Goldman Sachs: Agentic AI Demands Up to 130x More Power—and the Grid Isn't Ready

The Real AI Bottleneck

A Goldman Sachs Alternatives report published May 13 argues that the companies capturing roughly 90% of AI's profit pools today—chip designers, memory manufacturers, and semiconductor fabs—represent none of the physical bottlenecks that will determine whether AI can actually scale. Power generation, grid infrastructure, advanced cooling, and mission-critical services account for about 10% of AI-related earnings today, but 100% of the chokepoints.

The Agentic Energy Equation

AI agents use roughly 4x more computing tokens than standard chat interactions, and multi-agent systems use about 15x more. Goldman estimates agentic systems will be 60x to 130x more energy-intensive than the AI tools most people use today. Multiply that by persistent, always-on agents and aggregate infrastructure demand becomes exponentially larger than what current data centers were built to handle.

A 45 GW Power Shortfall

The U.S. faces a projected 45 gigawatt power shortfall for data centers by 2028, with 72 gigawatts of new capacity needed through 2030—the equivalent of 72 large nuclear power plants. Grid infrastructure upgrades and industrial cooling systems are the true investment bottleneck, not more GPUs.

A 600,000-Worker Gap

Nearly 600,000 jobs were advertised across skilled trades last year, while only about 150,000 new workers entered through apprenticeships. Before agentic AI adds a single gigawatt of new demand, the workforce needed to wire and cool that infrastructure is already deeply undersized.

Source: Fortune