Data Center Electricity Demand Surges: Could Account for One-Fifth of Total U.S. Electricity Consumption by 2035
The rapid development of artificial intelligence (AI) is reshaping the U.S. electricity landscape. According to forecasts from Goldman Sachs, Lawrence Berkeley National Laboratory, and other institutions, under the most aggressive scenario, data centers could consume 20% of total U.S. electricity by 2035—four to five times the current 4-5%.
Even as AI model developers announce efficiency improvements, underlying hardware demand continues to soar. Taking Moonshot AI's Kimi K3 model released in July 2026 as an example, while computational efficiency has improved, its larger architecture places higher demands on infrastructure such as high-bandwidth memory, which in turn drives strong demand for advanced chips and supporting systems from companies like NVIDIA, SK Hynix, and TSMC, further accelerating data center expansion.
Current Trajectories and Forecasts
Currently, data centers account for about 4-5% of U.S. electricity consumption. The latest underlying analysis shows that forecasts are being rapidly revised upward:
Lawrence Berkeley National Laboratory (2025 update): The reference scenario projects data centers will account for approximately 11.8% (649 TWh) of U.S. electricity by 2030, with a sensitivity range of 9.5-15.3%.
Electric Power Research Institute (EPRI, 2026 "Smart Power Supply"): 9-17% by 2030, and 10-20% by 2035.
Rhodium Group (high-growth scenario): 14% by 2030, and 18% by 2035.
BloombergNEF (BNEF): U.S. data center electricity demand will reach 106 GW by 2035, a 36% increase from previous forecasts, with nearly a quarter of new projects exceeding 500 MW in scale.
BNEF analyst Lloyd Arnold noted: "One out of every five kilowatt-hours generated in the U.S. goes to data centers. That's equivalent to the same amount of electricity originally used for electric vehicles, urban power supply, and other sectors."
Will Growth Slow Due to Grid Access, Regulations, and Natural Gas Constraints?
In the short term, grid access and transmission bottlenecks are the biggest constraints. Regional transmission organizations like PJM face massive interconnection backlogs. BNEF predicts that data center load in PJM alone could reach 31 GW by 2030, almost matching expected new generation capacity additions, leaving little margin. The North American Electric Reliability Corporation (NERC) has warned of increased risk of summer power shortages in multiple regions. Transmission upgrades and new line construction take years, and many projects face multi-year delays.
A further tightening of natural gas generator supply exacerbates the problem.The shortage of natural gas generators has further exacerbated the problem. Large gas turbines are essentially sold out through 2030. Bank of America analysts point out that data centers will increasingly rely on on-site gas engines/generators and behind-the-meter solutions. Currently, over 7.5 GW of data center projects are equipped with on-site power generation, with another 60 GW+ in the pre-construction phase. Co-location agreements such as Chevron and Microsoft's 2.67 GW natural gas power project in West Texas are accelerating.
Regulatory and political obstacles, especially in "blue states," add friction. Stringent environmental reviews, emission regulations, local zoning opposition, and occasional data center moratoriums have led to project delays or cancellations. Controversial projects and local restrictions have sharply increased. Growth is shifting toward more permissive jurisdictions (e.g., Texas/ERCOT, parts of the Midwest, the Southeast, and Ohio), where permitting is faster and existing industrial infrastructure is available.
Result: In the late 2020s, the growth rate of grid-connected data centers may partially slow or plateau, especially for large AI training clusters. Hyperscalers (Amazon, Google, Meta, Microsoft) are adapting by:
Developing behind-the-meter generation
Co-locating with existing or new power plants
Implementing demand response and workload shifting
Signing dedicated power purchase agreements (PPAs)
Overall capacity additions may continue, but without faster infrastructure solutions, timelines will extend and costs will rise.
The Urgency of Nuclear Power: Balancing Growth with Reliable Clean Electricity
Data centers—especially AI workloads—require highly reliable 24/7 baseload power. Intermittent renewables need significant storage or overbuilding to match this characteristic. Natural gas can fill the gap but faces emission scrutiny and the aforementioned turbine supply constraints. Nuclear power offers dispatchable, zero-carbon baseload electricity, perfectly aligning with hyperscalers' net-zero energy goals.
Tech giants have begun voting with their wallets:
Microsoft: Invested $16 billion in a 20-year PPA to restart Three Mile Island Unit 1 (835 MW, target 2027).
Meta: Reached agreements with Oklo, TerraPower (Natrium), Vistra, and Constellation to secure up to 6.6 GW of nuclear capacity by 2035.
Google: Committed to purchasing 500 MW from Kairos Power's fluoride salt-cooled reactor.
Amazon: Invested in X-energy's small modular reactors (SMRs) and signed co-location agreements.
Broader pipeline: Hyperscalers have committed approximately 9.8 GW of nuclear capacity through multiple agreements.Restarting existing units, increasing output, and emerging SMRs and microreactors are the fastest pathways. Policy support, streamlined NRC processes (e.g., DOE microreactor pilot projects), and co-location incentives are critical. Without accelerating nuclear deployment, the U.S. may face:
Slower data center growth (limiting AI/computing leadership)
Greater reliance on natural gas or coal plant retirements extended (increasing emissions and electricity prices)
Higher electricity rates passed on to consumers and businesses
Supply-Demand Balance: Demand vs. Available Power
Demand side: Surging and resilient. Improvements in AI models (e.g., Kimi K3's emphasis on memory) do not eliminate hardware demand but evolve it. Hyperscalers continue heavy capital expenditure.
Supply side: Lagging in the short term due to physical and regulatory bottlenecks, but responding through private sector innovation (on-site generation, co-location, nuclear PPAs). Utilities are significantly raising load forecasts (e.g., AEP expects 24 GW of new load by 2030, mainly from data centers).
By 2035, the high-end scenario of 18-20% is achievable if nuclear, natural gas, transmission, and renewables/storage scale simultaneously. Failure to accelerate reliable low-carbon resources like nuclear will likely force a lower growth trajectory or compromise reliability and climate goals.
Conclusion: Data centers consuming one-fifth of U.S. electricity by 2035 is possible under the current trajectory – but only if nuclear achieves meaningful scaling in the next decade. The tech sector is already acting; policymakers and regulators must match this urgency on permitting, transmission, and advanced nuclear deployment. Otherwise, grid constraints and supply chain limitations will self-impose a plateau – possibly at the expense of U.S. technology competitiveness.