Plugging in AI, Part 3: “The data center next door will lower your electric bill”

Two years ago, we wrote about the electricity gauntlet – the collision between AI's power demand and a grid that wasn't ready for it. The buildout has since accelerated faster than even the bulls expected. And public pushback is here.

Something clearly shifted this summer in public sentiment around data centers. Beyond the frenzied headlines, a few major signals stand out:

• On August 3, Governor Abbott ordered ERCOT and the Texas PUC to audit every data center in the interconnection queue – roughly 200 GW of requested demand – pausing development in the hottest market in the US.
• Last month, New York passed the first statewide moratorium on hyperscale facilities; dozens more states have proposed similar bills.
• Lenders now increasingly price community opposition as a formal credit risk.
• In March, Gallup found 7 in 10 Americans oppose data center construction in their area – more opposition than nuclear plants draw.

The reasons are varied but the theme is consistent: constituents don't see themselves as beneficiaries of the AI boom. The complaint cited most often is electricity bills going up.

So what would have to be true for the AI buildout to actually lower rates – not raise them more slowly, but lower them, for every household on the grid?

Against today's headlines that sounds impossible. It isn't. Utilities in Michigan, California, and Texas have already cut residential rates and credited data center load for making it possible. The hard part isn't proving it possible – it's making it the rule rather than the exception. Get that right and every mayor weighing a project has something real to offer their constituents. Get it wrong and "it'll lower your bill" is just a slogan.

Who will foot the electricity bill for the AI buildout?

For the first time in a generation, electricity prices are a national political issue. Rates have risen 42% since 2019, faster than headline inflation – driven by replacing aging equipment, wildfire mitigation, storm hardening, and rising gas prices. Utilities requested a record $31B in rate increases in 2025, double 2024, and another $18.6B in the first half of 2026 alone.

Data centers are generally not a major contributor, but they do create real costs. Where those costs land is a matter of market design. The clearest quantified case is PJM – the largest US grid operator, serving 67 million customers – where the independent market monitor pins $29.4B of recent capacity charges on data centers, almost all of it coming out of ratepayers' pockets.

Notably, those costs came from procuring firm capacity three years out against forecasts that are already proving wrong. Data center builders cannot honor their ratepayer-protection pledge under current rules, no matter how much they'd like to. That's a market-design failure, not an energy consumption problem. Fortunately, the fix is already in motion: PJM has asked FERC to let new large loads connect without their own capacity in exchange for being curtailed first in emergencies – and to stop counting uncovered data center demand in the forecasts that set capacity prices.

In short, electricity bills were climbing long before AI. Data centers are showing up to a system already under strain – and the scale is unprecedented. We're in the early innings of the largest private-sector infrastructure cycle in modern American history: adjusted for inflation and as a share of GDP, roughly the cost of the interstate highway system, the Apollo Program and the Manhattan Project combined.

Demand at that scale will move rates. The only question is which direction, and how far. The goal can’t be just to decouple AI's growth from ratepayer pain – it must be to make ratepayers the beneficiaries. That reduces to two questions: how well can we (1) get more out of the grid we already have, and (2) operate new compute loads flexibly?

(1) Sweat the grid we already have: the utilization opportunity

The U.S. electric system runs below its maximum capability during most hours of the year. That's partly by design – the grid has to survive extreme conditions and unexpected outages, not just operate efficiently under "average conditions". But it's also a consequence of how we've planned it: conservative static assumptions, fixed line ratings, fixed topology, and firm load. The result is an average load factor of just 53%. Our multi-trillion-dollar system is sized for the few peakiest hours of the year and sits half-idle the rest of the time. That creates an enormous opportunity.

Most of what ratepayers pay for is fixed infrastructure. So a large new load that pays its own way and stays off the peak doesn't raise rates – it spreads those costs across more kilowatt-hours and pushes average rates down. PG&E has spelled out the potential economics: 1–2% off residential bills per GW of new large load, and 10%+ from its full 10 GW pipeline.

Utilization is also the fastest solution. Out-building our way there isn't realistic: China added 543 GW of new capacity last year – roughly nine times the US – with its 93 GW of coal and gas alone exceeding everything America deployed from all sources combined. We won't win on brute-force supply; we have to out-optimize.

Grid-enhancing technologies – dynamic line rating, advanced reconductoring, topology optimization – can unlock real headroom at a fraction of the cost of new build, with essentially zero new ratepayer spend. These innovations depend on better modeling and study automation, and today's tools fall short: five-year queues running on decades-old solvers, serialized studies, siloed data between utilities and ISOs.

But speed is only half the fix. ERCOT's full queue holds roughly 474 GW of requests – more than five times the grid's record peak – and close to 90% of it is data centers. Nobody believes all of that is real, which is exactly why Texas is now auditing it by hand. Exelon's data center pipeline fell 40% the moment it began requiring transmission security agreements.

In other words: queue length is not the same thing as demand. Processing unverified requests faster only produces bad plans sooner – and, as PJM showed this year, households can end up paying for the phantom load in the meantime. A months-long manual audit is what you do when you don't have the software.

The north star: a granular, continuously updated grid model. Plans that update as projects firm up or fall away, co-optimizing across them. And the ability to surface latent capacity that static ratings hide.

(2) Reliable flexibility is the new capacity

The most powerful lever may be the design of the data centers themselves. AI could push the average cost per kilowatt-hour down if new loads were configured to operate off-peak.

A widely cited Duke University analysis estimated the existing U.S. power system could absorb roughly 98 GW of additional load if those loads curtailed an average of just 0.5% of the year. It popularized an important idea: not every data center needs firm access to every megawatt in every hour. Some training workloads can be time-shifted, some inference can move geographically, and batteries and on-site generation can cut grid withdrawals during constrained periods. A Johns Hopkins analysis of PJM found that connecting data centers as flexible rather than firm load would cut system costs by $15–16B per year – more than three times the savings from bringing new generation online.

This high-level analysis is only directionally right. A data center does not interconnect to an "average power system"; it interconnects at a specific point on a specific transmission network. That distinction matters enormously. System-wide generation headroom does not necessarily mean that electricity can be delivered to a particular data center without violating transmission reliability criteria. A transmission line hundreds of miles away may overload following the loss of another facility. A different contingency may bind during a period that has little relationship to the system's annual peak. In other words: Generation availability is not the same thing as transmission deliverability.

The real opportunity is optimizing across the whole solution space, transmission network and load together, to find the minimum-cost, minimum-time, reliability-compliant path to each incremental megawatt of AI load. That's what GridAstra has built, and why we believe grid intelligence will be one of the most important innovations of the AI infrastructure era.

Where we're spending time

We'd like to meet teams building:

• Grid intelligence and interconnection software: physics solvers, automated studies, and congestion management that surface capacity already in the ground and compress 5+ year queues.
• Flexibility orchestration: platforms coordinating power, cooling, batteries, and compute in real time – within a facility and across fleets – so non-firm interconnection is something hyperscalers can confidently sign.
• Cost-allocation infrastructure: the tooling that makes "beneficiary pays" work – measuring, verifying, and settling who drove which grid costs, so hyperscalers can credibly fund their own upgrades.
• Programmable grid hardware: solid-state transformers, networking and power electronics riding the 800 VDC transition, plus dynamic line rating, advanced conductors, and power flow control – turning capacity that's fixed into capacity that's controllable in software.

So, back to the opening question. Reliable, flexible interconnection has to become the default rather than the exception, and utilization has to be exhausted before new build gets approved. None of that requires a physics breakthrough – it requires better grid models, orchestrators, and planning tools operating at national scale.

Get that right and we can have both more data centers and lower bills. Get it wrong and we'll spend the next decade auditing queues and passing moratoriums while China adds nine times the power we do.

If you're working on any of this, please reach out – josh@innovationendeavors.com.