The 200-Gigawatt Spike and the 2035 Forecast
The numbers coming out of power forecasting models are startling, even for veteran infrastructure engineers. According to a report by TechCrunch, data centers are projected to consume one-fifth of all electricity generated in the United States by 2035. That represents a quadrupling of current power usage levels in less than a decade.
What is driving this extraordinary curve? The immediate catalyst is the exponential deployment of artificial intelligence workloads. Total U.S. data center capacity is on track to expand to nearly 200 gigawatts over the next ten years. For context, 200 gigawatts is comparable to the peak generating capacity of several major regional power grids combined.
The pace of these baseline revisions tells an even bigger story. Financial and industry analysts keep underestimating how fast hyperscalers build out server racks—a capital shift mirrored in Alphabet's massive cloud infrastructure spend. BloombergNEF revised its 2035 electricity demand forecast upward by 83% compared to its own estimates published just seven months earlier in December. They aren't the only ones adjusting their math on the fly. The Electric Power Research Institute (EPRI) has more than doubled its 2024 projections, while S&P bumped its electricity demand outlook up by more than a third between October and April. Energy modeling isn't just creeping upward—it is jumping by double digits every quarter.
Why AI Training and Inference Demand So Much Power
Traditional cloud data centers run relatively steady state compute loads. Web applications, transactional databases, and object storage consume predictable power curves with recognizable diurnal peaks. AI workloads behave completely differently. High-density GPU clusters pull maximum power continuously for weeks during heavy foundation model training runs, followed by constant high-load baseline demand for real-time inference, driving interest in intelligent token caching to optimize GPU efficiency.
Nearly half of the projected 200-gigawatt capacity across U.S. facilities will be consumed explicitly by AI training and inference tasks. The United States continues to retain the lion's share of world compute infrastructure. By 2033, forecasts show that the U.S. will host 64% of global AI chips as measured by total power demand.
Concentrating nearly two-thirds of the world's high-performance AI silicon inside domestic borders creates acute localization challenges. You cannot easily distribute massive multi-gigawatt facilities across rural countrysides without building high-voltage transmission lines to connect them. Power plants, transformer stations, and substations are being asked to deliver unprecedented baseline loads to concentrated geographical clusters that simply weren't built for megawatt-per-rack heat and power density.
Regional Grid Strains Across PJM and ERCOT
While power consumption numbers sound broad at a national level, grid stress is localized. Two major regional transmission operators bear the heaviest burden: the PJM Interconnection and the Electric Reliability Council of Texas (ERCOT).
PJM manages the regional grid serving 65 million people across 13 states and the District of Columbia, stretching from Virginia through Ohio and into Illinois. Because Northern Virginia has historically served as the world's densest data center hub, PJM finds itself at the epicenter of the energy strain. BloombergNEF projects that data centers will absorb an astounding 34% of PJM's total electricity supply in the coming decade.
Further south, ERCOT handles roughly 90% of the electric load in Texas. Driven by rapid data center construction, cheap land, and streamlined local permitting, ERCOT is expected to devote 22% of its total generating capacity to powering data infrastructure.
Having one-third of a major multi-state interconnection dedicated to server racks isn't just an operational quirk. It fundamentally changes how grid operators manage base load reserves, voltage stability, and emergency curtailment protocols during extreme weather events.
Surging Electricity Costs and Interconnection Queue Delays
The physical realities of grid capacity are crashing directly into the software industry's demand for rapid scale. Connecting a new 500-megawatt data center campus to the high-voltage transmission system requires extensive engineering studies, equipment procurement, and utility approvals that can stretch across years.
PJM struggled so intensely with the backlog of connection requests from both large power generators and massive energy loads that it froze its interconnection queue entirely for four years. While PJM reopened the queue to new generating facilities in April, the backlog created severe operational friction. The situation grew tense enough that American Electric Power (AEP), one of the nation's largest electric utilities, threatened to withdraw from the PJM interconnection altogether due to concerns over how load growth and cost allocation were being handled.
These bottlenecks directly affect power pricing. In PJM's territory, wholesale electricity prices jumped 76% over the past year. Data centers aren't shying away from these higher price points either. In PJM's most recent capacity auction, data center operators accounted for 38% of all cleared charges. Cloud providers and frontier AI labs are showing a willingness to absorb significantly higher power tariffs just to secure guaranteed access to the grid before their competitors do.
The Global Impact of Rising Compute Power Requirements
The U.S. market is the focal point of the current compute buildout, but energy demand from artificial intelligence is expanding globally. If aggressive AI adoption trends continue along their current trajectory, global data centers will generate an additional 1,935 terawatt-hours of electricity demand by 2033.
To put 1,935 terawatt-hours into perspective, that single incremental increase matches almost the entire annual electricity consumption of India—a country with a population of over 1.4 billion people.
This global appetite for energy puts software infrastructure directly in competition with municipal utilities, industrial manufacturing, and residential heating for limited electrical capacity. Finding power for the next generation of model training centers is no longer a routine site-selection line item for procurement teams. It has become a central strategic constraint for the entire technology industry. Grid stability, regulatory policy, and power line construction timelines will dictate the pace of AI advancement far more than chip architecture or capital expenditures ever will, reinforcing why network infrastructure and smart request routing are becoming primary competitive moats.
