Report Overview
"AI's Power Requirements Under Exponential Growth" is a research report published by the RAND Corporation. The report provides a comprehensive analysis of how artificial intelligence's exponential growth in computational resources is driving unprecedented power demands for data centers, and assesses the potential impact on U.S. competitiveness if these demands cannot be met.
Key Insight: AI data centers could need ten gigawatts (GW) of additional power capacity in 2025 alone, which is more than the total power capacity of the state of Utah. If exponential growth in chip supply continues, AI data centers will need 68 GW in total by 2027—almost a doubling of global data center power requirements from 2022.
Key Data Points
Key Insights Summary
Exponential Growth in AI Compute
AI's demand for computational resources has grown exponentially, driven by increasing training requirements and a rapidly expanding user base. Both developing and deploying frontier models take tens of thousands—and soon hundreds of thousands—of AI chips.
Unprecedented Power Demands
AI data center power demand grew tenfold over the last three years—from 0.4 gigawatts (GW) in 2020 to 4.3 GW in 2023. In 2025, total AI data center demand will likely reach about 21 GW of total power capacity.
Training Presents Unique Challenges
AI training requires large amounts of power capacity available at a single location. The compute used to train the most advanced models is rapidly increasing by about 4 to 5 times every year.
U.S. Infrastructure Challenges
The United States currently leads the world in data centers and AI compute, but unprecedented demand leaves the industry struggling to find the power capacity needed for rapidly building new data centers.
Geopolitical Implications
Failure to address current bottlenecks may compel U.S. companies to relocate AI infrastructure abroad, potentially compromising the U.S. competitive advantage in compute and AI and increasing the risk of intellectual property theft.
Renewable Energy Limitations
Making renewable energy sources suitable for AI data centers remains a challenge due to daily and seasonal variations. Current data center design requires power being available more than 99 percent of the time.
Content Overview
Document Contents
About This Report
An exponential increase in computational resources (compute) used for artificial intelligence (AI) training and deployment has recently enabled rapid advances in AI models' capabilities and their widespread use. The resulting unprecedented demand for AI data centers is already posing challenges for U.S. data center construction, primarily because it is difficult to find adequate power grid capacity.
This report provides two extrapolations of recent AI trends to assess future power needs, summarizes current bottlenecks for rapid data center construction, and discusses what a failure to resolve them could mean for U.S. competitiveness.
Summary
Larger training runs and widespread deployment of future artificial intelligence (AI) systems may demand a rapid scale-up of computational resources (compute) that require unprecedented amounts of power. We find that globally, AI data centers could need ten gigawatts (GW) of additional power capacity in 2025 alone, which is more than the total power capacity of the state of Utah.
If exponential growth in chip supply continues, AI data centers will need 68 GW in total by 2027—almost a doubling of global data center power requirements from 2022 and close to California's 2022 total power capacity of 86 GW.
Given recent training compute growth, data centers hosting large training runs pose a particular challenge. Training could demand up to 1 GW in a single location by 2028 and 8 GW—equivalent to eight nuclear reactors—by 2030, assuming that current training compute scaling trends persist.
Projecting Power Requirements for AI Data Centers
AI's demand for computational resources has grown exponentially, driven by increasing training requirements and a rapidly expanding user base. Both developing and deploying frontier models take tens of thousands—and soon hundreds of thousands—of AI chips.
For instance, xAI's Colossus supercomputer in Memphis, Tennessee, contains 100,000 AI chips and requires 150 megawatts (MW) of power—the generation capacity of around 55 modern wind turbines and the equivalent of about 53,000 U.S. households.
Total Power Requirements for AI Infrastructure
Continued exponential growth in demand would far outpace previous data center expansion. AI data center power demand grew tenfold over the last three years—from 0.4 gigawatts (GW) in 2020 to 4.3 GW in 2023.
In 2025, total AI data center demand will likely reach about 21 GW of total power capacity, more than a fourfold increase from 2023 and twice the total power capacity of the state of Utah.
Power Requirements for AI Training
AI training presents a unique challenge because it requires large amounts of power capacity available at a single location. The compute used to train the most advanced models is rapidly increasing by about 4 to 5 times every year.
Meanwhile, the energy efficiency of AI chips has grown much more slowly (only about 1.3 times per year), and data center Power Usage Effectiveness (PUE) has only moderately improved recently.
Challenges for Rapid AI Data Center Construction
To assess current challenges for rapidly expanding power generation and data center construction, we summarize common causes for delays cited in recent reports and analyses:
Insufficient Power Generation
A lack of power generation addition is increasing wait times for grid connections. For instance, in Virginia, the state with the largest share of data centers, connection requests now take between four and seven years.
Inadequate Transmission Infrastructure
Even when power is available, regions often lack transmission lines to deliver the power to sites suitable for data center construction. Projects to expand transmission infrastructure are difficult to coordinate and can take years.
Data Center Permitting Issues
Large data center projects face a variety of permitting challenges on the local, state, and federal level, limiting suitable sites and sometimes causing project delays and cancelations.
Supply Chain Delays
Data centers need a wide variety of inputs. Some, such as emergency power generators, now have waiting times of more than one year. Supply chain issues greatly delay data center construction.
Environmental Commitments
Data centers are subject to government regulations that limit their use of certain forms of energy. Given these constraints, compute providers' ability to procure sufficient power can trade off against environmental considerations.
Geopolitical Implications
Inadequate power may reduce the U.S. lead in AI compute. The United States leads the world in number of data centers and market share of compute providers. Although no direct estimates exist, this indicates that the United States likely hosts a significant majority of all AI chips.
However, the current challenges in power availability for rapid data center construction could reduce this advantage. An increasing number of U.S. companies are considering expanding their AI infrastructure to countries that offer more power availability, faster permitting, and additional financial incentives.
Compute is a primary enabler of AI progress; leading in compute enables leading in AI. Recent breakthroughs in AI have largely relied on a massive increase in compute used for training AI models. In other words, the success of a country's AI industry relies on access to specialized compute and the infrastructure needed to host it.
Potential Solutions and Future Research
The report identifies several areas for future research and potential solutions to address the challenges:
Quantifying Data Center Power Demand and Supply
- Model future increases in power grid supply and compare them with data center demand
- Assess factors that may reduce future power demand of AI data centers
- Continue studying potential bottlenecks to scaling of training and inference compute
Causes and Solutions for Data Center Construction Challenges
- Identify which state or federal regulations may limit the expansion of U.S. energy capacity
- Investigate opportunities for expanding power generation and distribution for data centers
- Develop and evaluate options for state and federal responses to energy shortfalls
- Assess the ability of private companies to meet energy shortfalls
Appendix and Methodology
The report includes detailed appendices covering the approach, methods, and sources used in the analysis, as well as limitations of the estimates.
Key methodological approaches include:
- Extrapolation of AI data center power needs based on continuing exponential growth in chip supply
- Modeling of power requirements for future AI compute clusters based on trends in training compute growth and efficiency improvements
- Analysis of current challenges for data center construction based on industry reports and publications
The report acknowledges several limitations, including the assumption that current exponential trends will continue, potential efficiency improvements that could reduce power requirements, and the possibility of decentralized training reducing the burden on single locations.
Note: The above is only a summary of the report content. The complete document contains extensive data, charts, and detailed analysis. We recommend downloading the full PDF for in-depth reading.