Baseload was never a kind of power plant

The Department of Energy has reached for emergency powers more than forty times since May of last year to stop coal plants from closing, and the justification offered each time is that these plants are baseload, that they form the floor beneath the grid, that letting them go dark puts the lights out. The generation data does not support any of it. The five plants operating under those orders produced 1.5 million megawatt-hours in the first quarter of this year against 4.3 million in the same quarter a year before, two of them produced no electricity whatsoever, one sat offline waiting on a repair, and the grid they were ordered to hold up went on functioning without them for three months. An emergency that generates no power is not an emergency about power, and the word carrying those orders is not describing a physical requirement of the electrical system so much as describing who has standing to get paid.

The scale of the departure is worth stating plainly because it is easy to skim past. Section 202(c) of the Federal Power Act lets the energy secretary order generation during a grid emergency, and from 2000 through the middle of last year the department invoked it in response to roughly twenty events, most of them hurricanes. Since May 2025 it has issued more than forty orders and stalled the retirement of at least 4.4 gigawatts of coal. The Energy Department has existed for close to fifty years and had never once ordered a generating unit to keep running past its retirement date until last spring, and it has now made a routine of it, which tells you the emergency being managed is not the kind that arrives with weather.

Baseload describes the flattest part of a demand curve

Here is the thing almost nobody says out loud, and it is not controversial among people who actually operate power systems. Baseload is a term for the load that sits on the grid at all hours before you layer on the variable demand that pushes the total higher. It is a description of a shape in a demand curve. It is not a category of power plant, it was never a category of power plant, and no physical law assigns it to coal or to nuclear or to anything else.

The reason we talk as though it were a category is an accident of engineering history. Coal and nuclear plants ran flat around the clock because they were bad at doing anything else, since ramping a large thermal unit up and down is slow, expensive, and hard on the metal. Their operators discovered that the cheapest way to run them was to never stop, and so they were dispatched against the flattest part of the demand curve and picked up the name of the thing they were serving. An inflexibility got renamed as a service. The Institute for Policy Integrity makes the point exactly, which is that grid operators have long met baseload with a narrow set of resources that happen to run constantly, but describing those resources as baseload resources does not establish that they are the only way, or the cheapest way, to serve that demand. Steve Holliday, then running National Grid, called the concept outdated back in 2015, though he softened considerably on nuclear later and I would not lean the whole argument on him.

What matters is the slide. A word that describes demand gets attached to a kind of plant, the plant inherits the necessity that properly belongs to the demand, and then anyone who wants to close the plant is arguing against physics rather than against a business model. That is a remarkable piece of linguistic engineering and it is currently load-bearing in federal policy.

MISO approved Campbell's retirement in 2022 and found no reliability problem

Watch it work on one plant. MISO reviewed the retirement of the J.H. Campbell coal plant in Michigan and approved it in March 2022, finding no reliability problem with the closure. Three years later the Energy Department declared that same plant a national emergency and ordered it to stay open, and it has renewed that order repeatedly ever since, most recently through the middle of next month. Nothing about the plant changed in those three years. The grid operator that studied it did not reverse itself. The only thing that changed was who was describing it and what they wanted the description to accomplish.

The states challenging these orders in the D.C. Circuit have made the obvious point, which is that grid operator data shows adequate capacity through at least May 2027, and that in the original round the relevant operators had neither asked for federal intervention nor identified any reliability risk tied to the retirements they had themselves approved. Oral argument was heard in May. Meanwhile the Sierra Club puts the cost of thirteen of these orders at roughly $235 million, paid by ratepayers, for plants that in several cases have generated nothing at all.

PJM capacity prices rose more than tenfold in two years

Capacity prices in PJM, the largest power market in the country, went from $28.92 per megawatt-day for the 2024 to 2025 delivery year to $269.92 the following year, then to the FERC-approved cap of $329.17, and the December auction hit the raised cap of $333.44 while falling roughly 6,500 megawatts short of the reliability requirement for the first time in the market's history. PJM's independent market monitor attributed 63 percent of the increase in the 2025 to 2026 auction to data centers, which works out to about $9.3 billion recovered from the bills of the 67 million people the market serves.

The detail that turns this from a statistic into a story is what that money is actually buying. In the December auction, data center load accounted for $6.5 billion of $16.4 billion in total capacity costs, and roughly $6.2 billion of that figure is tied to data centers that have not been built and may never be. The market monitor said as gently as an independent monitor can that the uncertainty in these load forecasts raises questions about what it even means to clear an auction against them. Nevada offers the cleanest illustration of the same dynamic, where NV Energy has fielded something like 22,000 megawatts of data center inquiries against a system peak of around 9,000 megawatts and has signed agreements for about 6 gigawatts, because developers shop the same project across many markets and only one of those inquiries becomes a building. Ratepayers are being charged today for firm capacity to serve demand that exists at present only as a spreadsheet, and the word justifying the charge is baseload.

Available generation fell because gas plants underperformed

The reliability argument collapses under its own evidence, which is the part I find hardest to forgive. In the most recent auction the amount of available generation fell, and NRDC's read of why is that it fell almost entirely because of gas plant reliability problems rather than because of retirements. The resources being defended as firm are the resources that did not perform. Wind and solar additions rose 75 percent in that same auction and are the reason PJM expects to get through this year without an unacceptable risk of blackouts, even though they remain a tiny 4 percent of supply, with 85 gigawatts still sitting in the queue waiting to connect and 37 gigawatts already approved but not yet under construction.

So the story we are being told is that variable resources threaten reliability and firm thermal resources protect it, and the actual auction data says the thermal fleet broke and the variable fleet showed up. The word survives the contradiction because words do not have to clear an auction.

Duke found 76 to 126 gigawatts of headroom in the system we already have

The most expensive consequence of treating baseload as a physical necessity is what it prevents anyone from taking seriously. Researchers at Duke's Nicholas Institute modeled the 22 largest balancing authorities, which together carry 95 percent of American load, and found they could absorb somewhere between 76 and 126 gigawatts of new demand with no new generation at all, provided that demand agrees to be curtailed between a quarter of a percent and one percent of the time. At a half percent, meaning fewer than 44 hours a year, the number lands near 98 gigawatts. PJM alone accounts for 18 gigawatts of that headroom, MISO for 15, ERCOT and SPP for 10 each.

Set that beside the 4.4 gigawatts of coal being held open by federal emergency and the comparison stops being close. The honest caveats matter and I would rather state them than have someone hand them to me, since the Duke team did not model transmission constraints, which bite hard in exactly the places data centers want to build, and they calculated against peak demand without crediting reserve margins, which cuts the other way and probably understates the headroom. Follow-up work suggests that even the mildest version of flexibility, where a facility simply disconnects during the costliest one or two percent of hours, would cut new gas construction by as much as 15 percent, which is roughly 12 gigawatts of plants nobody would need to build.

Flexibility does not lose this argument because it fails. It loses because a megawatt you never build cannot be put in a rate base and earn a regulated return on it, while a gas plant can, and the word baseload is what converts that preference into a technical requirement.

Three manufacturers control roughly 75 percent of large-frame gas turbines

If you want to know who benefits, look at who is sold out. Three companies, GE Vernova, Siemens Energy, and Mitsubishi Power, control roughly 75 percent of large-frame gas turbine manufacturing, and GE Vernova crossed 100 gigawatts under contract in the first quarter of this year with reservations effectively booked through 2030 and pricing on new orders running 10 to 20 points higher per kilowatt than it was three months earlier. A handful of firms now function as the gatekeepers of American electricity supply and nobody voted for them.

The detail that keeps this honest is the one the coverage tends to leave out, which is that only about 20 percent of GE Vernova's 100 gigawatts under contract is explicitly tied to data center load, while the other 80 percent belongs to conventional utilities, independent power producers, and industrial customers. The buildout being narrated as an AI story is mostly not an AI story. AI is the justification attached to it, in the same way baseload is the justification attached to Campbell, and in both cases the word is doing work that the underlying facts would struggle to do on their own.

Zhongwei runs on 80 percent green power at 0.398 yuan per kilowatt-hour

The useful thing about China here is not that it is doing better. It is that it is running the experiment American discourse insists cannot be run, and the results are sitting in Ningxia where anyone can look at them.

The model built at the Zhongwei cluster stacks solar connected directly to the data center by dedicated line, wind purchased through bilateral green power trading, storage for peak shaving, and the main grid underneath as backstop. The cluster runs more than 80 percent green power at 0.398 yuan per kilowatt-hour with efficiency down to a PUE of 1.1, and it now holds the second largest intelligent computing capacity among the country's ten national clusters. Read that stack again and notice what it is. It is firm capacity assembled from a portfolio rather than delivered by a single always-on machine, which is precisely what firm capacity has always actually been and precisely what the American conversation says a data center cannot run on.

The concept has a name there and the name is not baseload. It is 算电协同, compute-power coordination, and it entered the government work report on March 5 of this year as a national task, the stated goal being to make electricity strengthen compute and compute in turn strengthen the power system. Under 东数西算, the East Data West Computing program launched in 2022, more than 70 percent of new computing capacity has gone to the western provinces where wind and solar are abundant and were previously being curtailed for lack of anywhere to send them. The compute moved to the electrons. Nobody had to argue about what to call the electrons first.

I am not holding this up as a model to copy, partly because I do not think it transfers and mostly because the strain in it is real. Companies spent years choosing eastern sites anyway even while paying cost premiums of 30 to 50 percent, because latency and proximity to customers turned out to matter more than power prices. That pressure is intensifying rather than easing, since inference overtook training in data volume during 2025 and the ratio is heading toward three to one, and inference is the workload that cares where it sits. The western hubs may well be building for a version of AI that is already receding.

None of that damages the point. The point is that baseload is absent from the Chinese debate not because Chinese engineers have repealed thermodynamics but because the fight the word exists to win is not being fought there. If baseload described physics, it would show up in Ningxia. It does not show up in Ningxia. It shows up in a brief before the D.C. Circuit.

Where I land

The decisions that matter are not being made in the places generating headlines. On June 18 the Federal Energy Regulatory Commission issued show cause orders to all six of its jurisdictional grid operators, defining a large load as anything above 50 megawatts connecting above 69 kilovolts, and demanding they justify or rewrite their tariffs across five areas that include cost allocation designed to stop existing customers from subsidizing upgrades they did not ask for and, notably, entirely new transmission services built for large loads willing to be flexible. Flexibility is quietly winning at the commission at the same moment baseload is winning in the press releases, which is the clearest evidence available that the two words belong to two different conversations, one about engineering and one about money.

The White House produced a Ratepayer Protection Pledge in March with seven large technology companies signing on to cover the full cost of the power they use, and it is voluntary and carries no enforcement mechanism whatsoever. Ari Peskoe, who runs Harvard's Electricity Law Initiative, pointed out the obvious problem, which is that state commissions set electricity rates and the White House sets none of them. If you want to know who is actually going to pay for the AI buildout, the answer is being written right now in rate cases and tariff filings that will never be covered by anyone.

That is what makes the word worth attacking directly. Calling something baseload converts a scheduling problem and a cost allocation fight into a claim about the physical world, and claims about the physical world do not have to be defended in a rate case. They arrive as facts, they foreclose the alternatives before anyone has costed them, and they hand the bill to whoever is not in the room. The plants sitting idle under emergency orders this spring are the cleanest proof we are likely to get that the necessity was always rhetorical, because a plant that is genuinely holding up the grid does not spend three months producing nothing while the grid holds up fine without it.

‍