For years, artificial intelligence has been discussed as a new source of electricity demand: enormous data centers need increasingly large amounts of power to train and operate AI systems. A less familiar development is now emerging on the other side of the equation. AI itself is beginning to be developed as a tool for deciding how America should build, expand and operate the electric grid.

The most striking example arrived on September 1, 2026, when the U.S. Department of Energy announced an $11.5 million Office of Electricity project called Foundation Models for the Electric Grid: From Proof of Concept to Real-world Impacts, or GridFM 2.0. Led by Brookhaven National Laboratory, the project has an unusually ambitious objective: enable utilities to evaluate 1 billion potential grid scenarios in 24 hours. DOE says the effort could increase planning throughput by more than 10,000 times and make certain grid calculations more than 1,000 times faster than conventional approaches.

The important story is not simply that AI may make power-system calculations faster. It is that the economics of grid expansion could begin to change if planners can afford to explore vastly more possible futures before committing billions of dollars to physical infrastructure.

The grid has become a problem of too many possible futures

Electric utilities have always had to forecast demand, generation, weather, equipment failures and transmission constraints. But the planning problem is becoming substantially more complicated because several changes are occurring simultaneously.

Large AI data centers are appearing alongside advanced manufacturing facilities, industrial electrification, battery storage, renewable generation and new transmission projects. Their locations matter because electricity cannot simply be moved across the country without regard to transmission capacity. Timing matters too: a proposed data center that arrives two years earlier or later can create a very different infrastructure requirement.

DOE's July 2026 draft National Transmission Needs Study explicitly identifies data centers, expanding domestic manufacturing, large industrial loads and electrification as drivers of increased transmission requirements. The study also found that much of the nation's transmission congestion is concentrated in only about 5% of hours, suggesting that a grid can experience severe constraints without being continuously overloaded.

That last point is particularly important for AI-based planning. If congestion occurs only under particular combinations of weather, generation, demand and network conditions, planners need to understand a large number of combinations rather than one representative forecast.

Traditional planning methods can already perform sophisticated simulations. The problem is computational scale. Every additional assumption creates another branch in the decision tree: What if a proposed data center is delayed? What if three facilities arrive instead of one? What if a transmission line is unavailable during a heat wave? What if battery deployment is faster than expected? What if renewable output is unusually low? What if a large customer agrees to curtail consumption during grid emergencies?

The number of combinations quickly becomes enormous.

GridFM is attempting to change the economics of that calculation

Foundation models are associated in the public imagination with language, images and other forms of generative AI. GridFM represents a different use of the underlying idea. Instead of learning patterns from text, a grid foundation model can learn representations of electrical networks, operating conditions, system configurations and other power-system data and then apply that learned representation to multiple downstream analytical tasks.

A 2024 research paper titled Foundation Models for the Electric Power Grid, led by Hendrik F. Hamann and a large international group of researchers, proposed the concept as a way of using diverse grid data and topologies to address the increasing complexity and uncertainty of modern power systems. The researchers described both the potential and the weaknesses of applying foundation-model techniques to power grids, including the need for models that can generalize across different grid configurations and tasks.

The 2026 DOE project is therefore better understood as a transition from an interesting research concept toward utility-scale experimentation. Brookhaven says the three-year project will develop a next-generation GridFM capable of rapidly simulating scenarios involving new electricity loads. The project involves universities, national laboratories, technology companies and utilities, including MIT, the University of Illinois Urbana-Champaign, Stony Brook University, Cornell, the University of Texas at Arlington, GE Vernova Research, IBM Research, NVIDIA, Microsoft, AWS and several U.S. utilities and grid organizations.

That breadth is significant because grid AI cannot be developed effectively from generic AI data alone. Utilities possess operational information and knowledge about actual network behavior, while universities and technology companies contribute machine-learning methods and computing infrastructure. Brookhaven says the broader GridFM initiative involves more than 150 organizations and is being developed with open-source support through Linux Foundation Energy.

Brookhaven principal investigator Hendrik Hamann offered a revealing description of the project's objective: “AI is often seen as a growing burden on the electric grid, but through the Genesis Mission we have an opportunity to turn that challenge into an advantage.”

That reversal is the unusual part of the story. The same technological boom that creates enormous new electricity loads could also provide the computational tools needed to plan the grid required to serve them.

Data centers make the planning problem unusually difficult

AI data centers are not simply large versions of conventional electricity customers. Their scale can be enormous, their demand can change rapidly, and their geographic concentration can create localized grid constraints.

Lawrence Berkeley National Laboratory researchers Jessica Granderson, Ian M. Hoffman, Billie Holecek, Eliot Crowe, Sarah Josephine Smith and Natalie Mims Frick wrote in a 2026 article that U.S. data-center energy use is projected to double or triple from 2023 levels by 2028. Their research also points to an important characteristic that receives less attention than the headline electricity numbers: some data-center electricity consumption can be flexible.

Computational workloads can potentially be shifted in time or, in some circumstances, geographically. Facilities can also use batteries, modify supporting infrastructure and operate onsite generation. LBNL identifies four broad mechanisms for flexibility: computational load flexibility, changes to facility infrastructure operation, energy storage and onsite electricity generation.

This introduces a new variable into grid planning. A data center does not necessarily have to be treated as a perfectly rigid block of demand.

Imagine two otherwise identical 1-gigawatt data centers. In the first case, the facility requires its full load continuously and cannot respond to grid conditions. In the second, a portion of its computational workload can be delayed or moved, batteries can supply some electricity during short periods of stress, and the operator agrees to temporarily reduce grid consumption when specified conditions occur.

The two facilities have the same headline electricity requirement, but they do not necessarily impose the same infrastructure requirement. An AI-based planning system capable of evaluating enormous numbers of operating combinations could potentially quantify that difference rather than relying on a simplified assumption about peak demand.

The hidden connection between AI planning and grid congestion

This is where GridFM intersects with one of the least visible economic problems in the current data-center boom: congestion.

A transmission line does not need to be overloaded every hour to impose a major economic constraint. A network can have enough capacity most of the time while becoming constrained during particular combinations of high demand, low generation availability, weather events or outages.

DOE's 2026 transmission study found that congestion is heavily concentrated in a small number of hours. It also identified potential value from both regional and interregional transmission projects.

That creates an interesting optimization problem. Building a new transmission line is expensive and takes years. But not building it can also be expensive if congestion prevents new generation or large customers from connecting, forces expensive redispatch or causes electricity prices to rise during constrained periods.

More computational capacity changes the question from “Which transmission project should we build?” to something closer to “Across millions of plausible futures, which combination of transmission, generation, storage, flexible demand and operating rules produces the lowest overall system cost while maintaining reliability?”

That is a much richer question.

The economic value of GridFM therefore may not come from replacing engineers. It could come from allowing engineers to examine a much larger decision space before making decisions.

The billion-scenario target should not be confused with billion reliable predictions

There is, however, an important distinction between computational speed and predictive accuracy.

DOE's stated target of 1 billion scenarios in 24 hours is a project objective, not evidence that a deployed utility system can already perform that task with operationally validated accuracy. Brookhaven describes the work as a three-year research effort and says two utility deployments will be used to demonstrate the technology in real-world settings.

This distinction matters because power systems are safety-critical physical systems. A fast model that produces an incorrect answer can be worse than a slower conventional calculation.

Researchers are already working on precisely this problem. A 2026 study by Antonio Alcántara and Spyros Chatzivasileiadis proposed a “trustworthiness layer” for foundation models applied to power-system contingency assessment. The research used statistically valid uncertainty bounds and reported that its enhanced GridFM approach could provide substantially faster analysis than AC power-flow calculations in tested systems while also supplying uncertainty information.

The significance is broader than the particular benchmark. For an electric utility, an AI model should ideally answer not only “What do I predict?” but also “How confident am I, and under what conditions might this prediction fail?”

That requirement separates grid foundation models from consumer-facing generative AI. A chatbot can produce an imperfect answer that a human checks. A power-system planning model must ultimately fit into engineering workflows governed by reliability requirements, validation procedures, regulatory oversight and established simulation methods.

The other problem is not computation it is knowing which future is real

Even a perfect scenario engine cannot solve a forecasting problem if the assumptions going into it are unreliable.

This issue has become increasingly visible as proposed data centers proliferate. In June 2026, FERC Commissioner David Rosner warned that speculative large-load interconnection requests can clog queues and distort forecasts. He noted that prospective data-center projects may approach multiple utilities to identify attractive connection opportunities, creating the possibility of duplicate or non-real projects appearing in planning assumptions.

This is a crucial limitation for AI planning. A model can evaluate one billion scenarios, but it cannot automatically determine whether the billion underlying assumptions correspond to projects that will actually be built.

In other words, the future of grid planning has two separate problems: scenario generation and scenario credibility.

More computing power addresses the first. Better data, project-readiness standards, improved load forecasting and regulatory coordination address the second.

FERC's 2026 actions also illustrate why infrastructure planning is becoming more closely connected to economics. The Commission has emphasized cost-recovery agreements intended to protect other customers if planned large-load infrastructure is built but the customer fails to materialize. It has also directed attention toward technologies such as dynamic line ratings that can increase utilization of existing transmission infrastructure.

That suggests an important future role for AI: not simply deciding where to build new wires, but testing whether a proposed new line is actually the least-cost answer compared with better use of existing infrastructure, flexible loads, storage, generation siting or different operating arrangements.

Grid planning could become a portfolio problem rather than a construction problem

Lawrence Berkeley National Laboratory's June 2026 report Speed to Power: Solutions for Accelerating Large Load Connections, by Fredrich Kahrl and Natalie Mims Frick, identifies more than 40 potential approaches to large-load connection bottlenecks. The researchers organize them into five areas: load forecasting, interconnection, resource planning and procurement, markets and operations, and cost allocation and ratemaking.

The breadth of that list helps reveal why AI-based planning could be more consequential than a simple faster power-flow calculation.

A utility could potentially use a foundation model to compare combinations of physical and contractual solutions. One scenario might involve a new transmission line. Another could combine a smaller transmission upgrade with batteries. A third could use a flexible interconnection agreement. Another might pair a data center with nearby generation. Yet another might combine several measures and postpone a major transmission investment until demand becomes more certain.

FERC's June 2026 actions explicitly encourage some of these combinations. Rosner noted that flexible transmission services can reduce the need for network upgrades and generating capacity when large customers can limit their withdrawals, while studying electrically proximate generation and load together can sometimes reduce the grid impacts of connecting them separately.

AI could make such combinations easier to evaluate at scale.

The emerging value may be avoiding the wrong infrastructure

The most interesting economic implication of GridFM is therefore not necessarily that America will build the grid faster. It may be that America can become better at deciding what not to build.

Large infrastructure projects are usually justified using forecasts extending years into the future. But rapid technological change makes those forecasts unusually uncertain. AI efficiency improvements could reduce the amount of computing required per task. Data-center projects may be delayed or cancelled. Manufacturing demand could accelerate. Battery costs and deployment could change. New generation technologies could alter regional power flows.

Under these circumstances, committing to a single forecast can create expensive errors in both directions. Underbuilding can create congestion and delay economic activity. Overbuilding can leave customers paying for infrastructure that is underutilized.

LBNL's August 2026 update on electricity-rate designs for large loads highlights precisely this financial tension. The authors Natalie Mims Frick, Peter Cappers, Goksin Kavlak, Long Lam, Ryan Hledik, Jadon Grove and Sebastian Rotella note that meeting large-load demand creates risks including insufficient energy supply and underutilized investments that can affect other customers.

More comprehensive scenario analysis could therefore have value even when it produces no spectacular new transmission project. If millions of simulations show that a proposed investment only makes sense under a narrow set of assumptions, regulators and utilities may have a stronger reason to delay, redesign or condition the investment.

AI could make flexibility economically visible

Another potentially positive consequence is that flexibility may become easier to value.

Today, a large electricity customer can look like a fixed quantity in a planning spreadsheet: for example, a projected 1,000-megawatt load arriving at a specified date. But the physical reality can be more nuanced. Some computing jobs are time-sensitive; others may have scheduling flexibility. Batteries can cover short periods. Backup generation can alter the facility's grid dependence. Cooling systems have operating constraints. Workloads can sometimes be distributed between locations.

LBNL's research argues that these characteristics can make AI data centers unusually capable of participating in demand flexibility programs. The researchers caution that flexibility cannot replace the long-term need for additional generation, but they identify it as an important near-term tool for enabling new computing capacity while the grid expands.

If GridFM-style models can incorporate these characteristics into system planning, flexibility could stop being treated as an abstract demand-response concept and become a measurable infrastructure alternative.

That could produce a subtle but important change in how the economics of data centers are negotiated. Instead of asking only how much electricity a facility needs, utilities and regulators could increasingly ask how much certainty, flexibility and infrastructure commitment the customer can provide in exchange for faster interconnection.

The positive feedback loop: AI creates demand, then helps manage it

There is an unusual circularity developing in the American technology economy.

AI is helping drive electricity demand through data centers. That demand is exposing weaknesses in transmission planning, interconnection procedures and infrastructure forecasting. Those weaknesses are creating incentives to develop AI systems capable of analyzing the grid more rapidly. And those systems could, if validated, help determine how to accommodate the next generation of AI infrastructure.

It is not a guarantee of success. GridFM remains a research and deployment project rather than an established utility standard. The underlying models need extensive validation, and their usefulness depends on high-quality operational data, appropriate uncertainty estimates and integration with existing engineering and regulatory processes.

But the direction is notable. The conventional narrative treats AI and the electric grid as opponents: data centers consume electricity while the grid struggles to keep up. The emerging research suggests a more complicated relationship. AI can be both a new source of electrical demand and a new computational instrument for managing the consequences of that demand.

The practical breakthrough, if GridFM succeeds, may not be a futuristic autonomous grid. It could be something more mundane and potentially more valuable: giving engineers, utilities and regulators the ability to explore far more alternatives before billions of dollars are committed to steel, concrete, wires, transformers and generation.

That would represent a quiet shift in infrastructure economics. America's grid would still be built physically, one transmission line, substation and generator at a time. But the process deciding which infrastructure gets built could become dramatically more computational, probabilistic and scenario-driven.

In that sense, the most interesting AI application to the American power sector may not be another technology that consumes electricity. It may be an AI system that helps the people responsible for the grid decide how to spend the enormous amount of capital required to supply it.