North American Electric Reliability Corp., a nonprofit security watchdog for the power grid, is moving to draft new standards for large-scale artificial intelligence computing hubs that could lead to regulation of their operations. The NERC committee will begin the project on Wednesday and seek final approval by the end of the year.
A recent NERC white paper states that extreme power fluctuations, where power demand can fluctuate by hundreds of megawatts in an instant while training an AI model, are a “high probability and high impact” risk. This is a threat that could upset the carefully regulated power grid and cause uncontrollable cascading power outages.
The report cited a “significant reliability gap” posed by large data centers and called it “essential” to address this and other risks.
Steps to regulate the growing size and number of computing hubs represent a new and important challenge for the AI industry. Google, Amazon, Meta, Microsoft, Anthropic and others are developing gigawatt data center complexes. This requires as much electricity as a medium-sized or large city in America.
“The emergence of gigawatt-scale loads can dramatically increase the complexity of research, planning, and operations,” the paper states.
Due to rapid growth in AI, total peak power requirements for AI could reach 50 gigawatts by 2030, according to an analysis by the Electric Power Research Institute. Concerns that existing electricity demand and forecasts are driving up utility costs for households and other businesses are shaping the conversation around building AI. The risks to the U.S. power grid, already under stress from extreme weather events, are emerging for a wide range of regulators and standard-setting bodies responsible for ensuring the lights stay on.
The Edison Electric Association, which represents investor-owned electric utilities, asked NERC to require data centers to meet the types of engineering standards and protocols that govern generators and other major operators on the power grid.
“They should also be subject to similar regulatory requirements,” EEI said in a statement.
“As very large loads like data centers grow, planners and grid operators need a clear understanding of how those loads connect to and operate on the system,” said Todd Snitchler, CEO of the Electric Power Supply Association.
it happened in virginia
NERC highlighted a 2024 incident in Virginia where a power outage caused a sudden voltage spike. Controls at dozens of data centers were automatically disconnected from the grid and switched to backup power to protect sensitive computer chips and their cooling systems.
The grid operator suddenly faced an unexpected loss of approximately 1,500 megawatts of power demand.
The larger grid was able to continue operating. However, NERC said a similar incident involving a large cluster of very large data centers would create an unacceptable risk of cascading shutdowns and widespread outages.
The group issued a warning about these issues last September.
Some of the concerns identified by NERC are well known, including a lack of information from data centers and power companies about how computing hubs are designed and operated. This limits grid operators’ understanding of potential risks. This problem is further complicated by the fact that the speed of data center construction often outstrips the necessary power grid upgrades and preparation for the center’s power requirements.
NERC also cited a “lack of cooperation and coordination” between data center operators, power companies and grid operators. For example, according to the NERC Working Group, some data centers are not sharing adequate information about transitioning to backup power, which could impact grid operations.
NERC is moving relatively quickly to create a new standard by the end of the year. It will then require approval from the Federal Energy Regulatory Commission.
The accelerated schedule “underscores the urgency with which federal regulators and the broader industry are addressing the reliability risks posed by significant increases in load,” Jiecheng “Jeff” Zhao, chief engineer at Elevate Energy Consulting, said in a recent blog reporting on last month’s NERC technical conference. “Mandatory requirements are likely to be imposed” on data centers, Zhao wrote.
NERC says the “significant” risk is that the power of tens of thousands of computer chips will fluctuate rapidly as AI models are trained to understand and process vast amounts of information. When one training phase ends and the results are compiled, the power demand in a large data center can instantly plummet by as much as 300 megawatts, and then rise just as suddenly when training resumes.
NERC noted that data center operators typically do not tell grid operators that such sudden changes are coming or share the technical data that grid control room engineers need to protect their networks.
When extreme power fluctuations spread across the power grid, protective automatic controls can be triggered and power equipment and transmission lines can be shut down. The result is a cascade of equipment outages that “could lead to uncontrolled separations and cascading outages,” the NERC report said.
When data center utilization declines, it appears to the power grid that it no longer has a large power customer, said Julia Matevosyan, associate director and chief engineer at the Energy Systems Integration Group (ESIG), a nonprofit research group of grid experts. ESIG recently published a report documenting data center risks to the power grid that are consistent with the FERC Task Force’s analysis.
“If you only have one data center, that’s fine. But if you have multiple data centers in close proximity and they all get attacked, which means they all ‘disappear,’ the initial loss is much worse,” Matevosyan said in an interview.
ramp up, ramp down
In addition, disturbances can cause vibrations that can spread to nearby power plants on the power grid, damaging and, in extreme cases, destroying the drive shafts of electric turbines near data centers, Matevosyan said. When such fluctuations occur, “we don’t really know what the limits are,” she says.
The danger of sudden grid fluctuations was highlighted in an October blog by two senior managers at NVIDIA, a leading AI chip maker. NVIDIA’s Neeraj Srivastava and Harry Petty said the risks were documented in a joint study signed by 50 scientists from NVIDIA, Microsoft, and OpenAI.
“These erratic fluctuations, increasing and decreasing by hundreds of megawatts in seconds, pose a serious threat to the stability of the power grid, making power grid interconnections the primary bottleneck for AI scaling,” they wrote.
“We are at a critical inflection point. The industry can no longer rely on incremental improvements and requires fundamental architectural changes that can manage the power demands of modern AI,” they added.
The Data Center Coalition, which represents the largest AI technology companies and data center developers, declined to say whether it supports or opposes mandatory reliability standards.
Aaron Tinjam, the coalition’s vice president for energy, said in a statement that the coalition is “reviewing the recommendations from NERC’s Heavy Load Working Group and will continue to work with NERC, regional organizations, grid operators, and utilities across the country to provide education about the data center industry and ensure solutions that support both grid reliability and 21st century economic growth.”
Tinjam added that recent grid stability incidents related to data center operations in Virginia and Texas have led to increased collaboration and sharing of technical information between utilities and data center operators, helping to prevent new incidents.
“It is important to ensure that the requirements proposed by NERC are based on complete evidence and the technical capabilities and specifications of data center equipment,” Tinjum said.
Dion Harris, NVIDIA’s senior director of high performance computing and AI infrastructure, told E&E News last month that his company worked with Microsoft OpenAI to find a solution to the power fluctuation problem.
New features in NVIDIA’s cutting-edge Blackwell chip system will help mitigate sudden fluctuations in power demand caused by AI routines, reducing risks to grid stability, he said.
“Various solutions are being considered,” Harris said. “The important point is that we have enough information to confirm that there are no stability issues.”
NERC’s policy document states that voluntary cooperation on grid security is an option, but “directly enforceable NERC standards may be needed instead.”
