The AI buildout is bottlenecked by energy. But America has the electricity to power its data centers; the problem is getting it to them.
One of the most expensive projects in history is under construction in Abilene, Texas. This joint venture, Stargate, is the flagship of a bigger project by the same name led by OpenAI and Softbank, and is expected to cost well over $40 billion for a high-performance computing campus that will train new generations of AI models.
Stargate is just one major project in one of the biggest investment booms in history, driven by the belief that increasingly powerful AI models can deliver explosive economic growth. But it will require enormous amounts of electricity to work: Stargate is expected to draw 1.2 gigawatts, as much as 313,000 median American family homes, at peak load. A report by EpochAI and an energy research institute projected that total AI computing power would reach 100 gigawatts worldwide in 2030 if the 2025 growth rate stays steady. And data centers aren’t the only energy-hungry element of the AI revolution. The biggest battery manufacturing plants in the US draw energy at a rate of 115 megawatts, and the first phase of TSMC’s Arizona semiconductor plant will draw 200 megawatts.
Subscribe for $100 to receive six beautiful issues per year.
The primary bottleneck to this growth is the availability of electricity. But this doesn’t mean there is an energy shortage. Instead, the constraint is connecting the flood of new data centers and the plants to power them to the electric grid. Before any new piece of infrastructure can be connected, grid operators must study how it will change power flows around the grid and determine whether upgrades to the system are required. That process is significantly backlogged. Though the median power plant in 2005 waited less than 20 months for interconnection, this had jumped to 55 months by 2023.
The interconnection process wasn’t created for today’s world. Grids use an inflexible first-come, first-served queue that leaves some of the most valuable projects stuck behind less important ones. They also evaluate according to rigid conditions that don’t reward plants for being willing to cover their own power needs for short periods. To prepare for the AI age, grid processes need to change.
A power-hungry future
Estimates vary for how much power will be needed by the data centers and chip manufacturers of the future, but the heads of every major AI company agree that they will need more than they are currently able to get. Jensen Huang, CEO of Nvidia, has said that ‘every data center in the future will be power-limited’. Mark Zuckerberg said Meta ‘would build … bigger [AI training] clusters … if we could get the energy to do it’. And OpenAI CEO Sam Altman told Congress that ‘the abundance of [AI] will be limited by the abundance of energy’.
Regardless of how the data center boom plays out, there is a long-term shift towards electrification across the economy. Electricity can be converted into work instantly, unlike fuels, and with little energy loss. Electricity creates motion directly, while fuels must first be combusted in an engine. This is why electric vehicles can cost half as much to fuel even though electricity is more expensive than gasoline. The simplicity of electric motors also means that electric vehicles have half the lifetime maintenance costs of gas-powered ones.
Electricity also transmits information. Transistors switch on or off depending on the voltage applied to their gates, which allows circuits to perform logical operations. Radios, screens, and computers cannot run on gasoline alone.
Between 1990 and 2024, the price of electric motors declined by 97.5 percent. Power electronics fell 99.5 percent in price, processors built into devices by nearly 99.9 percent, and batteries 98.8 percent. As prices have declined, performance has improved. For example, the amount of energy a battery can store per kilogram has increased five-fold over the same period.
These trends allowed the progression from the Walkman to the iPod and then the iPhone, as well as enabling battery-powered cars, delivery vans, and bicycles. At the same time, the increased computing power built into devices is giving autonomy to newly electrified cars and trucks, robotic vacuum cleaners and mowers, and drones. The commercialization of humanoid robots may also be on the horizon. The future will be largely defined by technologies that run on electricity.
The grid bottleneck
This growth in demand is already straining power grids. Operators are increasingly forced to use expensive local plants because of what grid operators call congestion: cheaper plants are on the other side of transmission bottlenecks and there aren’t enough cables to get the electricity through. In the United States, the additional costs incurred because of grid inefficiencies like congestion ran to $11.5 billion in 2023, an increase of 45 percent from the year before.
ERCOT, the grid that provides 90 percent of Texas’s electricity (including for the plant being built in Abilene), forecasts that it won’t have enough power to meet demand in summer 2028. PJM is the US’s largest grid, both by the amount of electricity it provides and the population in its coverage area, serving an area between Chicago, New Jersey and North Carolina. In 2025, PJM was not able to buy enough future generating capacity to meet projected demand. MISO, another large US grid operating between Louisiana and Minnesota, concluded in one study that ‘resource adequacy risks could grow… absent increased new capacity additions’. PJM’s CEO put it more plainly: ‘We need capacity – a lot of capacity.’
Discussions about expanding electricity supply to power the future often become debates about which source is most suitable: gas, nuclear, solar, or something else. But these are a distraction. Far more fundamental is ensuring power can be efficiently delivered where needed.
When different generation technologies coexist, they can average out to a grid that is cheaper, more reliable, and less polluting than any single type alone. The question isn’t which technology to use, but what balance. But this question isn’t best answered in the marketplace of ideas. Instead, it should be addressed in the marketplace for electricity.
The market for electricity
In the US and Europe, power grids are largely liberalized: provided they can get the necessary permits, independent developers build power plants, and the local utility is required to connect those plants to the grid. In the United States, 88 percent of large-scale power projects currently in development are privately organized and funded.
The power market is run by the grid operator’s economic dispatch software. Each power plant tells the operator its costs, and the operator commissions the cheapest plants, accounting for transmission constraints. As load increases during the day, the operator keeps commissioning the next cheapest plant. The price of power is set at the marginal cost of the next cheapest plant, and that’s the price every power plant is paid.
Market prices signal to power plant developers about levels of supply and demand. In the same way, prices balance different energy sources based on the strengths and weaknesses of each. For instance, as more solar panels are built, the value (and therefore price) of power during the middle of the day, when the sun is shining most, adjusts downward. From December 2020 to September 2025, maximum solar output in ERCOT increased from 4 to 29.8 gigawatts. And from 2020 to 2025, the value of power at 1pm relative to the highest-priced hour decreased from 92.9 percent to 38.7 percent. As one technology type becomes overbuilt, prices reflect that and developers react accordingly.
The evolving daily price shape in response to the abundance of solar energy was a signal that the grid needed storage capacity, and power plant developers responded. From 2020 to October 2025, ERCOT went from having almost no battery storage to a combined battery discharge of 8.6 gigawatts. The same process has played out in California and many European markets.
One might assume that the price of electricity for consumers is dictated by market forces, like those that regulate supply and demand across different power plants. But to a large extent, it is not. Grid infrastructure, like large power lines, is generally planned by the grid operator, and the cost is passed on to consumers at a price approved by state and federal regulators. In one typical utility territory within ERCOT, the portion of regulated costs on the average residential customer’s bill has grown from 28 percent in 2002 to 40 percent in 2025.
Regulators of all major grids have set caps on wholesale prices in response to public outrage at price volatility. As a result, the generators needed to keep the grid reliable are sometimes unprofitable. Under normal market circumstances, generators would stop running when it wasn’t profitable to do so. Instead, regulators put further policies in place to prevent this. Markets for capacity mean that generators are paid not for energy itself but for committing to be available if there is extra demand. Must-run agreements pay unprofitable generators to keep running if they keep the grid reliable. These are negotiated bilaterally between generators and the grid operator, outside of any wider competitive process.
Policy choices also shift the equilibrium. For example, America’s Inflation Reduction Act gives a $30 tax credit for every megawatt-hour produced by qualifying renewables. Power plants that opt in, typically wind, paradoxically often offer their power at negative prices, making money from the tax credit even when they literally pay consumers to use them. At times, a large enough share of the market offers electricity at a negative price that electricity overall (not just from one supplier) can have a negative price. A similar dynamic is playing out in some European countries.
Despite these interventions, which make them less efficient, markets still find an equilibrium. The interaction of supply and demand creates prices that power plant developers use as indicators for what the market does or doesn’t need more of. If prices are high, new power plants enter. That’s why arguing about the best power generation method is overrated. Well-designed energy markets answer this question automatically. The real bottleneck is connecting to the grid at all.
Power struggles
xAI’s Memphis data center operated partially off-grid for months. When it first came online in 2024, it could reportedly draw only eight megawatts from the grid (enough to power a few thousand electric toasters). Rather than wait for its grid connection to be upgraded, xAI installed 422 megawatts of on-site gas turbines. Once transmission upgrades were completed, the project would shift to consuming grid power and the on-site generators would be used only for emergency backup.
Such off-grid generation is a temporary solution. Grid power is more reliable and, on average, cheaper. But thanks to the long queue of projects waiting for connection, 62 percent of data centers are considering off-grid solutions, either to get up and running faster or to improve reliability. Google is even exploring the possibility of a data center in space powered by solar panels. Grids connect to new generators and energy-hungry infrastructure only after studying how to do so using an engineering model that simulates power flows during peak load scenarios. If the new infrastructure would cause overloads on any parts of the system, then those elements need to be upgraded. The utility then needs to estimate the cost of those upgrades and build them.
The grid operators who run these studies are typically public bodies charging a small fee on electricity sales to fund their operations without seeking to make a profit. State utility commissions and the Federal Energy Regulatory Commission have determined that grid operators should run the interconnection queue as a first-come, first-served process that is essentially free.
That process is moving very slowly, creating a huge backlog of mostly phantom projects. Every US grid has more power plant capacity waiting to be connected than there are gigawatts of peak demand. The backlog also exists for the separate but similar queue of infrastructure not yet connected to the grid. Across ERCOT, there were 143.5 gigawatts of data centers seeking to connect as of October 2025. This compares to ERCOT’s highest ever total demand of 85.9 gigawatts in August 2024.
This system was designed for a different era. The last time electricity use grew five percent annually or faster was between the 1950s and 1970s, during the adoption of air conditioning, refrigerators, dishwashers, and washing machines. From 2005 to 2023, electricity use was almost completely flat. But as that lull in growth comes to an end, the grid will have to adapt.
Crossed wires
The main flaw of the interconnection process is that it uses a first-come, first-served queue. This means that high-priority requests can spend years stuck at the back of the line behind other less important ones.
Some requests are speculative, submitted by developers before they have ready customers. Some are duplicative, meaning a developer fishing for a good spot has requested to interconnect the same project in multiple locations. These are among the reasons that 72 percent of requests to connect submitted since 2000 were ultimately withdrawn. Other requests are high quality but small and low value. This creates a feedback loop: the harder it is for developers to estimate the cost to interconnect, and the longer they have to wait for an answer, the more speculative requests are made, further clogging the system.
Every major grid has a roughly similar process. The system impact study comes first. The operator models the grid during peak demand conditions and assesses the cost of any necessary upgrades. After reviewing those costs, the interconnection customer (the developers of the power plant or power-hungry infrastructure) decides whether to proceed. Almost 40 percent of withdrawals between 2020 and 2023 were at this stage. If the customer proceeds, the grid operator then determines how to physically connect the project to the grid. Again, it delivers a quote for these costs to the customer, who then decides if they want to proceed. If so, they sign an interconnection agreement, and the utility begins work on the upgrades.
There are some opportunities to automate these workflows, but each request takes time. The median delay from initial request to the interconnection agreement was 34.2 months for agreements signed in 2023.
Another major issue is restudies. Once a system impact study identifies required network upgrades, those upgrades are added to the grid model that’s used to study subsequent projects in the queue. If those upgrades are not built because the earlier project withdraws, any subsequent studies may need to be redone. In one extreme case, a restudy increased a 242-megawatt wind project’s network upgrade costs from $33.5 million to $99 million after the project was already operating.
The slowness of the interconnection queue adds to households’ energy costs. The clearest example of this is found in PJM, which has the slowest queue of the US grids. In 2025, PJM’s electricity suppliers paid $14.7 billion at auction to make sure they’d have enough power to supply their customers, a huge jump from $2.2 billion the year before. Meanwhile, many gigawatts of renewable power projects have been sitting in the queue for five years or more. If just 30 percent of those projects had been interconnected, the auction’s cost would have been 63 percent lower.
Grids have adopted many sensible reforms to the interconnection queue, such as requiring higher deposits to enter the queue and forcing developers to prove they have land to build on and to disclose if they’re making duplicative requests. The most significant reform has been transitioning to cluster studies, which study multiple projects in the same simulation. This reduces restudies because removing any one project has less impact on the overall solution. Clusters also allocate costs more fairly. This prevents a single project triggering a major transmission upgrade that renders it uneconomical. In cluster studies, the cost of these backbone upgrades are shared among all the projects that would benefit from them. While these were necessary reforms, the backlog has only grown as they fail to address the fundamental irrationality of the interconnection process itself: the inflexible first-come, first-served queue.
Tragedy of the commons
In the early 2000s, regulators wanted to level the playing field between new independent producers and incumbents. They worried that utilities might preference their own projects. As there was plenty of spare capacity at the time, the first-come, first-served system seemed both simple and fair.
But today, transmission capacity is a limited resource and independent power producers are thriving. This means there is much less concern around the power of traditional utilities. In response, many grids are proposing mechanisms to allow more valuable projects to jump the queue. MISO, PJM, and SPP, three large grids in the US, have proposed mechanisms to prioritize projects that are most viable or most necessary for the system’s reliability. But these mechanisms are band-aid solutions.
The flood of requests is a typical ‘tragedy of the commons’. Everyone is incentivized to spam the queue with requests. Auctions can fix this.
In the fishing industry, when tradable fishing quotas were introduced worldwide in the 1970s, what had been a mad dash for fish became an orderly and efficient process. Before the quotas, the whole season’s supply was caught in a few days of dangerous, non-stop fishing. Market share went to whoever bagged the fish fastest. Once quotas were introduced, fishers could time their catches with market demand instead of catching everything at the start of the season and then freezing it. And since more efficient fishermen made more money, they were willing to pay more, and quotas went to them rather than the fastest fishers, lowering prices for consumers.
Auctioning new grid capacity could bring similar benefits. The scramble for interconnection would be replaced with an orderly process in which the highest quality projects would get priority. Developers’ bids would reflect both the likelihood that the project will come online and the projected value of the project to the grid if so. Less viable and less valuable projects would be weeded out.
The simplest way to implement an auction would be to create small ‘fast-track clusters’ that receive expedited studies throughout the year. Projects in the regular annual cluster could bid to enter the fast track, and the highest bidders up to some preset number would be admitted. The fast-track proposals recently adopted by some grid operators follow this structure, except that admittance to the fast-track cluster is based on an administrative scoring mechanism, not developer bids. But administrative scoring mechanisms could never capture the subtleties that a developer’s bid would. For example, developers have private knowledge about a project’s chances to get permitted, like whether the site has trees with endangered bird nests or unhappy neighbors generating pushback.
Only connect
Another inefficiency of the current transmission system is that it’s built for peak load conditions that occur only a few hours per year. The rest of the time, much of the grid’s generation and transmission capacity is sitting idle. In 2024, 42 percent of ERCOT’s capacity went unused for half the time.
Building the network upgrades that allow for this very high peak capacity can take years. But new power plants must wait for that construction to be completed before they can join the grid, even though the equipment being constructed will rarely be necessary. A survey of power developers found that transmission construction was the top cause of project delays.
The solution is to let the power plant come online before the upgrades are constructed on the condition that it agrees to turn off during peaks. The power plant doesn’t contribute to the reliability of the system since it won’t be available when the system is under stress. But it will produce electricity the rest of the time, which will lower costs. This is known as energy-only service.
ERCOT has made this energy-only service, also known as non-firm transmission because it implies no guaranteed ability to export or import power, their default. They call it connect and manage. That’s a large part of why projects move through ERCOT’s queue faster than any other grid operator.
But despite being an option in all organized US grids, 87 percent of interconnection requests by capacity opt for the slower firm transmission service. That’s because non-firm service disqualifies power plants from the payments they could otherwise receive in exchange for being available at times of high demand. Other grids could adopt connect-and-manage if they eliminated this payment system and, like ERCOT, raised price caps to allow energy prices themselves to reflect scarcity value. Short of that, they could make non-firm service more popular by allowing generators to apply for firm service after being interconnected on a non-firm basis.
Can’t we just use less energy?
These proposed reforms would allow fast interconnection of new power plants. But if AI is a race with national security implications, they won’t be fast enough.
The fastest way to increase capacity for new large loads would be to decrease demand from existing sources. Demand reduction is the option energy planners have always reached for in a pinch. The concept of energy conservation originated in the late 1960s as the US power grid faced resource constraints for the first time in its history. Energy conservation became a national imperative. For the first time, utilities began promoting things like turning off lights and using less air conditioning. Christmas lights were framed as being wasteful and unpatriotic, and in 1973 the National Christmas Tree had only a single light.
Yet while short-term crunches can be fought with conservation, they are difficult to sustain, and increased energy efficiency can simply lead people to using more energy for other purposes. What’s more, the vast majority of electricity users are completely indifferent to the real-time cost to produce electricity, because they’re on fixed rates for terms of months or years. The wholesale cost of a clothes dryer load, which could consume four kilowatt-hours, could range from negative $0.12 to $20, but the retail customer always pays $0.64 at the average US residential rate of $0.16 per kilowatt-hour. This creates irrational behavior, like someone starting a load of laundry while the grid teeters on the edge of a brownout. If customers faced prices more closely linked to real-time costs, they would adjust their energy use accordingly. Prices would be more stable and, on average, lower.
Already, we’ve seen that power costs change throughout the day because of the increasing availability of solar power. Some power suppliers offer time-of-use rates that charge a different rate at each hour of the day to reflect what it costs to provide. As these are adopted more widely, the highs and lows of demand will increasingly follow the same daily pattern as wind and solar availability. Few customers think about their bill enough for small price differences to influence their behavior, but their power-hungriest devices, like their thermostat or EV charger, could schedule themselves to automatically run when the cost is lower.
Daily reconfiguration could be helpful. More helpful would be reducing demand during the few hours per year with extremely high wholesale costs, when the grid is under greatest stress. Even without time-of-use rates, customers can give their utility provider control of their most power-hungry devices, and the utility can turn them down during grid emergencies in exchange for a reimbursement on their bills. The US had 30.5 gigawatts of this capacity, called demand response, in 2022. But the headroom this gives is unreliable because participants may override the utility’s control, turning their air conditioning back up if they are too hot, for instance.
That’s what makes batteries, installed at customers’ homes, the ultimate load-managing tool. Electronic devices continue to work as normal in periods of high demand, just switching from grid to battery power. And batteries offset the demand of the entire home, not just one or two devices. Combined with rooftop solar, they can significantly reduce a home’s reliance on the grid, making room for new large loads. Perhaps the greatest advantage of all is that these at-home power resources don’t have to go through the interconnection queue; the installation process takes weeks to months, not years.
But devices like this free up relatively little grid capacity compared to the huge amounts that will be needed to power the AI revolution. They’re also relatively expensive, as the cost of installation has to be repeated on every home individually. Matching the demand of a single two-gigawatt data center, which could be handled by a handful of utility-scale projects, would require installing solar and battery systems at nearly 175,000 households.
The Data Center Trilemma
Most data centers have on-site generators as backup power for if the grid goes down because they need a stronger guarantee of reliability than the grid alone can provide. But this capacity could have another significant benefit.
Like generators, power-hungry infrastructure could come online on a connect-and-manage basis, interconnecting before all transmission upgrades are complete on the condition that it disconnects from the grid during peaks. Instead of turning off entirely, it would switch to their on-site backup power. And just like connect-and-manage lowers grid costs, flexible data centers would also lower costs for consumers by ensuring less capacity sits idle outside of peak usage.
Because grid peaks are brief and infrequent, data centers would have to disconnect for only a few hours per year. One study found that 76 gigawatts’ worth of new loads could be added (across an area covering most of the US) if these new loads were willing to disconnect during just 22 hours per year. And those 22 hours of disconnection aren’t consecutive. Each disconnection would last a few hours at most, meaning that data centers could rely on batteries rather than thermal generation. The National Renewal Energy Laboratory estimates that grid-scale, four-hour batteries can be built for $1,300 per kilowatt, as compared to around $2,500 for combined cycle gas turbines.
Disconnecting from the grid isn’t exactly ideal, but given the constraints, developers face a trilemma: data centers can be large, they can come online quickly, or they can receive firm grid service, but not all three. If large data centers want to come online quickly, they’ll have to be flexible.
The Powerful Shall Inherit
The data center boom has exposed the weakness of Western grids. But these are not the only new source of demand for electricity that we can expect in the coming years. Industrial electrification, hydrogen production, water desalination, heat pumps and air conditioning to cope with climate change, electric vehicles, and perhaps even electric aircraft are all likely to increase demand.
The twenty-first century will belong to those who make the most of these developing technologies. Without sufficient power, this will be impossible. And the main constraint to supplying power-hungry infrastructure is not power generation – it is getting that power where it is needed.
This is a blessing: insufficient power would be a much harder problem to solve. Instead, the solutions are simple. The queue of infrastructure that needs to connect to the grid should become more flexible, allowing fast-track slots to be auctioned so that the most valuable and viable projects can start sooner. And payment systems should reward both producers and large-scale consumers that are willing to disconnect for a few hours per year when demand is highest. These simple fixes can give America a power grid fit for the future.