Google backs a 1-gigawatt virtual power plant as AI strains the grid

The story

Google and energy-market platform Voltus have expanded their work on a virtual power plant in the PJM region to gigawatt scale, turning thousands of distributed devices and flexible electricity loads into a coordinated grid resource. Google describes the October 22 agreement as a 1-gigawatt, tenfold expansion of an earlier 100-megawatt deal. Voltus says the new commitment adds 1,070 megawatts of capacity across the Mid-Atlantic electricity market.

A virtual power plant is software-orchestrated capacity, not a new generating station. Voltus plans to combine resources such as commercial and residential batteries, smart thermostats, onsite generation and industrial processes that can temporarily reduce or shift consumption. When PJM faces a peak, the platform can dispatch participating assets as a portfolio. Customers remain connected to the grid and receive bill savings or direct payments for making their flexibility available.

That distinction is essential. The announcement does not mean Google has secured a continuous gigawatt of new electricity for its data centers, and the distributed assets will not all discharge or curtail at every hour. The relevant product is dependable flexibility during constrained periods. If Voltus enrolls enough assets and they perform as promised, PJM can meet part of a peak by lowering demand rather than starting another plant or waiting for a major transmission upgrade.

The scale is unusually large for a corporate-backed distributed-energy program. Google and Voltus announced their initial three-year, 100-megawatt arrangement in June. Google now calls the expanded portfolio one of the largest known distributed-capacity commitments by a hyperscaler. The territory matters too: PJM coordinates wholesale electricity across all or parts of 13 states and the District of Columbia, a region serving about 67 million people and a growing concentration of data centers.

Independent reporting places the deal inside a broader race to find capacity for AI infrastructure. Latitude Media reported that the first 100-megawatt portfolio was filled and oversubscribed, and that the companies expect the larger contract to run through 2040. It also noted the central execution question: virtual power plants remain a small share of the grid, so enrolling and retaining enough real assets to deliver gigawatts will be far harder than proving a 100-megawatt pilot.

The attraction is speed. New gas turbines, substations and transmission lines can take years to permit, manufacture and connect, while existing batteries and controllable loads can often be enrolled with software, metering and contractual work. Reuters reported in July that U.S. virtual-power-plant capacity rose 13.7% to 37.5 gigawatts in 2025, citing Wood Mackenzie, as utilities and technology companies looked for near-term responses to power demand from data centers.

The economics could also broaden who benefits from the AI build-out. Under Voltus's model, owners of distributed equipment earn revenue for participating rather than Google owning every asset itself. In theory, data-center demand helps fund controls and batteries at homes and businesses, while the resulting flexibility reduces stress for the wider system. Whether that produces lower system costs will depend on contract terms, capacity-market treatment and whether payments exceed the administrative and equipment costs borne by participants.

Important numbers remain undisclosed. Neither announcement provides a deployment schedule for the full portfolio, expected customer count, resource mix, price paid by Google or anticipated compensation per participant. Nameplate capacity is also not the same as accredited capacity: PJM will value each resource according to its availability and performance rules. Batteries have duration limits, thermostats depend on weather and customer comfort, and industrial loads cannot always be interrupted when prices spike.

INNOVOX analysis: the most consequential feature is not the headline gigawatt but the attempt to make flexibility bankable for a data-center customer. Google is effectively supporting a long-lived market for capacity assembled outside its campuses. If that demand gives Voltus confidence to finance enrollment and integration at scale, it may establish a repeatable model for converting dispersed equipment into infrastructure. The risk is that a large contract is mistaken for delivered capacity before the assets are recruited, accredited and tested.

What to watch next is measurable delivery. Voltus should disclose enrollment milestones, the share supplied by batteries versus demand reduction, geographic distribution and the capacity PJM ultimately accredits. Dispatch results during real grid events will show whether the portfolio responds at the promised scale and for long enough. The decisive policy test is whether this added flexibility accelerates new data-center connections while protecting reliability and avoiding higher costs for customers who do not participate.

INNOVOX analysis

The agreement makes grid flexibility part of the commercial infrastructure behind AI expansion. Instead of treating households and businesses only as electricity consumers, it gives their batteries, controls and adjustable loads a route into capacity markets. The model could spread some data-center growth benefits beyond data-center sites, but the announced scale is a procurement target, not proof that 1 gigawatt of dependable capacity is already operating.

What to watch

Watch for deployment milestones, the mix of batteries and curtailable load, PJM-accredited capacity, participant compensation, performance during peak events and evidence that the portfolio actually shortens data-center interconnection timelines without shifting costs to other customers.