The launch of SpaceX’s Starship in March this year is set to make space-based data centres a reality, according to journalists at The Economist.
The interest in orbital DPCs stems from a crisis in terrestrial infrastructure. According to Sightline Climate, between 30% and 50% of the capacity planned for commissioning this year may be delayed due to issues with permits, network connections, and local resident protests.
Several US states have imposed moratoriums on construction, while demand for electricity for AI continues to rise.
An alternative could be placing servers in orbit, where unlimited solar energy is available. In November 2024, Starcloud launched the Starcloud-1 satellite, the size of a refrigerator, equipped with a standard Nvidia H100 chip. It successfully trained the NanoGPT language model on Shakespeare’s works and ran Google’s open model Gemma.
Challenges
The primary obstacle remains the cost of launch. Currently, delivering a kilogram of cargo on Falcon 9 costs $3,400, and on Falcon Heavy, $1,500. Two other factors are equally important:
- specific power — how many watts of computing power a kilogram of satellite provides;
- satellite cost — in dollars per watt.
These parameters directly depend on the mass and efficiency of solar panels and cooling systems.
Another consideration is how radiation will affect the reliability of AI chips in orbit. To determine if the idea of orbital data centres is feasible, assessments of all these variables are necessary, noted The Economist.
Andrew McCalip, an engineer at the space startup Varda, created a web calculator that allows comparison of the costs of orbital and terrestrial data centres of a given capacity.
According to his calculations, building a terrestrial DPC with a capacity of 1 GW and operating it for five years would cost $15.9 billion. An orbital equivalent (with a launch cost of $500/kg, specific power of 37 W/kg, satellite cost of $22/W, and 98% illumination time) is estimated at $51.1 billion. Both sums exclude the cost of AI chips (GPUs), which would range from $15 billion to $30 billion and be the same for both options.
Solution
Starcloud, founded in 2024 specifically to implement the idea of orbital computing, has already obtained initial practical data. The experiment with launching the Starcloud-1 satellite based on Nvidia H100 provided information on chip performance in space and allowed for refinement of key parameters for future calculations.
The specific power indicator of 37 W/kg used in basic calculations is derived from Starlink satellites. However, AI satellites do not require heavy antennas for communication with Earth — only laser links with neighbours.
This means a much larger portion of the mass can be dedicated directly to computing. SpaceX CEO Elon Musk considers a specific power of 100 W/kg realistic, with the potential for 150 W/kg in the future. Starcloud plans conservatively for 70 W/kg.
A similar situation exists with satellite costs. For Starlink, it is about $22/W. An AI satellite without expensive communication components should be significantly cheaper.
Starcloud CEO Philip Johnston set the target at “less than $5/W” (excluding GPUs). With these achievable values (70 W/kg and $5/W), the cost of an orbital data centre drops to $16.7 billion — 5% more expensive than a terrestrial one.
Nevertheless, the main factor remains the cost of launch. The estimated $500/kg is about a third of SpaceX’s current expenses.
If SpaceX’s new fully reusable Starship rocket becomes operational, the price could drop to $100-200/kg. In such a scenario, the orbital centre would reduce to $12.1 billion — making it cheaper than the terrestrial option.
The first Starcloud satellite overheated, but a second launch this year will test a radiator with tenfold cooling efficiency. Preliminary data also shows that chips fail less frequently than calculated: with 5% failures instead of 9%, costs will decrease even further.
A pivotal event for the entire industry will be the 12th test flight of Starship in March. Its success will determine whether the era of cheap launches will arrive and if orbital data centres will become a viable alternative to terrestrial ones.
Earlier in February, Musk merged SpaceX and xAI into a single company valued at $1.25 trillion. The firm has applied for a license to create an orbital data centre from 1 million satellites.