
Running AI in space is no longer just an idea in research papers. Companies have already put AI chips in orbit, a commercial satellite with multiple GPUs is about to launch, and big tech firms are filing major plans with regulators. The reason is simple: AI data centers on Earth are running short of land and power. Space offers almost constant solar energy, so tech giants and deep-tech startups are now testing whether AI can run, and process data, directly in orbit.
Here’s who is leading the race and what still stands in the way.
Startups Leading the Way
- Starcloud: Starcloud is the best-known venture-backed company in the field. It sent its first NVIDIA H100-class system to orbit on a SpaceX Falcon 9 and has trained a large language model (LLM) in space. Next up is Starcloud-2, which is set to become the first commercial satellite running multiple GPUs in orbit.
- Lonestar Data Holdings: While others target orbit, Lonestar has gone to the Moon, where it has placed a data backup machine on the lunar surface. It has some of the clearest evidence of paying customers in the sector, and it focuses on high-security data storage and government contracts.
- Satlyt: After debuting at Startup Battlefield, Satlyt is building a platform that turns satellite constellations into virtual data centers. Its AI models run on the satellites themselves to analyse telemetry and spot anomalies. It is also working with NASA on streaming low-latency 4K video through relay satellites in lunar orbit.
- Aetherflux: Aetherflux is reported to be on track to put its first commercial compute node in orbit.
Big Tech’s Space Moonshots
- Google: Project Suncatcher. Google has reportedly launched its first prototype satellite carrying its Tensor Processing Units (TPUs). The satellite is about the size of a refrigerator and works like a single shelf in a data center rack. It is testing how AI chips survive launch, space radiation and the problem of cooling in orbit. Google also plans to fly prototype satellites with Planet Labs to test chips working together as a cluster, linked by laser connections between satellites.
- SpaceX: “Starmine”. SpaceX has reportedly asked the FCC for approval to launch a constellation of up to one million data center satellites. The project is said to be tentatively called “Starmine” (or “Star Mind”). The idea is to fly medium-sized, solar-powered satellites in tight formation so they can act together as one large AI computer. Some reports also say space computers co-designed with Nvidia are headed for mass production, though neither company has confirmed the details.
The Reality Check: What’s Holding It Back
Running AI in space has been shown to work, but several big obstacles stand between these experiments and real competition with data centers on Earth:
- The first real use is processing data in orbit: Sending everyday AI work from Earth into space doesn’t make economic sense yet. The more realistic near-term business is analysing data that satellites collect, right where they collect it. Only the useful results then need to be sent to the ground, which eases one of the biggest bottlenecks in satellite communication.
- Launch costs: Industry analysts estimate orbital data centers won’t make commercial sense until launch costs fall by about 90%, to roughly $200 per kg. That depends on fully reusable rockets such as SpaceX’s Starship.
- Heat and radiation: Getting rid of heat in space is surprisingly hard, because there’s no air to carry it away. Engineers are reworking hardware to shed heat effectively and to shield delicate AI chips from radiation.
Why It Matters
AI’s demand for electricity and land is growing faster than data centers on Earth can keep up. Space-based computing is a long-term bet. Most experts agree it won’t replace data centers on Earth anytime soon, but the progress in the past year shows it is moving from science fiction to engineering. Watch for Starcloud-2’s results, Google’s tests with Planet Labs and the progress of Starship. Together they will show how quickly orbital AI becomes practical.