SpaceX may adjust its data center strategy after major power and cooling setbacks

SpaceX adjusts its AI data center construction strategy: shift from speed to redundancy

According to The Information, SpaceX (NASDAQ: SPCX) is changing the way its AI data center is built, choosing to put redundancy in power and cooling systems above mere construction speed. The slight adjustment could put a brake on SpaceX's infrastructure expansion, putting it under huge pressure as the September 30 deadline approaches, involving a $920 million monthly computing power agreement.

Why "redundancy" is the current focus

This shift is a reversal of SpaceX's previous successful path. For more than a year, the company has worked hard to ensure that computing power is online, and redundant design of backup power supplies and cooling systems is often seen as secondary considerations. However, the wind direction has changed. Multiple SpaceX sites in Tennessee and Mississippi lacked backup cooling or power support for months. When the main system fails, the machine also stops. These system failures have affected the training process of SpaceX AI chat robot Grok.

Although relevant reports have not yet been fully verified, it is worth noting that if power and thermal design do pose challenges, this is not unique to SpaceX, but a common problem facing the entire industry.

Countdown to the US$920 million agreement

The redundancy fix comes as SpaceX is racing against time to meet its September 30 deadline agreement with Google. Under the agreement, SpaceX will provide Google with access to approximately 110,000 Nvidia GPUs at a monthly fee of $920 million.

The chips must be delivered to Google by September 30. Under the terms of the agreement, if it fails to meet the standards after a one-month grace period, Google has the right to withdraw from the cooperation, accept a smaller number of GPUs, or reduce payments proportionately. This is why SpaceX prioritized speed optimization in the early days: for every day the data center was shut down, revenue lost one point. Cleanview estimates that the annual revenue of AI data centers is approximately US$10 million to US$12 million per megawatt. So getting large sites online early means tens of billions of dollars in revenue over their lifetime.

SpaceX's spending further demonstrates the urgency of this matter. In the second quarter, its AI-related capital expenditures were approximately $15.8 billion, double the level in the first quarter. This brought total capital expenditures in the first half to US$23.6 billion.

In addition, as computing power capacity increased from 0.4 GW to 1.4 GW, and AI revenue in the second quarter increased by more than 210% year-on-year to exceed US$2.5 billion, the AI division still lost US$3.7 billion in the first half of the year.

Turbines, licenses and the Mississippi controversy

Reliability issues start with SpaceX's power source. Its Mississippi site provides power to Colossus 2, but the site's grid failed to supply power in time. To that end, SpaceX uses portable gas turbines, which can be used for up to a year under state law without an air permit.

As of July, Mississippi regulators had discovered 69 temporary turbines at the South Bay plant, double the number previously disclosed by SpaceX. The total capacity of the site is approximately 1775 MW.

In April this year, the National Association for the Advancement of Colored People (NAACP) filed a lawsuit, and residents filed a separate class action lawsuit complaining about air quality issues and turbine noise. The U.S. Department of Justice has defended SpaceX, saying cutting off power to Colossus 2 would pose a risk to AI innovation and national security.

Supply chain issues exacerbate this dilemma. Due to the delay in the arrival of 41 permanent equipment, SpaceX has had to use temporary turbines for a long time in Mississippi.

Changes behind management reorganization

SpaceX has begun to adjust its organizational structure. Several data center executives have left in the past few months, and the company has appointed Starlink executive Michael Nicolls to take charge of its AI infrastructure. SpaceX is pulling people from the rocket and Starlink projects to fill the gaps.

Elon Musk and his team are trying to take control of the hardware supply chain. The company is building a wafer foundry in Bastoop, Texas (Note: The original context here refers to the manufacturing of hardware facilities such as turbine blades, and is not a traditional semiconductor wafer factory. It is recommended to be understood as a precision manufacturing base). The base will produce turbine blades and guide vanes in-house. Musk said the move could advance the deployment of new gas turbines by as much as 18 months.

Musk has set clear goals for employees: He hopes SpaceX's AI computing power will reach 10 gigawatts by the end of 2027, much higher than the current approximately 1.4 gigawatts. Annual revenue is expected to be between $300 billion and $500 billion. Customers like Google both want production capacity on time and want those production capacity to run steadily.