SpaceX Adjusts Data Center Construction, Slows Expansion
SpaceX is systematically adjusting its approach to building AI data centers, with the pace of expansion at new sites likely to slow significantly. Insiders say the company will no longer follow the previous model of "building while operating, with redundancy added later," but will instead require that backup power and cooling systems be fully in place and tested before commencing operations.
Recently, Elon Musk has redirected resources from rocket and Starlink projects to ground facility construction. Wesley Salandro has taken over the data center team from Dan Rowland; Salandro has been with SpaceX for over seven years and previously managed Falcon and Dragon production. Rowland, who was involved with Tesla Dojo and reported directly to Musk, left shortly after the transition. Infrastructure, security, procurement, and finance leaders including Jake Palmer, Logan Beach, Tiffany Wilson, and Pablo Mendoza have also departed.
The Tennessee and Mississippi sites previously rushed to install equipment, resulting in some facilities lacking complete backup power and cooling for several months. The Macrohard site relied on over 100 mobile chillers, with availability falling below the internal target of 99.9%; the Mississippi site used temporary gas turbines due to supply chain delays affecting 41 permanent units, leading to power outages that interrupted model training.
To fill staffing gaps, the company has called for engineers from the Texas Bastrop and Florida Cape Canaveral rocket and Starlink teams to rotate for six to eight weeks, with over 1,300 people signing up and more than 300 officially entering the data center site in the past month. Meanwhile, the company is beginning to reduce its reliance on temporary power and cooling, and is reinforcing existing facilities, while expansion plans for locations outside Memphis, including new sites in Texas, have been postponed.
SpaceX reported that by the end of June, approximately 1.4 GW of computing power was online, with a target of exceeding 2 GW by year-end; AI infrastructure spending in Q2 was about $15.8 billion, with AI-related revenue around $2.6 billion. Google has agreed to pay about $920 million per month for approximately 110,000 Nvidia GPUs at full capacity, corresponding to a production milestone on September 30; Anthropic has locked in Colossus 1 computing power for about $1.25 billion per month. The company is also establishing a turbine blade and guide blade factory in Texas, aiming to shorten the gas turbine delivery cycle by about 18 months.
From a market mechanism perspective, this is a corrective action driven by events rather than a disappearance of demand. Buyers include Google, Anthropic, and SpaceX's own training needs, while sellers are SpaceX's rocket capacity, temporary power, and GPUs being deployed to ground facilities; funding continues to flow towards Nvidia chips, gas turbines, cooling equipment, and on-site engineering, but acceptance criteria have shifted from "launch first, then add redundancy" to "ensure redundancy before billing." Beneficiaries include contracted tenants sensitive to availability and suppliers capable of delivering permanent power equipment; those under pressure include short-term capacity built on temporary units and SpaceX itself, which must solidify availability before contract milestones.
On a supplementary level, the slowdown in ground expansion runs parallel to the narrative of orbital computing power: the company still plans AI satellites and Gigasat production lines, but recent contract fulfillment relies on the available uptime of the Memphis-South Haven cluster rather than on yet-to-be mass-produced in-orbit cabinets.
Source: Public Information
ABAB AI Insight
xAI converted the old Memphis factory into Colossus by using temporary gas turbines, mobile chillers, and rented cabinets to compress the construction cycle to a few months. This approach is expected to be seen as a dimensional reduction of the traditional multi-year data center cycle in 2024-2025; after merging with xAI in February 2026, SpaceX replicated this speed logic onto the public company's balance sheet, with Q2 capital expenditures rising to around $18.4 billion. Historical comparisons are clear: Tesla Dojo was long stuck on self-developed chips and factory coordination, and the replacement of Rowland indicates that Musk is not betting on "the same infrastructure team working twice as fast" but is instead transferring the procedural discipline from the Falcon production line to the data center.
The capital path is "buy time first, then buy structure." Money is initially directed towards Nvidia cabinets and quickly deployable gas turbines, then monetizing capacity through Google’s monthly lease of about $920 million and Anthropic’s monthly lease of about $1.25 billion; concurrently rotating engineers from Bastrop and Cape Canaveral effectively transforms the workforce inventory from rocket manufacturing into civil engineering and electromechanical strike teams. The motivation is not to slow down the AI narrative but to base contract billing on available uptime, as a single power outage could turn a lease from a growth asset into a contingent liability, thus shifting resources from "installing more cabinets" to "turning installed cabinets into deliverable power and cooling."
The comparison is not with traditional hosting providers but with how Amazon built its fulfillment network to solidify Prime delivery times and how Microsoft pre-purchased Azure power and cabinets to lock in OpenAI: both initially used aggressive investments to capture market share, then applied engineering discipline to underpin contracts. The industry phase has shifted from "who can stack gigawatts first" to "who can turn gigawatts into billable 99.9%." SpaceX is still in an expansion phase, but control has shifted from project manager-style rush work to the rocket production system taking over on-site processes, testing gates, and spare parts inventory.
Structurally, this is a reconstruction of the industry chain, with the mechanism shifting pricing power from installed capacity to available capacity. As grid access, permanent unit blades, and cooling redundancy become bottlenecks, temporary equipment can only account for capacity in advance but cannot lock in SLA early; those who can self-manufacture key rotating parts and move testing gates forward will determine whether the next round of leases can generate monthly income. Thus, the slowdown in ground expansion is not a retreat but a rewriting of the non-financeable speed gap into a deliverable infrastructure gap.
ABAB News · Cognitive Laws
Speed can capture market share, redundancy can generate rental income.
Contracts are priced not on installed capacity, but on uptime hours without outages.
Transferring production line discipline to a new battlefield is the true cross-border leverage.