America's biggest grid wants the right to switch data centers off

The grid pushes back

On Thursday, PJM Interconnection proposed a framework that would force data centers to run on their own backup generators when grid supply gets dangerously low [1]. PJM is the largest grid operator in the United States. It manages electricity for 67 million people, from Washington DC to Chicago [1].

The trigger is arithmetic. PJM runs a capacity auction, the market where it pays generators years in advance to promise power on the worst days. The most recent one hit its price cap of $325 per megawatt-day [1]. Even at the cap, PJM came up about 6.8 gigawatts short of the reserve it says it needs to avoid blackouts [1]. High prices are supposed to summon new power plants. They did not summon enough.

PJM cannot actually switch anyone off. It says it would need individual state governments to agree [1]. It also proposed a registry tracking where very large power users sit and how much they draw [1]. The plan is aligned with the White House’s Ratepayer Protection Pledge, a non-binding promise to keep data center costs off household bills [1].

Demand climbs regardless. The US Energy Information Administration expects national power use to rise from a record 4,195 billion kilowatt-hours in 2025 to 4,391 billion by 2027 [2].

Two-thirds of the queue is a mirage

Utilities have been asked for 1,066 gigawatts of new data center power in the US [3]. One gigawatt is roughly what a traditional nuclear reactor produces [3]. The consultancy Wood Mackenzie now expects grid operators and utilities to commit to about 28% of that [3]. The rest, it says, is phantom projects and long-shot pitches [3].

That queue is what everyone plans against, which is why questioning it moves real numbers. Texas paused new data center approvals on 3 August pending an audit [2]. The EIA promptly cut its forecast for Texas electricity growth in 2027 from 14% to 6% [2].

So build your own power

If the grid cannot say yes fast enough, the answer is to stop asking. Amazon is investing in a gas plant at its Pecos County data center in Texas [4]. Regulatory records show it would be permitted to release up to 33 million tons of carbon dioxide a year, more than any other power plant in the country [4]. Amazon says the site is “powered by new on-site generation that won’t raise electricity costs for Texas families” [4].

It will burn that gas the inefficient way. The plant relies on simple-cycle turbines, which throw away the waste heat [5]. Grid plants usually add a second steam turbine to capture that heat and emit less for the same power [5]. Even at half its permitted level, the Amazon plant would out-pollute 78 average US gas plants, on EPA figures [5]. The reason is speed. There is a years-long backlog for turbine orders and parts, so developers take what they can get, says Britt Burt of the energy research firm Industrial Info Resources [5].

Others are buying the fuel outright. Alpha Compute signed a binding term sheet for Pennsylvania land and about 1,800 acres of Marcellus gas rights, for $55 million [6]. It hopes to run a 200-megawatt campus there by the third quarter of next year [6]. Heron Power will spend $100 million turning a 286,000-square-foot warehouse south of San Jose into a transformer factory [7]. Its 5-megawatt units are the unglamorous hardware that connects a very large load to a grid; mass production is planned for late 2027 [7].

Elsewhere the shortage reads as an opening. Eskom’s chair says the South African utility has “power that we cannot sell” now that years of blackouts have ended, and is courting data centers [8]. In Europe, Mistral is pre-selling capacity it has not built yet. Multi-year enterprise commitments will underwrite 200 megawatts by the end of 2027 and a full gigawatt by 2030, the company says [9].

Who pays

Virginia became the first state to tax data centers on the electricity they consume, in late June [10]. State senator L. Louise Lucas expects up to $600 million in the coming year [10]. “How do we let the wealthiest companies in human history come to the commonwealth,” she said, “and not play by the same rules as everyone else?” [10]

The bills are already moving. Dominion’s fuel costs in Virginia are up 88% in five years, to a forecast $4.35 billion through June 2027 [11]. The utility now buys 23% of its supply on PJM’s wholesale market, up from 14% in 2021 [11]. Average monthly bills could rise as much as 13%, to $195 from $173 [11].

Local votes are turning too. Palm Beach county commissioners voted 5-1 last month to deny an expansion of a large data center [12]. It was proposed about 1,200 feet from an elementary school, and the hearing ran more than 12 hours [12].

The macro picture is stranger. Build-out spending is pushing up prices for electricity, chips, software and data center space, while corporate adoption stays uneven, which complicates the Federal Reserve’s inflation call [13]. The predicted job losses have not arrived either. Employment in the occupations most exposed is “largely stable”, Stanford’s Erika McEntarfer told the Guardian [14].

Also moving

Judge Donato gave Google one week to make third-party app stores easy to find and install inside Google Play, calling the current flow unacceptable [15]. Apple, facing the same kind of order, proposed taking 5% to 15% of purchases made through outside payment links, down from the 12% to 27% Epic called non-compliance [16]. The Ninth Circuit refused to halt the states’ social-media addiction case against Meta, ruling that Section 230 is a defense at trial rather than immunity from being sued [17]. Cisco guided fiscal 2027 revenue above estimates on AI networking demand; its shares rose 3% after hours and are up more than 56% this year [18]. And a public television station lost access to 50 terabytes covering 70 years of broadcast history after its cloud storage vendor dissolved [19].

The bottleneck nobody was watching

The under-covered story of the week is a chip nobody has been arguing about. Amazon Web Services told engineers to conserve processor cycles “at all costs” after an explosion in wait times for ordinary CPU server capacity, IEEE Spectrum reported [20].

CPUs are the general-purpose processors that run normal software; GPUs are the parallel chips that run AI models. Two years of shortage talk has been about GPUs and memory. But agents do not just think, they act, and acting runs on the CPU. AMD’s Madhu Rangarajan says seven of the eight stages in a realistic agent pipeline run entirely on the CPU [20]. Intel’s Souvik Kundu and researchers at Georgia Tech found each chip sits idle waiting on the other [20]. The GPU waits during tool calls; the CPU waits during model inference. Rescheduling the two can cut total time by up to 1.8 times under sustained load, they found [20]. One OpenAI model fired off as many as 300 actions an hour [20].

If you run agents in production, the number worth tracking is how long a task waits for a processor, not how much you spend on GPUs.