In a relatively short time, large loads have become a focus for utilities, policymakers, regulators, and retail customers. For example, according to the Smart Electric Power Alliance, state regulators approved 29 large-load tariffs in 2025, compared to just 14 between 2018 and 2024. In the summer of 2026, the Federal Energy Regulatory Commission (FERC) is expected to take action on the U.S. Department of Energy’s 2025 proposals to reform large load interconnection to the transmission grid.
The main reason large loads are grabbing so much attention these days is simple: there are a lot of them, and the speed and nature of their connection to the grid matter to utilities, retail customers, large customers that need electricity quickly, and the power system as a whole.
After years of flat electricity demand, large customers, such as data centers, are a significant driver of rapid load growth. Between 2025 and 2030, annual electricity demand is expected to grow at 2%-3.6% per year, compared with an average annual rate of 0.8% between 2019 and 2024. EPRI estimates that data center electricity use in the U.S. could account for between 9% and 17% of total electricity consumption by 2030, up from about 4% to 5% today.
The Interconnection Bottleneck
Another reason large loads are attracting so much attention is that many face long wait times before they can interconnect to the power grid. According to the Lawrence Berkeley National Laboratory (LBNL), projects that began commercial operation in 2024 spent an average of 4.5 years in interconnection queues, up from just 22 months in 2008.
A powerful tool to address the challenges of interconnecting large loads is a concept that sounds simple but is surprisingly complex in practice: flexibility. At a fundamental level, flexibility describes a load’s ability to adjust its electricity consumption in response to grid conditions. On a hot summer day, for example, when demand for electricity from air conditioning is high, a large, flexible load can temporarily ramp down its consumption, providing relief to a strained grid.
The details and definitions of flexibility, however, quickly become complicated, with different meanings across regions, markets, and industries. A utility may define flexibility very differently from a regulator, grid operator, or large load operator. The lack of a shared language and a definition of flexibility is not an abstract semantic problem. It is a meaningful barrier to securing faster, large-load interconnection times and grid support. It’s also a solution that is very much within reach with the right collaboration and cooperation.
“The technical solutions for flexibility exist. This is not a technology problem,” said Irene Danti Lopez, an EPRI senior team lead. “In our discussions, we found ourselves getting into a circular loop. The utility would come in and say it would value flexibility, noting that flexibility could unlock opportunities for faster interconnection. The utility would then ask the large load what flexibility it could provide, and the large load would respond that it was interested, but would first need to understand what the utility was looking for. As a result, the conversation would loop without converging on specifics.”
A Common Language for Flexibility
Establishing the shared language needed for flexibility is the aim of EPRI’s Flex MOSAIC™ project. Officially launched at CERAWeek in March 2026, Flex MOSAIC™ is a framework for uniformly classifying large load flexibility. The voluntary framework was developed through EPRI’s DCFlex initiative and resulted from collaboration among more than 70 utilities, system operators, regulators, hyperscalers, and technology providers.
Involving a broad set of collaborators in developing the framework was a necessary foundation for Flex MOSAIC™. “This is all about finding win-win solutions for utilities, grid operators, as well as data centers and other large loads so that we can accelerate speed to power, but also keep supporting power system reliability and operability,” Lopez said. “We don’t want one at the expense of the other, and we’re really trying to provide solutions to deliver both.”
The Flex MOSAIC™ framework defines flexibility based on a set of specific performance characteristics. These include how quickly a load can respond to a signal to ramp down consumption, how long that reduction can last, and how much power can be reduced or shifted. The performance characteristics are organized into five specific classes that utilities, large loads, and system operators can use and adapt to suit their local markets and conditions. The five classes are designed so that a utility or grid operator can look at a large load’s classification and immediately understand what to expect from it, just as they would assess any other resource on the system.
A Framework to Benefit Everyone
The definitions outlined in Flex MOSAIC™ provide critical certainty for large loads seeking interconnection and utilities and grid operators charged with operating a reliable grid. To understand how this shared understanding of flexibility benefits utilities and grid operators, consider the typical process they follow upon receiving an interconnection request.
Their first step in evaluating the request is to run a series of studies to assess the impact of the new load on the grid. Among other things, those studies assess whether existing transmission lines and other infrastructure can accommodate the new load without exceeding thermal or voltage limits. Upstream, studies are also conducted to assess whether sufficient generation capacity exists to meet demand during periods of grid stress.
Traditional planning approaches develop plans based on worst-case scenarios, where a large load runs at full power regardless of grid conditions. It’s an understandable approach for ensuring reliability. But it is also expensive and time-consuming because meeting peak demand often requires large grid investments that take years to complete.
A standardized flexibility classification is a necessary foundation. But it is only part of what utilities need to plan with confidence. Flex MOSAIC™ provides that necessary foundation by establishing clear, consistent definitions of what different classes of flexibility look like in practice. Equally important are the processes that verify that a large load can deliver what its classification promises, along with the controls grid operators need to activate flexibility when needed. Taken together, the framework and supporting processes allow utilities to treat a flexible, large load the same way they treat other system resources. For example, a utility reviewing an interconnection request from a Class C large load would have clarity about how to model it, what a realistic response could be, and how much weight to give to flexibility in its studies. This depends on a load’s performance being measured and validated against the framework’s definitions.
The benefits of a standardized framework for flexibility extend to the large loads seeking interconnection. Large loads that can demonstrate a credible flexibility classification give planners the predictability they need to accelerate interconnection. Rather than being treated as an inflexible contributor to rapidly increasing demand, large loads can also be assessed for what they offer to help the grid when support is needed. In some cases, a load’s flexibility classification may allow it to interconnect sooner, at a location that would otherwise be unavailable, or under terms that reduce the infrastructure investment required to serve it.
Standardized definitions of flexibility also allow owners of large loads to invest in enhancements that increase their flexibility. This information is needed by large load developers as early as possible. For example, a large load’s eligibility for a higher flexibility class may depend on its ability to modulate demand, as well as on whether it has on-site battery storage, backup generation, or other infrastructure built in during construction. Trying to add that capability later can be prohibitively expensive or, in some cases, impossible.
“If you need an energy storage unit on site to provide flexibility, a data center or factory builder needs to know that when they’re building,” Lopez said. “It will impact land requirements. It will impact permitting. If you try to add it later, it might be more expensive or simply not possible because you haven’t bought enough land and there are neighbors.”
Putting Theory into Practice
As Flex MOSAIC™ was being developed, EPRI and a group of partners conducted real-world tests to validate its core premise. In December 2025, EPRI, the utility National Grid, AI cloud provider Nebius, and the compute and energy management firm Emerald AI conducted a series of tests in the United Kingdom to evaluate whether a cluster of graphics processing units (GPUs) training AI within a data center could actually deliver the kind of flexibility the framework describes without compromising the computing work it was built to do.
To do that, the cluster was taken out of production and tested. The results: the cluster achieved 100% compliance with more than 200 requested power targets and ramp-rate constraints. Load reductions of up to 40% were delivered in under a minute, while service agreements for high-priority computing workloads were maintained even as power demand ramped down.
Several of the tests were designed to mimic unpredictable grid conditions, including unannounced, back-to-back grid events. This helped demonstrate that certain flexible, large loads can respond to repeated, operationally varied grid events, not just serve as a one-time demand-response asset. Besides demonstrating that a large load could provide flexibility to mitigate grid stress, the tests also showed that the cluster could respond to carbon-intensity signals, reducing emissions by more than 10% during the test period.
“What those results showed us is that high AI performance and grid responsiveness are not in conflict,” Lopez said. “A data center and other large loads can be real grid resources, and that changes what’s possible.
The tests provided a source of real-world operating evidence for the development of Flex MOSAIC™ and gave the framework’s performance definitions a foundation in real operating experience. A second set of six additional demonstration projects in the United States and Europe was announced in February 2026. Late in 2026, EPRI plans to conduct flexibility testing at the Aurora Project, a U.S.-based facility targeting nearly 100 megawatts of flexible AI load. That test will assess whether the flexibility demonstrated can be replicated at scale.
By October 2026, EPRI expects to publish a final version of the framework, along with practical guidance and tools to help utilities and large loads implement it. “The goal from the beginning has been to give everyone involved a common ground to work from,” Lopez said. “When utilities and large loads are speaking the same language, decisions get faster, interconnection gets easier, and the grid gets stronger. That’s a win for everyone, including the customers who depend on it every day.”
EPRI Technical Expert:
Irene Danti-Lopez
For more information, contact techexpert@eprijournal.com.
