Home

Starship will probably not reduce your launch costs by much

Starship is a large and fully reusable rocket, which means that when it's fully operational it could bring internal launch costs to SpaceX to a few hundred dollars per kg to LEO, or even lower. Unfortunately, that does not mean customers will ever see this price — the price will be set by the balance between supply and demand for LEO launch, and with the advent of large scale AI compute, demand is increasing rather dramatically, as launch offers the most scalable (and asymptotically cheapest) way to create new electrical power. SpaceX itself is leading in the development of orbital inference compute and will be its own first customer, so other customers for Starship are competing against the SpaceX AI satellite for a ride on Starship. That means they should expect to pay at least the net present value to SpaceX of an AI satellite per kg. How much is that?

Let's make a crude estimate. Suppose the AI satellite is able to package 70W of compute per kg (this figure is from their public AI satellite page), and suppose it lasts on orbit for five years. Suppose further that the AI satellite is able to produce $5 / kWh in revenue, which is about the rate at which the Colossus datacenter is currently monetizing electricity. Then each kg launched produces about $3000 in yearly revenue, or if you apply a ten percent discount rate (a number I just pulled out of a hat) over five years, a total present value of about $12,000. If you figure $50,000/kW in capex for the satellite, dominated by the compute, the net present value would be something like $8,500. In other words, if you are bidding less than $8,500/kg, it would be economically rational for SpaceX to turn you down and launch more of their own satellites instead. Obviously, this is very sensitive to the $/kWh revenue figure, but in order for launch cost to be in the hundreds of dollars per kg it would have to be lower by more than a factor of 10. The precise value of a kWh depends on what you think the long term trajectory of AI compute growth is; for my part, I don't think demand for intelligence will slow any time soon, so I don't think we will see cheap Starship launches for third party customers.