Elon Musk: Space team building data centers on the ground is "asymmetric warfare," with 10 gigawatts of compute expected to be deployable next year
Odaily reports: SpaceX released its third-quarter financial results after U.S. markets closed early this morning. During the investor call following the earnings release, Elon Musk discussed SpaceX's superintelligence (SI) business.
Elon Musk stated that people are now beginning to realize that the main constraint on SI deployment is actually power, cooling systems, and infrastructure—not chips. One thing SpaceX has done is apply the capabilities of a spaceflight team to the deployment of ground-based computing infrastructure—among the total mass of all payloads entering Earth orbit globally, SpaceX's market share is already close to 90%... The results this team has achieved are an order of magnitude higher than the combined performance of all other competitors worldwide in the space sector. Now SpaceX will devote this same team to the deployment of ground-based computing infrastructure, which is somewhat like having the New York Yankees play in a children's baseball league... The result is that SpaceX deploys power and cooling infrastructure far faster than anyone else, and may even at some point exceed the combined total of all other competitors.
Elon Musk added: "Recently, I have personally devoted a great deal of time to Terafab, often spending four days a week there, busy from morning to night getting Terafab's supercomputing cluster up and running. At the current pace, we will be able to add deployment of approximately 200,000 GB300 GPUs every month for the rest of this year, and we will maintain this pace next year as well. Next year's computing resources will mainly use the Vera Rubin platform. At present, I am quite confident that we can add 10 gigawatts of compute by 2027. That does not include the roughly 3 gigawatts of compute expected to come online before the end of this year, or rather 2.5 to 3 gigawatts. In other words, if by the end of next year we have less than 13 gigawatts of compute online, I would actually be surprised."
