The data center power debate is being argued in gigawatts. The binding constraint is time, and the unresolved question is who absorbs it.
By Alex Marshall
Almost every argument about data center load over the past two years has been an argument about size. How many gigawatts? How much of national generation? Whether the forecasts are inflated?
The Electric Power Research Institute now projects that data centers could account for 9 to 17 percent of US electricity generation by 2030, against 4 to 5 percent today. The interesting part of that finding is not the upper bound. It is the spread. An eight point range at national scale is not a forecast. It is a statement that the outcome has not been determined yet, and that it will be determined by decisions still being made.
Magnitude is the wrong axis to argue about. Two other things matter more, and neither of them is measured in gigawatts.
The reliability gap was always closed privately
"Five nines" means 99.999 percent availability. Roughly five minutes of unplanned downtime a year.
That has always been a facility standard. It has never been a grid standard, and no serious person in the power sector has ever claimed otherwise. Data published by the U.S. Energy Information Administration puts the average US customer at around two hours of interruption per year with major weather events excluded, and materially more when they are included. Distribution reliability is not the right comparison for a transmission connected industrial load, but the direction holds. The product the grid delivers and the product the digital infrastructure sector contracts for are separated by orders of magnitude.
That gap has always been closed on the customer side of the meter. Uninterruptible power supplies, redundant feeds, switchgear, standby plant. This is not new and it is not controversial.
What has changed is scale. At 20 megawatts, closing the reliability gap was a line item inside a construction budget. At 500 megawatts to a gigawatt, it is a capital structure decision. The same engineering problem, carried at ten or twenty times the size, stops being an engineering problem and becomes a balance sheet problem.
That is the shift the sector has not fully named. Resilience has been quietly reclassified from a design question into a financing question, and financing questions get answered by whoever is willing to own the asset.
The constraint is a delivery date
The second axis is time, and here the public data is unambiguous.
The 2026 edition of the Queued Up analysis from Lawrence Berkeley National Laboratory reports that for regions with available data, the median duration from interconnection request to commercial operation exceeded five years for generation projects that came online in 2025. Of the capacity that entered interconnection queues between 2000 and 2020, 13 percent had reached commercial operation by the end of 2025. Seventy five percent had been withdrawn.
Set that against the cycle time of the load. Compute capacity is procured in quarters. Generation and transmission are delivered in half decades. Those two clocks are not close to each other, and no amount of forecasting precision reconciles them.
This is why the scarce commodity in this market is not electricity. It is a delivery date.
Once a date carries a price, the economics reorganize around whoever can guarantee it. Margin accrues to the party that can compress the schedule, and over time the party earning that margin acquires the underlying asset. Ownership follows the constraint. That is the quiet structural story inside the buildout, and it is a story about counterparties rather than about kilowatt hours.
What the trilemma leaves out
The energy trilemma is usually presented as three competing objectives to be balanced: security of supply, affordability, and emissions. It is a useful frame and an incomplete one, because it omits the variable that is actually doing the work.
Any two of the three are achievable if you are willing to wait. It is compression of the schedule that forces the trade off. Time is not a fourth constraint sitting alongside the other three. It is the term that determines the exchange rate between them.
I have written about this as the Temporal Trilemma, and about the resulting scarcity of delivery dates as Speed to Power. [LINK: canonical definitions on alexmarshallenergy.com]
Why this matters to utilities and regulators
If the binding constraint is schedule rather than volume, then the questions in front of regulators change shape.
The useful question is not how much load is coming. It is on whose schedule it arrives, and who carries the risk if the schedule slips. Large load tariff design, cost allocation between new load and the existing ratepayer base, and the terms on which flexibility is offered in exchange for faster connection are all, in effect, negotiations over time rather than over energy.
Load flexibility is the clearest example. A customer willing to be curtailed is buying an earlier connection date with operational optionality. Whether that trade is priced well is one of the more consequential regulatory questions of the next three years, and it will be decided jurisdiction by jurisdiction rather than nationally.
On the book
I have set out this argument at length in Five Nines and Fast Power, published this year. The book examines what happens when a sector engineered around availability targets meets a power system planned around adequacy and cost, and how the resulting gap gets financed and ultimately owned.
I would rather use the comments here for the disagreement than for the book. So the question I would put to this community: in your market, is the queue the real constraint, or is it a symptom of something upstream in planning and cost allocation that a faster queue would not fix?
Alex Marshall is the author of Five Nines and Fast Power. He is Vice President of the COGEN World Coalition, a Council Member of the World Biogas Association, and a contributor to the iMasons Climate Accord Power Working Group. He writes at alexmarshallenergy.com.