Conference Presentation, Keynote, Fireside Chat
Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space
- Launch costs are projected to reach a break-even point for space data center viability when they decline to approximately $500 per kilogram, representing a tenfold reduction from current rates that is anticipated to be achieved by incoming launch vehicles.
- Starship is expected to drive launch costs down further to the range of $10 to $20 per kilogram, enabling a $100 billion capital expenditure to deploy an 88,000-satellite constellation capable of delivering roughly 20 gigawatts of new compute capacity.
- The proposed satellite architecture operates at two primary orbital tiers: a 400-kilometer altitude layer where satellites naturally de-orbit within months, and an unsynchronous "dawn-dusk" orbit capable of supporting terawatts of compute with low collision risk due to unpopulated conditions.
- Inference workloads are forecast to constitute approximately 99% of the total AI compute market within the next five to ten years, while large-scale training requiring structures around five gigawatts is estimated to remain impractical for at least 15 years.
- Technological adjustments include the development of chips designed to operate at higher temperatures (50 to 80 degrees) to reduce radiator surface area requirements by nearly half.
- The initiative is framed as a foundational step toward a Kardashev Type 2 or Type 3 civilization, with current capabilities focused on inference and future potential for space-based training.
- Collision risks in the near term are assessed as extremely low, supported by the rapid de-orbiting nature of the lower altitude layer and historical evidence of SpaceX managing 10,000 satellites without collision in low Earth orbit.
- The speaker seeks audience input on the timeline for space compute becoming cheaper than terrestrial alternatives, with options provided ranging from the next five years to never.