SpaceX makes a comprehensive adjustment to its AI data center construction model: expansion slows down, reliability is strengthened.

SpaceX is undergoing a systematic overhaul of its data center construction methods, a strategic shift that trades reliability for speed, marking a significant turning point in the infrastructure expansion path of Musk's AI business.

According to The Information, citing sources familiar with the matter, SpaceX's new management team is focusing on installing more backup power and cooling systems, demanding more comprehensive testing before data centers become operational, and working to reduce reliance on temporary power and cooling equipment. The pace of data center expansion may slow significantly.

This approach contrasts sharply with xAI's previous rapid deployment model of "building and operating simultaneously, with redundancy added later." Meanwhile, management is pushing forward with redesign projects in some existing data centers.

Last week, construction work accidentally affected a power line, causing a high-profile outage at SpaceX's large data center in Memphis, Tennessee.

SpaceX stated that the outage caused some Grok models to go offline and had a ripple effect on computing power leasing customers such as Anthropic and Google. Musk subsequently posted on the X platform that the company is "taking corrective measures to ensure that such incidents do not happen again."

As of press time, SpaceX shares rose 1.63% in intraday trading on Thursday. The stock initially surged over 3% in early trading following news of a $40 million satellite services contract from the UK government, before retreating somewhat.

In addition, according to the arrangements disclosed in the prospectus, SpaceX will have a lock-up period of 90 days after its IPO around September 9, during which up to 319 million Class A common shares will be eligible for sale, accounting for about 7% of the shares locked up for the 180 days.

Prioritizing speed can create hidden dangers.

SpaceX once considered the rapid construction of data centers as its core competitive advantage in the AI race.

xAI once claimed that its first data center was deployed with 100,000 GPUs in just 122 days—a record that Nvidia CEO Jensen Huang praised as a "superhuman" and "incredible" achievement in an industry where it usually takes years to complete a similar-scale construction.

To achieve this speed, xAI mobilized thousands of contractors and adopted a multi-process parallel operation approach to advance the construction: processes such as cooling pipe laying, electrical and network cabling are often carried out simultaneously in the same area, rather than waiting for the previous stage to be completed before starting the next stage.

SpaceX even emphasized its rapid construction capabilities as a key selling point in its IPO filing this June. However, this aggressive construction pace has come at a significant cost to reliability.

According to sources familiar with the matter, when the first two data centers went online, their redundancy was only a fraction of that of other hyperscale cloud computing service providers. This meant that a single point of failure could paralyze the entire computing cluster. In the early stages of the data centers' construction, frequent power outages forced AI researchers to suspend projects for hours and lose some model training progress.

Furthermore, because team members had been given a high degree of autonomy in procurement, the specifications of equipment within the data center varied, which further weakened system stability.

Major personnel reshuffle, Rockets team takes over

In late June of this year, Musk implemented a major overhaul of the data center management team, with SpaceX veteran Wesley Salandro replacing Dan Rowland as the team leader.

Rowland previously led infrastructure engineering work and at one point reported directly to Musk; before joining xAI, he worked on Tesla's Dojo supercomputer project, leaving the team shortly after Salandro took over. Salandro worked at SpaceX for over seven years, serving as Vice President of Production for the Falcon and Dragon rockets.

In the weeks that followed, several key personnel left the company.

Jake Palmer, who was in charge of physical infrastructure, Logan Beach, who was in charge of data center security, Tiffany Wilson, who was in charge of procurement, and Pablo Mendoza, who was in charge of finance, have all left the company.

Duke Energy site project engineer Aakash Verma moved to OpenAI in August, and site development manager Brent Mayo also left.

Liz Balke, a participant in the data center design, has now joined the Anthropic infrastructure team.

To fill manpower gaps, SpaceX has issued a call for volunteers to its rocket and Starlink engineers at Bastrop, Texas, and Cape Canaveral, Florida, for six- to eight-week rotational work assignments. More than 1,300 engineers reportedly applied, with over 300 officially joining the data center project in the past month.

Rebuilding order and slowing the pace of expansion

The new management team not only adjusted the technology roadmap but also restructured the organization. According to sources, SpaceX added several mid-level management positions to its previously relatively flat management structure, thereby reducing employees' authority over equipment procurement and on-site decision-making, which will objectively further slow down the construction progress.

Musk himself appears to support this new strategy. According to two sources familiar with the matter, he has been visiting the Memphis data center facility almost weekly in recent months, usually on Sundays, with his most recent visit being last Sunday.

Sources familiar with the matter said that although the new approach of building backup systems in advance and strengthening testing will inevitably slow down the overall construction process, SpaceX may not completely abandon xAI's previous accelerated construction model.

At the same time, striking a balance between rebuilding reliability and maintaining the pace of expansion remains a core challenge for this new team—especially given the rapid expansion of the computing power rental business and the increasing reliance on external customers.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.