SpaceX is adjusting its approach to building AI data centers, prioritizing backup power and cooling systems over rapid deployment speed, according to reporting from The Information. The strategic shift comes after the company experienced significant infrastructure setbacks that left multiple facilities without adequate redundancy.
For over 12 months, SpaceX had focused primarily on getting compute systems online quickly, treating backup power and cooling as secondary priorities. That approach has now reversed following extended outages at sites in Tennessee and Mississippi where backup cooling and power systems failed for months, causing primary systems to go offline and disrupting training for Grok, SpaceX's AI chatbot.
Racing Against a September Deadline
The reliability overhaul arrives as SpaceX faces a critical September 30 deadline tied to its agreement with Google. Under the deal, SpaceX committed to supply Google with access to approximately 110,000 Nvidia GPUs for $920 million per month. According to the agreement, Google may walk away, accept fewer GPUs, or reduce payments proportionally after a one-month grace period if the deadline is missed.
The urgency behind SpaceX's initial speed-focused approach becomes clear when considering the economics: AI data center revenue is estimated at $10 million to $12 million per megawatt per year, making early deployment potentially worth tens of billions of dollars over a facility's lifetime. SpaceX's capital spending reflects this pressure—the company spent approximately $15.8 billion on AI-related capital expenditure in Q2, double the Q1 figure, bringing first-half capex to $23.6 billion.
Infrastructure Challenges in Mississippi
The reliability issues began with power sourcing at SpaceX's Mississippi site, which supplies the Colossus 2 data center. The grid was unable to deliver electricity reliably, forcing SpaceX to rely on portable gas turbines. State law permitted the temporary turbines to operate for one year without air permits.
By July, Mississippi regulators identified 69 temporary turbines at the Southaven plant—twice the number SpaceX had disclosed. The site contained approximately 1,775 megawatts of capacity. In April, the NAACP filed a lawsuit, and residents initiated a separate class action suit citing air quality and noise concerns. The Department of Justice argued in SpaceX's favor, contending that power cuts would harm AI innovation and national security.
Supply chain delays compounded the problems. SpaceX waited an extended period for 41 permanent turbine units, forcing continued reliance on temporary equipment.
Organizational Changes and Future Capacity
SpaceX has restructured its data center operations in recent months, with several executives departing and Michael Nicolls, a Starlink executive, now overseeing AI infrastructure. The company is recruiting staff from its rocket and Starlink divisions to fill gaps.
To gain greater control over its supply chain, SpaceX is constructing a turbine manufacturing foundry in Bastrop, Texas. The company projects this in-house production could accelerate gas-turbine deployment by as much as 18 months.
Elon Musk has set an ambitious target for the AI division: reaching 10 gigawatts of compute capacity by the end of 2027, up from the current 1.4 gigawatts. Annual revenue at that scale is projected to reach between $300 billion and $500 billion, though customers like Google are now prioritizing both timely delivery and operational stability.
