SpaceX is recalibrating its AI data center build-out after reliability failures at sites in Tennessee and Mississippi, shifting priority from raw speed to power and cooling redundancy. The company is racing to meet a September 30 deadline tied to a $920 million-per-month compute deal with Google covering roughly 110,000 Nvidia GPUs; Google can reduce or cancel payments after a one-month grace period if capacity is not delivered.
The reliability push follows months during which some SpaceX facilities operated without backup cooling or power, causing primary system failures that interrupted training for Grok, the chatbot from SpaceXAI. In Mississippi, regulators found 69 temporary gas turbines at the Southaven plant feeding Colossus 2, twice the number SpaceX had previously disclosed. The NAACP and local residents filed lawsuits over air quality and turbine noise, while the Department of Justice argued that cutting power to Colossus 2 would threaten AI innovation and national security.
SpaceX's AI infrastructure spending reached about $15.8 billion to $16 billion in Q2, roughly twice Q1's level, bringing first-half AI capex to $23.6 billion. AI compute capacity grew from 0.4 gigawatts to 1.4 gigawatts, and AI revenue rose more than 210% to top $2.5 billion in Q2. The AI division still lost $3.7 billion in the first half. SpaceX has reshuffled data center leadership, placing Starlink executive Michael Nicolls in charge of AI infrastructure, and is building a Texas foundry to produce turbine blades and vanes in-house. Elon Musk wants SpaceXAI compute to reach 10 GW by the end of 2027, with projected annual revenue of $300 billion to $500 billion.
At the Goldman Sachs Communacopia and Technology Conference, CFO Bret Johnsen revealed a new unnamed AI hosting contract worth $1.11 billion per month starting December 1, or about $13.3 billion annualized. That adds to Anthropic's $1.25 billion monthly agreement through May 2029 and Google's $920 million monthly deal. Johnsen said AI compute deals are expected to exceed $3.4 billion per month by December, giving management 'even more conviction' in the $100 billion annual recurring revenue target for year-end 2026.
SpaceX is using only Nvidia chips across its data centers, and most hosting contracts are short-term—about 90 days with a 90-day option—so the company can reclaim compute for its own AI products. Longer term, SpaceX plans to launch its first orbital computing satellites next year using the V3 Starlink platform, with large-scale deployment in 2028. The orbital strategy depends on Starship reusability; a recent upper stage completed a precision ocean landing and was recovered. An upcoming Starship flight is expected to carry production V3 Starlink satellites.
