Elon Musk outlines a bold plan to run AI in space using satellite compute, sun-sync orbits and lunar factories, promising 100 GW a year past Earth's limits.
© whitehouse.gov
Elon Musk has floated a radical way to expand AI capacity: move the computations into space. The idea centers on satellites with on-board AI that would beam back only finished results from a sun-synchronous orbit aimed at low latency. He maintains this could become the cheapest way to generate what he calls an AI bitstream within three years, and within four years the fastest to scale. The timeline he sketches leaves little margin for hesitation.
The proposal then escalates. Musk argues it is getting increasingly difficult to secure accessible power on Earth for growing clusters, so he suggests launching a million tons of satellites every year. If each unit carried around 100 kW, his math points to an annual addition of roughly 100 GW of compute, supposedly without operating or maintenance costs.
His next step envisions building satellite factories on the Moon and using an electromagnetic accelerator—essentially a railgun—to push AI satellites to escape velocity without rockets. In the end state, he speaks about scaling beyond 100 terawatts of AI capacity per year and making real progress toward a Type II civilization on the Kardashev scale. The vision leaps from orbital hardware to lunar industry, reading less like a thought experiment and more like a timetable.
In a separate post, he added that if the Moon hosted factories, robots, and large-scale accelerators that closed the loop from production to deployment, the system could theoretically function without relying on traditional money, measuring its economy in watts and tons and effectively operating autonomously. Framing resources this way recasts the entire approach as an engineering-first plan.
Back in November, Musk had already suggested that thanks to effectively free solar energy and convenient radiative cooling in space, deploying and operating large AI systems in orbit could, within four to five years, be more cost-effective than equivalent Earth-based setups. He pointed to Earth’s twin constraints—rising electricity demand and the need to shed heat—which, as AI clusters scale, could run up against the limits of terrestrial infrastructure. The numbers he cites are dramatic; the unanswered steps in between are left for engineering to close.