One of the things that drew me to Google is the company’s deep engagement in and support for advanced energy technology innovation, market design research, and policy development. I’ll be continuing my own PhD research and staying close to the research community, and hope to share periodic updates here on our work, even if I’m not able to post very often.

In this spirit, I’m excited to share significant new research supported by our team on how large load flexibility can improve affordability, accelerate speed to power, and preserve reliability.

A new paper released yesterday, “Flexible Data Centers: A Faster, More Affordable Path to Power,” is the first publicly available study to combine real utility transmission system data, system-level capacity expansion modeling, and site-level capacity optimization to evaluate how flexibility can accelerate data center interconnections. The study was conducted by Camus Energy (led by Astrid Atkinson), Princeton University ZERO Lab (led by Jesse Jenkins), and encoord.

The findings are compelling. Specifically, the study found that combining flexible grid connections with a “bring-your-own-capacity” (BYOC) model in the country’s largest electricity market (PJM) can:

• Protect affordability: Flexible data centers contributed ~$733 million per gigawatt toward the costs associated with their incremental load, reducing net system cost increase by 96% compared to a scenario with the same volume of inflexible data centers.

• Preserve reliability: Grid power remained available for >99% of hours across all modeled data center sites, with on-site resources dispatched only 40-70 hours annually.

• Accelerate interconnection: This approach shortened the wait for grid power by 3 to 5 years compared to traditional timelines.

In many ways, this serves as a deeper, transmission-aware complement to our Duke University study from earlier this year, this time using a PJM-area utility’s actual transmission SCADA data and network models to evaluate not just generation constraints, but local transmission constraints as well. (For a cliffnotes version, Bloomberg, UtilityDive, and Data Center Richness covered the study yesterday.)

The conversation continues next week, when new modeling on large load flexibility in PJM by Duke University’s GRACE Lab (led by Prof. Dalia Patiño-Echeverri) will be presented in a webinar, alongside insights from EPRI, a former FERC commissioner, and a Google demand response expert (Tues, Dec. 9, 11am-12pm ET).

I hope you find this work valuable and look forward to ongoing collaboration with many of you in 2026!

It’s surprisingly challenging to distill these ideas into a few hundred words, but the crux is this:

The new electricity demand doesn’t automatically raise bills. A recent study showed that states where electricity use grew fastest often saw average household prices fall, while states with flat or declining use saw prices rise. The impact depends, in significant part, on how cleverly the existing grid is managed — and whether that new energy demand occurs at times when the grid is already strained.

Much of the grid sits idle most of the time. On a typical day, only about half of its capacity is in use, and the country’s most efficient gas plants run less than 60 percent of the time. Because the system must be able to handle its busiest hours — those moments during a heat wave or a cold snap when almost every air-conditioner, heater, factory and data center is running at once — those few peaks in demand end up determining everything, including the size of our power plants, the thickness of our wires and the bills we pay. Trim use during those peaks, and the cost savings ripple everywhere.

The solution is simple: Ask the largest power users to draw a little less from the grid during the limited hours when it’s most strained. They can do that by running briefly on batteries, using electricity generated on site or shifting workloads. Average Americans would never notice — emails would still send, chatbots would still respond and websites would still load — but the grid would breathe a little easier.

Inevitably, a significant amount of nuance has to be jettisoned for a piece like this, but I was glad to be able to incorporate a nod to the potential for data centers and other large loads to procure flexibility from off-site customers:

Some data centers that specialize in instant, on-demand services — like web search, streaming or chat — may find it impractical to dial back during peak hours. But even then, similar benefits can be achieved if data centers pay residential and business electricity consumers to shift their use during extreme peak hours with smart thermostats, networked appliances and battery systems that can adjust automatically. That could provide the same relief for the system while data centers compensate other users directly.

Of course, any discussion of this topic today would be incomplete without mentioning the U.S. Department of Energy’s recent Advanced Notice of Proposed Rulemaking (ANOPR), which effectively validates this approach:

The federal government is now validating this approach. In an extraordinary move, on Oct. 23 the Department of Energy directed the Federal Energy Regulatory Commission to propose a rule that would allow new data centers and factories to plug into the grid years sooner if they agree to lower their use during the rare hours when supplies run tight. The proposal has drawn strikingly broad, bipartisan support.

For most of you, this will read as a 101-level overview. Still, I hope it helps make these ideas — especially around system utilization and large load flexibility — accessible to a broader audience. If nothing else, we have this artistic depiction of “flexible” power infrastructure that ran alongside the piece:

Today, I presented on a webinar hosted by the National Association of State Energy Officials (NASEO) and the National Association of Regulatory Utility Commissioners (NARUC) on data center electricity demand and flexible load integration. As part of my presentation, I shared slides that I recently created to address what appears to be a significant gap between belief and reality on this topic.

There’s a widespread perception that data centers operate 24/7/365 at near-maximum demand, consuming electricity at a constant clip with little variability. That perception can drive utility planning, investment, and regulatory decisions. For example:

• Duke Energy, 2024 testimony to the NC Utilities Commission: "[Data centers] are operating at a consistent load factor 365 days a year, seven days a week, 24 hours."

• Resources for the Future, 2024: "They run very consistently. Let's say it has a 90 percent load factor."

• ARES Wealth Management Solutions, 2024: "Data center capacity calculated assuming a 90% load factor."

• Energy Futures Group, 2025: "A large load data center, since it is constantly active, will have a high load factor of 90-100%."

But this framing blends together distinct metrics — like load factor, capacity utilization, and uptime — that measure very different things. In doing so, it can make data center power demand appear more constant and closer to full capacity than it actually is. As a starting point, it’s critical to distinguish between load factor and capacity utilization rate (see Table 1).

Table 1: Terminology for Data Center Power Systems Planning

Untangling Load Factor, Utilization, and Uptime

Here are a few common points of confusion:

• Load Factor ≠ Utilization Rate. Load factor is the ratio of average demand to realized peak demand. But if a facility’s realized peak demand is only 80% of that facility’s rated capacity, and its load factor is 90%, then its true capacity utilization rate is only 72%.

• Load Factor ≠ Server Uptime. A facility with high electrical load factor may still have servers operating well below their compute capacity.

• Server Uptime ≠ "Five Nines" Uptime. The industry standard uptime guarantee (99.999%) is about customer-facing availability, not actual server utilization.

• Non-IT Load Variability Matters. Cooling and other infrastructure loads fluctuate and contribute to apparent consistency, masking underlying variation in IT load.

• PUE ≠ IUE. The predominant energy-related metric reported by the data century industry is Power Usage Effectiveness (PUE), but unfortunately, PUE indicates nothing about a given facility’s capacity utilization rate, which in the data center industry is sometimes called “Infrastructure Usage Effectiveness (IUE).”

These nuances matter, because they affect how we plan for data center power needs, how we size infrastructure, and how we assess the potential for flexible demand. As just one example, if you’re a utility regulator and your utility says it anticipates a new 500 MW data center will be interconnected in the next year, the only way to calculate that data center’s expected annual electricity consumption is if you know its expected capacity utilization rate (i.e., annual electricity consumption = annual capacity utilization rate * nameplate MW rating * 8760).

What the Data Say (and Don’t Say)

As Lawrence Berkeley National Lab noted in its 2024 U.S. Data Center Energy Usage Report:

The lack of primary performance and utilization data indicates that much greater transparency is needed around data centers. Very few companies report actual data center electricity use and virtually none report it in context of IT characteristics such as compute capacities, average system configurations, and workload types.

One of the few figures in the report that attempts to quantify utilization shows operational time for servers broken down by data center type — a rare attempt to characterize real-world data center operations (see Figure 1). But even this is only a proxy, not a true measure of capacity utilization.

Figure 1: Operational time of servers given data center type

Source: Shehabi, A., et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, Berkeley, California. LBNL-2001637

In the absence of clear public data, estimates vary widely (see Table 2). Some sources reference load factor; others conflate it with utilization or uptime. And as I’ve found, if you ask 10 people to explain data center capacity utilization rates, you’ll get 10 slightly different answers.

Table 2: Data Center Average Capacity Utilization Rate Assumption

Why Might Utilization Be Lower Than Expected?

There are several potential explanations for why data centers, including those supporting AI workloads, may operate well below maximum utilization. These factors highlight the operational realities that challenge the "always-maxed" myth:

• Maintenance and Redundancy. Many facilities are overbuilt to ensure uptime, with servers periodically taken offline for routine maintenance, software upgrades, or hardware replacements. Redundancy (e.g., backup systems or spare capacity) means not all resources are active simultaneously, creating built-in underutilization to prevent single points of failure.

• Nameplate ≠ Real-World Usage. The “nameplate” power rating of a chip, often based on its thermal design power (TDP), reflects the maximum theoretical power it could draw under extreme, fully saturated workloads. These ratings are used to inform the electrical system design, but in practice, chips rarely operate near those levels. Reaching 100% of nameplate power across an entire rack or cluster is not only unusual, it's extremely difficult to sustain, and often not desirable given reliability and thermal constraints.

• Inconsistent Workloads. Not every program or task in a data center demands full processing power around the clock. Usage levels vary based on factors like who the customers are, where they're located, the time of day, and the specific type of work being done. For instance, AI training may maintain a steady but occasionally bursty pattern at 80% server utilization, while AI inference (handling real-time user requests, like chatbots) may frequently drop to 40-60% due to fluctuating demand.

• Oversizing and Ramp-Up. Data centers may request and design for power capacity well above their expected near-term needs to maintain flexibility for adding servers or accommodating future workloads without waiting for another interconnection process. In practice, new facilities often operate well below designed capacity during the initial ramp-up period as additional servers are installed and workloads increase, sometimes for extended periods.

• Hardware Fragility at Scale. Hyperscale AI training clusters push GPUs to their physical limits, and fault rates rise accordingly. During a 54-day training of Meta’s Llama 3 405B model, using up to 16,000 NVIDIA H100 GPUs, Meta reported 466 job interruptions, with nearly 80% linked to hardware faults. GPU issues were the most common, including thermal stress and memory failure. Industry analyses suggest this could equate to an annualized GPU failure rate of ~9%, with cumulative risk exceeding 25% over three years. In large clusters, even a 1% failure rate can mean dozens of issues per day.

• Synchronous Sensitivity. In AI training processes that require all GPUs to work in perfect coordination, a failure in even one GPU can halt the entire operation, forcing a restart. Some companies employ diagnostic tools to quickly identify and address these issues, reducing overall downtime, but nevertheless the ripple effects from a single hardware glitch can still significantly disrupt cluster-wide performance and reduce average utilization.

Toward Better Data and Evidence-Based Planning

Understanding how data centers actually operate — including how much and when their capacity can be expected to be utilized — is foundational to how we plan our electricity infrastructure. Misunderstanding utilization rates and load shapes can lead to overbuilding and unnecessary investment, which ultimately raises costs for everyone (see: “Data Centers Could Make or Break Electricity Affordability”).

In fairness, not even the most sophisticated data center owner-operators necessarily know what their utilization rates and load shapes will look like in the face of uncertain market demand for AI services and a rapidly shifting competitive landscape. Amid such uncertainty, their preference is generally to maintain maximal optionality — overprovisioning capacity to handle potential surges or pivots. While this is an understandable commercial strategy, the practical result is that it contributes to power system overbuild, which is especially concerning to the extent that any costs are socialized to other ratepayers. Even if they're not, allocating scarce utility resources to support this headroom can divert them from other system needs, like reliability upgrades and proactive transmission expansion.

We’re still flying partly blind, and until we improve transparency — whether through benchmarking programs, voluntary reporting, PUC requirements, or academic-industry partnerships — policymakers and planners will be left to make consequential decisions based on assumptions rather than evidence.

##

Specifically, Google announced formal contracts with American Electric Power’s Indiana Michigan Power and Tennessee Valley Authority (TVA), committing to provide flexibility from its data centers to support resource adequacy planning. According to Google:

These capabilities… allow large electricity loads like data centers to be interconnected more quickly, helps reduce the need to build new transmission and power plants, and helps grid operators more effectively and efficiently manage power grids… [this is] an important step to enable larger scale demand flexibility, delivering grid reliability and cost-saving benefits in the places where these capabilities are deployed.

The emphasis on cost-savings is notable amid mounting concerns over electricity rate hikes, as highlighted by last week's Washington Post story, “The AI explosion means millions are paying more for electricity.” As I recently wrote here, if data centers are planned more intelligently, they could become one of our best defenses against rising rates by spreading existing fixed costs across more load.

Several factors make this a significant announcement, in my view:

• First-of-its-kind US agreements: These represent the first known contracts between a hyperscaler and U.S. utilities around AI data center flexibility. (There may be a select number of recent contracts with similar provisions, but they have not been publicized.)

• From operational flexibility to planning flexibility: This is the first known instance where AI data center flexibility appears to be explicitly integrated into utility planning processes, shifting from merely reacting in real-time to proactively shaping system design. Google is extending this flexibility beyond day-to-day operations — where they've already demonstrated success in regions like Oregon, Nebraska, and Taiwan — to the planning realm. This means committing to curtailment during peak periods to support resource adequacy, potentially reducing the need for new generation and transmission builds.

• Inclusion of machine learning workloads: Google notes that the flexibility provided includes machine learning workloads for the first time, as opposed to conventional CPU-oriented workloads. This marks an evolution from their previous focus on non-urgent tasks like video encoding.

• Definitive, long-term contracts: These are formal agreements over extended time periods with explicit commitments, rather than pilot projects, demonstrations, short-term agreements, or other informal arrangements.

• Time-bound flexibility: The agreements note that load shifting occurs during "certain hours or times of the year," underscoring the targeted and time-limited nature of the arrangement that aligns with grid needs without regular disruption.

Indiana Michigan Power (I&M) explained the agreement this way:

This offering will be used to reduce I&M's peak load in times of high energy demand… As a large load customer, Google's participation in this program boosts Indiana Michigan Power's ability to manage electricity demand during peak times. This helps lower overall energy costs, delivering savings to all I&M customers.

Interestingly, Google’s release and Michael Terrell’s post pointed to our Rethinking Load Growth paper, with Terrell noting:

Grid operators typically only utilize about 50% of available generating capacity. This is by design: they must plan and build enough power plants to meet the highest demand at any given time, but peak demand only occurs during a small fraction of the hours in a year. Research shows that even a small amount of flexibility for large energy loads, like ML, during peak times can reduce the need to build new power plants while accommodating new energy loads much faster - making more efficient use of existing generation. See this study for example.

It’s also important to point out the caveats. Google's release stated, "Data center demand flexibility is still in the early stages and will only be available at certain locations,” and I&M noted that the agreement requires approval by the Indiana Utility Regulatory Commission.

That said, such cautious framing isn't surprising, especially in the absence of clear incentives or regulatory guidelines in other jurisdictions (which may be changing, per Texas’ SB6 and SPP’s proposed nonfirm transmission services). What matters is that the capability is now proven — not only by Google, but emerging companies like Emerald AI and others (disclosure: I serve as an advisor to Emerald AI) — and within reach by other data center owner-operators, as Jigar Shah suggested today:

Even more importantly, Google’s flexibility agreements establish official precedent for definitive long-term contractual arrangements between data centers, utilities, and ISO/RTOs to "enable larger scale demand flexibility,” per Google’s release. Or as Terrell put it: "we hope this capability can enable the growth of AI to benefit more people while better optimizing the electricity system for everyone."

##

Steve Levitas is a former member of the NC Utilities Commission, appointed by Governor Josh Stein, and previously served as SVP for Regulatory and Government Affairs at Pine Gate Renewables. Tyler Norris is a J.B. Duke Fellow at Duke University, where his PhD research focuses on electric power systems.

—

It’s no secret that an anti-solar ideology has taken hold over segments of the American political right – so much so that Elon Musk himself now regularly uses his X account of 223 million followers to push back, echoed by GOP moderates like Senator Thom Tillis.

Yet even the more hardline voices of this movement have rarely pushed to terminate mature solar projects with fully executed contracts. To date, anti-solar advocacy has largely targeted earlier-stage projects seeking zoning approval at the county level.

But in an extraordinary development, a new commissioner on the North Carolina Utilities Commission (NCUC), Donald van der Vaart, broke with his fellow commissioners and voted to deny a key permit for a major, late-stage solar project. This is not just any project: it's a 275 MW solar facility procured by Duke Energy Progress under a well-established procurement program created by the Commission, at the direction of the state legislature, and fully compliant with an NCUC-approved resource plan. It would become the largest solar farm in the Carolinas.

Specifically, van der Vaart called for rejecting the project’s application for a Certificate of Public Convenience and Necessity (CPCN). CPCN applications for solar farms are rarely controversial, and it’s extremely unusual for them to be rejected; we are unaware of a single past instance of rejection for a project intended for bidding into Duke Energy’s solar procurement program, let alone an already-procured project.

According to recent polling by Conservatives for Clean Energy, an overwhelming majority of North Carolinians support the development of new solar and wind resources to help meet the state’s growing energy needs (see Figure 1). They recognize that a diversity of energy supply will allow the state to achieve a balance of reliability and affordability to address the inevitable need to retire aging, inefficient and dirty coal-fired power plants. And apart from their environmental and public health benefits, the zero-fuel cost of solar and wind provides an important hedge against the risk of natural gas and coal price volatility and supply constraints.

Figure 1: Poll of NC Registered Voters, March 2025

This all-of-the above approach was reflected in the order issued by the NCUC last fall directing Duke Energy to develop a mix of new solar, wind, battery storage, natural gas, and nuclear resources while implementing an orderly retirement of its coal plants, which on average are over fifty years old. Although the Commission’s plan was informed by a legislative directive to take “all reasonable steps” to achieve a 70% reduction in Duke’s North Carolina carbon emissions by 2030, the South Carolina Public Service Commission adopted a virtually identical plan even though that state has no carbon-reduction mandate. That’s because it found that a plan based on resource diversity was the most reasonable and prudent way to meet growing energy demand while protecting ratepayers.

In contrast to this reasonable, balanced approach, van der Vaart’s dissent represents a stark example of what some have called an “anti-solar mind virus.” His rationale relies on factual inaccuracies and misleading assertions, and his factual findings are unsupported by the official record. For example, he claims that solar failed to perform as expected during the June 23-25 heatwave – which is demonstrably false, as evidenced by the Energy Information Administration’s (EIA) Hourly Grid Monitor. In fact, peak solar production on June 23 was the second-highest recorded between June 20 and July 20, and on June 24 and 25 it was significantly above average over this period for both Duke Energy Progress and Duke Energy Carolinas (see Figure 2; van der Vaart also erroneously and sensationally asserts that NC’s grid experienced “a near-miss blackout” during the heat wave).

Figure 2: Electricity Generation in Duke Energy Progress

He further justifies his opposition by relying on a bill passed this summer by the NC General Assembly (SB 266) that would remove the state’s interim decarbonization standard. However, that bill was vetoed by Governor Stein and has no force or effect unless the Governor’s veto is overridden by the legislature [Update 7/29/25: today the NC General Assembly over-rode the Governor’s veto on SB 266]. Even if an override occurs, there is no evidence to suggest that would eliminate the need for solar procurement, let alone from one of the region’s least-cost projects. Such an asset would almost certainly be procured even without the state’s interim decarbonization standard, but a material delay could result in forfeiting federal tax credits, requiring ratepayers to pay more for the same power plant.

Finally, van der Vaart makes the erroneous claim that “solar facilities without battery storage increase reliance on dispatchable resources,” ignoring the fact that battery storage does not need to be paired on-site with solar plants to store the output of those facilities in support of system reliability – which is partly why the Commission previously approved Duke Energy’s own proposal to procure significant volumes of standalone battery storage.

None of this is to suggest that every solar project should automatically be approved or procured. On the contrary, competitive procurement within a Commission-approved resource plan exists specifically to identify the most cost-effective projects while ensuring system reliability. Nor does this mean solar is a perfect resource – it has obvious limitations, particularly for supporting resource adequacy during winter morning peaks. But it is well understood that solar production is strongly correlated with summer heat, since in the vast majority of cases, the hottest summer days are among the sunniest.

Nowhere is this clearer than in Texas, where solar output routinely surges during extreme heat events, precisely when electricity demand spikes due to widespread air conditioning use. During ERCOT’s record-breaking 2024 heatwave, when ERCOT set an all-time peak demand of nearly 86,000 MW on August 20, solar generation reached almost 20,800 MW – contributing roughly a quarter of midday power and keeping wholesale prices in check (see Figure 3). As the sun set and solar faded, battery storage picked up the slack, discharging a record 3,927 MW to help meet evening net load, which peaked near 71,000 MW.

Figure 3: ERCOT Peak Demand Record on August 20, 2024

As the CEO of its grid operator said at the time, “Over the last year, we’ve seen significant additions of energy storage resources, solar resources and wind resources, with a few additions also on the thermal side, the gas side. All of that has helped to contribute to less scarcity conditions during the peak periods of summer, like we experienced last year.” As a result, in 2025 ERCOT predicted “the lowest chance of power supply emergencies in years.”

Fortunately, as noted, van der Vaart’s fellow bipartisan commissioners didn’t buy his arguments and made the appropriate decision to grant the project’s CPCN, which can now proceed with construction to support the state’s electricity needs. Nevertheless, this event vividly illustrates a disturbing trend in ideologically motivated efforts to obstruct infrastructure development by exploiting local, state, and even federal permitting powers. If left unchecked, such tactics will exacerbate the country’s already-extreme supply constraints and limit our ability to accommodate new electricity demand.

##

In its latest auction, the largest U.S. grid operator cleared $329 per megawatt-day across its entire footprint — the FERC-approved price cap, and a 22% increase over last year’s already record-setting prices.

The result: a $16.1 billion capacity procurement, up 10% from the prior auction and more than 7 times the $2.2 billion cost of two years ago.

While capacity costs make up only a portion of total electric bills, PJM projects these price increases will lead to 1.5% to 5% bill increases for some ratepayers. And had it not been for an agreement to cap prices brokered with Pennsylvania Governor Josh Shapiro, PJM says the clearing price would have hit nearly $389/MW-day.

Figure 1: PJM’s Record Capacity Auction ($ billions)

According to PJM’s own Independent Market Monitor, the primary driver isn’t general demand growth or widespread resource retirements — it’s explosive data center expansion (emphasis added):

The basic conclusion of this analysis is that data center load growth is the primary reason for recent and expected capacity market conditions, including total forecast load growth, the tight supply and demand balance, and high prices.

… It is misleading to assert that the capacity market results are simply just a reflection of supply and demand. The current conditions are not the result of organic load growth. The current conditions in the capacity market are almost entirely the result of large load additions from data centers, both actual historical and forecast. The growth in data center load and the expected future growth in data center load are unique and unprecedented and uncertain and require a different approach than simply asserting that it is just supply and demand.

Specifically, the IMM found that data center load by itself resulted in a 174% increase in PJM’s last capacity auction clearing price (2025/2026), totaling $9.3 billion.

A Crisis for Affordability?

To be sure, data center load growth is by no means the sole driver of PJM’s capacity price spike. As I noted in my March 2025 Congressional testimony, PJM’s current market design imposes high barriers to new entry for generators, including a lengthy and costly interconnection process. These constraints, combined with tight reserve margins and generator retirements, have left the market vulnerable to sharp price swings even with moderate demand growth.

Nevertheless, data centers are a major contributing factor. An analysis by Bain & Company in October 2024 found that capital investments necessary to serve U.S. data center growth over the next decade will require utilities to generate 10% to 19% in additional revenue each year than previously forecast, contributing to “extraordinary” growth in electricity rates.

Similarly, an independent study in 2024 for Virginia’s Joint Legislative Audit and Review Commission (JLARC) found that a typical Dominion Energy residential customer could see up to $37/month in higher generation and transmission costs by 2040 due to data centers (real dollars). The report noted:

data centers’ increased energy demand will likely increase system costs for all customers, including non-data center customers, for several reasons. A large amount of new generation and transmission will need to be built that would not otherwise be built, creating fixed costs that utilities will need to recover. It will be difficult to supply enough energy to keep pace with growing data center demand, so energy prices are likely to increase for all customers.

For their part, state officials in Maryland are convinced that data centers are responsible for major new transmission costs. Here is Maryland’s People’s Counsel, David S. Lapp — appointed by the state Attorney General with advice and consent from the state Senate — writing earlier this month in The Baltimore Sun:

The real culprit behind any “supply” issue facing Maryland is the massive projections for the power needs of data centers outside of Maryland. That’s the demand problem that PJM is focused on, which is not Maryland’s problem to solve. If PJM doesn’t solve it correctly, it could impact Maryland, but it would be impractical and economically implausible for Marylanders to bear responsibility for that projected demand growth, not to mention fundamentally unfair to Maryland utility customers.

Unfortunately, this is arriving as electricity costs are already climbing. According to a new analysis by Power Lines, U.S. utilities requested or received $29 billion in rate increases in the first half of 2025 alone — more than double the same period last year and a record year for utility rate increases.

As Robinson Meyer warns in a new piece, “The Electricity Affordability Crisis Is Coming,” “virtually every trend is setting us up for electricity price spikes… Going into 2028, [policymakers] will need an actual plan to stabilize or cut electricity costs.” In testimony to the U.S. Senate Energy & Natural Resources Committee yesterday, Rob Gramlich identified eight separate causes of rising bills, including aging grid infrastructure, scarce generation capacity, severe weather, supply chain constraints, tariffs, termination of federal tax credits, permitting delays, and potential actions forcing uneconomic power plants to run.

Data Centers as Grid Rate‑Relief Valves

But here’s the twist: if we plan them intelligently, the very data centers driving costs today could become one of our best defenses against rising rates.

The reason is this: whether a new load triggers additional costs isn’t primarily about how much electricity it consumes, but when it draws power. Capacity markets, transmission upgrades, and even energy prices are driven by peak demand and congestion at specific times and locations. Give a facility the ability to trim consumption for a limited number of hours each year, and it can often sidestep expensive new generation and wires.

In other words, a small amount of well-timed flexibility — the ability to reduce load for a few dozen hours per year when the grid is most strained — can significantly reduce system-wide costs.

One of the clearest illustrations of this dynamic is the load duration curve, which Constellation Energy adapted from our report for its quarterly earnings call in May (below). What it demonstrates is that our existing power systems are built to serve extreme swings in demand during a small number of hours — specifically, more than 10% of the system is built to serve 35 hours/year of extreme peak load, on average.

Figure 2: Load Duration Curve for US RTO/ISOs, 2016–2024

Think of the power system like a bus that already runs its route whether the seats are full or half‑empty. Most of the costs — the bus, driver, and maintenance — don’t materially change with one more passenger. Flexible loads are the riders who hop on when there’s room, not forcing the operator to buy another bus. They fill empty seats, boosting total kilowatt‑hours served while leaving fixed costs virtually unchanged—so each kWh carries a smaller share of the bill.

That’s the premise of Rethinking Load Growth, where our first-order analysis found that PJM alone could accommodate 13 GW of new load without triggering major new capacity investments, provided the new load is flexible for 0.25% of its maximum potential annual uptime. Most curtailment events would last between 2-4 hours and would entail partial curtailment of the facilities. As we noted:

By strategically timing or curtailing demand, these flexible loads can minimize their impact on peak periods. In doing so, they help existing customers by improving the overall utilization rate—thereby lowering the per-unit cost of electricity—and reduce the likelihood that expensive new peaking plants or network expansions may be needed.

To offer a simplified estimate of potential cost savings from new data center load, consider that Lawrence Berkeley National Lab’s 2024 Data Center Energy Usage report forecasted up to 580 TWh of U.S. data center electricity consumption by 2028, on the high end. Roughly one‑third of our electric bills represent the marginal cost of generating the next kilowatt‑hour, on average, while about two‑thirds pays for fixed costs, including power plants, grid costs, and overhead that doesn’t change when the lights are off. At an average retail electricity rate of 13.2 cents/kWh, 580 TWh of data center load would generate approximately $51 billion in annual contributions toward existing fixed costs of the U.S. power system.

A New Paradigm for Load Planning

Echoing this point, experts at Brattle and Clean Air Task Force note in a new report this week that large load flexibility “can address both resource adequacy and transmission adequacy concerns by quickly connecting loads to power without requiring extensive transmission or generation investments and without raising concerns over unfair cost allocation.”

This is easier said than done, in part because U.S. utilities have rarely if ever planned the power system for flexible loads, building on a century of load planning premised on a “duty to serve” and cost-of-service ratemaking (COSR). As the North American Electric Reliability Corporation (NERC) Large Loads Task Force states in a just-released report:

Large loads interconnecting to the grid typically request firm utility service during the interconnection process. This means that the utility is required to provide the transmission infrastructure capable of serving the load’s peak demand request at all times. Concepts have been explored [23] in some areas around mandatory load curtailments by the ISO or utility during stressful grid conditions, which might allow for TPs to assume a lower peak demand for the load and potentially reduce the transmission buildout exclusively needed to support the load. Utility controllability of loads is usually not accounted for at the transmission planning stage. However, many loads are still expected to obtain firm transmission service to ensure reliable electric service delivery, necessitating transmission buildouts to support the load and an increase in energy supply during both normal and emergency system operations.

That’s why emerging efforts like PG&E’s FlexConnect program, Southwest Power Pool’s proposed non-firm transmission service (illustrated below), PJM’s proposed co-location service options, Texas’ new SB6 law, EPRI’s DCFlex initiative, Google’s demand response participation, and startups like Verrus, Emerald AI, and Gridcare are so significant — among others.

As Google’s Brian George explained at FERC’s technical conference on resource adequacy last month (emphasis added):

“We have reduced or shifted our loads in response to system events. We also have the ability to shift loads across regions based on demand at a given time. The trick that we’re trying to figure out now alongside those like EPRI and everyone else, is how do we shift this from an operational approach to more of a planning approach. I’ve heard the Duke study referenced several times today and I think there’s a lot of opportunity there. We have the operational capability, but how do we transition that to more of a planning approach – and that encompasses things like backup generation and demand response."

SPP’s Proposed Nonfirm Transmission Services

Answering Rep. Castor’s Question

“Will this help us address the growing concern with additional costs to consumers — to other ratepayers — as AI data centers demand more energy?” That was the verbatim question posed to me by Ranking Member Rep.  Kathy Castor on the House Energy & Commerce Energy Subcommittee during my Congressional testimony.

Load flexibility isn’t a silver bullet, and there are a variety of important measures regulators can implement to protect ratepayers and rationalize the large load interconnection process, as outlined in Johns Hopkins University’s recent data center playbook and Harvard Electricity Law Initiative’s March 2025 paper.

But on Rep. Castor’s core concern, my answer then and now is yes: integrating data centers on flexible terms doesn’t just mitigate rate shocks, it can spread the costs of the power system we’ve already built across more kilowatt‑hours and offset upward pressure on bills.

In short, the same server farms now blamed for record capacity prices can become the grid’s pressure‑relief valve — provided that grid planners incorporate flexibility from the start and hold operators to it.

##

Related resources:

• “From Bottlenecks to Breakthroughs: Rewriting the Grid Planning Playbook in the Southwest Power Pool and Texas.” Arushi Sharma Frank. Center for Strategic & International Studies. https://www.csis.org/analysis/bottlenecks-breakthroughs-rewriting-grid-planning-playbook-southwest-power-pool-and-texas

• “‘AI Fast Lanes’ for an Electricity System to Meet the AI Moment.” Costa Samaras. https://www.cmu.edu/work-that-matters/energy-innovation/ai-fast-lanes-electricity-system-meet-ai-moment

• “PG&E tries to prove that a big utility can innovate.” Volts with David Roberts. July 2025. Available here.

• “Orchestrating Power for Data Centers with Brian Janous of Cloverleaf.” Build, Repeat - A Paces Podcast. Available here.

• “The Key to Unlocking More Data Center Power.” Data Center Richness. May 2025. https://datacenter.libsyn.com/the-key-to-unlocking-more-data-center-power

• “Can we add dozens of giant new data centres to the electricity grid?” The Energy Gang. May 2025. https://www.woodmac.com/podcasts/the-energy-gang/can-we-add-giant-new-data-centers-to-the-grid/

• “How Load Flexibility Could Unlock Energy Abundance.” Energy Capital with Doug Lewin. April 2024. Available here.

This past week, the North Carolina State Senate passed S.B. 261, seeking to overturn a critical component of the state's energy framework by repealing its interim electricity decarbonization standard. The bill also grants Duke Energy a major concession, allowing it to increase electricity rates charged to consumers to secure profits before new baseload power plants become operational.

A central justification for the bill offered by its lead sponsor is that it will support nuclear power development. As Duke Energy’s former North Carolina president and current Senate Majority Leader Paul Newton stated at the legislative hearing, “[Nuclear] is what we need more of on our grid… The model will pick natural gas and nuclear resources every time when allowed to do so, because they have the least cost.”

However, the available evidence suggests the opposite. Analysis by North Carolina's consumer advocacy agency, the Public Staff, concluded the bill would slash anticipated nuclear capacity additions by 50% by 2035, in favor of more natural gas power plants (see Figure 1 below).

Why does this matter beyond North Carolina? Because among all U.S. regions, the Carolinas are uniquely positioned to support a national revival of nuclear power after decades of stagnation, owing to a more nuclear-favorable regulatory environment. However, despite its promise, new nuclear remains expensive in the U.S., in significant part due to an eroded supply chain, depleted workforce, and lost construction expertise. Georgia’s recent completion of two AP1000 reactors at Vogtle was an important step toward rebuilding these capabilities, albeit at a high cost to ratepayers. But without consistent state and federal support, nuclear development is likely to stall.

North Carolina’s interim decarbonization standard, established by bipartisan compromise under H.B. 951 in 2021, required a 70% reduction in electricity sector CO₂ emissions by the early 2030s. To meet this target, Duke Energy proposed, and the North Carolina Utilities Commission approved, plans for 600 MW of new nuclear capacity by 2034. This made North Carolina perhaps the only state besides Georgia with a utility commission-approved resource plan explicitly including new nuclear.

Instead of reinforcing this progress, S.B. 261 undermines the very policy enabling medium-term nuclear growth. As nuclear expert Jim Hopf noted, this is “another example of how new nuclear construction depends on decarbonization policies. Things like repealing the IRA's clean power supports, and this example of a state delaying decarb targets, make it hard to economically justify new nuclear construction.” Indeed, as I recently testified to the U.S. House Energy & Commerce Energy Subcommittee, preserving the Inflation Reduction Act’s technology-neutral tax credits and the Department of Energy’s Loan Programs Office is essential to support nuclear growth (my written testimony is here, and the relevant video segment is here at 1:26:23).

Does this mean no new nuclear reactors will ever be built again in the Carolinas? Not necessarily. The bill leaves intact the state’s standard requiring a carbon-neutral electric power sector by 2050, which may still drive nuclear development in the 2040s. But repealing the interim standard significantly weakens the rationale for nuclear deployment in the 2030s.

Perhaps even more troubling is the opaque and rushed manner in which S.B. 261 was crafted. The original 2021 law resulted from a multi-year process involving broad stakeholder engagement, independent analysis, and public debate, culminating in a major bipartisan compromise. Two extensive resource planning proceedings followed, involving thousands of pages of comments and testimony, hundreds of hours of hearings, and months of work by the utility commission.

In contrast, S.B. 261 was introduced and passed by the NC Senate within 72 hours, without stakeholder input, bipartisan engagement, or publicly accessible analysis. Remarkably, Senate leadership privately commissioned modeling from the Public Staff to justify their position, yet this taxpayer-funded analysis remains hidden from public scrutiny.

Such hasty policymaking undermines business confidence and investment stability – precisely the conditions required to attract private capital to long-lead, capital-intensive nuclear projects. If state leaders can overturn key policies without transparency or deliberation, investors will naturally question the state’s regulatory reliability.

The state’s Democratic governor is expected to veto the bill, and it remains to be seen whether the GOP-controlled legislature will secure enough votes to override it. Regardless of the outcome, this episode is a reminder of how fragile America's nuclear future remains and highlights a fundamental contradiction among nuclear advocates on the political right. Restarting America's nuclear industry hinges on clear and consistent policy support, for which the most compelling rationale is precisely the one S.B. 261 erodes: decarbonization. Policymakers nationwide should heed this cautionary tale – without stable and ambitious climate policies, America's nuclear renaissance remains precarious at best.

Subscribe now