GridCARE index pegs hidden US grid headroom at up to 120 GW

A different question about AI's power problem

Most of the debate about powering artificial intelligence has been about how many new power plants and transmission lines the United States needs to build. GridCARE, a grid software startup that came out of Stanford, is asking a narrower question: how much spare capacity is already sitting unused on the grid that customers have paid for? On Oct. 9 the company said it had qualified 5 gigawatts of additional capacity on the existing U.S. grid over the past 12 months, working with six utilities. It also released a Power Acceleration Index, which estimates that today's grid could support another 100 GW to 120 GW of demand and produce $270 billion to $350 billion in net benefits for customers.11

The company first presented the index at a summit in San Francisco on Oct. 8 and describes it as a national benchmark built on standardized assumptions.1112 CEO Amit Narayan framed the idea directly. In his telling, the industry has been asking how much power America must build, and GridCARE wanted to know how much power already exists in the grid that has been built.16

The headline number is large. Whether it holds up depends on assumptions about flexibility, location and regulation, and those deserve as much attention as the number itself.

How the 120 GW figure is built

The argument is simple. Power lines and substations are sized for peak demand, which happens only a few hours a year, so for most of the year they have room to spare.11 GridCARE cites a December 2025 Stanford study that found some Western utilities used as little as 18% of their transmission line capacity even at peak.12 The company's website says the U.S. grid is less than a third utilized even in its most congested areas.18

The catch is flexibility. Large customers such as data centers and factories could connect sooner if they agree to cut their power use when the grid is under strain.11 The 100–120 GW estimate assumes "moderate qualified flexibility." GridCARE itself says that actual capacity at any site still requires network- and site-specific studies.12 Daily Energy Insider added that the amount of new demand that can be served will vary with local grid conditions, weather, equipment outages and each customer's flexibility requirements.11

GridCARE's main claim is about what it calls "qualification." A data center saying it can reduce load is not enough, in the company's view. A utility needs proof that meets reliability standards: how much load a location can take under normal conditions and during outages, how often the site will have to cut back, how fast, and for how long.12 The utility then checks those findings in its own planning tools.12

The projects behind the numbers

The 5 GW claim rests on a handful of utility partnerships, and they are not all at the same stage. The most developed is with Portland General Electric in Hillsboro, Oregon. There, GridCARE qualified more than 400 MW of data center capacity: 80 MW available in 2026, and the rest to be energized by 2029 at the latest.11 Trade coverage says those projects had faced a seven-year waitlist.16 Latitude Media reported earlier that the Hillsboro work produced flexible interconnection agreements for five data centers.2

GridCARE says it has qualified more than 2 GW with several utilities in the Pacific Northwest.11 In New York, it worked with National Grid on up to 650 MW.11 Southern California Edison studied one constrained circuit in the Inland Empire and found up to 20 MW of additional flexible capacity.20 Work with South Carolina's Santee Cooper is still exploratory: the two are looking for ways to serve large loads using existing infrastructure.11

Coverage of the New York project describes it in slightly different terms. GridCARE's announcement says it "qualified" 650 MW.12 National Grid CEO Zoë Yujnovich, speaking to investors in May, said GridCARE had identified up to 650 MW of potential flexible connection capacity.15 Daily Energy Insider used similar wording, describing "potential capacity" for customers willing to accept flexible terms.11 The gap between "identified potential" and "qualified and energizing" matters. Readers should treat the 5 GW total as a pipeline at different stages of maturity, not as 5 GW of data centers already running.

The gap between the two headline figures is also large. Five gigawatts in a year is about 4% of the company's upper national estimate. The rest of the 120 GW is a modeled result from the index, not something that has been shown in the field.

Independent research points the same way

GridCARE's estimate is not an isolated claim. In early 2025, researchers at Duke University's Nicholas Institute studied 22 large balancing authorities that together serve about 95% of U.S. peak load. They found that 76 GW of new load could be added if those loads were curtailed 0.25% of the time, 98 GW at 0.5%, and 126 GW at 1%.26 Average curtailment events would be short, about 1.7 to 2.5 hours depending on the scenario. In nearly 90% of curtailment hours, at least half of the new load would keep running.26 Lead author Tyler Norris said the team was surprised by how much room it found across many balancing authorities.25

GridCARE's 100–120 GW falls within the Duke range, close to the 0.5% and 1% scenarios, and the company has cited the Duke work itself.20 Other studies support the general idea. A Camus Energy, encoord and Princeton analysis found that a flexible 500 MW data center in PJM could connect three to five years sooner than an inflexible one while still drawing grid power more than 99% of the time.1 Modeling of Texas's ERCOT grid by Relae found that data center demand response could eliminate forced load shedding even with 40 GW of new data centers, at an average curtailment of 5% of demand during fewer than 1% of hours.29

These studies agree that the grid has real headroom outside peak hours. They disagree on how much of it can be safely used at specific sites, and on what terms.

Where the skeptics push back

The strongest criticism of national headroom estimates goes to that site-level question. Energy analyst Brian Handshy argues that the Duke study is being misused. He points out that its authors called it a first-order estimate, did not model transmission constraints, and said it was unclear whether data centers could actually provide the assumed flexibility.22 In his view, system-wide headroom answers a planning question. It does not tell you how much firm capacity a specific project can claim at a specific connection point.22 He also says that 44 hours of downtime a year means 99.5% uptime, which falls short of typical data center service commitments.22

This criticism does not undercut GridCARE's business. If anything, it supports it. GridCARE sells site-by-site, utility-verified qualification, which addresses the locational gap Handshy describes.12 The company has also pointed to batteries and on-site generation as ways for data centers to stay flexible without shutting down computing work.3 That partly answers the uptime concern. The weak point is the national index. Its 120 GW comes from a screening model, not from adding up qualified sites.

Utility partners have also been clear about the limits. PGE's Larry Bekkedahl said the tools gave the utility confidence to connect loads of 50 MW to 500 MW, but added: "We do not solve the 2-GW data center problem."9 That matters because the largest AI campus proposals, such as Meta's Louisiana project, which Entergy said could need up to 2.2 GW, are exactly the loads that flexible connections cannot handle alone.9

The ratepayer argument

GridCARE's other big claim is about electricity bills. It estimates that if the value of using existing capacity were passed back to customers, household bills could fall 7–9%, or $130 to $170 a year.11 Daily Energy Insider noted that these are projections from the index, not guaranteed savings. Any bill credits would need tariff and regulatory approval.11 Whether existing customers actually come out ahead will depend on how utilities set up these arrangements and whether regulators allow the savings to flow through.11

The timing helps the company. Utility research firm PowerLines reports $23.1 billion in electric and gas rate increase requests in the first three quarters of 2026, and $4.5 billion in the third quarter alone, the highest third quarter on record.16 The economics behind GridCARE's pitch is that selling more power over wires that are already built spreads fixed costs over more sales. That logic is sound. It matches the company's earlier modeling, which found that a flexible 1 GW data center could lower rates 5% across customer classes in a mid-sized utility territory.1 But the company's own caveats make clear that turning that into lower bills is a regulatory decision, not something the software does.

The takeaway

The best reading of the index is that it is a credible order-of-magnitude figure that independent research supports, but it is not a forecast of what will be built. The evidence that the grid has room beyond its rare peak hours is strong, and the 100–120 GW range lines up with the Duke findings. The evidence that this room can be turned into reliable, financeable connections is real but still limited: a few hundred megawatts are actually energizing in Hillsboro, and much of the rest is at the identification or study stage.112

For AI developers trying to get power quickly, flexible interconnection already looks like the fastest route for mid-sized sites. Startups are building businesses on exactly that, and Axios notes that unused capacity is the quickest and most efficient source of new power for hyperscalers.14 For gigawatt-scale campuses, new generation and transmission are still unavoidable. The 120 GW figure could ease some pressure on the U.S. power buildout, but it does not replace it.