Google Is Paying to Power Its Data Centers With Your Thermostat

Google signed a three-year agreement with Voltus, a virtual power plant operator based in San Francisco, to finance 100 megawatts of distributed energy capacity inside PJM Interconnection, the largest power grid in the United States. Announced on June 2, the deal has Voltus aggregating batteries, smart thermostats, and other flexible devices from households and businesses across PJM’s 13-state footprint, paying those participants directly, and delivering the resulting capacity to the grid on Google’s behalf. The plant is due online in 2027.

Whether the architecture delivers the enrolled capacity its target requires is the part no signed contract can settle.

Google Funds 100 Megawatts It Doesn’t Own

Voltus calls the structure Bring Your Own Capacity (BYOC), a framework in which a data center operator finances the flexibility potential of communities near a new facility, delivering that contracted capacity to the local utility rather than building generation or waiting in the interconnection queue. Voltus launched BYOC in September 2025 with Cloverleaf Infrastructure, a data center developer; Google is the first hyperscaler to sign on.

The mechanics: Google funds the program. Voltus recruits participants from homes and businesses in PJM, pays them to allow their devices to respond to grid signals, and dispatches those distributed energy resources (DERs) when demand spikes, adjusting thermostats and discharging home batteries to free up accredited capacity. Voltus already has more than 7.5 gigawatts of DERs connected across nine North American wholesale markets. In February 2026, it extended its residential pipeline through a partnership with Octopus Energy US, a retail electricity provider, adding household-level flexible assets across PJM, MISO (Midcontinent Independent System Operator), New York, and California. According to Wood Mackenzie data cited by Voltus, the VPP market grew more than 33% in the prior year, with PJM and ERCOT leading in disclosed offtake capacity.

What BYOC lets Google avoid is a generation investment. Building a gas plant or solar farm takes years and hundreds of millions of dollars before a single electron reaches a data center. Interconnection queue timelines in PJM run to 2030 and beyond for most new generation. The BYOC approach delivers accredited capacity within the three-year deal window, funded by Google but drawn from equipment already installed in people’s garages and living rooms.

In March 2026, Google disclosed it had integrated 1 gigawatt of demand response capacity into long-term energy contracts with multiple US utilities, covering machine learning workloads that could be shifted during grid stress. Amanda Peterson Corio, Google’s global head of data center energy, told Utility Dive the reason Google prefers paying third parties rather than flexing its own hardware:

The cost of capital in the data center, of our chips, can be billions and billions of dollars of hardware that only gets utilized to our customers if it’s running.

She was describing the cost of server downtime. Whatever Voltus charges households to adjust a thermostat, it is considerably less than what Google loses when a GPU cluster goes offline.

PJM’s Supply Crunch, Priced in Billions

PJM Interconnection serves roughly 67 million people across 13 states from Illinois to the District of Columbia through a capacity market in which generators are paid for agreeing to be available during peak demand. Those prices have climbed in a way the market wasn’t designed to absorb.

Sources: IEEFA (2024/25 through 2026/27 delivery years); PJM Interconnection December 2025 auction release (2027/28 delivery year).

Data center demand drove most of the increase. Monitoring Analytics, PJM’s independent market monitor, found data centers accounted for $6.5 billion, or 40%, of the $16.4 billion in capacity costs from the December 2025 auction, which set rates for the 2027/28 delivery year. Across PJM’s three most recent base capacity auctions combined, data center-related load represented 45% of $47.2 billion in total costs. The December 2025 auction was the first in which the entire PJM footprint fell short of its reliability requirement.

On the supply side, PJM processed more than 170,000 MW of new generation interconnection requests since 2023 but still had 30,000 MW in a transition queue as of late last year. PJM projects its summer peak load will grow at 3.6% annually, reaching roughly 222 GW by 2036, with data centers driving the bulk of the increase. Maryland filed a complaint with the Federal Energy Regulatory Commission (FERC) over roughly $2 billion PJM is charging the state for grid upgrades tied to data center growth, one indication of how broadly those costs are spreading to ratepayers.

Those numbers explain why Google is financing a VPP. Waiting for generation is slow and getting slower.

Duke’s Headroom Formula

The theoretical case for BYOC rests partly on a study published in February 2025 by the Nicholas Institute for Energy, Environment and Sustainability at Duke University, detailed in a Data Center Dynamics analysis of US grid headroom. Researchers found the existing US grid could absorb up to approximately 100 gigawatts of new flexible data center load if those loads curtailed demand during fewer than 1% of annual operating hours, averaging about two hours per curtailment event. The study introduced the concept of curtailment-enabled headroom: the additional load a grid can absorb using existing capacity with brief, modest usage reductions.

US grids are sized for absolute peak demand, not average consumption. The hottest summer evening, when air conditioning, industrial loads, and consumer electronics all converge, defines the ceiling. That ceiling is reached for a small fraction of annual hours. A load willing to pull back during those windows uses almost none of the incremental generation or transmission infrastructure the grid would otherwise require. The Brattle Group, an economics consultancy, estimated in analysis cited in the Voltus-Google announcement that better utilization of existing grid resources through VPP-type solutions could save US consumers more than $100 billion over the next decade.

Not all data center loads are equally curtailable. Training a machine learning model can be paused for two hours without meaningful business cost, but live user queries cannot be deferred. As AI inference grows as a share of total data center power, the curtailable fraction shrinks. A follow-up study from Duke’s Nicholas Institute, published in February 2026, found that modest flexibility with 20% of a data center’s total demand flexible cuts projected new natural gas construction by nearly 20% and reduces average data center electricity prices by up to $3 per megawatt-hour. Google has addressed its own curtailable workloads through demand-response contracts with Entergy Arkansas, Minnesota Power, DTE Energy, and other utilities, which can shift those machine learning loads when PJM runs tight. The Voltus deal doesn’t touch Google’s internal load. It finances the flexibility of the surrounding grid.

The Households Voltus Still Needs to Enroll

California’s EV Charging Enrollment Gap

The VPP’s performance depends on people actually joining. A study published in early 2026 and examined at the UC Berkeley Energy Institute blog on managed EV charging enrollment tested participation across the full eligible population at Peninsula Clean Energy (PCE), a California utility, rather than among self-selected volunteers. Researchers randomized all 12,174 eligible EV-owning households into experimental groups at different monthly incentive levels.

• 1% enrolled when offered no financial incentive
• 4.6% enrolled at $40 per month, roughly 15% of an average participant’s electricity bill
• At the $40 level, the enrolled group was too small to produce statistically detectable effects on aggregate electricity consumption

Managed EV charging differs from the thermostat-and-battery flexibility Voltus aggregates, and California’s EV-owning population is not a demographic proxy for PJM’s 13-state footprint. But the mechanism, paying households to cede some control of their electricity equipment, is the same. Neither Google nor Voltus has disclosed the per-household payment rate they plan to use in PJM, and that figure will largely determine how close enrollment comes to the 100-megawatt target.

The Gallup Backdrop

A separate challenge sits beneath the enrollment economics.

A Gallup poll conducted March 2-18, 2026, across 1,000 US adults, found 71% oppose the construction of an AI data center in their local area, with 48% strongly opposed and only 27% in favor. It was the first time Gallup had asked Americans specifically about data center construction, and opposition landed higher than for nuclear power plants, 71% vs. 53%. Roughly 47% of Americans opposed AI data centers in a late-2025 Gallup comparison, a 24-point shift in a matter of months. Among opponents, half cited excessive resource use, with 18% specifically naming energy consumption and another 18% water use. About one in five cited higher utility bills. Roughly 69 jurisdictions across the US had enacted moratoriums on data center construction by the time the poll was conducted.

Voltus’s pitch to households is direct: get paid for devices you already own, give up minimal control, and support local grid stability while data center demand grows around you. The program hasn’t made a single payment offer to a household in PJM yet. The 2027 launch is when that sequence begins, and when the Gallup numbers start to meet the Voltus pitch.

Regulation Is the Other Lever

Google and Voltus are working inside PJM’s voluntary capacity market. Alongside it, regulation is arriving with less patience.

Texas enacted a law requiring large electricity users, including data centers, to curtail demand or switch to backup power during grid emergencies. In the broader US, proposals have circulated that would allow new data centers to connect to the grid years ahead of standard interconnection timelines if they commit to reducing their draw when the system approaches its ceiling. Both represent mandatory versions of the BYOC premise: data center flexibility as the price of connection.

The larger PJM event is an emergency capacity auction that President Trump and a bipartisan group of 13 governors pushed the grid operator to hold. PJM confirmed in April 2026 it is targeting up to 15 gigawatts of new generation through a bilateral negotiation process running from September 2026 through March 2027, in which data center companies would bid directly for 15-year generation contracts and pay for new baseload power. PJM’s emergency auction follows the same policy logic as the voluntary approach: data centers should finance the grid capacity their growth consumes.

Microsoft, Amazon, and Meta face identical interconnection bottlenecks in PJM. All three have been signing nuclear power deals and building natural gas plants to supply their own generation, a strategy that sidesteps the queue but at far greater capital cost and carbon exposure. Voltus CEO Dana Guernsey said in September 2025 the company was in active conversations with multiple hyperscalers and expected additional BYOC announcements. The June 2 Voltus-Google press release framed the deal as an "industry-leading scalable blueprint" for data center capacity procurement, a model other large loads could follow once the enrollment numbers come in. One hundred megawatts per year is modest against the several hundred megawatts a hyperscale campus typically requires. The significance is in the contract structure it proves out.

The VPP is due online in 2027, and that is when Google’s wager on household participation becomes a data point.