Tech
The Grid Writes Rules for AI
Two announcements landed within a week of each other, and they describe the same problem from opposite ends. One sets out how the largest computing sites must behave once they are plugged into the grid. The other asks the people who build and run the energy system how much electricity those sites will end up needing. Read together, they show an industry writing rules for loads whose size is still being argued over.
A Rulebook for the Biggest Loads
On October 15, NERC announced that its first package of standards for large computational loads passed the final stakeholder ballot (source: Utility Dive). The standards apply to sites with at least 50 megawatts of total connected load served through equipment at 100 kilovolts or above, and several buildings can count as one site. NERC's earlier proposal set the line lower, at 20 megawatts and 60 kilovolts (source: POWER Magazine).
The package has three parts. The first requires site owners to provide modeling data on how their equipment trips, rides through disturbances and recovers. The second requires operators to follow operating instructions from their transmission operator, balancing authority or reliability coordinator, with exceptions where an action cannot be physically carried out or would breach safety, equipment or legal requirements. The third requires owners to coordinate protection settings with neighboring entities and to record breaker sequences, fault recordings and disturbance data (source: POWER Magazine). NERC's board also approved two new registration categories on October 8, Computational Load Owner and Computational Load Operator (source: POWER Magazine).
The Incidents Behind Them
The standards follow a pattern of data center loads leaving the grid all at once. In July 2024, successive faults in the Eastern Interconnection caused the near simultaneous loss of roughly 1,500 megawatts of data center load because of customer side protection and controls. Frequency rose to 60.053 hertz and returned to 60.0 within four minutes, and NERC concluded that the event showed how large scale load tripping can meaningfully affect grid frequency (source: POWER Magazine). A separate January 2026 review looked at 26 cryptocurrency mining events in ERCOT between January 2023 and September 2025. Each involved a drop of more than 100 megawatts, and facilities lost between 17 and 95 percent of the power they had been drawing (source: POWER Magazine). A Level 3 alert in May listed seven actions grid entities should take while permanent rules were developed, but those actions were not mandatory standards (source: Utility Dive and POWER Magazine).
The package leaves a notable gap. It sets no minimum voltage and frequency ride-through requirements, which are being handled through separate standard authorization requests, and NERC owes a work plan for additional computational load standards by March 1, 2027 (source: POWER Magazine). NERC also states that the new obligations do not convert firm service into non-firm service (source: POWER Magazine).
A Forecast That May Still Be Too Low
How much load these rules will end up covering is the question the second announcement addresses. A survey of 1,832 energy decision-makers in 78 countries, commissioned by the Adipec conference and run by Kekst CNC between August and October, found that 81 percent expect AI driven electricity consumption to exceed existing projections. Thirty two percent expect it to be significantly higher, and 3 percent expect it to fall short (source: The National). The survey reflects industry expectations and is not an independently verified demand forecast (source: The National).
The projections respondents were measuring against are already large. The IEA expects data centers to consume about 950 terawatt hours of electricity by 2030, roughly double the 2025 level (source: The National). NERC's own forecast, published in January, has summer peak demand growing by 224 gigawatts over ten years, a 24 percent increase from 2025 (source: Utility Dive). About a quarter of respondents believe the industry is unprepared for AI driven demand, and grid build and connection delays rank among the biggest obstacles they named (source: The National). Forty three percent expect technology and AI companies to gain the most influence over the energy sector, compared with 15 percent for governments and regulators (source: The National).
Where the Two Meet
The rulebook and the survey meet at the same site. A 50 megawatt threshold means the standards will reach a growing share of new campuses if demand lands above forecast, and each of those campuses brings the same question of what happens to the grid when a large load trips. The standards address that question at the level of one site. The survey describes the number of sites still to come. We have uploaded the location data for nearly 1,500 operating US data centers to Panthaion, which gives a baseline view of where operating capacity sits today. You can explore the dataset directly alongside the new standards to see how much of the country's computing load already sits on the grid.
What This Means If You Pay an Electricity Bill
The standards are about reliability. Their purpose is to keep a sudden loss of data center load from disturbing the frequency that every customer on the grid depends on, so the benefit to a household is fewer disruptions when a large load trips. They do not set out who pays for new transmission or generation, and that question is being settled elsewhere, in utility rate cases and state proceedings. The survey result matters for that reason. If demand outruns projections, grid investment has to outrun them as well, and utilities generally recover that investment through rates. A forecast that lands too low is one a household can end up paying for later.
What to Watch Next
NERC expects its Board of Trustees to consider the standards in the fourth quarter of 2026, after which they go to FERC. A July FERC order requires NERC to file the standards and registration criteria by December 31, 2026 (source: POWER Magazine). The ride-through standards and the March 2027 work plan will show whether the gap in this first package gets closed, and the next round of demand forecasts will show whether the survey respondents were right.
Want to contribute your thoughts or findings in data and AI with climate, energy, or nature? You can find every past edition of Data in Energy at panthaion.org. Send us a DM or email to vmarsland@projectcoeus.com.
Sources: