Elon Musk — "In 36 months, the cheapest place to put AI will be space”

The case for orbital data centers begins with a mismatch: Musk says chip output is growing exponentially while electricity output outside China is roughly flat. Dwarkesh challenges the economics by noting that energy is only 10–15% of a data center's total cost of ownership, while GPUs are harder to service in space and may depreciate faster. Musk's answer is that access to power becomes the binding constraint as compute scales, and he predicts that space will become the cheapest place to put AI within 36 months, perhaps 30. These are forecasts, not demonstrated operating results.

Musk's proposed advantage is nearly continuous solar generation without clouds, seasons, night, or atmospheric losses. He estimates that a given panel could produce about five times more power in space and argues that avoiding batteries makes the total system up to ten times cheaper. Dwarkesh pushes on land availability and asks whether orbit is mainly a way around permitting. Musk agrees that permitting and the pace of construction matter, while also arguing that orbital deployment is the only path to much larger energy scales. The exchange does not quantify the full launch, maintenance, communications, or replacement costs that would need to be compared with the claimed savings.

The ground-based alternative also runs into industrial bottlenecks. Musk says utility interconnection studies can take a year and cites turbine blades and vanes as a constrained part of gas-turbine production. His xAI Colossus 2 site used privately built power generation, but he argues this is not easy to generalize because power plants and their equipment must be sourced. He estimates that the United States averages about half a terawatt of electricity use and uses this scale to emphasize how difficult it would be to add data-center demand comparable to national consumption.

Solar manufacturing is presented as another possible response, though Musk says US tariffs and limited domestic production make near-term deployment harder. He states that SpaceX and Tesla are targeting 100 gigawatts per year of solar-cell production, and argues that orbital cells can be cheaper to manufacture because they need less glass and lighter framing. He also claims that Chinese solar cells cost about $0.25–$0.30 per watt and that space could reduce the effective cost by removing battery requirements. The interview treats these as Musk's estimates; it does not provide independent cost data or a full system-level comparison.

The scale of the ground-based challenge is illustrated by Musk's account of bringing one gigawatt online for xAI: the team had to combine many turbines, work through permitting problems in Tennessee, move the project across the state line to Mississippi, and build high-voltage lines and a power plant. He argues that people underestimate the difficulty of producing enough electricity for data centers. This account supports his claim that power availability, rather than chip output alone, may limit near-term centralized compute.

Musk argues that a data center's power requirement is much larger than the GPU nameplate load. Networking, CPUs, storage, peak cooling, and reserve capacity for servicing generators all add to the generation requirement. He estimates that 110,000 GB300 systems need roughly 300 megawatts once those factors are included, and 330,000 systems need about one gigawatt. The example is meant to show why chip output can outpace the ability to power centralized clusters, particularly in hot locations such as Memphis.

The constraints extend through the energy supply chain. Musk identifies turbine blades and vanes as a bottleneck for gas generation, saying only three companies cast them and that they are heavily backlogged. He says SpaceX and Tesla may need to make these parts internally. Solar is another route, but he points to tariffs, land, permits, batteries, and the pace of domestic cell production as obstacles. He states that both companies are targeting 100 gigawatts of solar-cell production per year.

Musk predicts that within five years, annual additions of AI capacity in space could exceed the cumulative total on Earth, with a few hundred gigawatts per year and potentially approaching one terawatt before rocket-fuel supply becomes limiting. Dwarkesh translates 100 gigawatts into roughly 10,000 Starship launches, or about one launch per hour for a year. Musk compares that rate with commercial aviation and says a fleet of 20 or 30 ships could potentially support it if each were reused about every 30 hours. These are ambitious projections whose feasibility depends on launch cadence, orbital operations, and the power systems discussed earlier.