U.S. Power Grid Threatens AI Growth, Flex CEO Warns

The United States risks throttling its own artificial intelligence ambitions because the electrical grid cannot physically support the technology’s surging energy demands, according to Revathi Advaithi, chief executive of global supply chain and manufacturing giant Flex.

Advaithi delivered the assessment during a keynote conversation with Fortune executive editorial director Diane Brady at the Most Powerful Women Summit in Santa Barbara, California. Her warning shifts the focus away from the semiconductor shortages and software breakthroughs that typically dominate discussions about America’s AI competitiveness, pointing instead to a much older, heavier physical constraint: high-voltage transmission infrastructure built largely in the mid-20th century.

A 2025 study from the University of California, Berkeley projects that data centers will consume approximately 11.8 percent of total U.S. electricity generation by 2030. That figure represents a massive acceleration from historical baselines. For context, the entire U.S. residential sector currently accounts for roughly 38 percent of national power consumption. Concentrating nearly 12 percent of the country’s electrical output into specialized computing facilities requires localized grid capacities that most American municipalities simply do not possess.

The mismatch between hardware development cycles and civil engineering timelines forms the core of the bottleneck. Advaithi noted that developing a new advanced microchip takes roughly five years. Building the generation plants, substations, and high-voltage transmission corridors required to power the factories running those chips takes decades. The U.S. permitting process alone—spanning environmental reviews under the National Environmental Policy Act, state utility commission approvals, and local zoning battles—routinely delays major transmission projects by seven to ten years before construction begins.

Flex occupies a unique vantage point on this problem. As one of the world’s largest Electronics Manufacturing Services providers, the company builds hardware for major technology brands and operates dozens of massive manufacturing campuses globally. Advaithi oversees supply chains that directly feed the companies building AI servers, giving her visibility into both the demand signals from hyperscalers and the physical limitations of the facilities expected to produce and house their equipment.

Her critique extends beyond government regulators to the technology industry itself. Data center operators have historically favored proven, standardized cooling and power distribution architectures over experimental designs, prioritizing uptime guarantees over rapid innovation. This risk aversion is understandable—a single hour of downtime at a major cloud facility can cost millions of dollars—but it slows the adoption of higher-efficiency power systems that could ease the burden on regional grids.

The fundamental architecture of alternating current power delivery has remained largely unchanged since Thomas Edison and George Westinghouse established the first commercial grids in the 1880s and 1890s. While generation sources have diversified to include nuclear, wind, and solar, the physical mechanism of stepping voltage up for long-distance transmission and stepping it down for local use relies on aging transformers and copper lines. Much of the U.S. high-voltage transmission network was constructed in the 1960s and 1970s with a 50-year lifespan, meaning a significant portion is already operating past its intended engineering life.

Advaithi predicted the power deficit will persist throughout the next decade unless federal and state authorities drastically compress approval timelines for next-generation distribution technologies. She emphasized that safety concerns inherently slow progress in electrical infrastructure, as failures carry immediate consequences for public health and economic stability, unlike software bugs that can be patched remotely.

Hyperscale operators including Microsoft, Amazon, and Meta have attempted to bypass these grid constraints by signing direct power purchase agreements with nuclear plant operators and funding small modular reactor startups. However, these private workarounds address only a fraction of the broader national capacity shortfall facing the wider AI ecosystem, which includes smaller developers, enterprise adopters, and the manufacturing base that supports them.

Federal energy regulators are currently reviewing proposed reforms to interconnection queue processes, which dictate how new power generation projects connect to the existing grid. As of late 2026, more than 2,000 gigawatts of proposed generation and storage capacity remain stalled in regional transmission organization queues across the country, waiting for engineering studies and regulatory clearance.