AI Is Running Out of Planet

Elon Musk built Colossus, his Memphis data center, in four months. To power it, he brought in portable gas generators, dozens of them. The facility consumes as much electricity in a year as 200,000 homes.

The International Energy Agency estimates global data center demand could surpass one terawatt-hour before the end of the decade, more than double what they consumed in 2022. Musk put it plainly in an interview: AI faces three bottlenecks as it scales. Chips, transformers, and electricity generation. Chips can be manufactured faster. Transformers, with enough money, too. Electricity is a different problem. No grid was built to absorb that kind of growth, and expanding one takes decades and hundreds of permits.

Musk's answer is to skip the problem entirely.

In January, at Davos, he said the lowest-cost place to run AI will be space, and that this would be true within two years, three at most. SpaceX filed plans with the FCC for a constellation of up to one million satellites that would function as an orbital computing network, powered by solar energy and connected through laser links.

Solar panels on the ground average a capacity factor of around 24%. In orbit that number exceeds 95%, with no day/night cycle and no atmospheric losses. No land, no permits, no water for cooling. The problem is the price: building and launching a 1-gigawatt orbital data center would cost over $50 billion, roughly three times the equivalent on Earth.

Starcloud, a YC-backed startup, already trained the first language model in space using an Nvidia H100 aboard a satellite launched on a Falcon 9. It was NanoGPT, trained on the complete works of Shakespeare. Google announced Project Suncatcher, with TPU chips on satellites planned for 2027. China has been launching experimental constellations for years, and even Europe has its own program.

The energy problem is real, and the most important players in the industry are taking it seriously. If launch costs keep falling with Starship and orbital infrastructure matures, space-based data centers could move from speculative bet to a genuine part of the answer. For now, the race to scale AI has no clean ending.