If you’ve been paying attention, you’ll have seen that Elon and the boys are about to put datacentres into space. It’s not sci-fi: SpaceX is targeting its first AI-compute satellites for 2027.
It’s largely happened already if you stretch your definition of a datacentre a bit to include the compute and capacity on board every Starlink satellite. There are 11,156 of those in orbit right now, and counting.
Why space wins
SpaceX’s planned Starmind constellation is going to dwarf most terrestrial installations. That’s largely because space has none of the same zoning and regulatory hurdles, and a dawn–dusk SSO (sun-synchronous orbit) gives you near-constant sunlight. In its FCC filing, SpaceX asked to launch up to one million satellites, and reckons that launching a million tonnes of them a year would add about 100 GW of AI compute.
This is a major boon. Megawatts’ worth of compute can go up on a single Starship launch: each Starmind “AI1” satellite is essentially an Nvidia Vera Rubin rack drawing about 120 kW, and Starship V3 can lift more than 100 tonnes to orbit. No batteries, power-wheeling, political ass-kissing or environmental impact assessments can get in the way.
In an increasingly compute-starved economy, this is a crazy leap forward, and it will increasingly be the way we use our AI systems. Meanwhile, neo-Luddism, the very real power draw of datacentres and fundamental misunderstandings of how they’re cooled keep perpetuating the “datacentres use so much water” myth. For the record, US datacentres use about 0.2% of the country’s water, and the newest chips are cooled in closed liquid loops that push direct water use towards zero.
Demand is outrunning the grid
Whether or not you like AI, or the unprecedented industrial build-out to satisfy its users (half of which is just me prompting Claude like an evil mastermind late on a Tuesday night), the demand is insane. And it increasingly can’t be met by datacentres built close to residential areas or hooked up to old and failing power grids.
945 TWh is roughly what all of Japan uses in a year.
Source: IEA, Energy and AI
The lead times for turbines, cooling systems, racks and the GPUs themselves are only getting longer. Order a heavy-duty gas turbine from GE Vernova today and it won’t arrive until 2031.
So even if it costs ten times more to send a system into space, if you can use it today instead of in four years, operators like Elon will launch it long before it’s strictly economical, and recoup the cost later by selling the compute or building great products with the capacity they have on hand. Look at Anthropic, which is paying SpaceX $1.25 billion a month, and up to $84.5 billion through 2029, for capacity on its Colossus supercomputers. Compute is the business now.
In summary
Space is about to get a lot of GPUs, and we are all going to benefit.
