10 Million GPU Data Centers on Earth and Mars

xAI is completing a 200,000 GPU data center in Memphis. They completed 100,000 GPUs two months ago and were adding another 100,000 GPUs.

300 Megawatts of power is supplied by natural gas powered turbines.

Water cooling needs about one million gallons per day.

xAI and Tesla are planning to build a one million GPU data center by the end of 2025 or early 2026. Broadcom is working with Meta and other hyperscalers to build three data centers in 2027 with one million GPUs.

One million GPU data centers would cost about $150 billion. A Ten million GPU data center on Earth could cost $1-2 trillion. The additional costs for sending the GPUs to Mars could become trivial if constructing the power and saving hundreds of square kilometers of land purchases could be saved.

Elon Musk and SpaceX plan to send 1000 SpaceX Starships to Mars every two years. This would bring 250,000 to 350,000 tons to Mars every two years. This would mean 5,000 Teslabots in each Starship with Teslabot payload.

An Nvidia B200 Server weighs about 130 kilograms. A SpaceX Starship with advanced raptor engines could take about 300 tons in each rocket. This would be about 3000 Nvidia GPU servers.

Tesla is making Dojo training chips which would be comparable to Nvidia chips. Dojo 2 should be roughly comparable to Nvidia B200s and Dojo 3 to the next generation Rubin chip.

If SpaceX, Tesla and xAI were designing for lower mass with comparable performance, then getting Dojo 3 (2026 chips) down to 25% of the weight would enable 1000 Starships to transport the 10 million GPUs to Mars along with automated Teslabot construction.

SpaceX would also need to bring solar power to generate 15 gigawatts of solar power on Mars. Mars has 43% of the solar power per square meter. Instead of needing one square kilometer of tilted solar panels in the USA to generate 42 megawatts, one square kilometer would generate about 15 megawatts of power. This means 1000 square kilometers of solar power generation would be needed on Mars.

SpaceX could either choose to deploy at the poles of Mars where there is 24 hour sunlight or at the equator where there are valleys and other terrain.

Solar towers with concentrated sunlight could work better since mainly mirrors would be needed. Mirrors produced on Mars would remove most of the weight that would have to be transported to Mars. Mirrors would be relatively simple.