Introduction
The United States power system is experiencing a demand shock driven by Artificial Intelligence and the data centers AI relies on. Data centers are already accounting for 31 GW during periods of peak demand, a figure that is expected to double to 66 GW by 2027. Additionally, there are projections of data centers consuming nearly a fifth of all U.S. electricity generation by 2030. These data centers effectively behave as a continuous industrial-level baseload as opposed to fluctuating like domestic consumption.
Virtual Power Plants (VPPs) are not designed to generate a sustained energy supply but instead are used to either increase supply or decrease demand during periods of peak demand and grid stress. In this area, they’ve acted as one of the most cost-effective tools and a significantly cheaper alternative to “peaker plants” to manage peak electricity demand.
Policymakers and private sector leaders alike have recognized this potential. The U.S. Department of Energy has set a target of 80–160 GW of VPP capacity by 2030, supported by federal incentives and growing utility and private-sector adoption. However, it remains uncertain whether VPPs, which have thrived as tools for short-duration peak management, can scale to support the continuous, high-load electricity requirements of AI data centers. This report evaluates to what extent VPPs can serve as a partial solution to that electricity demand growth.
What Are Virtual Power Plants?
Virtual power plants are a collection of often thousands of decentralized energy sources, including residential batteries and home solar panels, and controllable electrical loads such as electric vehicles, smart thermostats, and electric water heaters, that are linked via software that allows utilities to collectively dispatch them similarly to how they would a single traditional power plant. The software monitors both local energy supply and demand and, during peak demand periods, can either draw power from batteries and solar panels or lower demand by, for example, adjusting smart thermostats or time-shifting EV charging to prevent brownouts and blackouts. Participants of VPPs are then compensated by either cash payments or lower electricity bills, creating an incentive structure that gives utilities access to flexible supply without devoting large amounts of capital to build additional power plants.
The two-pronged allure of virtual power plants is that they act as a cheaper alternative to the aforementioned “peaker plants,” which are fossil-fuel generators that utilities only activate during periods of peak demand, and are a cleaner alternative. In many cases, VPP capacity is typically 40% to 60% cheaper than that of natural gas peaker plants and utility-scale batteries and produces fewer emissions.
However, VPPs continue to rely on voluntary participation and customer behavior, which is often difficult to predict. Additionally, VPPs heavy reliance on solar power and battery storage leads to it sharing those reliability concerns, such as performance varying due to weather conditions and battery resources having finite duration.
Policy and Real-World Application
Policy support has also risen in the United States. The DOE’s national target represents a goal of tripling the current capacity of 30–60 GW, a figure that varies based on how demand response is counted. This increased capacity is projected to be enough to serve 10–20% of peak electricity demand and save an estimated $10 billion annually as they replace the need for those fossil-fuel-based peaker plants. Furthermore, federal policy tools, including incentives from the Inflation Reduction Act and financing through the Department of Energy’s Loan Programs Office, have further encouraged the adoption of EVs and smart appliances and therefore the expansion of the asset base that VPP’s need to operate.
In California, VPP-style demand response programs have been repeatedly deployed during heat waves to avoid rolling outages, while pilot programs in Texas have shown that residential batteries can provide much-needed short-duration grid support during peak summer conditions. However, these examples also reinforce that VPPs are most effective as peak-shaving and contingency resources rather than continuous energy providers.
The Scale Problem
As previously mentioned, AI’s strain on the grid is due to its high rates of both growth and load factor. AI workloads already account for 15-20% of data center electricity consumption and are the fastest-growing source of energy demand, accounting for around 50% of all new U.S. electricity demand growth in 2025. This growth is concentrated in large, continuously operating facilities, which is not the load type VPPs typically address.
AI’s energy demands therefore face a mismatch in both scale and duration compared to VPP resources. For example, a highly visible utility-scale VPP deployment in the United States is Google’s 100 MW demand response partnership in the PJM region, developed via a three-year agreement with Voltus, recognized as the largest current US VPP operator with 7 GW of capacity. Combined with the DOE and congressional initiatives, this seems to show a promising overlap of public and private interest in VPP expansion as a way to aid in grid flexibility. However, a single large AI data center campus can require a continuous draw of 100-300 MW, showing that one campus alone can require up to three times the capacity Google’s VPP project can supply. Furthermore, VPP’s shorter bursts, while ideal for flexibility, mean a traditional 100 MW VPP cannot provide the energy supply equivalent to the continuous demand of a 100 MW AI campus, as its couple of hours of output cannot be sustained for 24/7 requirements.
Additionally, modern GPU clusters already consume between 50–150 MW, and future frontier AI training clusters are expected to exceed 250 MW per site. Furthermore, several proposed AI-focused campuses are now targeting grid connections of 1 GW or more. These loads place them in a category above traditional commercial loads and on a level akin to large power stations.
The Location Problem
Geographic and grid-level constraints further complicate VPP’s answer as a possible solution. VPP assets tend to be focused in California and other Southwestern states. Conversely, data centers tend to be increasingly concentrated in regions with favorable transmission access, land availability, and tax incentives, especially Virginia, Ohio, Georgia, and Indiana. This mismatch between where VPP resources exist and where data center demand is growing most rapidly is one of the key constraints limiting VPPs as a near-term solution.
The region most strained by data center growth is PJM, where a large number of generation projects are waiting in interconnection queues, meaning grid connections can be delayed for several years. These delays in supply creation while demand skyrockets have led to PJM's 2025/2026 Base Residual Auction capacity prices to increase by over 800% year over year, from $28.92/MW-day to over $269/MW-day. Additionally, PJM implemented a price cap/floor system, setting a ceiling of $325/MW-day for the 2026/2027 and 2027/2028 auctions. However, that limit did not bind in practice, with 2026/2027 seeing $329 and $333 for 2027/2028. These higher prices improve the economics of demand-response and VPP participation.
Partial exceptions exist, with Texas representing a rare overlap and opportunity. In ERCOT, high solar penetration, rapid battery deployment, and flexible retail markets have created an environment where VPP-style aggregation can scale more quickly than in many other regions. However, ERCOT’s relative grid isolation from eastern interconnections also limits the extent to which surplus flexibility can be transferred to data center-heavy regions in PJM.
Conclusion
Virtual Power Plants are and will continue to be a highly effective tool and a more cost-effective alternative to pricier peaker plants for dealing with periods of peak demand. Policy support and recent private deals are further evidence that they will play a growing role in the grid, but VPPs are designed as short-duration resources to be deployed during times of grid stress rather than sustained energy supply sources.
AI data centers, however, are not marginal peak loads but continuous, industrial-scale consumers operating at 100–300 MW per site, with future facilities likely to require more than a gigawatt. VPPs depend on voluntary participation, have limited duration, and operate on a far smaller scale. Ultimately, this means VPPs can address the edges of the problem, but they cannot act as a replacement for the underlying generation required and are best understood as a secondary tool within a much larger buildout problem.