Elon Musk is targeting 15 to 20 gigawatts of compute capacity for SpaceX data centers by the end of next year, but he says memory semiconductors are the project’s biggest bottleneck. Speaking on a SpaceX conference call, Musk said demand for memory chips is rising about 200% year over year, while manufacturers are increasing unit volumes by only about 20% annually. The resulting shortage could support higher prices and continued capacity expansion by memory producers. The planned power draw would approach twice New York City’s roughly 10.4-gigawatt peak summer demand and would require vast quantities of DRAM and high-bandwidth memory to support AI models, training data and cloud services. Musk also said on August 4 that SpaceX would use Nvidia hardware exclusively going forward, citing the Vera Rubin architecture. Nvidia shares rose 3.4%, while AMD fell more than 8% despite record revenue of $11.5 billion, a 50% annual increase that exceeded the $11.3 billion Wall Street consensus. AMD’s data-center revenue doubled to $6.7 billion, or 58% of sales versus 42% a year earlier, and net income of $2.3 billion exceeded the $1.7 billion forecast. CEO Lisa Su said she has "tremendous respect" for Musk and reaffirmed guidance for server revenue to grow more than 80% in the second half of 2026. AMD continues to win major agreements with OpenAI, Meta and Anthropic, including an Anthropic deal covering up to 2 gigawatts of compute on AMD’s newest chips, but both Nvidia- and AMD-based facilities face the same memory constraint. Micron Technology may benefit from the shortage, while AMD’s valuation and the market’s reaction to Musk’s endorsement highlight the premium investors place on sentiment in AI infrastructure. Sustained memory scarcity could slow data-center construction, raise costs for cloud providers and pressure enterprise customers, making memory manufacturing an increasingly important part of the next phase of the AI trade.