Elon Musk: Space Computing Power Will Eventually Account for 100% of Total Power
Elon Musk posted that computing power in space will ultimately round up to 100% of total computing power. This aligns with his recent statements about moving AI reasoning to orbit while keeping training on the ground: the terrestrial power grid, land, and cooling represent hard limits, while solar power in orbit is about five times that on the ground, providing sustainable power without the need for cooling towers due to radiation heat dissipation.
SpaceX refers to its first-generation orbital computing satellite as AI1 or Starmind: it has a wingspan of about 70 meters, wider than a Boeing 747, with an average computing load of about 120 kW and a peak of 150 kW, a density of about 70 kW per ton, operating at an altitude of about 600 kilometers, with a load equivalent to a single NVIDIA GB300 cabinet on the ground. The chip compartment is replaceable, initially planned to use NVIDIA chips, with a long-term goal of collaborating with Tesla on radiation-resistant Terafab chips. They have designed the Vera Rubin NVL72 for orbit with NVIDIA, aiming for a test launch in Q4 2027 and scaling up by 2028. The company has applied to regulators for up to 1 million computing satellites; the Starship can be reused, with the premise of launching cabinets into orbit.
Musk has suggested a steeper curve: within about 30 to 36 months, orbit will become the most economical place for AI deployment; within five years, annual new launches could exceed the cumulative stock on Earth; from Earth, about 1 terawatt annually, and on the Moon, due to the lack of atmosphere and one-sixth gravity, satellites can be launched using electromagnetic accelerators, potentially reaching 1,000 terawatts annually. The IPO documents list orbital data centers as unverified or non-existent technologies, with a 100 GW timeline that "may be difficult or impossible to determine," with estimates of needing about 10,000 Starship launches to reach the 100 GW scale. In an all-hands meeting in August, he stated that training remains on Earth while daily reasoning moves to space, with the Memphis cluster still being the current source of gigawatts.
Mechanically, what is bought is the right to expand computing capacity free from the queue at ground substations, while what is sold is the traditional data center path that must acquire land, connect to high voltage, and build cooling in states with electricity. The event-driven aspect is that the leader has written the share prediction as a limit, applying the same growth function to Starship and solar satellites. Funding flows from ground tenant contracts and IPO fundraising into satellite production lines and launches; benefiting is SpaceX, which simultaneously controls launch capacity and constellation operations, while cloud vendors and power companies that bid per kilowatt for ground cabinets are under pressure. The 100% figure is an asymptote, not a quota for a specific year; delays, maintenance, radiation, and debris will determine which reasoning can go into orbit and which must remain on the ground.
Source: Public Information
ABAB AI Insight
Musk has framed the geography of computing power as a physical law: Earth's electricity and heat dissipation hit a wall first, while space's light and vacuum have yet to encounter one. A 120 kW satellite equals one cabinet; one million satellites is what he aims for as a grid. This statement serves to transform SpaceX from a rocket company into a computing logistics company—launch frequency determines new gigawatts, and Starlink experience dictates whether a million-satellite constellation can fly. The IPO risk factors have already categorized the same issue as "technology unverified"; the 100% on Twitter is a direction for engineers and shareholders, not a delivery contract.
The capital path is "ground rental cash—orbital hardware capital expenditure." Anthropic and Google rent Memphis in exchange for monthly fees, which are then used to build 70-meter wingspan flying cabinets. NVIDIA is first on board because the self-developed radiation-resistant chips are still on the Terafab blueprint. If Starship fails once, the orbital gigawatts will shift according to the launch window. The lunar electromagnetic cannon is a long-term option that removes Earth launch from the freight structure; currently, the constraints remain the cost of getting into orbit and on-orbit maintenance.
This is analogous to how AWS moved computing from data centers to the cloud, and oil moved from land to offshore platforms: both are geographic shifts after original site costs explode. The difference is that there are no maintenance trucks in orbit; a single satellite failure is a write-off. The industry phase has shifted from who can get electricity to who can deliver both electricity and chips to 600 kilometers. Whoever first turns a 120 kW satellite into an insurable, remotely replaceable payload product will define the next generation of reasoning pricing.
Structural judgments belong to technological substitution. Pricing power shifts from land and electricity prices to launch costs and on-orbit power density. The mechanism is that training requires low latency and high bandwidth, suitable for the ground; reasoning requires scale and electricity prices, suitable for solar and radiation panels. When 100% occurs, ground cabinets will not disappear; they will just no longer determine marginal supply. The first determinant of share will be Starship flight frequency, not model parameters.
ABAB News · Cognitive Laws
- The upper limit of computing power is ultimately written where electricity and heat dissipation can go.
- When one satellite equals one cabinet, launch frequency is the expansion speed.
- Training stays on the ground, while reasoning has the opportunity to move into orbit.