Plugging It In Is Harder Than Building It — What an Interconnection Queue Actually Decides

113 Gigawatts Are Standing in Line

In Japan, connection review applications asking to attach large storage batteries to the power grid accumulated to a cumulative 113 gigawatts by March 2025. A gigawatt is 1 billion watts, and this amount is in the same digit range as the electricity used at the moment the entire country uses electricity the hardest during the year, namely the annual peak demand. Around that time, the large storage batteries actually connected to the power grid and running totaled 0.23 gigawatts altogether (DOI: 10.2139/ssrn.7087944, preprint).

The two numbers do not have one representing a part of the other. Applications piled up that much, and around the same time, the total amount of storage batteries actually attached to the system, meaning the power grid, was that much. It means there was that large a gap between applications and connection.

The spot where this gap arises is the interconnection queue. As seen in the final section of the primer "Transmission, Distribution, Transformers — The Road from Power Plant to Wall Outlet", the interconnection queue is a line where applications submitted to connect newly built facilities to the power grid line up in order and wait for review. In this article, "connection" refers to what is otherwise called grid integration, the act of attaching new facilities to the system. The queue is similar to a bank number ticket. Unlike a bank, however, what remains by my turn can change depending on what the person ahead did at the counter. The reason for that is seen in the third section.

The process by which the line grew remains in data. The same study collected data published by Japan's Organization for Cross-regional Coordination of Transmission Operators from 2018 to 2025 to build a research panel for the first time, meaning data stitching the same items together across time points. Connection review applications for storage batteries increased 16-fold in 2 years, and by the end of 2025, 86 percent of connection review applications were for storage batteries. Where the line crowded was not the north rich in renewable energy, but the Tokyo region where demand is clustered.

Japan decided to raise the cost of standing in line. Starting January 2026, new applications must prove land use rights with documents. Starting April 2026, grid-scale storage batteries must pay double the deposit, and if they withdraw their application themselves, they lose that money. The aforementioned paper is a pre-registered evaluation design that pins down in advance the evaluation method to see the effect of this system, meaning deciding from the beginning what and how to measure before seeing the results. It wrote down in advance as the threshold to catch a collapse in applications of 42 to 77 percent, and also wrote down in advance what judgment to withhold if a certain result comes out. Whether it had an effect has not come out yet.

There is a reason this experiment is precious. Because there is no place observed after raising the cost of standing in line, this question has only been debated with models. And the most spotlighted model changed its answer as it changed its premises. As the aforementioned Japanese paper introduces in its abstract, that model is the model by Johnston, Liu, and Yang that analyzed the PJM electricity market region in the eastern United States. This model answered "yes" at first, but in its 2026 revised edition, it changes its conclusion that fees ultimately barely move completed capacity (secondary citation — the original model could not be verified. DOI: 10.2139/ssrn.7087944). The reason is ironic. It is because the projects that clog the line the most are simultaneously the projects with the most value.

Permitting Took 250 Days

Slow local government permitting is often blamed as the cause of long lines. In Massachusetts, wind and solar projects that received local permits took an average of 250 days to be permitted, and that duration was growing longer year by year (DOI: 10.3389/fsuep.2026.1771870, peer-reviewed). Even though local permitting is the most common approval method in the US, there had been no data measuring how long it actually took. So the authors built their own dataset that even included projects that were ultimately canceled, and this is the first attempt to measure that duration.

Projects that received permits and projects with smaller capacities had shorter durations, while canceled projects and projects with larger capacities took longer. However, the authors reported this only as a correlation. This data cannot tease apart whether projects were canceled because they took too long, or whether they dragged on a long time because they were bound to fall through anyway.

At 250 days, that is eight months, which is not short. But the time required for interconnection is on a completely different scale. An American study proposing that the grid side pre-select sites to attach AI data centers describes the typical interconnection process used as a benchmark—where developers choose a site, apply, and wait in line for review—as taking 5 to 8 years (arXiv:2605.14714, preprint, synthetic grid modeled after the Texas grid). Another study opens its problem statement by noting that securing interconnection capacity can take longer than building the facility itself (arXiv:2608.29359, preprint). Both studies did not measure this time directly but cited it as a starting point for discussion.

Permitting is eight months and interconnection is multiple years, so the numbers are on different orders of magnitude. However, the two figures differ in both their targets and measurement methods. One is the actually measured average wind and solar permitting duration in a single state, Massachusetts, while the other is the length of the entire interconnection process cited as a benchmark by a data center study. Even so, when the orders of magnitude differ this much, cutting the permitting time in half leaves a very strong possibility that several years will remain untouched. The bottleneck may lie in the structure of the line rather than the review speed. This is an inference by this article, which has placed the two studies side by side.

The Spaces Left Over Change Once the Person Ahead Connects

Before looking at the structure of the line, let us revisit the two layers of the path electricity travels. As seen in Electricity from Scratch, the transmission network is the thick road carrying large amounts over long distances, while the distribution network is the thin road dividing power finely down to neighborhoods and buildings. The two layers differ in both why they get congested and how much it costs to expand them. The studies in this paper also look at different layers, which will be specified each time.

One identical question applies to both layers: Grid capacity is the amount this grid can accept further while leaving the currently laid wires and transformers as they are. It is like the maximum occupancy sign in an elevator. To carry more people, you must either do construction to increase the occupancy or change the way people ride.

Applications for connecting distributed generation to the distribution network are processed mostly in the order they arrive, everywhere in the world. Distributed generation refers to small power generation facilities scattered and connected close to where the power is used, such as rooftop solar panels. This first-come, first-served approach is underpinned by the assumption that no matter who connects first, the total amount will be the same once everyone is connected.

In the distribution network, this assumption can break down. When a preceding facility connects, the voltage across various points on that distribution line changes, and therefore the remaining capacity available to a subsequent business varies depending on the preceding order (DOI: 10.2139/ssrn.7287446, preprint, simulation using a standard test network). The authors verified this by swapping the order across two standard test networks. A standard test network is a virtual power grid model that researchers use together so they can compare results under the same conditions. Penetration rate is the ratio representing how large the capacity of the power generation facilities attached to that network is compared to the amount of electricity that network uses. If it exceeds 100 percent, it means more can be generated than the network uses, and in this experiment, the penetration rate was varied from 46 to 499 percent.

The results split into two depending on the penetration rate. When the penetration rate was low, the order made no difference at all, and only after exceeding about 120 percent did the order begin to determine the results. Above that, the differences were large. The best-arranged order connected an average of 22.4 percent and a maximum of 67.6 percent more capacity in one test network compared to first-come, first-served. In the other test network, it was an average of 9.7 percent and a maximum of 17.3 percent. This means the magnitude of the effect varies depending on the shape of the network.

This paper also places weight on the solution side. Because changing the order pushes someone backward, the authors bounded the positions to which a business could be delayed to half of the total number of positions. Even so, the total performance loss stopped at about 1 percent. However, they also added the limitation that near the threshold where the order begins to produce an effect, this gain was not robust.

What to take away from this is not the number 22.4, but the property underlying it. At least in this distribution network simulation, every time a single case connected, the margin left for the next person changed. This means the queue may not be a line of mutually unrelated applications.

Layered onto this is the problem of the value of standing in line. A prior Japanese study called this situation "A Free Option That Can Block the Power Grid" in its title. An option is the right to choose whether or not to use something later. If the cost of standing in line is cheap, getting in line itself becomes an option close to free. Even for a project with neither land nor capital, if you just pull a ticket in advance, you retain an early number for when your circumstances improve later. The measure Japan introduced in 2026 targets this exact point.

At the same time, it is hard to blame the applicants. For power generation developers, how much money will go into connection is uncertain from the outset. A study proposing to redesign connection procedures and transmission planning summarizes its own proposal as replacing highly uncertain grid reinforcement costs and costs caused by congestion with a fixed connection fee (DOI: 10.2139/ssrn.3936998, preprint). Read in reverse, this means those costs are currently highly uncertain. In a line where you do not know the price, getting in place first and figuring out the math later is rational for an individual.

Why Not Just Build More Power Lines?

A major reason lines get congested is that the path is narrow. As we saw in Electricity from Scratch, congestion is a state where lines cannot carry all the electricity we want to send because it hits their limit. This is like cars coming to a halt on the way home from work because there are not enough lanes. So the natural retort is that we should simply widen the path, and in reality, that is often the right answer. The problem is that widening paths can take longer than building the facilities themselves.

In November 2021, voters in Maine, USA, voted down a transmission line project that sought to bring hydroelectricity from eastern Canada to New England. Eastern Canada is one of the premier hydroelectric regions in the world and is well-positioned to send electricity to the northeastern United States, but the state the lines needed to pass through refused them (DOI: 10.1108/oxan-db268663, a policy briefing published in an industry journal).

There are two main reasons why widening paths is slow. The first is people. When researchers surveyed the positions of stakeholders and local residents regarding grid reinforcement to connect a new nuclear power plant in southwestern England, three different strands of discourse emerged (DOI: 10.1068/a43401). The points where the discourses diverged were four: whether current procedures actually guarantee public participation, whether the grid operator can be trusted, whether the power system should be centralized or decentralized, and what transmission lines hung from steel towers—namely, overhead transmission lines—would leave behind in the region. Of the four, two were about procedures and trust, one was about what kind of power system to move toward in the future, and only one was about the impact the power lines would leave on the region.

The authors of the same study note that the fact that residents' positions are spread across multiple directions shows the limitations of explaining local opposition through the term NIMBY, which means "Not In My Back Yard." And while they recommend citizen participation early in the planning stage, they add that as long as the future shape of the nation's power system remains undecided, there are limits to that recommendation.

A study that qualitatively organizes and analyzes the positions of parties involved in cross-border transmission line conflicts between the US and Canada points to a similar place. To build consensus before a lawsuit, core stakeholders must participate, trust must be addressed, and the discussion must be narrowed down to the problem of this transmission line. What the authors considered most decisive was designing a process in advance to build consensus during the early stages of development (DOI: 10.1111/nejo.12054, a qualitative conflict assessment).

The second is timetables. Different types of transmission investments have different times from decision to actual operation—known as lead time—and planning models must handle those differences. When demand is uncertain and candidates with different lead times are mixed together, a method that reflected those differences stage by stage in the planning phase produced better decisions than other methods (DOI: 10.3390/en11092429, simulation on a modified Garver 6-bus test network). A bus refers to a node in power grid calculations where lines meet and generators or loads attach.

Cutting Back and Connecting

There is also a way to connect without widening the roads. Curtailment is the act of dialing down power generation to a lower level than the generation facilities are capable of producing, due to grid conditions. It is like the wind blowing while you deliberately slow down the wind turbine's blades.

Which is cheaper between reinforcing the grid and curtailment has been calculated. A study that ran actual load and generation data from three distribution networks in Finland found the break-even point where the cost of reinforcing the grid when adding more solar power meets the cost of curtailing output occasionally (DOI: 10.1049/gtd2.12936, Monte Carlo simulation + Finnish measured data).

The answer was "both." Curtailment was a cheap method up to the point of raising hosting capacity by 13 percent in rural areas, 7 percent in suburban areas, and 8 percent in urban areas. Beyond that point, the compensation costs paid out for the curtailed energy exceeded construction costs. The authors organize this in chronological order. Curtailment is the cheapest tool available right away, and it defers grid investments until hosting capacity reaches 118 percent, 106 percent, and 97 percent, respectively. You buy time with curtailment, and once you pass that point, construction becomes a realistic choice.

Curtailment is not always a waste either. In a grid mixed with highly volatile renewable energy and thermal power that cannot easily change its output, curtailing a portion of the renewable energy can reduce total generation costs. However, because renewable energy has almost no fuel cost, curtailing it excessively makes it expensive. The authors' conclusion goes a step further. Because consumers, thermal power companies, and renewable energy generators do not benefit equally, leaving the decision to power generators results in a level of curtailment that is suboptimal for society as a whole. Therefore, that level must be determined by the grid operator (DOI: 10.2139/ssrn.2441095, preprint, stylized analytical model).

How much is actually curtailed can be seen in hourly data from the California grid from May 2014 to July 2022 (DOI: 10.2139/ssrn.4294917, preprint). Four statistical methods all pointed in the same direction. As the share of solar and wind power rose, the curtailment rate rose; it also rose when the share of nuclear power, which cannot easily change its output, increased. Conversely, as demand increased, the curtailment rate fell, but that effect flattened out at about 60 percent of peak load.

If you subtract the last sentence of the same paper, you end up reading these results backward. Even though the share of wind and solar is increasing, the impact of curtailment on the cost per unit of electricity was moderate. Curtailment was a headache in California during that period, but not yet large enough to break the business.

However, looking at the average misses something. A study analyzing connection conditions in renewable energy zones targets this exact spot (DOI: 10.21203/rs.3.rs-3412892/v1, preprint). Because wind power's production at its peak is 2 to 4 times the average, and solar power's is 4 to 11 times, curtailment becomes inevitable once market share reaches 50 percent. The crux is what comes next. When you build one last additional turbine, the extra curtailment rate caused by that single turbine is at least 3 times the overall average curtailment rate.

New market entrants look at the average curtailment rate when joining, but the burden they actually place on the grid is much greater than that. That does not mean the authors suggest blocking entry. The authors argue for a compromise. They suggest determining capacity based on the last turbine, while guiding entry levels based on the average.

There is also a counterexample to the answer that transmission lines should simply be built quickly. A model study calibrated to UK onshore wind calculates that if overhead transmission lines can be built fast enough, the desirable level of curtailment for Scottish wind hardly changes. However, the subsea interconnector proposed by the UK is expensive. Therefore, the efficient level of curtailment actually rises, resulting in a point where adding more Scottish wind becomes unprofitable (DOI: 10.2139/ssrn.7009282, preprint, closed-form analytical model). Whether widening the road is the answer depends on which road you widen and at what cost.

Agreeing to Be Curtailment-Ready in Exchange for Getting Connected First

If you make a single promise, you can move to the front of the line. Flexible interconnection is an approach where you promise in advance to reduce your output during times the grid cannot handle it, and in exchange, you get connected first. It is similar to registering without a waitlist by attaching a condition that you will give up your seat during peak season. While curtailment in the previous section is something that happens after you are connected, a flexible interconnection is a contract made before you hook up.

What is discussed in the United States as an option in this direction is energy-only interconnection. It is a method of connecting a generator without guaranteeing that its electricity can be carried all the way to the consumer at all times. This method can avoid reinforcement work that attempts to clear congestion before connection, but in turn, the risk of curtailment can grow (DOI: 10.2139/ssrn.6870718, regulatory review paper, preprint).

What this regulatory review confirms is the most important warning in this article. Even within the United States, interconnection review rules and procedures vary greatly from jurisdiction to jurisdiction. The authors also pinpoint where those differences come from. The institutional design for ensuring adequate electricity supply is different, real-time transmission operation methods are different, and state-level power procurement practices are different. U.S. regulatory authorities have preferred leaving regional discretion over having the federal government set uniform rules, which has widened the differences even further. If it varies this much within a single country, reading these systems into another country is even more dangerous. This final sentence is not the conclusion of the paper, but an inference of this article.

Therefore, none of the systems appearing in this article should be read as the state of affairs for the entire world. Japan's deposit system, America's jurisdiction-by-jurisdiction interconnection rules, and Britain's transmission line siting conflicts all play out on top of different systems.

The question left by the same paper is also significant. The threshold itself that determines when reinforcement work is required on the grounds of congestion can become a barrier blocking entry into the market. That is why the authors raise the question of whether the power market's competition policy should extend beyond the spot market to cover market entry, which interconnection rules now dictate.

The same idea emerges on the electricity consumption side. It is a method of connecting first and adjusting the limit of electricity that can be sent and received in real time. A study that applied this method to a data center consuming hundreds of megawatts—in other words, hundreds of millions of watts—reveals an honest tension. The limits set from the outside change moment by moment, but the AI training workloads running inside must not be interrupted. The authors report that placing an on-site battery as a buffer between them significantly increases the workload that can be committed a day in advance. The case study was run by applying Australian measured data to the IEEE 39-bus test network (arXiv:2605.14105, preprint, simulation using a standard test network).

On the distribution network side, there is research that treats the risk of curtailment itself as an interconnection condition. Allowing curtailment increases hosting capacity, but how often and how severely that curtailment happens is a risk for the connecting party. The authors included both a condition that limits rare but extremely severe curtailment cases and a condition that caps the number of times the grid intervenes. Even with those added, hosting capacity could be significantly increased while guaranteeing those two things (arXiv:2604.20185, preprint, simulation). The value of flexible interconnection ultimately comes down to how often you get curtailed, and whether that can be written into a contract is the actual point of contention.

And the Loads Lined Up

So far, the ones standing in line have mostly been the electricity makers. Now, the electricity users take their turn in line. In this article, load means, as defined in the primer "Alternating Current and Frequency — A Giant Merry-Go-Round the Whole Country Pushes Together", the electric power that the power grid must supply at that exact moment, and a large load refers to a major customer like a data center that uses a large amount of such power in one place.

Because the U.S. interconnection queue for generators is a public record, anyone can see who has applied to connect what size plant and where. Equivalent records for large-load customers applying to connect with hundreds of megawatts are not public. A policy brief pointing out this asymmetry argues that federal regulators should apply the public standards they have built for generators over the past 30 years to the load side as well. It cites state-level precedents from Texas and Georgia in the U.S., and points out a path to accomplish this by amending tariff sheets rather than passing new laws (DOI: 10.2139/ssrn.6614823, preprint policy brief).

Half of the line is invisible. In "How Much Electricity Does AI Use? — Why Every Source Gives a Different Number", the numbers for the same target diverged depending on what was counted and up to what point. The queue problem happens upstream of that. One side of the line is not counted at all.

Whatever is connecting, the grid asks the same question. Can we connect more here? As seen in the primer "Transmission, Distribution, Transformers — The Road from Power Plant to Wall Outlet", hosting capacity varies by location. When connecting small- and medium-sized data centers to a distribution grid, some locations hit the current limit first, some hit the voltage fluctuation limit first, and some hit the limit due to both acting together (arXiv:2606.01407, preprint, simulation). The same study also shows that adding batteries, dispatchable distributed generators, and reactive power compensation devices to support voltage allows the three to complement each other in different ways, actually increasing the hosting capacity.

The limits of hosting capacity are not just about heat and voltage either. If a large data center drops off the grid or switches to its own backup power when the voltage briefly wavers, a massive demand vanishes from the grid in an instant. A study incorporating this behavior into interconnection planning reports that calculating the connectable capacity based only on normal operating states can lead to an overestimation. Increasing the margin on the uninterruptible power supply side, which provides temporary electricity during outages, restored the hosting capacity, and this condition changed not just the total amount, but even how much to distribute and connect at which points (arXiv:2609.03030, preprint, IEEE 118-bus simulation).

How much location determines the outcome is shown by the New England simulation seen in the primer. When calculation was run 1,560 times by only changing the locations of placing 1 gigawatt of solar power and 1 gigawatt of data center, only 14 combinations passed without line overloads (arXiv:2509.01778, preprint, simulation using a standard test grid). Same facilities, same capacity, yet the outcome diverged depending on where they were placed. A large load is a matter of grid hosting capacity before it is a matter of the user's side.

So Who Pays?

There is a study that unpacks the large-load rate schedules of the United States—documents detailing the rates and contract conditions electric utilities apply to large customers—from their original sources (DOI: 10.2139/ssrn.7250402, preprint). The researchers extracted and coded 93 clauses across 27 policy cases involving electric utilities, states, and regulatory proceedings. This study views large-load rate schedules as public contracts established in advance to prepare for cases where a guest does not arrive or uses less than expected.

The problem setup is concrete. Electric utilities must make irreversible investments before data centers start running. If that business is delayed, reduced, or canceled, the question is who pays for the unused facilities.

The authors coded the clauses as "present," "absent" only after searching documents to the very end, and "unverifiable," and did not convert unverifiable items into weaker protective measures. In a validation test where 40 core cases were recoded with the help of a model, the judgment agreement rate was 92.5 percent, the proportion that matched even in exact order was 85.0 percent, and the weighted kappa coefficient was 0.508. The authors emphasize that this audit is a reproducibility audit checking whether repeating the same procedure yields the same judgments, not a coder reliability audit testing whether different people judged separately and matched results.

The results showed both sides. Protective measures were strong, but the variance across locations was large. A test re-analyzing all unverifiable clauses as absent lacked sufficient data to draw a conclusion, failing to prove that the rankings remained stable under that assumption. However, it did show why current records cannot line up the country in a single row.

The final sentence of this paper is where this article lands. Large-load rate schedules determine whether the losses when speculatively built facilities go unused will remain with large customers, shift to utility investors, or be distributed among ordinary ratepayers. The procedures surrounding connection determine not just when it attaches, but who pays when things go wrong.

And the demand forecasts that form the basis of planning are off to begin with. Although the data is from before the data center boom, there is a study that compared 11,416 summer peak demand forecasts produced by 185 US planning regions from 2006 to 2022 against actual records, spanning horizons from 1 to 10 years ahead (DOI: 10.2139/ssrn.7411248, preprint). Forecasts 1 year ahead had no one-sided error, meaning no bias. However, the further out they looked, the larger the bias grew. 10-year-ahead forecasts written before 2016 exceeded the actual peak by an average of 12.4 percent, and 78 percent of those were over-forecasts.

This bias was a stable property across planning regions. The correlation of average errors between the first and second halves of the sample was 0.81, meaning regions that made many mistakes early on made many mistakes later as well. By the authors' estimation, planners treated under-forecasting as up to 2.7 times more costly than over-forecasting. That is understandable since they carry a supply obligation. However, the authors draw a line, stating that the explanation ends there. This preference alone does not explain how forecasts change. Once planners recognized their own errors, they overcorrected, and successive revisions targeting the same year moved in opposite directions.

Stopping here would mean translating this paper in completely the opposite way. As demand growth resumed after 2015, the direction of the bias flipped. Therefore, applying adjustments tailored to the earlier period straight across would have made the later forecasts even worse. The authors' recommendation is different from cutting forecasts uniformly. It is to set ranges of uncertainty by time horizon, adjust them to each institution's past records, and check whether the regime has shifted before applying mechanical corrections.

The answer to whether the line gets shorter when the load side becomes flexible is conditional. In a study that plugged flexibility into a system expansion model, if data centers could delay or relocate tasks, investment and operating costs fell by 3 to 21 percent (arXiv:2604.05376, preprint, simulation). What the authors emphasize is the fact that the effect is inconsistent rather than the magnitude. Results varied significantly depending on location, the scope of flexibility, and the load state at the time, and increasing flexibility did not necessarily reduce the required generation capacity.

There are also studies that saw significant effects. This is the Texas synthetic grid study seen earlier. A synthetic grid is a virtual power network built to mimic the characteristics of a real system, and this study used one with 2,000 buses. The method involves pre-determining the scope of flexibility data centers must maintain and having the grid side pre-certify candidate locations that match it. The attachable candidate sites expanded by 9 to 17 percent at 1 gigawatt and 19 to 21 percent at 2 gigawatts, while the median wholesale price across all hours of the day remained nearly unchanged. The authors note that bypassing major transmission reinforcements this way takes 12 to 18 months until electricity first flows. Even taking a conservative view compared to the normal developer-driven process of 5 to 8 years, this is 3.5 to 4 years shorter (arXiv:2605.14714, preprint, synthetic grid simulation).

The counterargument of whether one could just build their own power plant and avoid the grid has also been calculated. There is a study that priced and compared nine on-site power generation options available to AI data centers on a whole-site basis (arXiv:2608.08170, preprint, simulation). Megawatt-hours are a unit measuring the amount of electricity used, just like watt-hours. Combined-cycle gas turbine power is 47 dollars per megawatt-hour based on what comes out of the generator, but when calculated based on the electricity used by the entire site, it becomes about 114 dollars. The whole-site calculation included actually measured AI training loads, the cost of fuel delivered to the site, two tax credits from US tax law, and reserve power rates relying on the grid. Under the same conditions, grid electricity was 92 dollars. A four-hour battery had a physical ceiling of up to about 18 percent of annual energy.

The authors' summary is brief. Nothing beats the wires. The authors view on-site generation not as a means to lower costs, but as a way to buy access and scale, concluding that most current investments are targeting the wrong items.

What Exactly Is the Queue Deciding?

Wind power plants and data centers sit at exact opposite ends of the electricity supply chain. One side builds, and the other side consumes. Yet when you place the development processes of the two industries side by side, both hit the same roadblocks: securing land, permits, grid connection, and local community opposition (DOI: 10.2139/ssrn.6506945, preprint). Drawing lessons for the data center construction boom from what wind power has endured over 30 years, the authors recommend information disclosure standards, benefit-sharing agreements with local communities, and social acceptance research. They do, however, add the caveat that structural differences between those who build and those who consume lead to distinct policy implications.

Therefore, the statement "the queue is the problem" is only half true. Before it is a device that delays, the queue is the procedure that decides who pays for grid reinforcement costs. Turning what we have looked at so far into questions leaves us with five.

① What is the price of standing in this line? If the price is cheap, projects lacking both land and funds grab a ticket (DOI: 10.2139/ssrn.7087944). Raising the price would seem to clear out the line, but as that Japanese paper reports, a revised model analyzing PJM in the U.S. answered that fees barely move completed capacity. Japan is running that experiment, and the results are not yet out.

② What did the person ahead of you do before leaving? In distribution network simulations, the order of connection changed the total volume, and that effect only appeared once penetration exceeded roughly 120 percent (DOI: 10.2139/ssrn.7287446). At least under those conditions, the queue is not a line of unrelated applications.

③ Should we widen the road, or use less? In Finland's distribution network, curtailment was a cheap tool up to a certain point, after which construction was better (DOI: 10.1049/gtd2.12936). And when the cost of widening the road is expensive, curtailment actually increases (DOI: 10.2139/ssrn.7009282).

④ Can the hours we cannot use be written into the contract? The real value of flexible connection is how often and how severely power is trimmed, and whether it can be written down as a guaranteeable condition is the core issue (arXiv:2604.20185). In the U.S., those rules vary by jurisdiction (DOI: 10.2139/ssrn.6870718).

⑤ Who pays when things go wrong? Large load rate schedules can leave losses with customers, shift them to investors, or distribute them among general ratepayers (DOI: 10.2139/ssrn.7250402). And long-term demand forecasts reliably miss the mark across planning zones, while the direction of that error can flip (DOI: 10.2139/ssrn.7411248).

A review paper summarizing the burden AI data centers place on the grid divides that burden into three time scales: long-term planning and connection, short-cycle operation and markets, and real-time behavior (arXiv:2509.07218, preprint, review). The final scale was covered in "The Few Seconds a Generator's Spinning Mass Gave Away for Free — What Gets Harder as Grid Inertia Falls". This article looked at the first scale.

Nothing flashy happens on the first scale. There are no blackouts and the frequency does not wobble. Applications pile up, reviews fall behind, costs are recalculated, and certain projects are simply canceled quietly. The gap between Japan's 113 gigawatts and 0.23 gigawatts also widened that way. That gap is not an accident; it is the result of procedure.

Attaching something to the power grid begins as an engineering problem and ends as an accounting problem. Calculation provides the answer for how many amperes a wire can endure. Calculation does not provide the answer for who will pay the cost of upgrading that wire to be thicker. Physics determines the former, while the rules for lining people up determine the latter. And right now, those who want to generate electricity and those who want to consume it are standing in that same line together.

Source Papers

Papers published on arXiv, SSRN, or Research Square are marked as "preprint" and whether they were subsequently published in an academic journal was not separately verified. This field relies heavily on simulations over standard test networks, which is noted in the text each time.

• "The Free Option to Congest the Grid: Japan's Battery-Storage Interconnection Queue" (2026, preprint, pre-registered evaluation design) — DOI: 10.2139/ssrn.7087944 — 113 gigawatts of cumulative battery storage applications in Japan versus 0.23 gigawatts connected. Pre-registered evaluation of the 2026 institutional reform.
• "Permitting and siting of wind and solar in Massachusetts" (2026, peer-reviewed) — DOI: 10.3389/fsuep.2026.1771870 — Local government permitting takes an average of 250 days and is trending longer.
• "Flexibility-Aware Framework for Efficient Planner-Initiated Siting of Data Center" (2026, preprint, synthetic grid simulation) — arXiv:2605.14714 — 2,000 buses in Texas. First power-on in 12 to 18 months through pre-certified flexibility ranges.
• "Minimizing Grid Interconnection Capacity Requirements for AI Data Centers" (2026, preprint, simulation) — arXiv:2608.29359 — Calculates that on-site solar, batteries, and task deferral can reduce required connection capacity by 6 to 13.3 percent depending on the budget.
• "Connection Order Matters: Path-Dependent Hosting Capacity in DG Interconnection Queues" (2026, preprint, standard test network simulation) — DOI: 10.2139/ssrn.7287446 — Distribution grid connection order changes total hosting capacity, triggering only above roughly 120 percent.
• "Generator Interconnection, Network Expansion, and Energy Transition" (2021, preprint, 2-node simulation) — DOI: 10.2139/ssrn.3936998 — Proposal to replace uncertain grid reinforcement and congestion costs with fixed connection fees.
• "Opposition threatens US-Canada transmission lines" (2022, industry policy briefing) — DOI: 10.1108/oxan-db268663 — Rejection of transmission lines in the November 2021 Maine referendum.
• "Discourses of Energy Infrastructure Development: Transmission Line Siting in the UK" (2011, peer-reviewed) — DOI: 10.1068/a43401 — Three-pronged discourse of UK transmission reinforcement and limits of the NIMBY concept.
• "Conflict Assessment in Energy Infrastructure Siting: Northern Pass" (2014, peer-reviewed, qualitative conflict assessment) — DOI: 10.1111/nejo.12054 — Designing consensus-building procedures early in development is most critical.
• "Transmission Expansion Planning under Uncertainty for Investment Options with Various Lead-Times" (2018, peer-reviewed, test network simulation) — DOI: 10.3390/en11092429 — Planning that explicitly incorporates lead-time differences yields better decisions.
• "PV hosting capacity improvement: economic comparison between curtailment and network upgrade" (2023, peer-reviewed, Monte Carlo + Finnish field measurements) — DOI: 10.1049/gtd2.12936 — Curtailment is cheaper until raising hosting capacity by 13, 7, and 8 percent, after which construction is better.
• "Economic Curtailment of Intermittent Renewable Energy Sources" (2014, preprint, stylized model) — DOI: 10.2139/ssrn.2441095 — Curtailment can reduce costs, but if excessive it becomes expensive, and leaving it to operators leads to selections below optimum.
• "No Alarms and No Surprises: Dynamics of Renewable Energy Curtailment in California" (2022, preprint) — DOI: 10.2139/ssrn.4294917 — As penetration rises, curtailment rates increase; as demand rises, they decrease. Moderate impact on levelized costs.
• "Marginal vs average curtailment of renewables in Renewable Energy Zones" (2023, preprint) — DOI: 10.21203/rs.3.rs-3412892/v1 — Marginal curtailment rate is at least 3 times the average. A compromise proposal to guide capacity by the marginal rate and entry by the average.
• (Title unconfirmed — the bibliographic title field was contaminated with abstract text, making it impossible to secure the original title. The DOI resolves normally) (2026, preprint, closed-form analytical model) — DOI: 10.2139/ssrn.7009282 — UK onshore wind correction. Expensive subsea interconnections increase efficient curtailment.
• "A Review of United States Energy-only Generator Interconnection Service Policy" (2026, preprint, regulatory review) — DOI: 10.2139/ssrn.6870718 — Connections without deliverability guarantees avoid front-end construction but increase curtailment risk.
• "Battery-Assisted Operation of Hyperscale AI Data Centers under Connect-and-Manage" (2026, preprint, standard test network simulation) — arXiv:2605.14105 — Buffering conflicts between real-time connection limits and non-interruptible workloads using on-site batteries.
• "Risk-Aware Hosting Capacity Analysis for Flexible Load Interconnection" (2026, preprint, simulation) — arXiv:2604.20185 — Increasing hosting capacity while limiting curtailment tail risk and intervention counts.
• "The Missing Half of the Queue: Load-Side Interconnection Queue Disclosure" (2026, preprint policy brief) — DOI: 10.2139/ssrn.6614823 — U.S. generator queues are public records, but large-load queues are not.
• "Hosting Capacity Assessment and Enhancement for Edge Data Centers" (2026, preprint, simulation) — arXiv:2606.01407 — Hosting capacity varies by location and limiting reasons are current, voltage, or a mix of both, which can be expanded using batteries, distributed generators, and reactive power compensation.
• "Hosting Capacity Assessment of Data Centers with Voltage Ride-Through Capability" (2026, preprint, IEEE 118-bus simulation) — arXiv:2609.03030 — Static planning alone can overestimate connectable capacity.
• "Grid Transmission Evaluation for Solar Deployment and Data Center Growth" (2025, preprint, standard test network simulation) — arXiv:2509.01778 — IEEE 39-bus. Out of 1,560 power flow calculations, 14 combinations avoided overload.
• "Who Pays When the Data Center Never Comes? U.S. Large-Load Electricity Tariffs" (2026, preprint) — DOI: 10.2139/ssrn.7250402 — Coding 93 tariff clauses. Tariffs determine whether downside risk falls on customers, investors, or ratepayers.
• "Accuracy and bias of long-horizon load forecasts used in utility capacity planning" (2026, preprint) — DOI: 10.2139/ssrn.7411248 — 10-year-ahead forecasts exceeded actuals by an average of 12.4 percent, but inverted after 2015.
• "To Defer or To Shift? The Role of AI Data Center Flexibility on Grid Interconnection" (2026, preprint, simulation) — arXiv:2604.05376 — The effect of flexibility is inconsistent and increasing it does not necessarily reduce generation capacity.
• "Beyond the Grid: On-Site Power Technologies for AI Data Centers" (2026, preprint, simulation) — arXiv:2608.08170 — Natural gas combined cycle at $114 versus grid power at $92 on a site-wide basis. Four-hour battery cap is about 18 percent.
• "Hyperscale Data Centers and Wind Energy: Cross-Sector Lessons" (2026, preprint) — DOI: 10.2139/ssrn.6506945 — Two industries at opposite ends of the supply chain hit the same bottleneck.
• "Electricity Demand and Grid Impacts of AI Data Centers" (2025, preprint, overview) — arXiv:2509.07218 — Organizing grid burdens across three time scales: long-term planning, short-term operation, and real-time behavior.