Press resource · Following the research
Working press resource. Studies are posted as they come out, whatever they find. Each entry links to the primary source. Last verified August 10, 2026. On deadline: 850.973.7687
I am trying to follow the research wherever it leads. My argument is not that data center demand pushes prices up or down. It is a principle that holds either way: companies should not be allowed to externalize costs while internalizing profits. Cost means total cost: economic, environmental, social, and quality of life. Data centers and utilities are today's example. The principle is the point.
"Both sides in this debate reach for the number that flatters their case. I am less interested in whether the average went up or down than in a question the average cannot answer: when a data center triggers a billion dollars in new capacity, who pays for it. That question survives every study."
This page is where I track credible studies as they come out, whatever they find. When the research is strong, I say so. When it complicates my thinking, I say that too. The goal is an honest read, not a scoreboard.
Entry one · July 2026
Researchers Asa Watten, John Bistline, and Geoffrey Blanford estimate that data centers caused average retail electricity rates in the United States to fall modestly between 2015 and 2024. Their reasoning is economies of scale. A power system carries large fixed costs. When durable demand grows, those costs spread across more kilowatt-hours, so the average unit price falls. Their preferred estimate is that a doubling of data center capacity lowers residential prices by about 3.5 percent. In the states where capacity grew fastest, they find residential rates fell around 6 percent.
Two of the three authors are affiliated with the Electric Power Research Institute. EPRI is an independent, nonprofit research organization. It is not owned by utilities, and I am not suggesting the work is compromised. It does collaborate closely with utilities, and readers deserve to know where research comes from. I include this as disclosure, not to discredit. The study is careful, and I take it seriously.
The mechanism is real. Spreading fixed costs over more usage can lower average prices. That is standard economics, not spin. The analysis is methodical, and its central finding for the 2015 to 2024 period belongs on the table. The study also went looking for something I care about and did not find it. The authors tested whether data centers were shifting costs onto residential customers during this period, and their data show no systematic cost-shifting between customer classes. I take that finding at face value. It is a fair result for the years they studied.
First, this measures a period that is mostly behind us, before the demand that started this whole debate. The study covers 2015 to 2024. ChatGPT did not launch until late 2022, so most of the window predates the AI surge entirely. More important is the kind of growth the study captures. New demand lowers average prices only when the system has slack. EPRI's authors are explicit about the reverse: if anticipated load from data centers materially exceeds realized consumption, "utilities risk overbuilding capacity whose fixed costs would be spread across fewer kWhs, reversing the mechanism we identify." The study describes a grid with room. That is not the grid Florida is about to have. The build-out now underway, running to 2030, is exactly that reverse case: concentrated, fast, and requiring new generation and transmission. So the study does not contradict my concern. It describes the conditions under which its own finding flips, and those are the conditions we are entering.
Second, a falling average is not the same as a fair allocation. The study measures average rates across a state. My argument is about who pays for the specific new capacity a single large customer triggers. Even if the average falls, the question of whether the party that caused a new line or plant paid for it, or passed it to everyone else, is a different question. An average cannot answer it. Only looking at who bears the specific cost can.
Third, the forward-looking research points the other way. A March 2026 Dallas Fed analysis projects that data centers will raise the electricity component of inflation through retail prices, by a small amount now and more by 2030, with slower renewables growth making it worse. The same analysis notes that current connection fees tend to be modest, leaving existing customers on the hook for grid expansion. I do not cite this to claim prices will rise. I cite it to show the evidence genuinely points in more than one direction, which is the honest state of the question right now.
EPRI's own researchers recommend the policy I am arguing for. On supply constraints, they write that price increases "can be avoided entirely by delaying interconnection or allocating the inflated incremental cost of grid expansion to rates paid by data centers." They name the states already doing it: Kansas, Michigan, and Delaware, which have pursued minimum bills or long-term contracts from new large loads. This is worth stating plainly because the study is being cited as a rebuttal to cost allocation, and its discussion section recommends cost allocation.
Whether average rates rose or fell, the question I care about does not change. Did the party that triggered the new capacity pay for it, or did others. That is a question about allocation, and it survives whatever the average does.
Entry two · July 2026
On June 30, 2026, PJM stakeholders endorsed a backstop capacity procurement mechanism. The mechanism is capped at 555 dollars per megawatt-day. The proposal was submitted by the Data Center Coalition together with a coalition of utilities and cooperatives. That group includes Dominion, Duke, Exelon, PPL, Duquesne Light, Buckeye Power, East Kentucky Power Cooperative, Old Dominion Electric Cooperative, and Southern Maryland Electric Cooperative.
The cap sits at 555 dollars per megawatt-day. PJM's most recent capacity auction cleared at 333.44 dollars per megawatt-day. The cap is a ceiling on the backstop's average cost. It is set above the last price the regular auction produced.
The mechanism bills the large loads that drive the need for procurement. The cost is not spread across all ratepayers. Load-serving entities bill those loads directly. This is cost allocation by cause. The party that triggers the new capacity pays for it.
In the same process, PJM stakeholders did not reach consensus on any Connect and Manage curtailment option. No version cleared the vote. PJM staff instead proposed a large load registry to help state regulators set their own load-reduction priorities.
In its July 2026 letter, the PJM Board stated that new large loads should bear the costs they cause, and then named the limit on its own authority. PJM does not have jurisdiction to allocate retail costs directly to individual data centers, so state action is essential, with PJM supporting those efforts through a new Large Load Registry and its settlement and billing data. The largest grid operator in the country, covering thirteen states and the District of Columbia, has now said the allocation decision belongs to state legislatures and state commissions. That is where every entry on this page has been pointing, and it is what makes Florida's October 1 tariff filings a state-level decision with no federal backstop.
This is the allocation question in practice. When a large load triggers new capacity, someone pays for it. The endorsed mechanism assigns that cost to the load that caused it. That is the principle this page keeps coming back to.
Entry three · July 2026
The first two entries on this page look at published research. This one looks at sworn testimony in a proceeding that is still open, and it is the entry where I concede the most.
Duke Energy Florida filed a petition on April 22, 2026 asking the Florida Public Service Commission to approve a large load tariff. Docket 20260064-EI. Intervenors filed testimony opposing it on June 24. Duke filed rebuttal on July 8. Hearing is August 25 and 26.
I read the petition, the intervenor testimony, and the rebuttal. The rebuttal is the strongest utility-side case I have encountered on the question this page is about, and parts of it are better than the arguments I had been answering. What follows is what I think holds up and what I think does not.
Duke's rate design witness, Steven Wishart, makes a point in rebuttal that is more difficult than the version in his direct testimony.
Incremental cost and average embedded cost move in opposite directions over time. A newly built asset carries its highest revenue requirement in year one and declines from there as depreciation reduces the investment the utility earns a return on. Wishart puts that decline at roughly 2.6 percent per year over a 35-year book life. Meanwhile the system's average embedded cost tends to rise, because the utility keeps adding new plant.
The consequence: a customer paying embedded rates across a long contract may contribute more revenue over the life of the assets than incremental pricing would have collected from the facilities originally built to serve it. Comparing a customer's first-year incremental cost against today's embedded rates is a snapshot, and snapshots of two lines moving in opposite directions are not a good guide to where they end up.
This is not a fairness argument. It is an arithmetic argument, and it deserves a numerical answer rather than a principled one. I do not have that answer, and I have not seen one in the record. Wishart offers an illustrative analysis for a 100 MW customer concluding that GSD-1 revenues recover incremental cost over the contract term. Whether that illustration holds for a 1,000 MW customer requiring new generation is exactly what a hearing is for.
I have described the gap between what state cost-of-service review reaches and what federal review reaches. On shared transmission, Wishart's rebuttal complicates that.
He points out that FERC Order No. 2003 considered a "but for" or participant-funding standard, under which the customer whose request triggers an upgrade pays for it, and declined to adopt it as the general rule for non-independent transmission providers like Duke. FERC's stated concern was that identifying which facilities satisfy a "but for" test is inherently subjective and could permit discriminatory cost allocation. FERC generally preserved a crediting approach instead, on the reasoning that network upgrades benefit the system as a whole.
So the federal order most often cited in these arguments does not support direct assignment of shared network costs. It supports direct assignment of customer-specific interconnection facilities, which is what Duke already proposes at 100 percent, paid in advance.
Wishart also surveyed large load tariffs at AEP Ohio, Dominion Energy Virginia, Evergy, Wisconsin Electric, Ameren Missouri, Indiana Michigan Power, Consumers Energy, Santee Cooper, and Northwestern Energy. By his account, none directly assigns shared transmission network upgrades on a "but for" basis, and most recover generation through embedded rates. Evergy's incremental generation rider was removed in settlement. Florida Power & Light's Incremental Generation Charge is, in his reading, the outlier rather than the model.
Two things follow. Direct assignment of shared costs is not the national trend, and I should not imply it is. And a distinction matters that I have not always drawn carefully enough: dedicated facilities and shared network facilities are different questions with different answers.
Duke's protection through 2027 is more substantial than I credited.
Benjamin Borsch, who runs resource planning for Duke Energy Florida, testified that the company has no large load customers under signed agreement, none in its Ten-Year Site Plan, nothing in the load forecast suggesting one is imminent, and that a combined cycle unit to serve a very large customer would take six to eight years to site and build.
Matthew Chatelain testified that a 2024 rate settlement freezes base rates through 2027, and that if Duke incurs costs to serve a large load customer during the freeze that existing rates do not cover, those costs fall to shareholders. He committed that Duke will not seek to adjust rates during the settlement period even if that pushes the company below its authorized return band. Duke also committed to filing a large load rate schedule to take effect January 1, 2028, and asked the Commission to write that obligation into its order.
Taken together, that is a real near-term protection, and anyone arguing that Florida households face imminent cost shifting from Duke's tariff should account for it.
The near-term protection is a settlement that expires. It runs through 2027. The question this page asks is about what happens after that, when new capacity is actually built and a rate schedule that does not yet exist determines who pays for it.
Duke's answer is that the next rate case is the right forum, with a full evidentiary record and an actual customer. That is a reasonable regulatory answer. It is also a deferral, and the thing being deferred is the specific question the statute asks. Section 13.08 of the proposed tariff states that large load customers take service under existing GSD-1 or GSDT-1 rates until a new schedule is filed and approved. The tariff establishes the framework now and sets the price later.
Reasonable people can disagree about whether that sequencing satisfies a statute requiring utilities to file tariffs by October 1, 2026 that reasonably ensure each large load customer bears its own full cost of service. That is the disagreement in front of the Commission, and I do not know how it will rule.
Wishart cites Georgia Power to make a point against strict incremental pricing: large load customers can contribute positively to a utility's embedded costs and reduce rates for everyone else. He is right that they can.
What is striking is how Georgia handled it. Under an approved settlement, incremental revenues from large load customers must reduce a typical residential customer's monthly bill by at least $8.50 during 2029 through 2031. The benefit is written into the settlement as an enforceable number.
That is the whole argument on this page, executed. Georgia's regulator did not assume large loads would lower residential bills. It made the contribution explicit, measurable, and binding. Whether costs get shifted or benefits get shared, the answer should appear somewhere a regulator can enforce it, not in an expectation about how the numbers will work out.
Duke's rebuttal changed my view on three things: the arithmetic of incremental versus embedded cost over an asset's life is a real argument, FERC Order No. 2003 does not support "but for" allocation of shared network costs, and Duke's protection through 2027 is genuine.
It did not change the question. When new capacity gets built to serve a very large customer, someone pays for it, and the answer should be written down where it can be enforced. Georgia wrote it down. Florida takes this up at hearing on August 25 and 26.
I have no view on how the Commission should rule.
Entry four · August 2026
On June 18, 2026, FERC issued show cause orders under Section 206 of the Federal Power Act to all six regional transmission organizations and independent system operators. The PJM order is PJM Interconnection, L.L.C., 195 FERC 61,211 (2026), Docket EL26-67-000. The MISO order is 195 FERC 61,212 (2026).
As ordered, responses are due 60 days from issuance, on or by August 17, 2026. Each grid operator and its transmission owners must either show cause why the tariff remains just and reasonable without the identified reforms, or state what tariff changes would remedy the concerns. Generation adequacy informational reports were due separately on July 20, 2026, and interested entities may respond within 30 days of those filings.
Requests to hold a proceeding in abeyance, for up to 90 days, were due August 3, 2026. A granted request could extend the August 17 date for that region, so August 17 is the deadline as ordered rather than a date that necessarily applies everywhere. Confirm whether an abeyance was granted for a given region before citing the August 17 deadline for it.
This is a docket to watch, not a result to report. The orders set a date and a question. What any operator files, and how FERC rules on it, cannot be read from the orders themselves and is not stated here. What the orders do establish is that the allocation question this page keeps returning to now runs on a federal clock as well as a state one.
Every study on this page links to its primary source, verified before publishing. The FERC show cause response deadline and the PJM and MISO docket numbers were added from primary sources on August 10, 2026. The framing is deliberately economic. The question throughout is cost causation and cost allocation: who triggers the buildout, and who pays for it. Reporters and editors are welcome to reach out for a source interview.
Contact: Dr. Mark R. McNees, Florida State University, mmcnees@fsu.edu, 850.973.7687