US Energy Department keeps 760 MW Eddystone plant running as PJM points to data centre load growth
A sixth emergency order for one Pennsylvania plant shows how tight the gap between grid capacity and data centre demand has become.
The US Department of Energy has ordered the Eddystone Generating Station in Pennsylvania to keep operating through 20 November 2026, extending the life of a plant that was originally due to retire on 31 May 2025, Data Center Knowledge reported on 25 August 2026. The order, numbered 202-26-40 and issued on 21 August under Section 202(c) of the Federal Power Act, is the sixth such directive the department has applied to the facility.
Eddystone is owned by Constellation Energy. Units 3 and 4, each rated at 380 MW for a combined 760 MW, can run on natural gas or oil. The plant sits within PJM Interconnection, the grid operator serving Pennsylvania and a large part of the eastern United States.
The report ties the extension to PJM's reliability outlook. PJM projects load growth of about 25 GW, of which 15 GW, or 60 percent, is attributed to data centres. At the same time, roughly 17 GW of fossil-fuel generation retirements have been announced. Looking further ahead, PJM forecasts 32 GW of load growth by 2030, with 30 GW of that expected to come from large loads such as data centres. The system set an all-time peak of 168,158 MW on 2 July 2026.
Darryl Lawrence, Pennsylvania's consumer advocate, was quoted saying that PJM's system is fragile during extreme weather conditions. The report characterises the situation as one where rising demand, generator retirements and delays in bringing new resources online are converging, leaving the grid operator dependent on plants that were scheduled to close.
Why it matters
A federal emergency power being used for a sixth time on one plant shows how thin the margin has become between electricity supply and demand in the largest US grid region. Data centres are now the dominant driver of forecast load growth in PJM, which changes how utilities, regulators and developers negotiate. Expect more scrutiny of large-load interconnection, more pressure on data centre operators to offer flexibility, and continued delays for projects that assume power will simply be available. The physical grid, not chip supply, is increasingly the pacing constraint on AI capacity.
A federal emergency power being used for a sixth time on one plant shows how thin the margin has become between electricity supply and demand in the largest US grid region.
- Projected near-term load growth25 GW
- Data centre share of that growth15 GW
- Announced fossil-fuel retirements17 GW
- Forecast load growth by 203032 GW
Figures: PJM figures as reported by Data Center Knowledge
What you can learn from this
- Capacity and energy are different problems. A grid must match supply and demand every second, so what matters during a heatwave is available capacity in megawatts at the moment of peak, not total energy over a year. A plant that runs rarely can still be essential if it is available on the few hours that decide reliability. This is why an ageing gas-and-oil station that was scheduled to close can be ordered to stay open. Peak demand figures, such as PJM's 168,158 MW record, are the numbers that drive these decisions.
- Reserve margin and why retirements bite. Grid operators plan to hold generation capacity above expected peak demand by a margin that covers forced outages and forecasting error. When announced retirements approach the size of forecast growth, that margin shrinks, and every plant closure becomes a reliability question rather than a business one. New generation takes years to permit and build, so the timing mismatch cannot be closed quickly. The Eddystone order is a stopgap for exactly that gap.
- Emergency orders as a tool. Section 202(c) of the Federal Power Act allows the Department of Energy to direct a generator to operate during an emergency, overriding a planned retirement for a defined period. Orders are time-limited, which is why they are renewed repeatedly rather than issued once. Understanding this mechanism explains why the same plant keeps appearing in the news. It also shows that regulatory intervention is now part of the data centre power story, not just markets.
- Data centres are inelastic loads. Many industrial customers can reduce consumption when prices spike, but data centres serving latency-sensitive workloads run flat out around the clock. A gigawatt of AI capacity therefore behaves like a permanent new city on the grid. That is why forecasts single out large loads and why operators are exploring ways to make some computing, such as batch training, shiftable in time. Load flexibility is becoming a design requirement rather than a nice-to-have.
- Facility demand exceeds IT load. The power a data centre draws from the grid includes cooling, power conversion losses and lighting on top of the servers themselves. Power usage effectiveness, or PUE, expresses that ratio, so a 100 MW IT load at a PUE of 1.3 draws about 130 MW. When forecasts quote gigawatts of data centre load, ask whether the figure is IT load or facility load, because the difference is material at grid scale.
We teach this
How to use this in practice
- Read your grid operator's live dashboard. Find the real-time demand and capacity page for the grid you live on, whether PJM, ERCOT, AEMO, National Grid ESO or another operator, and record today's forecast peak, the current demand and the available reserve. Note the fuel mix as well if the page shows it, because that tells you what is actually meeting the peak. Done looks like three numbers in a note, with the date, that you can compare against the same page in a week.
- Compute facility demand from IT load. In a spreadsheet, take three hypothetical IT loads of 50, 100 and 500 MW, apply PUE values of 1.2, 1.3 and 1.5, and calculate the resulting grid draw in each case. Add a column converting the annual energy to megawatt-hours. Then add a row for a PUE of 1.1 to see how much cooling efficiency is worth at 500 MW. Done looks like a nine-cell table that makes the gap between IT load and grid demand obvious.
- Draw the supply-demand gap. Sketch a simple bar chart or diagram with forecast load growth on one side and announced retirements on the other, using the PJM figures from the report, and mark where emergency orders sit in the picture. Label each bar with the year it applies to so the timing mismatch is visible. Done looks like a one-page visual that explains why a retired plant is still running.
- Write a load-flexibility plan for something you run. For a home lab, a small server or a personal cloud project, list every workload and mark which could be delayed by an hour, a day or a week without harm. Estimate what share of total consumption is shiftable, then set one workload to run on a timer during your grid's off-peak hours. Done looks like a short table with a percentage at the bottom, which is the same exercise data centre operators are being asked to do at scale.
Sources
- DOE Keeps Eddystone Power Plant Online Amid Data Center Demand Surge — Data Center Knowledge
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report