A heat wave forces the AI grid reckoning early

Climate & Energy 4 min 68 sources

A record heat wave pushed the biggest US grid to the edge, and the federal government ordered old fossil plants to run flat out — as regulators scramble to plan for the data-center demand still coming.

Key takeaways

A record heat wave pushed the biggest US grid near its 2006 peak, and the government ordered old fossil plants to run flat out to cope.
The deeper problem is AI data centers: their electricity demand is arriving faster than power plants and grid wires can be built, and regulators are scrambling to plan for a shortfall two years out.
The cost of filling that gap late — with expensive gas plants and pricier clean-energy contracts — spreads across the bills of the 65 million people the grid serves.

The AI electricity crunch stopped being a forecast this week and became an order. A record heat wave baked the eastern United States. In response, the US Department of Energy told the country’s biggest grid operator to run every fossil-fuel plant it could [6]. It also told the operator to be ready, as a last resort, to switch off data centers — to keep everyone else’s lights on [6]. It is the clearest sign yet that demand from artificial intelligence is arriving faster than the grid can grow to meet it.

The heat wave that broke the record

PJM Interconnection — the grid operator for 13 states from the mid-Atlantic to the Midwest, plus Washington, D.C. — expected Thursday’s electricity use to break its all-time summer peak of 165,563 megawatts, a record standing since 2006 [6]. A megawatt-hour is roughly the power an average US home uses in a month, so this is a whole region’s air conditioning switching on at once.

The heat was extreme. The National Weather Service warned of heat indices up to 115 degrees Fahrenheit across the central and eastern US through the July 4 weekend [6]. On Friday, forecasters put Washington, D.C. — heading toward 105 degrees — in the hottest 1.1 percent of places on the planet, alongside the Sahara and the Gobi Desert [57].

Other grids strained too. The New York grid operator expected a peak near 32,410 megawatts; ISO New England braced for “exceptionally tight” conditions and 25,850 megawatts of load [6]. Con Edison asked New Yorkers to limit air conditioning and large appliances between 2 and 10 p.m. [6].

The order to burn everything available

On Tuesday, citing the “emergency caused by the expected load stress,” Energy Secretary Chris Wright ordered PJM to maximize output from a specified list of plants [6]. Those plants are all fossil-fuel burners, including many that can run on either natural gas or fuel oil [6]. In a heat emergency, the grid reaches for whatever it already has — and what it already has is old, dirty, and dispatchable.

The same order approved something new. As a last resort before cutting power to households, PJM can now direct utilities to curtail data centers and other large customers that have their own backup generators [6]. Curtailment normally means switching off wind or solar the grid can’t use. Here it flips — the grid may switch off the AI data centers themselves rather than dim everyone’s lights.

The demand nobody has built for yet

The heat wave is the acute crisis. The chronic one is data centers. PJM has been warning for months that it faces a supply shortfall later this decade. Three forces drive it at once: old power plants retiring, too few new ones getting built, and demand rising fast — much of it from data centers running AI [3].

On Tuesday, PJM stakeholders approved a “backstop procurement” plan to cover an expected shortfall about two years out [3]. In plain terms: utilities, and possibly the data centers themselves, would ask PJM to buy a set amount of guaranteed capacity in a one-time auction outside its normal process [3]. The average cost is capped at $555 per megawatt-day, and large customers get billed for it [3]. PJM’s board aims to file the plan with federal regulators this month [3]. The coalition behind it includes Dominion Energy, Duke Energy, Exelon and the Data Center Coalition — the users and the utilities negotiating who pays to keep the lights on [3].

The bill lands on ordinary customers

Two quieter stories this week show where the cost goes. New gas plants look cheap on paper but aren’t. Pipelines, fuel and storage can add 30 percent to a project’s true cost — costs regulators often overlook when they approve them, energy group GridLab argued [5]. And as US clean-energy tax credits phase out, analysts expect the price of long-term power contracts — the deals that finance new solar and wind — to rise [2]. “That missing money has to come from somewhere to make the project pencil,” said Josh Price of research firm Crux [2].

For anyone with an electricity bill inside PJM’s footprint — about 65 million people — this is the thread that matters. A data center runs whether it’s 70 degrees or 105, and the capacity to serve it gets built by everyone. The cost of building that late, in a hurry, with old plants and expensive contracts, spreads across the whole bill.

The efficiency lever, going the other way

One more move ran against the grain of a straining grid. The US Department of Energy proposed a rule to “permanently end” longstanding energy-efficiency requirements for home appliances, calling the mandates a “scam” [43][38]. Efficiency standards are the cheapest way to cut peak demand — a more efficient air conditioner draws less power exactly when the grid is most stressed. A consumer advocate said the proposal would create “an obstacle course of restrictions” bogging down the program [43]. Loosening the rules that hold demand down, in the same week a heat wave is testing whether supply can keep up, is a choice worth noticing.

02 · Lesson · why it matters

The strain lands on whoever can't leave

When a big new user plugs into a shared system, the cost it creates rarely lands on the newcomer — it gets spread onto everyone who was already there.

Two orders that told the truth

Read this week’s emergency order carefully and it tells you who the grid protects, and in what order. When electricity ran short, the government did not tell the data centers to power down first. It told the old power plants to burn harder. And it kept a switch in reserve to cut the data centers only as a last resort — after households, after everyone else, had already been protected.

That sounds like the newcomer bearing the risk. Look again. The data centers signed contracts that let them keep running. The plants ordered to run flat out are the old ones — dirtier, more expensive, some already scheduled to close. And the cost of all of it, the emergency dispatch and the last-minute capacity, doesn’t get billed to the AI companies that created the surge. It gets spread across the bill of everyone on the grid.

The newcomer makes the mess; the incumbent cleans it up

Here is the pattern, and it runs far past electricity. A big new user plugs into a shared system — a road, a water supply, a school district, a power grid. The user is large, well-funded, and got there by choice. The system was already carrying everyone else: the households, the small businesses, the plants that were quietly keeping the lights on.

The newcomer’s arrival tightens the system. There’s less slack, higher peaks, more strain. But the newcomer rarely pays for the strain it caused. It can’t easily be made to — it negotiated hard, it’s mobile, it can threaten to build somewhere else. So the cost of the tightening lands on the party that can’t leave. That party absorbs it not because it’s fair, but because it has nowhere else to go.

Economists have a plain word for this: an externality — a cost created by one party that gets paid by another. This week you can watch it happen in real time. A data center’s electricity habit runs the same at 105 degrees as at 70. The capacity to serve that steady, enormous appetite gets built by everyone. It gets built late and in a hurry, with old plants and pricey contracts. The extra cost spreads across 65 million people’s bills — a little more each month, for years.

Why the numbers stay quiet

A cost-shift like this survives because it’s hard to see. No one sends the household a bill labeled “data center.” The grid operator negotiates a one-time purchase of guaranteed power. It caps the cost at $555 for each megawatt held ready per day, then spreads it across large customers, who pass it down. It arrives as a slightly higher rate — folded into a number you already pay, blamed on “rising costs” in general.

The party that created the cost is named nowhere on the invoice. That’s what makes it durable. A visible cost gets fought; a spread-out, unlabeled one just gets absorbed. The skill worth carrying past today is learning to ask, whenever a shared system tightens: who created this strain, and who is actually paying for it? The two are almost never the same party — and the gap between them is where the money quietly moves.

The choice hiding as a fact

It’s tempting to read this as nobody’s fault — demand rose, the grid strained, bills went up, that’s just how it works. But every step was a decision someone made. Someone decided data centers could commit huge load without first proving the power to feed it existed. Someone decided the cost of the shortfall would be socialized across all ratepayers rather than charged to the users who caused it. Someone decided, this very week, to loosen the efficiency rules that quietly hold demand down.

None of those are laws of physics. They are arrangements, and each one sets who gets to create a cost and who has to swallow it. They look like plain fact only because they were settled before this week, out of sight. An arrangement can serve the party who wrote it and still keep the lights on for everyone else — both are true. But ask, always: what looks like a natural cost here, and is it really just a cost that got moved?

What no single seat can see

The AI company sees a lease and a deadline, and a bill it isn’t asked to pay. The grid operator sees an emergency order and a shortfall to cover. The regulator sees a rule to sign. The family in the 105-degree apartment sees a rate that crept up and a reason they were never told. Each is telling the truth from where they sit. None can see the whole shape of who paid for whom.

We are all somewhere on this grid — creating the strain, absorbing it, or standing between. The cost that moved from the newcomer to the incumbent isn’t a villain’s doing. It’s what happens when a shared system meets a user big enough to negotiate its way out of the bill, while everyone else is too dispersed to notice they picked it up. Seeing that doesn’t tell you who’s right. It tells you to hold your certainty more loosely — because the cost you’re paying was often created by someone you can’t see. The part of the system you can see is almost never the part that decides.

03 · Lab · your turn

Who Pays for the Gap

Rehearse a shared system straining under a big new user, and feel the cost get shifted onto the people who can't leave.

04 · Hope · carry this

A mismatch of clocks is the fixable kind of problem — the whole grid we lean on today was once built years ahead of the demand that filled it. People who see the gap early can still close it in time.