PJM Interconnection is twenty-eight years old. It operates the largest electrical grid in the United States. From a campus in Valley Forge, Pennsylvania, it supplies electricity to 67 million people across 13 states and the District of Columbia. Every five seconds, day and night, it makes thousands of small decisions about which power plants generate, which transmission lines carry the load, and which substations step down the voltage for homes on the other end. It does this so reliably that almost no one knows it exists. That is the job. The grid that hides itself from view is the grid that’s working.
PJM was born in 1997, the year the Federal Energy Regulatory Commission (FERC) ordered the wholesale electricity market to open. It grew up in the era that followed: cost-of-service regulation, integrated resource plans, and the slow, disciplined addition of capacity matched to demand under stable growth. From roughly 2005 to 2022, American electricity demand remained mostly flat. PJM’s role during those years was to keep the lights on inside a system that largely stayed in place. It did the job, got good at the job, and built its identity around the job.
But the world that created PJM is no longer the world PJM inhabits. The organization is only just starting to acknowledge this. That realization came in a press release.
***
On the afternoon of December 17, 2025, PJM announced the results of its annual capacity auction — the three-year-ahead commitment where generators agree to supply power for a future summer, with the price determined by the gap between supply and demand. The result is best understood through three numbers: what generators were paid, how much capacity PJM needed, and how much it actually got.
The price first. The auction settled at $333.44 per megawatt-day. As recently as the 2024/2025 delivery year, the same auction had cleared at $28.92 — roughly a tenth of the current figure. Then prices surged. The 2025/2026 auction settled at $269.92, and the auction after that would have gone higher still had Pennsylvania’s governor not intervened to negotiate a price ceiling — the cap. This December’s auction hit that cap, as did the one before it. Without the cap, the December auction would have cleared at roughly $530.
Now the shortfall. A PJM capacity auction is a market: the grid operator calculates how much capacity it needs to meet its reliability standard, generators submit offers to supply that capacity, and PJM accepts offers from cheapest to most expensive until either supply meets demand or the cap stops the bidding. This time, the cap stopped it. PJM needed more than 152,000 megawatts to meet its standard. It secured 145,777. The auction came up roughly 6,500 megawatts short of what the grid required.
The reliability standard itself is a buffer above expected peak demand. PJM’s rule is that procured capacity must come in at twenty percent above forecasted peak load — enough margin to keep the lights on through extreme weather, transformer failures, and the rest of what can go wrong on a summer afternoon. This auction cleared at 14.8 percent above peak load instead of the required 20. The buffer had shrunk by roughly a quarter.
It was the first time in PJM’s twenty-eight-year history that the entire region failed to meet its reliability requirement. The auction aimed to cover an additional 5,250 megawatts of peak demand, of which 5,100 — ninety-seven percent — was due to data centers. Many of those data centers had not yet been built; they had submitted applications, signed contracts, and secured land. The grid was being asked to commit now to serving buildings that may or may not come online by summer 2027.
The press release used cautious language. The chief operating officer, during a media call, stated that the system would be “very close” to the reliability standard and that the lights would not go out. He repeated this twice. PJM’s rules require an investigation when reliability targets are missed by more than one percentage point. They had missed by 5.2. The investigation started that week.
Inside the institution, a question that had been building for two years had now been spoken aloud: whether the architecture PJM operated under was still adequate to the world it operated in. The institution now had to decide whether to recognize that the answer was no.
***
Here is the math, briefly, because the institution’s recognition turns on it. American electricity demand is growing exponentially. After two decades of flat consumption—when light bulbs got more efficient, factories moved overseas, and most homes already had the appliances they needed—demand started rising again in 2022 and then accelerated. Total US electricity use is expected to increase by about twenty-five percent by 2030. Data centers alone are projected to nearly triple their power consumption in the next decade, from around 40 gigawatts today—roughly equivalent to 20 nuclear power plants—to 106 gigawatts by 2035. Doing the math, that’s 53 nuclear power plants. This is BloombergNEF’s December 2025 estimate, which is thirty-six percent higher than their estimate from just eight months earlier. The forecast keeps getting revised upward, and those revisions are still too low.
Supply, on the other hand, is rising linearly and is now bending downward. In 2025, the United States added 53 gigawatts of new utility-scale generation, the most since 2002, with another 86 gigawatts planned for 2026. These numbers are impressive, but not impressive enough. Most of the new capacity comes from solar and battery storage. Now, I’m a fan; don’t get me wrong. But solar panels produce electricity only when the sun is shining — about 25 percent of the hours in a year. Batteries don’t generate electricity at all; they store it for later use. Data centers, by contrast, run twenty-four hours a day, seven days a week. The 86 gigawatts listed on paper is, in actual round-the-clock equivalent, closer to 25 gigawatts. The Department of Energy estimates the country needs about 100 gigawatts — the equivalent of 50 nuclear power plants — of new generating capacity by 2030 to keep up with rising demand. About half of that growth is driven by data centers alone. Planned additions will fall well short. In the last quarter of 2025, American utilities, revising their long-term plans, halved the wind and solar capacity they intend to build by 2035. The reasons include technical issues, expiring tax credits, and anticipated regulatory shifts. The result is the same regardless of the cause: when two curves diverge the way supply and demand are diverging right now, the math eventually catches up with you. And bites you.
The divergence has now become visible in PJM. ERCOT in Texas is two to three years behind. The Midwest is several years behind that. The shape of the next decade is being set—in slow motion—region by region. Train wreck comes to mind.
Five things bind, and they bind together.
Equipment first. A large power transformer — the device that boosts electricity to high voltage for transmission and then steps it down before entering the substation outside your neighborhood — takes between eighteen months and four years to produce. Globally, the factories that manufacture them operate at about ninety-eight percent capacity. There is essentially no surge capacity to meet the rising demand. The specialized steel used in the cores is made by only a handful of mills worldwide. The skilled workers who wind the copper and assemble the units have an average age of fifty-five. American utilities have ordered transformers for projects they aim to start in 2030 and beyond. Some will arrive on time, but many will not.
Then the queue. Every developer who wants to connect a new power plant or a data center to the grid must first wait in a queue. PJM has processed over 170,000 megawatts of these requests since 2023 — enough to power most of the country. About 57 gigawatts have completed studies and secured or been offered interconnection agreements. Many of those have stalled in the next stages: permitting, supply chain, and financing. The federal regulator has been working to reform the queues since 2023, and the reform is beginning to succeed. But the queue still progresses more slowly than the line behind it grows.
Then transmission. Even after a power plant gets approved and built, the wires that carry electricity from the plant to the load can take a decade or more to permit and construct under current rules. Each state has its own siting authority, and each county along the route has its own opinion. PJM’s board approved an $11.8 billion transmission expansion plan in 2024. Multi-year construction timelines mean those lines do not deliver electrons until well after the data centers they were meant to serve have already announced themselves online.
Then people. Linemen, transmission engineers, transformer manufacturing workers, substation electricians—every skilled worker needed for the grid is aging faster than apprenticeships can replace them. These trades are highly skilled, and apprenticeships take years, but the demographics of new entrants do not match the scale of the country’s buildout. Without people, the wires won’t go up, and the substations won’t be maintained, no matter what the equipment supply or regulatory framework permit.
Then politics. When supply can’t meet demand, prices go up. Residential customers in PJM’s footprint saw electricity rates increase by as much as twenty percent in the summer of 2025 — increases confirmed by the grid operator’s independent market monitor, mainly caused by data center demand. Of the $47.2 billion in capacity costs from PJM’s last three auctions, $21.3 billion — about forty-five percent — is linked to data center load forecasts, much of it for facilities not yet built. The cost of the new infrastructure is shared among all customer classes under regulatory rules designed for stable, predictable demand. Those rules are now producing outcomes — residential customers paying for capacity upgrades driven by data centers they don’t use — that the political system will not tolerate indefinitely.
It is worth saying plainly that the politics here challenge both ideological tribes. Conservative voters who believe in free markets are currently defending a cost-allocation system that socializes industrial infrastructure costs onto residential ratepayers — the opposite of what market discipline requires. Progressive voters who believe in corporate accountability are viewing the same system through climate and corporate-power lenses that hide the simpler truth: the rules were created for an era of predictable demand and now lead to outcomes they were never meant to produce. The most direct way to fix the cost-allocation pathology — making industrial customers pay for industrial infrastructure — is a market discipline approach, yet neither tribe currently embraces it. This explains, in part, why the rooms where the actual decisions get made are quiet. The fight has yet to find its political vocabulary.
The five binding constraints do not stack; they cycle. Equipment lead times increase project costs. Higher project costs lead to more rate cases—the formal proceedings by which a utility asks its state regulator for permission to change what it charges customers. More rate cases result in higher residential bills. Higher bills generate political pressure. Political pressure increases regulatory uncertainty. Regulatory uncertainty raises the cost of capital. A higher cost of capital drives up project costs. The loop closes, and once it has closed, it carves itself deeper with each turn.
***
Meanwhile, in a rented house in a college town, my 22-year-old son — who is graduating today — opens his electricity bill and does what 22-year-old college students do when faced with a financial shock: he calls home. His bills, lately, have been higher than ever, even though he and his roommates have been using less heating and air conditioning, not more. The math is supposed to work the other way. He has done what any cash-strapped kid would do, and his bill has gone up anyway.
What he does not realize is that his home state hosts the largest data center market in the country. The cost of building substations, transmission lines, and generation capacity to feed those data centers is being spread across every customer of his utility, both residential and commercial, under regulatory rules written when the biggest customers were factories that hired local workers and paid local taxes. The data centers do not do either of those at the same scale as the load they draw. The bill he has in his hand is the first installment on an infrastructure buildout he did not vote for, will not directly benefit from, and cannot opt out of.
He is not alone. His landlord opened the same envelope, which means that if her costs go up, she’ll likely raise the rent for the new tenants next year — tenants who will face higher rent and higher utility bills. His friends in the apartment complex across the parking lot got the same notice. His mom and I, in the next county, received the same notice on the same day. Our electric bill is rising as well. Not so much shocking to us as frustrating. I suspect our next stage in the cycle, come 2027 or 28, is “pissed off.”
Public support for new data center construction near residential areas has fallen sharply in every state where major projects have been announced. In Virginia, local elections in November 2025 largely centered on this issue — the Spotsylvania Board of Supervisors race was described in local media as a referendum on the data center industry, and the Prince William County Gainesville supervisor race featured both candidates competing on who would oppose development more aggressively. Similar local political dynamics are visible in Ohio and Indiana, where moratoria and zoning disputes are mounting, though the electoral patterns are less clear. Georgia is more divided — some rural communities are welcoming projects, while others are resisting them. Yet, nobody in the data center industry currently sees this as the existential political challenge it’s becoming. PJM, however, recognizes this — quietly, through stakeholder discussions, even as press releases continue to assure customers that the power will stay on.
***
PJM grew up in a particular world. So did every utility commission, every state public utility commissioner, and every federal regulator now sitting in the rooms where the next decade of American electricity will be decided. The American electric grid was reshaped through a series of decisions made between 1992 and 2010. Wholesale markets were established. The regional grid-operator system was created. Cost-of-service regulation was institutionalized. The integrated resource plan became the standard planning document. From roughly 2005 through 2022, electricity demand remained relatively flat. Efficiency gains absorbed economic growth. The system worked. The phrase “we always figure it out” became the operating assumption for every American utility executive, every state utility commissioner, and every federal regulator. It became true because it had, for years, been true. Not anymore.
The cement set during a time when figuring things out was straightforward. The current period is not that time. The same regulatory framework that brought fifteen years of stable rates and reliable power now produces forecasts off by thirty-six percent in just eight months, plans that cut wind and solar capacity in half even as demand increases, and rate cases that shift costs onto customers who can least afford them. The system isn’t failing because anyone in particular is failing at their job. It’s failing because the assumptions it was built on are no longer valid, and the people running it cannot easily see that from inside the system that taught them what to see.
The most significant sign that the architecture was beginning to recognize itself appeared in late 2025 when PJM’s Independent Market Monitor — a watchdog responsible for keeping the grid operator honest — issued a recommendation that would have been unthinkable five years earlier. The monitor stated that new data centers should be required to bring their own generating capacity instead of relying on the existing grid. The grid wasn’t designed to handle loads of this size and concentration on this timeline. Asking it to do so was creating the very cost-allocation issues that residential customers were protesting. PJM’s board responded by initiating an emergency stakeholder process — the Critical Issue Fast Path — to review the proposal. As of this writing, no decision has yet been made.
This is what an institution looks like at a turning point: still using the procedural language of the world it was built for, while the actual world it operates in pulls at its seams. The conversations that could change the structure move slowly. Federal transmission siting authority would help; it requires new legislation. Restructuring cost allocation would also help; it needs political consensus to assign costs to the customers who drive them, despite the lobbying power of those customers. Demand-flexibility authority — the right of a grid operator to refuse new connections when the grid cannot reliably support them — would be beneficial; it requires the regulatory framework to recognize that the grid is no longer a passive carrier responding to demand but an active agent that must shape that demand. Each of these changes necessitates the institution to revise its core assumptions about itself. Institutions cannot implement this quickly.
What probably happens is this.
Sometime between 2027 and 2029, two regional grids hit hard walls. PJM is already there, in slow motion. ERCOT in Texas is not far behind. New interconnection requests get rejected at scale. Data center construction slows down or shifts to other regions — including Texas, Wyoming, Mississippi, Louisiana, North Dakota, and increasingly, Norway and the Gulf states. Hyperscalers accelerate behind-the-meter generation: their own gas plants, restarted nuclear reactors, and occasionally renewables paired with batteries. Residential rates increase faster than wages. State politics turns ugly around utility commissions. A grid event in one region — a heat wave plus a transformer failure, and AI load — triggers emergency federal intervention. The federal response is partial; full federal action lacks political consensus.
By 2036, the grid that operates as one coordinated system has functionally split into three or four regional systems with substantially different rate structures, reliability standards, and demand-shaping authorities. AI capacity has moved to wherever electricity can be assembled. Some data centers planned for 2026 will never be built. Some that are built will sit partly idle, waiting for transmission upgrades that did not arrive on time. Residential bills will be substantially higher. The political settlement that emerges from this period will determine the structure of American electricity for the next thirty years or more.
A smaller probability suggests things will worsen. A widespread grid failure during a 2027 or 2028 heatwave could trigger blackouts across multiple states. Public anger may accelerate political reforms, leading to regulatory changes based on new principles. An even smaller probability says everything goes smoothly — small modular reactors come online faster than expected, AI efficiency gains reduce per-token energy by an order of magnitude, the equipment manufacturers expand quickly, transmission siting reforms succeed, and workforce pipelines meet demand. Multiple unlikely events must align perfectly, which is a rare combination. Your chances of winning the lottery might be better.
***
Like my last article on the future of AI, the leverage is mostly in the boring places, and the order matters more than the elegance.
First-stage moves — the ones the system can currently absorb — are the rule-level interventions. The federal interconnection queue reform is already underway and is beginning to distinguish between speculative and legitimate requests. Load-forecast disclosure standards — requiring utilities to publish their methodology and assumptions when submitting integrated resource plans — would close the gap between displayed plans and actual operations. PJM’s 2026 long-term load forecast revision, which included stricter scrutiny of large loads, is the immediate step.
The next set sits one altitude up. State-level demand-flexibility tariffs — allowing utilities to offer flexible rates to data-center customers in exchange for authority to curtail them during stress periods — are in place in some states and may expand. Investment in transformer manufacturing and the workforce pipeline that supports it pays off over years, not quarters, but it must begin now to influence the decade ahead. The bring-your-own-capacity proposal that PJM’s market monitor submitted for new large loads is the model: a rule change the existing stakeholder process can accept because it proceeds within the current architecture rather than being imposed from outside.
Second-stage moves — the ones the system cannot currently absorb but that the first-stage moves are designed to make receivable — are the structural ones. These include a federal transmission siting authority, restructured rate allocation that charges infrastructure costs to the customers responsible for them, and updated reliability standards that reflect the actual operating characteristics of AI data center loads. The deepest move — reform of the entire integrated resource planning architecture from a passive common-carrier model to an active demand-shaping authority — is the one the system cannot yet consider structurally. This is because it would require regulatory institutions to recognize that their own foundational assumptions are part of the problem.
Skip the first-stage moves and try to deliver the second-stage ones directly, and the regulatory immune system reads it as an attack on its legitimacy. The system entrenches. The order is important because the first-stage moves rebuild the foundation on which the second-stage moves can land. This is not a hedge; it reflects the operational consequence of interpreting the current signals as they are, not as we want them to be.
None of this is sexy. You won’t find influencers talking about it on YouTube. All of it is within reach of ordinary democratic action — state lawmakers passing demand-flexibility authority, state public utility commissioners enforcing disclosure requirements, federal regulators completing reforms they’ve already begun, congressional committees approving workforce investments, and ordinary citizens showing up at rate-case hearings. The decisions that determine who pays for the AI transition’s electricity bill will be made in rooms most Americans will never enter, by people most of us will never meet, on dockets most reporters ignore. The leverage exists. It just isn’t glamorous.
***
PJM Interconnection is twenty-eight years old. It is one of the most important institutions in American life, yet almost no one outside the industry knows its name. It was created to keep the lights on inside a particular world, and it succeeded. Now, that world has shifted. The institution has begun to notice. The noticing looks like a press release that uses careful wording and an investigation that begins quietly—like a stakeholder process initiated on an emergency timeline or a market monitor making a recommendation that would have been unthinkable five years ago. This is what early recognition looks like in an institution that cannot fully admit what it is just beginning to see.
The lights will stay on—in most places, most of the time. What is changing is who pays for them. The electrons cost what they cost; the wires cost what they cost; the transformers cost what they cost; the cost has to be allocated to somebody. The decisions being made right now in state utility commission hearings, in federal regulatory dockets, in congressional energy committees, and in the boardrooms of the equipment manufacturers will answer that question. They are being pressed into cement that will harden over the next decade. Most Americans are not yet paying attention, because the lights are still on and the bills have not yet broken anyone’s budget.
But they will, though not all at once. Region by region, year by year, rate case by rate case. By 2036, electricity will be a much larger share of household and industrial costs than it is today, and the political settlement around who pays will look very different from anything America has tried before. PJM, the institution that quietly held the lights on for twenty-eight years, will either become something different by then or it will be replaced by something different. Both possibilities are open. Neither is comfortable. The country that figured it out the last time, with stable demand and slow build, will need to figure it out a second time, with exponential demand and a substrate that cannot easily revise itself. The first time was easier. The second time is happening now, whether we’re ready for it or not.
***
US vs EU
I’ll compress this: Europe is two-to-five years behind the US on data center buildout in absolute terms, but the more important fact is that Europe is making a smaller and structurally different bet — one that trades scale and speed for sovereignty, sustainability, and regulatory coherence. The grid pathologies I described above exist in Europe too, but they’re concentrated in specific hubs (Dublin, Amsterdam, Frankfurt) and not — yet — generating the system-wide political reckoning that PJM is starting to face.
The interesting question is whose bet ages better? The US bet is large enough to break its own grid and is increasingly displacing demand to other regions, including Norway. The European bet may end up structurally dependent on US compute regardless, which is exactly the dependency the EU’s sovereignty framing is designed to prevent. Both bets have failure modes. The US failure mode is grid collapse and political backlash. The European failure mode is permanent compute-dependency on a system whose values they don’t share.
Neither is comfortable. Neither is certain. This is what complexity looks like.
Yeah, there's that--the local annoyances. At scale, the cost doesn't just go up for everyone; the lights could well go out intermittently, albeit not for everyone at once. Annoyance at scale.
Recently visited family just outside of Scranton PA and saw a lot of “NO DATA CENTER” yard signs… residents are saying these centers pollute the water, hike up electricity costs, are obnoxiously loud, and release a terrible odor into the air… thoughts?