Training a large language model today costs between $50M and $200M in cloud compute. The primary cost driver isn't hardware it's electricity. And electricity costs are rising because the world is running out of places to put data centers.
The Power Problem
Every AI training run produces heat. Every joule of compute becomes a joule of heat that must be removed. On Earth, cooling a data center typically consumes 30-40% of total power meaning for every 100 watts of GPU compute, you need 40 additional watts just for cooling. This Power Usage Effectiveness (PUE) ratio typically ranges from 1.2 to 1.5 at the best facilities.
In orbit, PUE approaches 1.0. Heat is radiated directly into space at virtually zero cost. The thermal sink is infinite and free.
Solar Power: Unlimited and Uninterrupted
At 550km altitude, STELLAR satellites receive approximately 1,400 W/m² of solar irradiance uninterrupted by clouds, seasons, or grid capacity constraints. A single satellite with 200m² of solar panels (standard for commercial satellites) generates ~280 kW continuously.
Compare this to terrestrial data centers, which must negotiate utility agreements, build substations, and compete for limited grid capacity in a market where AI training is already straining infrastructure in Virginia, Oregon, and Singapore.
The Economic Case
Conservative modeling shows orbital AI training achieving cost parity with terrestrial cloud by 2028, with cost advantages emerging by 2030 as constellation scale increases. The crossover happens when:
1. Satellite manufacturing costs continue declining (20-30% per year, following SpaceX and satellite industry trends) 2. Launch costs continue declining (Starship targets <$100/kg to LEO) 3. Terrestrial electricity costs continue rising due to AI demand
For the largest training runs $100M+ experiments the economics favor orbit even before satellite manufacturing costs reach parity.
What This Means for AI Development
The implication is stark: the most advanced AI systems of 2030 will not be trained in Virginia data centers. They will be trained in orbit, where power is free, cooling is free, and the only constraint is compute density per kilogram of satellite mass.
STELLAR is building the infrastructure that captures this value. The question isn't whether orbital AI training happens it's whether Western AI labs or state actors get there first.