US Data Center Power Demand Forecast to Consume 20% of US Electricity by 2035

It is widely assumed that the primary bottleneck to the artificial intelligence revolution is silicon—the sheer availability of advanced graphics processing units from companies like Nvidia and AMD. But the evidence increasingly points to a much more fundamental physical constraint: raw electricity. According to a new, heavily revised forecast from the research firm BloombergNEF (BNEF), data centers in the United States are on track to consume a staggering 20% of the nation's total electricity by the year 2035. This shift threatens to reorder the priorities of the national grid and test the limits of utility infrastructure.[1][2]

The projection represents a massive upward revision of the technology industry's anticipated energy footprint. Currently, data centers account for approximately 5.9% of total US power consumption. The BNEF report estimates that this figure will nearly triple over the next decade, pushing data center power needs to roughly 194 gigawatts. To put that immense scale into perspective, a single gigawatt is roughly equivalent to the maximum output of a traditional nuclear reactor. Achieving the 2035 forecast would mean the data center industry alone would require the continuous electrical output of nearly 200 nuclear power plants operating around the clock.[1][5]

This 194-gigawatt figure is 83% higher than the 106-gigawatt projection BNEF issued just months earlier in December. The rapid escalation in the forecast does not stem from facilities currently being built, but rather reflects the sheer volume of new, gigawatt-scale AI campuses entering the early stages of the development pipeline. As hyperscalers race to secure land and power for these massive projects, traditional electrical architectures and baseline grid assumptions are quickly becoming inadequate to handle the concentrated load.[2][6]

The mechanism driving this unprecedented surge is the fundamental architecture of modern AI hardware. Rack-scale solutions built specifically for AI training and inference—such as Nvidia's Vera Rubin and AMD's Instinct architectures—are becoming significantly more power-hungry with each successive generation. Nearly half of the projected 194-gigawatt capacity will be dedicated entirely to these intensive AI workloads, and the United States is expected to host 64% of the world's AI chips by power demand in 2033.[1][2]

The consequences for the broader US energy grid are profound and potentially disruptive. Lloyd Arnold, a BNEF analyst and co-author of the report, noted that by 2035, one out of every five units of energy generated by every coal plant, gas plant, and solar farm in the country will be routed directly to data centers. This is the exact same finite energy pool required to power growing cities, domestic manufacturing, and the rapidly expanding fleet of electric vehicles, setting up a fierce competition for baseline power generation.[6]

Meeting this exponential demand will require an unprecedented physical expansion of infrastructure, and current grid capabilities are already showing severe signs of strain. The record amount of data center capacity ever connected to the US grid in a single year stands at 7.1 gigawatts. Even if utility companies manage to sustain that record-breaking pace every single year for the next decade, BNEF calculates there would still be a 19-gigawatt supply shortfall by 2035 under their base-case scenario.[1][2]

This looming shortfall introduces significant execution risk for the artificial intelligence industry. Mark Daly, BNEF's head of technology and innovation, has pointed out that the US is currently a highly challenging environment for data center construction. Developers face intense, cross-industry competition for specialized labor and high-voltage power equipment. Lead times for critical components, such as the high-power transformers required to step down grid voltage for server use, are now stretching up to five years.[2][3]

Furthermore, the financial burden of this massive grid expansion is raising urgent questions about consumer impact and ratepayer equity. In regions like the PJM interconnection, tightening supply and rising data-center demand have already contributed to tens of billions of dollars in increased utility costs. If state regulators do not adjust how new infrastructure costs are allocated, everyday households and traditional businesses could face continued upward pressure on their monthly electricity bills to effectively subsidize the grid upgrades required by hyperscalers.[4]

Despite these mounting logistical and financial hurdles, the development pipeline continues to swell. While 15 state legislatures have considered temporarily banning data center development to protect their local grids and ratepayers, the broader national trend remains focused on aggressive expansion. The tech industry is now actively exploring alternative energy strategies, including co-locating data centers directly at nuclear power plants or investing in advanced geothermal and natural gas generation, to bypass congested utility queues and secure the power needed to maintain the pace of AI advancement.[3][4][6]