EPRI finds data centers cut retail prices 3.5% per doubling, but $6.3B risk looms

Here’s the counterintuitive headline hiding in plain sight: from 2015 to 2024, average retail electricity prices decreased by 3.5% for every doubling of data center capacity, according to new research from the Electric Power Research Institute (EPRI). That finding runs directly against the most common worry in the AI era. Americans have been bracing for the idea that more data centers automatically means higher utility bills.

The EPRI researchers backed the claim with data from the Federal Energy Regulatory Commission (FERC) and retail revenue from the U.S. Energy Information Administration (EIA), and they didn’t stop at the national number. On a statewide level, the decrease was about 6%. The logic is grounded in how electricity pricing works, which is different from most consumer goods markets. In a typical product market, prices track the cost to produce it. Electricity is stranger: prices are largely about cost recovery, meaning fixed costs get spread across the load and the kilowatt hours consumed. As more electricity gets used, fixed costs are divided among more consumers and more energy, which can lower the average retail price.

This also helps explain why “electricity anxiety” may not match what has happened in the recent past. EPRI’s working paper points to economies of scale as a major driver, but it also acknowledges that load increases from higher data center usage can bring more generators online, and many generators are becoming more energy efficient. Think of it like this: if demand rises enough to change the supply mix, the system can become more efficient, and that efficiency can flow through to retail rates. For 2015 to at least 2024, the research suggests data center demand pushed in that direction.

Now for the catch that makes this story matter to boards and finance teams: the paper’s “it lowered prices” trend is explicitly not guaranteed to continue. EPRI coauthor Asa Watten, an EPRI researcher, frames the risk as a potential reversal driven by the difference between planned capacity and actual demand. The fear is not just “AI uses more power.” It’s what happens when the grid builds capacity expecting data center consumption that never materializes at the scale projected.

If that mismatch happens, Watten argues, data center investments could increase prices in a way that they did not in the past. He describes the core mechanism like this: data centers create fixed costs. If customers for those data centers are fewer than expected, then the “denominator” that spreads those fixed costs gets smaller. Fixed costs remain fixed, but fewer people and fewer kilowatt hours absorb them, which can lift retail prices. In plain terms for decision-makers: if you build the power pipeline for a boom that then disappoints, the bill does not disappear. It gets reallocated across remaining load.

This is where the policy and grid planning reality intrudes. Fortune notes that the largest determinant of future electricity prices will be whether the AI buildout meets the hype. Goldman Sachs had projected that the AI infrastructure buildout would increase electricity costs by 6% between 2026 and 2027, and an additional 3% by 2028. Meanwhile, grid operator PJM, described as the largest power grid operator in the country, projected in a report this week that a $6.3 billion increase in consumer electricity costs over the next three years can be mostly attributed to increased data center power demands. In other words, one credible research track says “prices fell historically,” while another planning track warns “costs could rise soon,” depending on how the grid and demand align.

There are already signals that the story may be evolving regionally. The source points out Virginia, the state with the most data centers, where residential electricity prices have increased by more than 13% in the last year, citing data from the EIA. That doesn’t automatically mean the national relationship will flip, but it does show how quickly local rate outcomes can diverge from national averages when infrastructure and retail dynamics get stress-tested.

So what does this mean for executives trying to underwrite risk in an environment full of AI uncertainty? It means the question is not whether data centers use electricity. The question is whether electricity systems can match demand efficiently enough, at the right pace, without spreading fixed costs too thin. It also intersects with investor mood. The article notes a debate around an AI bubble and whether it will pop, with signs that investors are becoming skeptical of AI’s promise. It cites share price declines for Tesla and Alphabet after both companies announced an increase in AI capital expenditures. If investors are rethinking capital intensity, that can bleed into whether hyperscalers actually need every bit of capacity they plan.

That skepticism has a concrete edge from the business side. Mark Cuban warned on the All-In podcast that “a lot of data centers…are going to be turned into pickleball courts” because hyperscalers may assume AI adoption will continue increasing, but AI will become cheaper to use due to increased power efficiency. The premise is that the capacity created through data centers may not be necessary if utilization rises faster than power demand per unit of computing. Whether you agree with that framing or not, it connects directly to Watten’s mechanism: the grid and the fixed cost base are not designed for “maybe” demand.

And yet there is a counterweight. Watten does not like to speculate on what the future of AI holds, but he does emphasize that energy efficiency generally should keep improving. He also points to the possibility that efficiency-increasing electrification like more electric vehicles and electric heat pumps could reduce household energy costs in a way that could happen independently from an AI boom. If that happens “done well,” he says, more electric cars could mean prices are also going down, or at least not going up, with positive spillovers to neighbors.

For peers making capital allocation decisions, the strategic takeaway is blunt: past price declines per unit of data center growth do not automatically protect you from future rate risk. The EPRI findings show how efficiency and load growth can lower retail prices when the system scales well. But PJM’s $6.3 billion consumer-cost forecast and the statewide evidence in places like Virginia highlight how quickly the relationship can tighten, especially when grid investments, fixed costs, and actual demand timelines stop lining up. In the AI era, the best underwriting is not a single number. It is a stress test for mismatch.