SpaceX Restructures AI Data Center Team
Elon Musk has recently adjusted the team responsible for building AI data centers at SpaceX, replacing several leaders and bringing in executives from the rocket and Starlink satellite internet business to take over key positions. The adjustments come after existing facilities in Tennessee and Mississippi exposed civil engineering and operational reliability issues.
According to sources familiar with the matter, SpaceX AI prioritized speed of deployment during the early construction phase, resulting in some data centers operating for months without backup cooling and power systems. Such lack of redundancy can make single-point equipment failures, power fluctuations, or cooling failures more likely to lead to service interruptions.
Personnel changes include former Vice President of Falcon rocket and Dragon spacecraft production Wesley Salandro, and Logan McConnell, head of product support operations at the Starbase launch site, joining the AI infrastructure team. SpaceX has also brought in an engineer from Neuralink to assist with civil engineering matters and continues to recruit for related positions.
Jake Palmer, the former head of physical infrastructure at SpaceX AI, left at the end of July; reports also indicate that core members Zach Wells and Pablo Mendoza have recently left the team. The personnel changes, combined with engineering issues, have increased uncertainty regarding the launch timelines of new campus projects, including those in Texas, aside from the Memphis hub in Tennessee.
Current public reports do not disclose the specific addresses of the facilities in Tennessee and Mississippi, IT load sizes, actual downtime occurrences, financial losses incurred, or whether it has affected existing model training and inference services. The reports also do not show that SpaceX, SpaceX AI, or Musk have publicly responded to the engineering issues and organizational adjustments, so the information primarily comes from anonymous insiders.
Standard designs for data centers typically handle transformer, UPS, diesel backup power, cooling units, and power path failures with redundancy architectures such as N+1 or 2N; if the reports of "months without backup cooling and power" are true, it indicates that the project accepted a higher-than-industry-standard availability risk in pursuit of deployment speed. Whether this violates specific construction standards or contract SLAs cannot be determined from the existing public information.
In market mechanisms, the buyers of AI training and inference clusters are model developers, cloud platforms, and enterprise customers, with continuous operation relying on stable power, cooling, network, and data center operations. The involvement of SpaceX management, which has experience in launches, satellite networks, and large-scale manufacturing, indicates that the company is attempting to repair the infrastructure delivery chain with high reliability engineering capabilities; beneficiaries may include suppliers providing backup power, liquid cooling equipment, grid access, and engineering contracting, while those under pressure are construction and operation service providers who trade extreme compression of timelines for orders but cannot provide reliable redundancy.
Source: Public Information
ABAB AI Insight
SpaceX has long relied on a "rapid iteration—field testing—design correction" engineering approach to advance Falcon, Starship, and Starlink: rocket prototypes can expose issues during test flights, which can then be improved in the next round of manufacturing; however, data centers are continuous operation assets, and training clusters, customer tasks, and grid access cannot withstand the same frequency of trial and error. Bringing in rocket and Starlink managers does not mean directly copying the rocket-style experimental culture to data centers, but rather attempting to transplant supply chain, manufacturing rhythm, remote operations, and fault response capabilities into facilities management that emphasizes availability.
The contradiction in capital pathways lies in the return logic of AI infrastructure: securing power, land, and GPUs early can lock in customers and computing revenue sooner, but the heavy asset returns of data centers depend on long-term stable operations. The lack of backup power and cooling may reduce initial capital expenditures and shorten construction periods, but it will increase downtime, equipment wear, defaults, and later renovation costs. If SpaceX AI rapidly expands before the project is sufficiently engineered, the reliability costs that should have been reserved for the operational phase will return to the balance sheet in the form of rework, delays, and higher financing risks.
This contrasts sharply with the routes taken by Meta, Google, and Microsoft in building ultra-large-scale data centers: they typically accumulate years of experience in building their own data centers, standardizing modules, ensuring power redundancy, and establishing global operational systems before replicating scale in new campuses. SpaceX AI is attempting to quickly cut into the AI computing demand window, and its current stage resembles the early cloud computing expansion or the infrastructure race of cryptocurrency mining in the 2010s—first competing for power and computing positions, then addressing long-term unit economics and stability. The reported uncertainty in the pace of projects in Memphis, Mississippi, and Texas indicates its transition from "rapid construction" to "engineering governance."
Essentially, this is a reconstruction of the industrial chain: the core assets of AI competition extend from model capabilities to sustainable operational physical systems, where power capacity, cooling redundancy, grid connection permits, construction quality, and maintenance response collectively determine the number of sellable computing hours. Model companies can accept short-term functional errors, but cannot tolerate long-term stoppages of training clusters due to physical facility failures; thus, data centers are no longer just containers for servers, but constraints on AI commercialization capacity. Whoever can close the loop on speed, reliability, and financing will hold the pricing power of infrastructure.
ABAB News · Cognitive Laws
Software can be hot-fixed, but power outages can only lead to downtime.
Rapid construction wins orders, but redundancy determines whether delivery can be made.
The upper limit of computing power is defined by chips, while cash flow is defined by reliability.