $1,000–$2,000 per kilogram to below $200 per kilogram), which is expected to take about a decade. Therefore, SpaceX’s space data center project is more of an “option” rather than a guaranteed success. Whether Musk’s ambitious plan will work depends on whether the Starship can significantly reduce launch costs.
This analysis explains SpaceX’s new strategy, the challenges of ground-based data centers, the pros and cons of space-based solutions, and the progress of startups in a way that is understandable to non-financial professionals.