Musk Claims SpaceX is Accelerating the Expansion of Computing Facilities
SpaceX CEO Elon Musk responded to The Information, stating that the report was inaccurate and that the company is rapidly accelerating. The original post suggested that the expansion of ground-based AI data centers was slowing down. He did not provide new capacity figures in this response.
The company is also advancing the orbital data centers launched by Starship. The Starmind project plans to begin deployment and scale next year, with a maximum declaration of 1 million units. CFO Bret Johnsen stated at a Goldman Sachs conference on September 10 that ground data centers aim to achieve 5 to 10 gigawatts by 2027, with orbital capacity following. At the International Astronautical Congress, it was noted that each satellite has about 250 kilowatts, with approximately 10 clustered together, and within each cluster about 10 terabits per second.
On the ground side, there are already two Colossus facilities in Memphis and Southaven, Mississippi, with a third named Minihard. By the end of June, the computing power was 1.4 gigawatts, up from 0.4 gigawatts a year ago. In the second quarter, AI infrastructure spending was approximately $15.8 billion. Google pays $920 million monthly at full capacity, corresponding to about 110,000 NVIDIA graphics processors, with a delivery deadline of September 30.
The Information previously reported that some sites in Tennessee and Mississippi had been without backup power and cooling for months. Macrohard has used over 100 mobile cooling units, with an online rate below the internal target of 99.9%. In Mississippi, the delay of 41 permanent units has caused temporary gas turbines to run beyond their expected duration. Jake Palmer, the head of infrastructure, left in July, and Kevin Holland, a veteran of Starship, took over, with the task of ramping up the next six phases of computing power by the end of the year.
One facility in Memphis is said to exceed 2 gigawatts. Musk himself is living in a trailer on-site to monitor the construction. He also mentioned that the company is building its own turbine blade factory in Texas, reducing the lead time for new gas turbines by up to 18 months.
The buyers are Google, which purchases computing power monthly, and self-use model training, while the sellers are the power grid, gas turbines, and NVIDIA chips. The event was driven by the denial of the "slowdown" report, with funding continuing to flow into construction and orbital launches, rather than being paused. Beneficiaries include chips, turbines, and Starship launches, while the local power grid and the online rate supported by temporary units are under pressure.
Source: Public Information
ABAB AI Insight
SpaceX integrated with the original xAI in February 2026, converting the Memphis Colossus from a model company's server room into a rocket company's deliverable. Google registered $920 million monthly in June for about 110,000 graphics processors, with September 30 as the full capacity milestone. The Information reported that backup power and cooling have not kept pace, while Musk denied the slowdown with "rapidly accelerating"; the two sets of facts contrast speed and online rate, not whether operations have stopped.
The capital path is divided into two segments. The ground aims to achieve 5 to 10 gigawatts by 2027, while the orbital capacity will utilize Starmind to connect with Starship's transport capacity. Kevin Holland was transferred from Starship to manage the six phases due by the end of the year, effectively using the launch team to manage the server room. The Texas turbine blade factory has brought the 18-month turbine lead time into its own facility. The declaration of 1 million units and the structure of 10 units per cluster shifts the power bottleneck from the Tennessee grid to the cost of entering orbit.
A comparison can be made to Meta building a gas power plant to supply its AI cluster, and Microsoft purchasing nuclear power from Constellation to restart the Three Mile Island unit. The difference is that SpaceX simultaneously possesses the launch vehicle. The situation is stuck in ground delivery reliability reports while the orbital plan is still in the parallel phase for next year's debut.
This is a restructuring of the supply chain. Computing power, which used to be purchased from the power grid and hosting providers, is now sourced from its own turbines, chip contracts, and rockets. The acceleration denial pertains to the halt of operations, not the denial of the absence of backup systems. The orbital data center frames the same gap as a launch demand.
ABAB News · Cognitive Law
Denial of slowdown does not equate to backup power being in place.
Monthly rental computing power turns the construction period into a contract.
When the ground grid is insufficient, rockets are framed as outlets.