Other organizations have also revised their projections upward. Lawrence Berkeley National Laboratory estimates data centers could consume approximately 11.8% of U.S. electricity by 2030 in its reference case, while the Electric Power Research Institute projects a range of 9% to 17% by 2030 and as much as 20% by 2035. Rhodium Group similarly forecasts data centers could reach 18% of U.S. electricity demand by 2035 under a high-growth scenario.
The rapid increase is being driven by continued investment in AI infrastructure rather than slowing hardware demand. While AI model developers continue improving efficiency, newer systems often rely on larger architectures and more advanced memory technologies. The report points to Moonshot AI's Kimi K3 model as an example, noting that although it improves computing efficiency, its expanded architecture increases demand for high-bandwidth memory, sustaining the need for advanced hardware from companies including Nvidia, SK Hynix and TSMC.
BloombergNEF expects nearly half of future data center capacity to support AI training and inference workloads, with the United States remaining the dominant market. By 2033, the firm projects the country will host 64% of global AI chips based on power demand.
Despite the strong outlook for demand, electricity infrastructure is emerging as one of the industry's biggest constraints. Regional transmission systems are struggling to connect both new power generation and large data center projects, creating lengthy interconnection queues that could delay future expansion.
BloombergNEF forecasts that data centers alone could account for 34% of electricity demand across the PJM Interconnection by 2030, while ERCOT, which serves most of Texas, could dedicate 22% of its generating capacity to data centers. PJM has already experienced significant congestion after pausing new interconnection requests for several years, and electricity prices within the system have climbed sharply as demand has outpaced available supply.
Supply challenges extend beyond transmission. Large natural gas turbines remain in limited supply through the end of the decade, prompting developers to pursue alternative approaches such as on-site generation, behind-the-meter power systems and co-located energy projects. According to Bank of America analysts, more than 7.5 gigawatts of data center projects with on-site generation are already under construction, with more than 60 gigawatts in pre-construction.
Regulatory requirements and permitting timelines are also influencing where new facilities are built. Environmental reviews, zoning disputes and local opposition have delayed projects in some states, while developers increasingly favor regions with faster permitting processes and existing industrial infrastructure.
As electricity demand accelerates, nuclear energy is becoming a larger part of hyperscalers' long-term power strategies because it can provide continuous carbon-free electricity. Microsoft, Meta, Google and Amazon have all announced nuclear-related energy agreements or investments intended to support future AI infrastructure, while the broader industry has committed to nearly 9.8 gigawatts of nuclear capacity across multiple projects.
BloombergNEF estimates that if AI adoption continues on its current trajectory, data centers worldwide will create 1,935 terawatt-hours of additional electricity demand by 2033, nearly matching India's current annual electricity consumption.
The report concludes that data centers consuming roughly one-fifth of U.S. electricity by 2035 is achievable under aggressive growth assumptions, but only if generation capacity, transmission infrastructure and reliable power sources expand quickly enough to keep pace with AI-driven demand.
This analysis is based on reporting from Energy News Beat.
Image courtesy of Microsoft.
This article was generated with AI assistance and reviewed for accuracy and quality.
