Abstract:
SpaceX is adjusting the construction method of its AI data center, giving higher priority to the reliability of power supply and cooling systems and pre-operation testing. The new plan may extend the construction cycle of the data center, reflecting that the company began to modify its previous expansion model that focused on speed and computing power after the merger of xAI.
From rapid operation to reliability construction
SpaceX plans to complete more backup power and cooling systems and conduct more comprehensive testing before the data center is launched, while reducing the use of temporary power and cooling equipment. Some existing facilities may also be redesigned.
This adjustment is different from xAI’s model when building the Colossus data center in Memphis. The project once installed and activated 100,000 GPUs in 122 days. In order to compress the construction period, the construction team simultaneously promoted the cooling pipeline, power and network projects. The data center was operational before some backup systems were completed, and temporary power sources such as gas turbines were used for a long time.
Rapid construction brings operational risks. Frequent power outages in the early days forced AI researchers to suspend projects and lose some model training progress. On September 3, construction activities affected the power lines of a data center in Memphis, causing some Grok models to go offline and affecting some external computing customers. Musk later said the company was taking corrective measures to prevent similar incidents from happening again.
Rocket Engineer took over and many members of the original data center team resigned
This shift in strategy follows a management reshuffle. SpaceX completed the acquisition of xAI in February this year, and the rocket manufacturing and operations team began to take over the data center business at the end of June. Wesley Salandro, who has more than seven years of SpaceX experience and was responsible for the production of Falcon rockets and Dragon spacecraft, succeeded Dan Rowland to lead the relevant team.
SpaceX also recruits personnel from the rocket and Starlink departments for rotational support. Among more than 1,300 volunteers, more than 300 engineers have joined the data center project in the past month. At the same time, more than a dozen people in charge of the project have recently resigned, involving departments such as infrastructure, security, procurement and finance.
The new management has added middle management positions and tightened the authority of site personnel to purchase equipment and change construction plans. These measures may slow down construction but help unify equipment specifications, control costs and reduce single points of failure. As SpaceX begins leasing computing power to external customers such as Anthropic and Google, the impact of data center outages has expanded beyond Grok's own business, and the importance of reliability has increased.
Comments