Every AI company on Earth is fighting the wrong war.

They think the constraint is power. It is not. The constraint is heat.

xAI holds permits for 41 natural gas turbines in Mississippi drawing over a gigawatt to run Colossus. A significant fraction of that energy does not run computation. It runs cooling. Chillers. Fans. Water. Heat exchangers. The infrastructure required to prevent silicon from destroying itself with its own waste heat. Every hyperscaler faces the same wall. Data center electricity is projected to double to 945 terawatt-hours by 2030 per the IEA. Transformer lead times run three years. Grid queues stretch four. Every new gigawatt of cooling requires permits from communities that do not want it.

This is why Musk announced on March 21 that 80% of Terafab’s planned output will go to space.

The reason is one equation.

P equals epsilon sigma A T to the fourth.

Stefan-Boltzmann. In vacuum, radiation is the only cooling mechanism. And it scales with the fourth power of temperature. Raise a chip from 80 degrees Celsius to 120 and radiated power per square meter nearly doubles. Not because you added hardware. Because physics did the work for free. The D3 chip being designed for Terafab’s orbital satellites will run hotter than any terrestrial processor specifically to exploit this. At 100 degrees Celsius with emissivity 0.90, a satellite rejects 100 kilowatts of waste heat through roughly 100 square meters of radiator. No water. No fans. No grid. No permits.

SpaceX confirmed this number on March 22. The 100-kilowatt AI Mini Satellite render shows a radiator of approximately 100 square meters, described as “quite small relative to the solar panels.” Energy collection requires area. Heat rejection barely registers.

On Earth, cooling scales linearly. Double the compute, double the infrastructure, double the resistance. In vacuum, the heat sink is the cosmic microwave background at 2.7 Kelvin. It is infinite, free, and already built. The Sun delivers 1,366 watts per square meter continuously in orbit with zero weather and zero night. But energy was never the real bottleneck. The real bottleneck was always thermodynamics demanding that waste heat go somewhere, and on a planet with an atmosphere, somewhere does not scale.

In vacuum, somewhere is everywhere.

Terafab targets one terawatt per year of compute. The entire United States grid runs roughly 0.5 terawatts. The orbital mass required demands tens of thousands of Starship flights annually, which skeptics call absurd until they learn propellant costs run single-digit millions per launch while the US industrial oxygen market alone exceeds fourteen billion dollars a year. The only bottleneck is Starship reusability, and Falcon 9 booster B1067 just flew for the 34th time on March 30.

The company pursuing this operates the only reusable orbital rockets in existence, flies more than ten thousand satellites with autonomous collision avoidance executing 300,000 maneuvers per year, and is designing chips purpose-built for vacuum thermal regimes.

The race to artificial general intelligence will not be won by the company with the most GPUs. It will be won by the company that solved cooling. And there is exactly one that plans to manufacture its own silicon, launch it on its own rockets, and reject its waste heat into infinity at the fourth power of temperature.

T to the fourth. Remember that exponent.