Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects

Overview

This 2025 arXiv preprint (arXiv:2509.07218) by Xin Chen, Xiaoyang Wang, Ana Colacelli, Matt Lee and Le Xie is a broad review of AI data-center infrastructure, load characteristics and grid impacts. The study first describes the components (IT, power delivery, cooling, auxiliaries) and the load patterns of model preparation, training, fine-tuning and inference.

It then analyses challenges on three timescales: long-term planning and interconnection, short-term operation and markets, and real-time dynamics and stability. Proposed solutions span grid-side measures (forecasting, dynamic modelling, ride-through requirements, demand-response design), data-center-side measures (storage, efficient cooling and hardware, grid-aware scheduling) and end-user measures. The paper also summarises regional demand projections for the U.S., China and Europe.

The figures are striking: global data centers consumed around 415 TWh in 2024 (about 1.5% of global demand), projected to reach around 945 TWh by 2030, with AI identified as the primary driver. Traditional racks run at 7-10 kW, while AI racks range from 30 kW to over 100 kW, averaging more than 60 kW in dedicated AI facilities. Citing EPRI, IT accounts for 40-50% of total load, cooling 30-40% and other facilities 10-30%; in AI data centers IT typically exceeds 60%.

Inference accounts for approximately 60% of total AI energy use at Google versus 40% for training, and may reach up to 90% of a model's lifecycle energy. An ERCOT study shows that a loss of 2,700 MW of large electronic loads can drive system frequency up to 60.4 Hz, a level at which conventional generators may fail to ride through. Proposed specifications for data-center UPS ride-through suggest maintaining operation down to 50-70% of nominal voltage with resynchronisation within one second.

Key points

1. Global data centers consumed around 415 TWh in 2024 (about 1.5% of global demand), projected to reach around 945 TWh by 2030, with AI the primary driver.
2. Rack power density: traditional 7-10 kW/rack versus AI racks from 30 kW to over 100 kW/rack, averaging more than 60 kW in dedicated AI facilities.
3. Cited EPRI shares: IT 40-50% of total load, cooling 30-40%, other facilities 10-30%; in AI data centers IT typically exceeds 60%.
4. Inference accounts for approximately 60% of total AI energy use at Google versus 40% for training, and may reach up to 90% of a model's lifecycle energy.
5. An ERCOT study shows a loss of 2,700 MW of large electronic loads can drive frequency up to 60.4 Hz, where conventional generators may fail to ride through.
6. Proposed UPS ride-through specification: maintain operation down to 50-70% of nominal voltage with resynchronisation within one second.
7. Challenges are treated on three timescales: long-term planning, short-term operation and real-time dynamics.

Using it in design

• For AI-campus grid studies, use the default rack-density bands (7-10 kW traditional, 30-100+ kW AI).
• Use the IT/cooling/other energy splits and the training versus inference split in energy forecasts.
• Add a ride-through specification field for UPS (minimum voltage %, resynchronisation time).
• Warn when total campus electronic load on one substation is large enough to matter for frequency (multi-GW load-loss scenarios).

Who should read it

Engineers and planners working on grid interconnection, demand forecasting and power system design for AI data centre campuses.

Related standards

• IEEE 1547

Publication details

Authors

Xin Chen; Xiaoyang Wang; Ana Colacelli; Matt Lee; Le Xie

Publisher

arXiv preprint (Harvard University et al.)

Year

2025

Document number

arXiv:2509.07218

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2509.07218

Licence

CC BY-NC-ND 4.0 — share unaltered with attribution for noncommercial purposes

Citation