After Refinery Capacity Holds at 18,027, One Utilization Point Is 180.27 Thousand Barrels

For the week ending September 25, 2026, the U.S. Energy Information Administration published a refinery utilization rate of 92.5 percent. Two weeks earlier the same series read 96.8 percent. The natural reading of a 4.3 point fall is that something came out of service: a unit down, a turnaround started, a plant idled. That reading is wrong, and the release itself supplies the correction. Across both weeks the capacity figure in the denominator sat at 18,027 thousand barrels per day and moved by zero. Every point of the decline came from the numerator, and the numerator is not crude oil.

This is a definitional fault line rather than an arithmetic one. The percent has a published formula, and that formula names a numerator and a denominator that are each narrower and wider than the headline suggests. Follow both terms to the documents that define them and 92.5 percent stops being a statement about how full American refineries are. It becomes a statement about how much of three input classes went into one kind of unit, measured against a capacity figure from which two kinds of downtime have already been subtracted.

The Column Header Settles the Denominator

Start with what the table calls its own columns. The weekly refinery table publishes four of them in a row: “Crude Oil Inputs | Gross Inputs | Operable Capacity | Percent Operable Utilization”. The fourth header carries the answer that is easiest to skip. It says percent operable utilization, and the capacity series feeding it is published under the title “Weekly U. S. Operable Crude Oil Distillation Capacity (Thousand Barrels per Calendar Day)”.

Two words in those labels decide the number. The capacity concept is operable, which is wider than the plant that was running. The unit is the calendar day, which is already a reduced figure rather than a physical ceiling. The U.S. Energy Information Administration, Weekly Inputs & Utilization definitions page states the formula outright: “Represents the utilization of all crude oil distillation units. The rate is calculated by dividing gross inputs to these units by the operating/operable refining capacity of the unit.”

Check it against the published values in the U.S. Energy Information Administration, Weekly Inputs & Utilization table, released September 30, 2026. Gross inputs for the week to September 25 were 16,670 thousand barrels per day; operable capacity was 18,027 thousand barrels per calendar day; 16,670 divided by 18,027 is 0.92472, which rounds to the published 92.5. Repeat the division on the week to September 18: 16,951 over 18,027 is 0.94031, published as 94.0. And on the week to September 11: 17,449 over 18,027 is 0.96794, published as 96.8. The formula reproduces every print to the tenth. The numerator is gross inputs, not crude.

One Slash, Two Capacity Concepts

That formula sentence contains a slash, and the slash is where the fault line runs. It offers two capacity concepts in one phrase — operating, then operable — and they are not the same size. The column header settles which one the published rate uses, but each concept is worth counting out, because the distance between them is the distance between a plant that is running and one that could be.

The same page defines the wider term in three parts: “The amount of capacity that, at the beginning of the period, is in operation; not in operation and not under active repair, but capable of being placed in operation within 30 days; or not in operation but under active repair that can be completed within 90 days.” Count the semicolons. Branch one is capacity in operation. Branch two is idle capacity that is not under repair and could be started inside 30 days. Branch three is idle capacity that is under active repair, where the repair can finish inside 90 days.

The entry does not stop there. Its next sentence gives the arithmetic and the units: “Operable capacity is the sum of the operating and idle capacity and is measured in barrels per calendar day or barrels per stream day.” Operating capacity is branch one; idle capacity is branches two and three. The slash therefore spans a one-branch denominator and a three-branch denominator that differ by the idle term, and the two produce different percentages from identical inputs. A refinery under active repair that will finish inside the 90-day window sits outside branch one and inside branch three, which means its capacity stays at the bottom of the published fraction while its inputs leave the top. An earlier piece on how capacity utilization is a ratio before it is an indicator made the same point about a different release.

The Numerator Holds Three Input Classes

Now the top of the fraction. The definitions page defines gross inputs as “The crude oil, unfinished oils, and natural gas plant liquids put into atmospheric crude oil distillation units.” Three input classes, and only one of them is crude oil. The other two are refinery-made intermediates and gas-plant streams. A summary that treats 92.5 percent as a crude-processing ratio has attached a three-stream percentage to a one-stream commodity.

The table keeps the distinction visible. It publishes a separate crude oil inputs column, and the page defines that term as “The total crude oil put into processing units at refineries.” Two columns, two populations, and the percentage is computed from the wider one. The arithmetic consequence is fixed and easy to carry: the denominator across these three weeks is 18,027, and 18,027 divided by 100 is 180.27 thousand barrels per day for each percentage point. Every 180.27 thousand barrels per day of unfinished oils and gas plant liquids in the numerator is one full point of published utilization that is not crude.

That is the size of the term, not a claim about its value in any given week. The week-level crude split sits in a column this piece deliberately leaves alone, for the reason given in the final section. What holds regardless is direction: a crude-only ratio over the same denominator cannot exceed the published rate, and falls below it by the non-crude volume divided by 180.27.

Calendar Day Is Already a Marked-Down Maximum

The denominator's unit does the rest of the work. The capacity series is published in thousand barrels per calendar day, and the U.S. Energy Information Administration glossary entry for barrels per calendar day draws a hard line between that unit and the physical ceiling. The unit is “The amount of input that a distillation facility can process under usual operating conditions.” The glossary then lists six limitations the calendar-day figure takes out, and the last two are downtime: “the reduction of capacity for scheduled downtime due to such conditions as routine inspection, maintenance, repairs, and turnaround; and 6. the reduction of capacity for unscheduled downtime due to such conditions as mechanical problems, repairs, and slowdowns.”

The ceiling has its own term in the same glossary. Barrels per stream day is “The maximum number of barrels of input that a distillation facility can process within a 24-hour period when running at full capacity under optimal crude and product slate conditions with no allowance for downtime.” The calendar-day figure is that stream-day figure with the six listed limitations subtracted, and the glossary says so by pointing at the stream-day entry from inside the calendar-day entry.

The consequence is easy to lose. Because scheduled and unscheduled downtime are already out of the denominator, 92.5 percent is not 92.5 percent of what the plant could do on a perfect day. It is 92.5 percent of what it is expected to do on an ordinary one. The published rate therefore sits closer to its own 100 mark than a stream-day ratio would, and a reading near 100 is not by itself evidence of a physical limit. It is evidence that inputs matched the usual-conditions estimate.

The same logic runs the other way for the weekly series. Capacity held at 18,027 across the weeks to September 11, September 18 and September 25. Gross inputs fell from 17,449 to 16,670, a drop of 779 thousand barrels per day. Divide 779 by 180.27 and the result is 4.32 percentage points; the published rates fell from 96.8 to 92.5, which is 4.3 points. The entire move is numerator. Readers who treat a weekly utilization decline as a capacity event are reading a term that stood still.

Two Other Agencies Count the Same Plants

Refineries are counted twice more in federal statistics, under capacity concepts that are not the calendar day. The Board of Governors of the Federal Reserve System, G.17 Capacity and Capacity Utilization methodology names its capacity concept “sustainable maximum output, the greatest level of output a plant can maintain within the framework of a realistic work schedule after factoring in normal downtime and assuming sufficient availability of inputs to operate the capital in place.” That is not an engineering maximum and not a 24-hour throughput figure. The same page states how the rate is built: “Monthly utilization is derived by dividing a production index by a capacity index.” Two index numbers, not two barrel counts.

Those estimates lean on a survey run by another agency, which the methodology page records: “In manufacturing, most capacity indexes are based on responses to the Census Bureau’s Quarterly Survey of Plant Capacity; these industries account for a bit less than 75 percent of total industry capacity.” That survey asks the plant itself. The U.S. Census Bureau, Quarterly Survey of Plant Capacity Utilization describes its own collection in one line: “The QPC collects value of actual production, estimated value of production at full capacity, and estimated value of production under emergency conditions.” Its utilization rate is a ratio of weighted actual production value to weighted full-capacity production value — dollars over dollars, as the respondent estimates them, with a third concept for emergency conditions standing alongside.

So one refinery feeds three rates in three units. EIA divides barrels of three input classes by barrels of operable calendar-day capacity. The Federal Reserve divides a production index by a capacity index built on a realistic work schedule; total industry utilization for August 2026 printed at 76.3 percent on that basis, in the Board of Governors of the Federal Reserve System, G.17 Industrial Production and Capacity Utilization release. Census divides a value of output by a value of output at full capacity, as the plant judges full capacity. None of the three is mistaken. They answer different questions, and only the first is a barrels-per-day statement.

Concrete Framework

Four steps turn the published percent back into something comparable.

Step one. Read the column header before the value. The header is Percent Operable Utilization. That one word fixes the denominator as operable rather than operating capacity, which means idle units inside the 30-day and 90-day windows stay at the bottom of the fraction. A unit returning from repair raises the rate by raising inputs, not by raising capacity.

Step two. Convert points into barrels before interpreting them. Divide the published capacity by 100. Across these three weeks that is 180.27 thousand barrels per day for each point. Any commentary calling a one-point move large or small should be restated in barrels first; a 4.3 point fall is 779 thousand barrels per day of inputs, which is a different sentence from the same fall described as refineries losing 4.3 percent of their capacity.

Step three. Separate numerator moves from denominator moves. The weekly capacity term is flat within a month. If capacity prints the same as last week, the whole change in the rate is inputs. Only when the capacity figure itself steps does the denominator carry part of the move, and that step is a survey update rather than a weekly observation. The same release day can therefore carry one term that is an observation and another that is a vintage, a pattern this blog traced when one EIA release day carried two different measurement bases.

Step four. Keep the percent inside its own agency. A 92.5 percent EIA rate and a 76.3 percent G.17 rate are not comparable levels, because one divides barrels of three input classes by operable calendar-day capacity while the other divides a production index by a sustainable-capacity index. For a cross-agency read, compare each series against its own history rather than against the other series. The same restriction holds for any dollar-based figure from the Census survey.

What to Watch Next Week

Three observable things over the next two weekly releases, each checkable in the same table.

The capacity term. Watch whether the operable capacity column still prints 18,027. Within a month it holds; across a month boundary it can step, because it comes from a monthly survey rather than a weekly count. The first weekly print carrying a new capacity figure is the one in which part of the rate change is genuinely a denominator change, and it is the only kind of week where a utilization move can fairly be described as capacity moving.

The gap between the two input columns. The table publishes crude oil inputs beside gross inputs. Watch the difference between them rather than either level. Divide that difference by 180.27 and the result is the count of published utilization points that are not crude. If the difference widens while the rate holds, crude runs fell and unfinished oils and gas plant liquids filled the space.

Turnaround entries. If units enter active repair during October, the operable definition says where they land. A unit whose repair can be completed within 90 days keeps its capacity in the denominator throughout. The expected signature is then a falling numerator against a flat denominator: the rate drops, 18,027 stays, and the drop converts to barrels at 180.27 for each point. A week that breaks the pattern, with capacity falling too, is the interesting one: something left the operable count altogether.

What Would Invalidate This

Three things, and two of them are gaps left open on purpose.

The crude-input column was read and not used. The combined weekly table publishes a crude oil inputs column beside gross inputs. Its week-level values did not reconcile consistently across repeated reads of the page, so no value from it is quoted or charted here. Read directly from the release file, those values turn the non-crude share of the numerator into a measured quantity rather than the arithmetic bound used above, and the crude-only correction can then be stated rather than bounded.

The stream-day comparison is unverified. The U.S. Energy Information Administration, Refinery Capacity Report Table 1 carries a barrels-per-calendar-day block and a barrels-per-stream-day block for the same distillation units, which would turn the glossary contrast into two numbers. That table was read, and its two national totals could not be assigned to the two column groups in an order consistent with the glossary, under which a stream-day figure cannot fall below the calendar-day figure for the same units. No value from it appears here. A checker should go there first and read the column group headers before the values.

The formula sentence has a slash in it. If the operable column were ever computed on operating capacity instead, the arithmetic above would stop reproducing the published prints. It reproduced them to the tenth for three consecutive weeks, which is the evidence offered here; it is not a claim about other periods, or about regional tables where the capacity term is smaller and a single unit carries more of it. Three weeks of one month also share a capacity vintage by construction, so they test the formula rather than the estimate underneath it.