SpaceX Overhauls Data Center Build-Out, Potentially Slowing Expansion
The Takeaway
- SpaceX’s new data center strategy prioritizes reliability over rapid build-out.
- Frequent outages plagued xAI’s initial fast-built data centers.
- Rocket engineers now lead data center team, replacing veteran staff.
Elon Musk is famous for his “move fast” management philosophy, which he demonstrated most starkly when he built new data centers for his AI startup in record time two years ago. But a new team of rocket engineers Musk recently installed to run his data centers is taking a very different approach.
SpaceX is changing how it builds the giant facilities that house AI servers in a way that might slow the building of new campuses, according to people familiar with the situation. The new management team is looking to install more backup systems for power and cooling, and to more thoroughly test those systems before the data centers start operating, instead of adding them later on, as the AI company was doing, two of the people said. SpaceX also wants to reduce the use of temporary power and cooling systems in its data centers, one person said.
In some cases, the new management is redesigning portions of existing data centers, the people said.
The new approach marks a shift from the one xAI—Musk’s AI company, now part of SpaceX—took when it initially began building its giant Tennessee data center campus, Colossus, in 2024. At the time, Musk chose to prioritize speed and computing capacity over building out the infrastructure needed to increase reliability, the people said.
But in late June, Musk shook up the management of SpaceX’s data center team, replacing data center veterans with people from SpaceX’s rocket launch team. (SpaceX acquired Musk’s xAI startup earlier this year.)
The new management of the data center team told some staff when they took over this summer that SpaceX engineers had been brought in to improve reliability at the company’s facilities in Tennessee and Mississippi, two people said.
Early on, frequent outages forced the company’s AI researchers to put projects on hold for hours at a time and lose some training progress, one person said. Reliability may be even more important for SpaceX now that it is renting out computing capacity to other companies, such as Anthropic and Google.
Reliability had improved earlier this year but there are signs the management overhaul could be causing new problems. Construction work affected a power line and caused a high-profile outage at one of SpaceX’s massive data centers in Memphis, Tenn. last week, according to a person familiar with the situation. The outage knocked some of SpaceXAI’s Grok models offline and had ripple effects also hurting some of its compute customers, SpaceX said.
A spokesperson for SpaceX did not respond to a request for comment.
Slower Build-Out
Musk appears to support the new policy. In recent months, he has visited the facilities in Memphis on a nearly weekly basis, stopping by most Sundays, two people said. He most recently visited the site this past Sunday, they said.
In the past, he repeatedly touted his company’s ability to rapidly build data centers as a key advantage in the AI race. SpaceX, which acquired Musk’s xAI startup earlier this year, even cited the speed factor in its initial public offering paperwork in June.
SpaceX has said the AI unit brought its first data center online with 100,000 graphics processing units in just 122 days—an unprecedented timeline in the industry, where it can take years to build a data center of that size. Nvidia CEO Jensen Huang has praised the team’s efforts as “superhuman” and “unbelievable.”
To move quickly, xAI brought on thousands of contractors and built different parts of the facility in parallel, with crews installing cooling pipes, pulling electrical and networking cable, and carrying out other work simultaneously, often in the same areas of the facility, rather than waiting for completion of one task before starting the next phase, the people said.
But the speed has come with some trade-offs. Many big cloud firms set up their data centers in phases to ensure things go smoothly and the center can operate reliably, one person said.
SpaceXAI’s process included bringing the minimum power and cooling infrastructure for the data center online before finishing backup systems, and then adding some of those backups months later, people familiar with the matter said. The company has relied on temporary power and cooling systems, such as dozens of gas turbines, since the data centers opened, one person said.
When they first came online, the company’s first two data centers had only a fraction of the redundancy of data centers other hyperscalers operated, the person familiar with the design said. That left them vulnerable to outages, because a single equipment failure could take entire clusters offline, the person said.
Individuals on the data center team also had the authority to approve changes to construction plans and equipment, and they could buy equipment as needed rather than optimizing for costs, the people said. At times, that meant that there was a mix of different parts used in the data centers, rather than a uniform supply of racks and servers, one person said. That could also undermine reliability, another person familiar with the situation said.
To be sure, while building backup systems earlier and doing more testing is likely to slow the construction process, SpaceX probably won’t abandon xAI’s accelerated construction model entirely, one person said.
Data Center Exodus
SpaceX completed its acquisition of xAI in February, at a time when the startup was heavily burning cash. A few months after the deal closed, SpaceX leaders held information-gathering meetings with members of the data center team, one of the people said. (SpaceX held similar meetings at xAI’s office in Palo Alto, Calif., shortly before dozens of engineers left the startup.)
Over the past few weeks, more than a dozen leaders on the data center project have left the team, replaced with SpaceX engineers and leaders with manufacturing and operations backgrounds outside data center construction.
The exits came shortly after SpaceX’s Wesley Salandro took over leadership of the team from Dan Rowland, who had led the infrastructure engineering and at one point had reported directly to Musk. Salandro has worked at SpaceX for more than seven years and was vice president of production for the Falcon and Dragon rockets. Rowland, a leader on Tesla’s Dojo supercomputer project before he joined xAI in 2024, left shortly after Salandro took over.
Along with Salandro, SpaceX leadership issued a call for some of its rocket and Starlink engineers in Bastrop, Texas and Cape Canaveral, Florida to volunteer to help with the data centers for six to eight week rotations, two people said. More than 1,300 engineers volunteered and over 300 SpaceX engineers have joined the data center project over the past month, they said.
Within a few weeks, Jake Palmer, who led physical infrastructure for SpaceXAI, left. So did Logan Beach, head of security for data centers; Tiffany Wilson, head of procurement for data centers; and Pablo Mendoza, head of data center finance. Aakash Verma, SpaceXAI’s project engineer for its Duke Energy site also left to join OpenAI in August, along with Brent Mayo, who was the manager of site builds at SpaceXAI. Liz Balke, who helped design the data centers, now works on the infrastructure team at Anthropic.
SpaceX leaders have also added more middle managers in recent months to what was once a relatively flat organizational structure, two people said. This means workers have less autonomy to approve equipment orders or make decisions onsite which can slow down the build process, the people said.
There could be more shake-ups to come. After last week’s data center outage, Musk wrote on X that the company was “taking corrective action to ensure this does not happen again.”
