The Cost and Network Limits of Space-Based AI Compute
The paper concludes that while low-Earth orbit data centers could potentially support AI inference, the prohibitive costs of launch, power, and cooling combined with network limitations make them uncompetitive with terrestrial facilities for training frontier-scale large language models.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the world's most powerful computers as giant, humming cities built on the ground. These "AI data centers" are like massive factories where billions of tiny workers (chips) race to solve complex puzzles, teaching artificial intelligence how to write, draw, and think. But these factories are hungry. They gobble up huge amounts of electricity, need massive air-conditioning systems to keep from melting, and require vast stretches of land. As these AI factories get bigger, they are hitting a wall: there isn't enough power, space, or money to keep building them on Earth.
So, some dreamers are looking up. They wonder: what if we built these factories in space? In the low-Earth orbit (LEO), just a few hundred miles above us, the sun shines bright and constant, and the cold vacuum of space could act as a giant freezer. This idea suggests we could launch thousands of satellites, each carrying a slice of a super-computer, to form a floating AI brain. But before we pack our bags for the stars, we need to ask: is this actually cheaper and faster, or is it just a sci-fi fantasy? To answer this, we have to look at two main things: the physics of getting heavy metal into space and staying there, and the "network" problem—how fast these floating computers can talk to each other.
This paper, written by Kees van Berkel, takes a hard, mathematical look at the idea of "Orbital AI." It compares a hypothetical 1-gigawatt (1GW) AI data center floating in space against a similar one sitting on Earth. The author uses a mix of physics calculations and computer network models to see if the space version can keep up. The study focuses on two big questions: How would a space-based AI perform at training giant language models compared to Earth? And how will future tech changes affect this gap?
The paper finds that while putting AI in space might work for simple tasks like answering questions (inference), it is likely a terrible idea for the heavy lifting of training new, massive AI models. The main culprit isn't the cost of launching rockets or the lack of solar power; it's the network. On Earth, computers in a data center are connected by a super-fast, tightly woven web of cables (called a Clos network) that lets them shout information to each other instantly. In space, the author suggests we would have to use laser beams to connect satellites in a giant, floating mesh.
The analysis shows that this space mesh is much slower and has far less "bandwidth" (the amount of data that can cross at once) than the Earth version. To train a massive AI model, thousands of computers need to share information constantly. In the space scenario, the author's models suggest this sharing would be so slow that the computers would spend most of their time waiting, rather than thinking. The paper estimates that training an AI in space could take hundreds of times longer or cost hundreds of times more than doing it on Earth, simply because the "conversations" between satellites are too slow.
The author also points out some hidden costs that aren't in the price tag. Launching millions of tons of satellites would create a lot of space junk and could pollute our atmosphere when old satellites burn up on reentry, potentially harming the ozone layer. While the paper admits that technology might improve—perhaps with super-powerful lasers or smarter satellite arrangements—it concludes that with current and near-future tech, the dream of training frontier-scale AI in space within the next few years is not credible. The math suggests that for the heavy work of teaching AI, the ground is still the best place to be.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.