# Energy and Artificial Intelligence: The New Grid Bargain > New Jersey's Data Center Fair Share Act will let a data center move up the interconnection queue by paying to install heat pumps, rooftop solar, and batteries in other people's homes, with aggregators bundling those households into verified virtual power plants. PJM's new 15-year backstop contracts and ERCOT's ride-through rules ask for the same thing in different currencies, and New York's statewide moratorium shows what happens where no such trade exists. Content type: article Source URL: https://www.agentpmt.com/articles/energy-and-artificial-intelligence-the-new-grid-bargain Markdown URL: https://www.agentpmt.com/articles/energy-and-artificial-intelligence-the-new-grid-bargain?format=agent-md Updated: 2026-07-22T23:11:03.268Z Author: Pancakes Tags: AI Agents In Business, AI Powered Infrastructure, AgentPMT, News --- # New Jersey Will Let Data Centers Buy Grid Access by Paying for Heat Pumps Governor Mikie Sherrill signed the Data Center Fair Share Act on July 7, and it contains one of the odder bargains in American energy policy. A data center that wants onto New Jersey's grid ahead of the queue can get there by buying strangers a new water heater. The mechanism is more sensible than it sounds. Under the law, a developer funds household upgrades across the state: swapping inefficient electric-resistance space and water heating for heat pumps, putting solar on roofs, adding batteries. Companies then bundle those households into virtual power plants, and utilities verify that the resulting capacity will actually be available by the time the data center comes online. That verified capacity buys priority in the interconnection queue, the multi-year line every large electricity customer waits in before it can plug into the grid. Canary Media, which reported the signing, says the program could lower bills for participating households by hundreds to thousands of dollars a year, though that figure is a projection, not a result. None of it is operating yet. New Jersey's Board of Public Utilities has a year to write the standards, utilities get 180 days after that to submit program proposals, and the first households could see equipment installed around mid-2028. The same law creates a new rate class so data centers cover their own energy use and infrastructure costs, following Minnesota, Oregon, and Virginia. California, Colorado, Illinois, Pennsylvania, and New York are weighing versions of the upgrade program. What makes it interesting is that it is a trade rather than a tax. A tax takes money and sends it somewhere. This converts a developer's capital into dispatchable capacity that already sits behind meters in existing homes, paired with the automated energy management software an aggregator uses to call on it. No new transmission, no new turbines, no five-year build. And the capacity has to be measured and confirmed before it counts toward anything, which is the detail that makes the rest of this month's news legible. ## The demand math that made the trade necessary Utilities do not hand out queue positions out of goodwill. They are staring at load growth that their planning models were not built for. BloombergNEF nearly doubled its forecast for US data center power capacity in 2035, to 194 gigawatts, which would be close to a fifth of all US electricity consumption. Treat that as a forecast rather than a measurement: it moved that far in seven months, and BNEF's own Mark Daly told Latitude Media the firm's build-rate methodology may be too generous given how hard it has become to hire people who can do sophisticated electrical work and to get power equipment on site. Even discounted, the direction sets up a sequencing problem. Generation and wires arrive years after the load wants to connect, and the load is impatient in a way that a gas turbine order book is not. That is why capacity that already exists, sitting in water heaters and home batteries and flexible industrial load, is suddenly worth trading a queue position for. It is the fastest megawatt on the board because it is already installed. One step removed from the substation, the same measurement question lands on anyone running AI workloads. Grid operators are insisting on verified numbers before granting access; a lot of software teams still scale agents without knowing what a single run costs, which is how a promising automation turns into a surprise invoice. AgentPMT's test bench takes the grid's posture: every workflow run itemizes the cost of each tool call and the model tokens behind it, and you can swap the model, run the identical workflow again, and compare cost against output quality before committing to scale. That is right-sizing rather than guessing, and it is the same discipline utilities are now applying to load, described in more detail in [Cost Attribution for Agent Work: Every Dollar Gets a Name](https://www.agentpmt.com/articles/cost-attribution-for-agent-work-every-dollar-gets-a-name). ## PJM turned the same bargain into a market New Jersey's law would be a curiosity if the idea were not showing up somewhere else every few days. PJM, the grid operator covering thirteen states and Washington DC, ran a capacity auction that hit its price cap for the third consecutive time, at a cost of $16.4 billion, and still fell short of its own reliability requirement. An auction that pays the maximum and still cannot buy what it needs has stopped functioning as a price signal. Canary Media reported that this round brought almost no genuinely new capacity to the table, a fraction of what the prior five auctions produced between them. So PJM stakeholders approved something else: a reliability backstop procurement, in which data centers and the utilities serving them directly fund new generation and battery capacity under 15-year contracts. The first auction is targeted for September, pending sign-off from PJM's board and federal regulators, and the contract length is the actual news. Fifteen years is a financing instrument. Developers cannot get steel in the ground on a one-year clearing price, and duration is what converts a data center's appetite into someone else's power plant. It also creates an obvious way to get fleeced, and the Natural Resources Defense Council is saying so out loud. Tom Rutigliano, the group's senior advocate for climate and energy, warns that utilities could step forward as buyers on behalf of large loads without binding those loads to pay, and that the backstop could become "potentially a way to funnel a lot of money to projects that are already in the works." His standard is blunt: states "have to make damn sure there's a data center that pays for it, or \[utility\] shareholders pay for it." That is a design question about tariffs, not an argument against building compute, and it is answerable with contract language. ## Behavior became a hardware requirement The third version of the bargain does not involve money at all. It asks what the load will do. In February 2025, a high-voltage line malfunctioned in Virginia and roughly forty data centers, about two gigawatts between them, simultaneously flipped to backup power to protect their own equipment. Each individual decision was correct. Together they handed the grid operator a sudden load-loss event to manage. A data center's instinct during a voltage dip is to disconnect; at that scale, self-protection becomes somebody else's reliability incident. ERCOT's interconnection rules now say the opposite: loads of 75 megawatts or more have to ride through disturbances, staying connected or restoring at least 90 percent of their power draw within seconds. That is a behavioral requirement written into the terms of access, and it has a price tag. Crusoe Energy announced this month that it will deploy 5 gigawatts of ON.energy's uninterruptible power systems across its campuses starting this year. The equipment is a shock absorber in a cabinet, a battery, two power conversion systems, and a transformer sitting between the data center and its power source, decoupling the sharp swings of GPU training runs from the grid on one side and shielding the site from external faults on the other. "This technology protects our infrastructure, it protects our generation assets, and it aligns with where grid requirements are heading," said Chris Dolan, Crusoe's chief data center officer. Three levers, then, all buying the same thing. New Jersey sells access for contributed capacity, PJM sells it for contributed capital, ERCOT sells it for [compliant behavior](https://www.agentpmt.com/articles/ai-and-the-environment-texas-makes-data-centers-pay). A mid-year developer survey published by Latitude Studios and sponsored by Bloom Energy, which sells on-site power and therefore has a stake in the answer, points in the same direction: developers are putting money into grid support and resiliency, and they increasingly expect to bring their own generation. Flexible load only works if something upstream can decide which work to slow, shift, or shed, and can show afterward what that decision cost. That is a workload orchestration job, and it is where AgentPMT operates: hard spend caps per agent connection with per-tool and per-vendor restrictions, schedules and webhook triggers so a workflow fires when conditions are right instead of the instant somebody asks, an approval node you can drop into a workflow when a human should sign off before an expensive step runs, and an audit feed that records every action down to the request and the response. It does not dispatch megawatts. It governs and meters the work sitting on top of them, which is the same accountability the grid is now writing into its own rules. The fleet-level version of that argument is in [From Assistant to Workforce: Operating an Agent Fleet](https://www.agentpmt.com/articles/from-assistant-to-workforce-operating-an-agent-fleet). ## Where there is no bargain, there is a pause New York took the other road. On July 16, Governor Kathy Hochul highlighted a one-year hold on state environmental permits for new hyperscale data centers, the first statewide moratorium in the country. "Progress shouldn't arrive with a higher utility bill, depleted water supply or noise pollution, which is why I signed the nation's first ever statewide moratorium on hyperscale data centers," she said. Read the rest of the order and the pause looks less like a wall and more like a drafting table. During the year, the state will produce a Generic Environmental Impact Statement assessing energy demand, water use and quality, and air quality from data center construction and operation, the most systematic artificial intelligence environmental review any state has attempted. New York is trying to answer how AI affects the environment inside its own borders with enough precision to write rules from. Alongside it comes a Community Investment Framework covering local infrastructure, child care investment, prevailing wage standards, project labor agreements, local hiring, and apprenticeships, with a formula communities can use to open negotiations. A pause is what a state reaches for when it has no agreed price for the trade and no way to verify what it would be buying. The pressure behind that decision is real and organized. On July 18, 142 rallies took place across 42 states in a coordinated day of action. TIME reports that hundreds of cities, towns, and counties have now enacted bans or moratoriums, and that dozens of major projects were delayed or canceled in the first quarter of this year in part because of organized local opposition. In Coweta County, Georgia, resident Steve Swope told TIME his concern is his well: "Our wells will either run dry or they'll become polluted and caked with sediment and mud." No state has a settled method for pricing the environmental effects of artificial intelligence at this scale, so the complaints that land are the specific ones, and they name practices rather than an industry. Most Virginia localities negotiate data center proposals under non-disclosure agreements, which makes it impossible for a resident to know what was promised. Supply costs socialized through tariffs land on households that never signed anything. Siting large cooling loads in water-stressed basins is a choice a developer makes and can unmake. Each of those has a fix, and states that build the fix get projects; states that cannot get moratoriums. Amazon's public posture shows what the answer looks like when someone bothers. Brandon Oyer, the company's director of energy and water strategy, described siting decisions that now favor grid locations where added load helps existing infrastructure rather than straining it, which is what drove investments in Jackson, Mississippi and South Bend, Indiana. Project Rainier, its Indiana campus, runs on outside air nearly year-round and falls back on water cooling only a handful of days. The company has also signed the White House pledge to cover its own electricity production costs rather than pass them to other ratepayers. Disclosed numbers, offered before anyone subpoenas them, are what sustainable AI development looks like at the operating level. For teams doing that work in software, AgentPMT's [Global Energy & Power Grid Data](https://www.agentpmt.com/marketplace/energy-access-production) and [Climate, Environment, and Land Data Hub](https://www.agentpmt.com/marketplace/climate-environmental-data) products give an agent the grid and climate feeds a siting or reporting workflow needs. Three dates carry this forward. PJM's backstop auction is targeted for September, pending federal approval. New Jersey's utility board has a year to write its standards, with utility proposals 180 days later. New York's impact statement runs the length of its pause. All three are asking the same thing of anyone who wants to build: show what the load draws, show how it behaves when the grid stumbles, show what it gives back, and let someone else check the arithmetic. That is a higher bar than writing a check, and a more durable one. At the seam between energy and artificial intelligence, the developer who can produce those numbers on demand is the one getting connected first. * * * ## Sources - New Jersey law will let data centers pay for home energy upgrades, Canary Media - PJM's old way of getting power built isn't working. Has it found a fix?, Canary Media - BNEF nearly doubled its forecast for US data center power demand, Latitude Media - Crusoe's latest deal targets power electronics for grid reliability, Latitude Media - Community Backlash to AI Data Centers Is Growing Across the U.S., TIME - Data centre backlash goes national as protests hit 42 US states, The Next Web - Governor Hochul Highlights First Statewide Moratorium on New Hyperscale Data Centers, New York Governor's Office - Inside Amazon's approach to data center sustainability: A conversation with Brandon Oyer, Latitude Media - Survey: As community scrutiny grows, data center developers turn to on-site power, Latitude Studios (sponsored by Bloom Energy)