Virtual power plants and the AI load crunch: the only capacity that arrives in months

In 50 words: US virtual power plants enrolled 38 GW in 2025, up 21 percent, with residential battery enrolment up 153 percent. The DOE targets 80 to 160 GW by 2030 — 10 to 20 percent of peak demand and roughly $10 billion in annual savings. Brattle puts 400 MW of VPP capacity at $2 million a year against $43 million for equivalent gas.

The constraint on AI is no longer chips. It is interconnection.

A data centre developer in Northern Virginia can wait up to seven years for a grid connection. In Dallas-Fort Worth the queue runs to 2027 or later. Gas turbines are backlogged through at least 2028 as manufacturers sell out their order books. Large power transformers, the unglamorous iron that every one of these projects needs, have seen lead times double to as much as five years. Transmission lines take a decade or more.

Set that against the deployment cycle of the industry demanding the power. A hyperscaler decides to build, and expects to be training models in that building within 18 to 24 months. There is no supply-side resource that can be procured on that timeline — with one exception. Capacity that already exists, sitting behind millions of meters, waiting to be aggregated.

That is the case for the virtual power plant, and 2026 is the year it stopped being a demonstration project.

Table of contents

• What a VPP actually is
• The numbers as of 2026
• Why speed beats everything else right now
• The cost comparison is not close
• The data centre angle: 300 hours a year
• Where VPPs genuinely fall short
• What has to change for the DOE target to land
• Read-across for India and other markets
• What to watch next
• Frequently asked questions

What a VPP actually is

A virtual power plant is an aggregation of distributed energy resources — home batteries, commercial storage, smart thermostats, EV chargers, water heaters, backup generators, rooftop solar — coordinated by software so that, from the grid operator's perspective, the fleet behaves like a single dispatchable resource.

The word "virtual" is doing real work. Nothing is built. The batteries are already in garages, the thermostats already on walls. What the VPP adds is the control layer, the market registration and the contractual arrangement that lets thousands of individually trivial resources be bid into a wholesale market or called on for capacity as one unit.

This is distinct from, though related to, the DERMS platforms utilities use to manage distributed resources. A DERMS is the utility's operational tool for grid management. A VPP is a market participant. In practice, the same hardware often sits underneath both.

Two dispatch modes matter:

• Load reduction. Thermostats set back two degrees across 100,000 homes, EV charging paused, industrial process shifted. The grid sees this as supply because it removes demand at the moment of scarcity.
• Export. Home and commercial batteries discharge into the grid. This is the mode growing fastest, because a battery is a far more precise and reliable resource than aggregated thermostat behaviour.

The numbers as of 2026

US virtual power plants enrolled roughly 38 GW of capacity in 2025, a 21 percent increase on the prior year, per Ohm Analytics. The composition of that growth is the interesting part: residential battery enrolment rose 153 percent year over year. Data centre-driven demand is credited with a 33 percent jump in VPP deployments.

The Department of Energy's Pathways to Commercial Liftoff work targets 80 to 160 GW of national VPP capacity by 2030. At that scale VPPs would supply roughly 10 to 20 percent of peak demand and save on the order of $10 billion a year in grid costs. At current growth rates the low end of that range is achievable; the high end is not, without policy change.

For scale: 38 GW is comparable to the entire installed generating capacity of a mid-sized European country, assembled without a single planning permission, transmission upgrade or environmental review.

Why speed beats everything else right now

The strongest argument for VPPs in 2026 is not that they are cheap, though they are. It is that they can be delivered inside the window during which the demand actually materialises.

Two deployments make the point concretely. California enrolled over 750 MW of customer-sited storage into flexible programmes, adding roughly 500 MW between January and October 2025 alone. Ontario stood up a 90 MW residential VPP by enrolling 100,000 homes in six months.

Six months. Against seven years for a Northern Virginia interconnection.

This inversion is new. For most of the history of the power sector, the fastest way to add firm capacity was to build a gas peaker. That is no longer true, and it is not true for reasons — turbine order books, transformer supply, interconnection queue length — that will not resolve quickly.

The cost comparison is not close

Brattle Group analysis puts the cost of 400 MW of VPP capacity procured for resource adequacy at roughly $2 million a year. The equivalent capacity from new gas generation plus the associated grid upgrades comes to about $43 million.

That is not a marginal improvement. It is a factor of twenty.

The reason is straightforward: the capital asset is already paid for. A homeowner bought the battery for backup power and bill savings. The VPP operator is renting a few hundred hours a year of its availability, not financing its construction. RMI's modelling suggests VPPs could cut total system generation costs 20 percent by 2035 on that basis.

The obvious objection is that this only holds while VPPs remain a thin slice of the system. Rent enough of the fleet, often enough, and homeowners will demand compensation approaching the value of the service rather than a modest annual payment. That is a real effect, and it will compress the advantage. But a factor of twenty leaves considerable room for compression before the economics stop working.

The data centre angle: 300 hours a year

The sharpest insight in the current literature is about how little flexibility is actually needed.

RMI's illustrative scenario finds that 14 GW of continuous data centre demand could be accommodated using roughly 300 hours a year of VPP-enabled flexibility. That is about 3.4 percent of the hours in a year.

The reason is that grid scarcity is extraordinarily concentrated. A power system is not short of capacity most of the time; it is short during a few dozen hours of extreme heat, extreme cold or unusual outage coincidence. If a data centre can be firmed through those specific hours — by aggregated distributed resources discharging, or by the facility itself curtailing non-urgent compute — the rest of the year takes care of itself.

This reframes the interconnection problem. The developer does not need seven years of grid buildout to serve a constant load. It needs a credible answer for a few hundred hours. That answer can be a VPP contract, on-site storage, or a flexible-load commitment — and increasingly it is a combination, with hyperscalers moving from passive offtakers to active grid participants offering load flexibility in exchange for faster connection.

It also connects directly to the on-site storage strategies data centres are already deploying. The two are complements: batteries in the building handle ride-through and short scarcity events, the VPP contract handles system-level capacity obligations.

Where VPPs genuinely fall short

An honest assessment has to acknowledge what aggregated distributed capacity cannot do.

• Duration. Most residential batteries hold 10 to 14 kWh usable and are also serving the homeowner's own backup and arbitrage needs. A VPP is a two to four hour resource. It does not solve multi-day cold snaps or the long-duration storage problem.
• Firmness and consumer override. The homeowner can opt out. Participation rates during a called event are high but never 100 percent, and they degrade if events are frequent or badly timed. Grid operators must derate accordingly, and the derate is not always well characterised.
• Locational value. A VPP is only useful for a constrained data centre if the enrolled resources sit behind the same constraint. Aggregating batteries across a whole state does little for a single congested substation in Loudoun County.
• Measurement and verification. Proving a negative — that demand would have been higher without the intervention — depends on baseline methodology, and baseline gaming is a real problem in demand response markets.
• Market access. FERC Order 2222 opened wholesale markets to distributed aggregation, but implementation across the RTOs has been slow and uneven, with participation models that vary enough to make multi-market scaling genuinely difficult.

Anyone claiming VPPs remove the need for firm generation and transmission is overselling. The accurate claim is narrower and still powerful: they are the cheapest and fastest way to cover the peak hours, which defers and shrinks what has to be built.

What has to change for the DOE target to land

Three things separate the 80 GW low case from the 160 GW high case.

Market rules. Order 2222 implementation needs to actually finish, with participation models that a national aggregator can work with rather than fifty bespoke integrations. Aggregators currently spend disproportionate effort on regulatory engineering rather than customer acquisition.

Compensation that reflects value. If a VPP saves $41 million against the gas alternative, paying the aggregate fleet $2 million captures very little of that value for the people whose assets provide it. Household participation payments in most programmes remain modest. Raising them is both fair and the fastest route to enrolment growth.

Default-on enrolment. The single largest determinant of VPP scale is whether battery and thermostat owners are enrolled by default with an opt-out, or must opt in. Ontario's 100,000 homes in six months reflects programme design, not unusual enthusiasm.

Read-across for India and other markets

India does not have a residential battery fleet to aggregate, so the direct VPP model does not transfer. But two adjacent things do.

The flexible-load logic applies immediately. India's peak problem is concentrated in evening hours, and the resources that could be aggregated are commercial and industrial — captive diesel, cold storage, water pumping, EV charging depots. That is closer to classical demand response than to a battery VPP, and it needs the same market plumbing.

The second read-across is timing. India is deciding now how distribution-level storage gets funded and who controls dispatch — including, as of July 2026, DISCOMs being permitted to spend PM Surya Ghar incentive balances on battery storage. Whether those assets end up as utility-controlled islands or as an aggregatable fleet is a design decision being made today, largely by default, and it will determine whether India has the option of VPP-style aggregation in a decade.

What to watch next

Watch whether a hyperscaler signs a VPP capacity contract as a condition of interconnection rather than as a sustainability gesture. That is the transition from voluntary to structural, and it is the signal that utilities have accepted flexible capacity as a substitute for wires.

Watch residential battery enrolment growth. At 153 percent year over year it is the fastest-moving component, and it is the one that turns VPPs from a load-reduction resource into a genuine dispatchable one.

And watch the first serious failure. A large VPP that underdelivers during a major scarcity event will do more to shape market rules than any amount of successful ordinary operation.

Frequently asked questions

What is a virtual power plant?

A virtual power plant is a software-coordinated aggregation of distributed energy resources — home and commercial batteries, smart thermostats, EV chargers, rooftop solar and flexible industrial loads — that is bid into electricity markets or called for capacity as though it were a single power plant. Nothing new is built; existing customer-owned assets are coordinated and compensated.

How much VPP capacity does the US have?

US virtual power plants enrolled roughly 38 GW in 2025, up 21 percent year over year, with residential battery enrolment rising 153 percent. The Department of Energy targets 80 to 160 GW by 2030, which would meet 10 to 20 percent of peak demand and save around $10 billion annually in grid costs.

Are virtual power plants cheaper than building gas plants?

For resource adequacy, substantially. Brattle Group analysis puts 400 MW of VPP capacity at roughly $2 million per year against about $43 million for equivalent new gas generation plus associated grid upgrades. The advantage exists because the underlying assets are already financed by their owners; the VPP rents availability rather than funding construction.

Can virtual power plants power AI data centres?

Not continuously, but they can cover the hours that matter. RMI modelling indicates 14 GW of continuous data centre demand could be accommodated with roughly 300 hours a year of VPP-enabled flexibility, because grid scarcity is concentrated in a few dozen extreme hours. VPPs work as a firming and interconnection-acceleration tool alongside on-site storage, not as a replacement for baseload supply.

Why are VPPs suddenly growing so fast in 2026?

Supply-side alternatives have become slow and scarce. Gas turbine order books are full through at least 2028, large transformer lead times have doubled to as much as five years, and data centre interconnection queues run to seven years in Northern Virginia. A residential VPP can be enrolled in six to twelve months, which is the only timeline that matches how quickly AI-driven demand is arriving.

What are the main limitations of virtual power plants?

Duration is the biggest: most residential batteries provide two to four hours, so VPPs do not address multi-day scarcity. Participation is voluntary and never complete, so capacity must be derated. Value is locational — aggregated resources only relieve a constraint if they sit behind it. And wholesale market access under FERC Order 2222 remains unevenly implemented across regions.

---

Researched and drafted with AI assistance; reviewed and edited by the named author within 24 hours of draft. Capacity and cost figures are as published by DOE, RMI, Brattle Group and Ohm Analytics as of mid-2026. See our editorial standards and AI disclosure.

Related reading: Battery storage for data centres · DERMS and distributed energy management · Long-duration storage in 2026 · Battery storage hub · What resource adequacy means