Report 095 · Energy Storage
Every headline about American battery storage is written in gigawatts. A battery takes two numbers to describe, and the one everybody quotes is the less interesting of the pair. The other is sitting in the same federal spreadsheet, one column over. Add it up and the national fleet is not four hours deep. It is 2.75, and it is really two different fleets wearing one number.
On 7 August 2026 the Energy Information Administration published a piece with a headline number that got picked up everywhere: US battery storage capacity has averaged 70 percent growth over the last three years. The article is short, accurate, and worth reading. Here is its arithmetic, in its own units.
Six capacity figures, all in gigawatts. Not one megawatthour, and no mention of duration anywhere in the piece. That is not an EIA failing, it is a convention, and the whole industry writes this way. It is still a problem, because for a battery the gigawatt number alone does not tell you what the thing does.
A gas turbine can be described by one number because it will keep producing as long as you keep feeding it. A battery cannot. It has a power rating, how fast it can push, and an energy rating, how much it holds. Divide the second by the first and you get duration, which is the figure that answers the question most readers think they are asking.
EIA has said this itself, plainly, in a 2022 explainer:
We calculate a battery's duration by using the ratio of energy capacity (measured in megawatthours [MWh]) to power capacity (in MW).
And in the same piece, its definition of what duration means: "A battery's average duration is the amount of time a battery can contribute electricity at its nameplate power capacity until it runs out." That article reported that units built through 2020 averaged about 3.0 hours.
So the second number exists, EIA collects it, and EIA knows exactly what to do with it. The question is what it says now, four and a half years and roughly 45 gigawatts later.
The source behind the August article is the Preliminary Monthly Electric Generator Inventory, Form EIA-860M. The June 2026 workbook is a public download of about 14 MB. Its Operating sheet carries a column labeled Nameplate Energy Capacity (MWh), sitting a few columns to the right of Nameplate Capacity (MW). Nothing has to be estimated. You just have to add up the column nobody adds up.
Filtering that sheet to Technology = Batteries, as of June 2026:
That ratio is 2.754 hours.
Two checks before anyone leans on it. My 51.83 GW is EIA's "nearly 52 GW," and 43.6 plus 8.3 is 51.9, so the power side reconciles. Adding up the Planned sheet the same way gives 13.98 GW for the rest of 2026, 26.03 for 2027, and 13.79 for 2028, against EIA's published 14, 26 and 14. The parse is reading the same file EIA is describing.
So: 52 gigawatts, and 142 gigawatthours. If a headline is going to carry one number, the second is arguably the more honest choice, because it is the one that scales with how long the lights stay on.
Here is where this stops being a units complaint and becomes an engineering one. 2.75 hours is a real average and a misleading one, because the distribution underneath it is not a hump around 2.75. It is two spikes.
Rounding each unit's own duration to the nearest tenth of an hour, the three most common values in the operating fleet are:
Grouped into bands by power capacity: under 1 hour is 2.64 GW (5 percent), 1 to 2 hours is 13.47 GW (26 percent), 2 to 4 hours is 11.66 GW (23 percent), and 4 to 8 hours is 23.76 GW (46 percent). Above 8 hours there are four units and 20 MW, which rounds to nothing.
Nearly as many units are one-hour batteries as are four-hour batteries. Those are not the same product with a different sticker. A one-hour battery is a fast-twitch machine built to hold frequency and ride out events measured in seconds to minutes. A four-hour battery is a shift-the-evening-peak machine. They carry different cell economics, different thermal loads, different cycling behavior, different degradation curves, and they get bid into different markets. EIA's 2022 explainer found the same split by application: in 2020, more than 40 percent of energy capacity could do both grid services and load shifting, about 40 percent did load shifting only, and about 20 percent did grid services only.
Averaging a one-hour fleet and a four-hour fleet into "2.75 hours" produces a number that describes almost none of the actual hardware. It is the statistical equivalent of averaging a sprinter with a marathoner and reporting a middle-distance runner who does not exist.
This is the part I did not expect. The Planned sheet of the same workbook, the one that produces the "54 GW coming" figure, has no energy capacity column. Twenty-three columns: entity, plant, location, nameplate megawatts, summer and winter capacity, technology, fuel, prime mover, planned operating month and year, status, latitude, longitude. No megawatthours.
So the 54 GW in the pipeline cannot be converted into gigawatthours from this file. Not because nobody bothered to sum it, but because the number is not published there. Whether the coming fleet skews one-hour or four-hour, whether the next 26 gigawatts add 26 or 104 gigawatthours of stored energy, is not answerable from the monthly inventory the growth headlines are built on.
Which means the most-cited storage statistic in the country is a power number extrapolated forward with no energy number attached. If the mix shifts hard in either direction, the gigawatt trend line will not so much as flinch.
I help design the AI battery-cycling systems for a veteran-owned (HUBZone) energy-storage integrator; I do not own it and earn nothing from this link. Full policy here. I am flagging it because duration is not an academic parameter to me, it is the constraint the scheduler lives inside, and you should know I have a hand in that work when you read me insisting on it.
A cycling algorithm is, at bottom, a bet about the next few hours. On a one-hour asset the decision space is tiny and the penalty for being wrong lands immediately: commit the energy in the wrong five minutes and there is nothing left when the event you were holding for actually arrives. On a four-hour asset you can be wrong for an hour and still recover. Same chemistry, same control software, completely different risk posture, and the whole difference is that ratio in the spreadsheet column.
The same gap shows up in how people talk about backup. "52 GW of batteries" gets read, by nearly everyone outside the industry, as though the grid now has 52 GW of something you can lean on. What it has is 142 GWh, and how long that lasts depends entirely on how hard you pull. Drawn at full rated power, the whole national fleet empties in under three hours. That is not a criticism of the fleet, which is doing precisely the job it was financed to do. It is a criticism of a unit convention that lets a reader believe a duration was specified when it never was.
Nameplate energy is not usable energy. The MWh in that column is what the cells are rated to hold when new. Real projects install more cells than the contract delivers, precisely so the delivered figure survives degradation, which I went through in a PJM reference design that installs 1,030 MWh to deliver 800. So 142 GWh is an upper bound on nameplate, and usable energy at the meter today is lower.
Duration is not capacity credit. Grid planners already know all of this and do not pay in either gigawatts or gigawatthours. They pay in accredited capacity, where a four-hour battery in PJM is credited at 59 percent of nameplate and a ten-hour battery at 78. Nothing here contradicts that. It is the same physical fact arriving through the market-rules door instead of the datasheet door.
Preliminary data is preliminary. EIA-860M states plainly that it publishes preliminary monthly estimates that are later corrected without individual acknowledgment. The 2.754-hour figure is my calculation on the June 2026 release as published, not an EIA statistic, and a later vintage will move it. My expectation is that it drifts slightly longer, since 46 percent of installed power already sits in the four-hour band.
Storage is the one generation-adjacent asset class where a single capacity figure is structurally incomplete, and it is the one we habitually describe with a single capacity figure.
The fix costs nothing. The number is already collected, already public, already one column over. Quote both, or quote the hours. When you read a storage announcement that gives you megawatts and no hours, the right response is not skepticism about the project, it is recognizing that you have not yet been told what the thing is.
And when someone does give you both, look past the average. A national fleet that is 26 percent one-to-two-hour assets and 46 percent four-hour assets has no typical member. It is two industries sharing a unit.
Scope note: all figures cover utility-scale battery storage reported to EIA, meaning generators of 1 MW or greater at plants meeting the survey threshold. Behind-the-meter and residential storage is not in this dataset and is not counted anywhere above. "Nameplate energy capacity" is a rated value for new cells rather than measured deliverable energy, so every duration derived from it is a nameplate duration, not a field-verified one. The 2.754-hour fleet average, the duration bands, the per-unit modes and the planned-capacity sums are my own calculations on the June 2026 preliminary release and are not published EIA statistics; the underlying power-capacity totals were reconciled against EIA's own published figures before use. Preliminary monthly data is revised in later releases. This is general technical analysis, not procurement advice for any specific project.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.