The AI buildout bottleneck is shifting from memory and chips as agentic AI takes center stage, according to Goldman Sachs (NYSE: GS).
The advancement of agentic AI faces "critical physical bottlenecks," the investment bank's analysts said in a May 13 report, citing limited data center capacity, a projected drop in U.S. power output, a shortage of skilled grid workers, limited land and long supply-chain times for components like steel.
Goldman says agentic AI is 60-130 times more energy-intensive than AI chatbots, citing a June 2025 Korea Advanced Institute of Science & Technology study.
The agentic AI buildout through 2030 would require the U.S. to create 72 gigawatts of additional power, equivalent to 72 large nuclear power plants, Goldman Sachs analysts said. This is in addition to 760,000 grid workers and 207,000 experienced transmission and distribution workers, they said.
The market is too focused on data centers and chips and has yet to recognize these bottlenecks, Goldman said. They cited a 9x difference in earnings before interest and taxes between chip manufacturers, memory and server companies and power, component, and data center service companies.
"Many investors are still looking to replicate past successes in data centers, missing the critical chokepoints that will define the next phase of growth," they said.
Goldman Sachs' post comes even as memory and chip companies such as SanDisk Corp. (NASDAQ: SNDK), Seagate Technology Holdings (NASDAQ: STX), and Micron Technology Inc. (NASDAQ: MU) have been among the best performers this year, posting high triple-digit percentage gains so far this year.