THE BARATELLI INSTITUTE
· Mentoring at Scale
Home
THE BARATELLI INSTITUTE
· Mentoring at Scale
United States commercial crude sits about one per cent above its five-year average. In the same country, in the same week, the Strategic Petroleum Reserve reads its lowest since November 1982. Both are true and they are measuring different things.
Inventories are the only oil series that records what actually happened rather than what someone estimates is happening. Production is surveyed, demand is modelled, capacity is judged — but a tank either holds the barrels or it does not. That makes stock data the closest thing the market has to a ledger, and it is why a reader who checks one number a week should check this one.
It is also the series most easily misread, because there are at least three of them and they do not agree. This page carries all three side by side rather than choosing one.
Prepared by The Baratelli Institute · publication date October 24, 2026. Every figure below carries its own measurement date, which is set by whoever measured it and is often older than this page.
The three stock series that matter, each with the date it was measured and the body that measured it.
| Series | Figure | Measured | Source | What it carries |
|---|---|---|---|---|
| US commercial crude oil stocks | 412.5 million bbl | week ended October 16, 2026 | EIA Weekly Petroleum Status Report, Table 4, published October 21, 2026 | DREW 3.1 million on the week, from 415.6 million in the week ended October 9. The unrounded pair reads 412,517 thousand barrels against 415,577, a decline of 3,060 thousand barrels. 🔴 THE ROUNDING DISCLOSURE IS RETAINED, AND THIS WEEK IT PAYS OUT IN THE OTHER DIRECTION. The prior entry recorded that Table 4's published change of -8.6 million did not reproduce from the difference of the two published levels, which gave 8.5, because the EIA rounds levels and differences independently, each from the unrounded series. This week the two agree: 415.6 minus 412.5 is 3.1, and the unrounded 3,060 rounds to 3.1 as well. That agreement is a coincidence of this week's decimals and not a correction of last week's gap. A reader who takes the two figures reconciling as evidence that the earlier disagreement was an error has drawn the wrong conclusion from the easier week. Independent rounding produces agreement most weeks and disagreement some weeks, and the disclosure is worth keeping in the weeks it is not needed. 🔴 THE DRAW RUN IS NOW THREE WEEKS LONG, AND THE PRIOR ENTRY'S ACCELERATION CLAIM IS REFUTED BY THE VERY NEXT PRINT. Two entries ago this row recorded three consecutive builds of 3.0 and 0.9 million and labelled the shape as decaying. The entry before this one then recorded draws of 3.2 and 8.6 million and wrote: 'the draws are accelerating where the builds were decaying.' The third draw is 3.1 million. 3.2, then 8.6, then 3.1 is not acceleration; it is one large week between two ordinary ones, and the large week was the one that invited the pattern. The refuted sentence is retained and labelled rather than rewritten, because what it got wrong is more useful than what it got right. It carried its own hedge in the following line — 'the Institute is recording that and not extrapolating it, for the same reason the SPR row below declines twice to call two equal observations a rate' — and the hedge is what held. The noun was still wrong. Naming a shape across two observations is already an extrapolation, however carefully the next sentence declines to project it, and a reader takes the word 'accelerating' as the finding and the disclaimer as boilerplate. The lesson this row now carries forward: describe the prints, and let three numbers be three numbers. 🔴 THE YEAR-ON-YEAR COMPARISON IS WITHHELD THIS WEEK, AND THE GAP IS IN THE INSTITUTE'S READ RATHER THAN IN THE RELEASE. Against the 426.0 million of two years ago the level is 13.5 million, or 3.2%, below. The twelve-month comparison is NOT printed, because Table 4's year-ago column advances with the week and this week's year-ago column for commercial crude was not captured in the Institute's read. The prior entry's base was 423.8 million, which is the year-ago figure for the week ended October 9; beside a level dated October 16 it would be a percentage computed across two different calendars. That is precisely the Two-Clock failure this reference exists to avoid, and it is the failure mode that looks most like diligence, because the resulting sentence reads as an updated comparison and the figure in it is real. A YEAR-AGO BASE IS ITSELF A WEEKLY MEASUREMENT. It is not a property of the level it sits beside, it is not a constant, and it cannot be carried forward one row at a time while the level above it advances. The scope of the gap is stated rather than left to be inferred, because a partial read is a provenance state and the reader is owed it: the two-years-ago column WAS read for commercial crude, for the SPR and for gasoline, and the year-ago column was read for the Alaska in-transit line named below. It was not recorded for the commercial crude, Cushing or SPR lines. So three comparisons stand withdrawn on one ground — the year-on-year above, the SPR row's year-on-year below, and the Cushing premium decomposition further down — and each states that ground where it is used rather than pointing at a neighbour, because a withdrawal that defers to another row's reason dies when that reason is lifted. The cost of the admission is three comparisons a reader might want. The cost of the alternative was three cross-calendar percentages that would have read as findings. 🔴 AND THE PRIOR ENTRY'S CENTRAL FINDING CANNOT BE ADVANCED, WHICH IS A SHARPER CONSEQUENCE THAN A MISSING PERCENTAGE. The last entry's longest block decomposed both premiums crossing zero in a single week, and its conclusion was that the twelve-month reading was the more extreme in both weeks for OPPOSITE reasons — 420.3 million being the lower base one week and 423.8 million the higher the next. That finding is about the relationship between two bases over time, and it requires both bases on the current clock. With one leg unread it is neither confirmed nor refuted this week; it is simply not measurable, and it is retained and labelled as an open reading rather than restated as a current one. Its warning survives the gap intact and is the part worth carrying: a reader who takes a premium turning into a deficit as evidence about this year has attributed to the numerator what the denominator did. 🔴 The five-year-average comparison remains DROPPED from this row for a FIFTH consecutive release cycle, and the grounds are now MAINTAINED RATHER THAN RE-MEASURED, which is a weaker position than the prior entry's and is labelled as one. Last cycle the withdrawal moved from read-and-unavailable to READ AND REFUSED: the Highlights document was read, and its narrative prose block — the paragraphs carrying the five-year-average comparison — was dated SEPTEMBER 11, 2026, inside the same file as tables dated October 9, 2026. The Highlights prose block was NOT re-read for this print. So the honest statement of the ground has two branches and the figure is withheld under both. If the prose block is still on its September 11 clock, the gap against the October 16 tables beside it is thirty-five days, five weeks, and wider than when the refusal was first entered — the refusal stands and strengthens with every cycle that does not advance it. If the prose block has since advanced, the comparison is available and the Institute has not read it, in which case the correct description is not a refusal at all but an unread figure. The two branches differ in what they say about the publisher and agree in what they permit the Institute to print, which is nothing. Recording that distinction rather than carrying the stronger claim forward is the same discipline the SPR row below applies to its own withdrawal, and the reason is the same: a withdrawal is only as current as the measurement its ground rests on, and a ground inherited unmeasured is an assumption wearing the costume of a finding. Re-reading the Highlights prose date is carried as the open item for the next cycle. Cushing diverged from the country for a THIRD consecutive week, in the same direction as last week and at under half the width. The NYMEX delivery point BUILT 987 thousand barrels to 27.6 million, a 3.7% build against a 0.7% DRAW nationally — about four and a half points apart, where last week's 7.7% against 2.0% was nearly ten. Three weeks of a delivery point building into a drawing country is the observation; the narrowing is recorded without a shape being named for it, per the acceleration refutation above. 🔴 THE CUSHING PREMIUM DECOMPOSITION IS WITHHELD ON THE READ GAP NAMED ABOVE, AND THE PRIOR ENTRY'S ARITHMETIC IS RETAINED AS MEASURED ON ITS OWN WEEK. Last week this row separated a 20.9% premium over the year-ago 22.0 million into roughly eight and a half points of tank and roughly three and a half points of the base falling underneath it, and that separation is exactly why the premium cannot be restated here without both bases on the current clock: the quantity of interest is how much of the premium is the base. Printing a new percentage against last week's 22.0 million denominator would attribute the whole of this week's build to tightness and hide the one term the prior entry existed to isolate. The level is printed, the premium is not, and the reason is the finding rather than an apology for its absence. The draw sits against a refinery run whose two input measures moved in OPPOSITE directions, which is the week's most easily misread detail. Gross inputs to refineries FELL to 16,860 thousand bbl/d from 16,871, while crude inputs specifically ROSE to 16,649 from 16,615. Both are published and both are correct: gross inputs include non-crude feedstocks, and the gap between the two measures narrowed from 256 to 211 thousand bbl/d, so about 45 thousand bbl/d of non-crude feedstock came out of the run while 34 thousand bbl/d more crude went into it. A reader quoting 'refinery runs fell' and a reader quoting 'crude runs rose' are both right and are describing different quantities. More crude run is crude that does not accumulate, so the crude-input direction is consistent with the draw — but only the direction: 34 thousand bbl/d across seven days is about a quarter of a million barrels against a draw of 3.1 million, under a tenth of it, and the rest is regional and in trade. 🔴 THE SUPERSEDED CAPACITY CLAUSE REVERTS, AND A CLAUSE THAT REVERTS IS WORTH MORE THAN ONE THAT HOLDS. The entry before last recorded three consecutive weeks of operable capacity 'not moving' at 18,027 thousand bbl/d. The last entry superseded that clause because capacity rose 17 to 18,044. For the week ended October 16 capacity reads 18,044 AGAIN — unchanged — so the original attribution is available once more and the whole of the utilisation change this week is the run rate. Both prior versions are retained and labelled. The instructive part is that a clause can be correct, superseded, and correct again inside three weeks, which is a reason to state what a figure rests on every time rather than to carry a standing premise: a reader who took either version as the settled state of the world would have been wrong in one of the two following weeks. 🔴 A UNIT TRAP ON UTILISATION, AND THIS WEEK THE TWO MEASURES AGREE IN DIRECTION — WHICH IS WHEN THE TRAP IS MOST DANGEROUS. Table 2's WEEKLY utilisation is 93.4%, DOWN from 93.5%. The Highlights document's Refinery Activity block reports utilisation for the FOUR WEEKS ENDING at 93.0%, DOWN from 93.2%. Last week they pointed opposite ways and the disagreement was its own warning. Now both fall, and a reader who sees two falling numbers will be tempted to treat them as one measurement confirmed twice. They are not. They remain a single weekly observation and a four-week mean, the mean is still shedding the higher weeks of September, and agreement in direction is not corroboration between measures that cannot be compared. They must never be compared with one another, and a figure from one must never be substituted for a missing figure from the other. 🔴 THE DISTRICT PICTURE IS A GULF COAST DRAW AGAIN, AND THE DISTRICTS DO NOT SUM TO THE PUBLISHED TOTAL. PADD 3, the Gulf Coast, fell 5,041 thousand barrels to 229,318, against a national draw of 3,060 — so once again one district drew more than the whole country, offset by builds elsewhere. PADD 1 rose 1,137 to 8,550; PADD 2 rose 635 to 107,639; PADD 4 rose 786 to 23,496; PADD 5 fell 578 to 43,513. Two arithmetic disclosures, both of them the independent-rounding property above showing up at district level. The five districts sum to 412,516 thousand barrels against the 412,517 printed as the national total, and the five district changes sum to -3,061 against the published -3,060. Each is a one-thousand-barrel gap on figures the publisher rounds separately, and neither is a mistake in the release or in this row. A reader who adds the districts and comes up one short has found the rounding, not an error. And the within-district detail is the part a national figure cannot show: Cushing's 987 thousand barrel build sits INSIDE PADD 2's 635 thousand barrel build, so PADD 2 EXCLUDING Cushing FELL about 352 thousand barrels. The district that contains the delivery point rose; the rest of that district drew. A reader given only the national figure would take this week as a broad drawdown, and a reader given only PADD 2 would take the Midwest as building. Excludes the SPR. Provenance note, carried forward because it is a live publisher change rather than a figure: the EIA's combined Weekly Petroleum Full Report PDF was DISCONTINUED effective September 23, 2026, and its path now serves a discontinuation notice. This row is read off Table 4 on the publisher's live weekly path directly for that reason. The Institute reads the table, states which table, and does not treat a stale sibling as evidence about what was published. 🔴 THE SECOND ROUND OF PUBLISHER CHANGES HAS NOW TAKEN EFFECT, AND THE PRIOR ENTRY'S FORECAST OF ITS SCOPE IS BORNE OUT. The last entry recorded, on the release's own clock and a week ahead of the event, that effective October 21, 2026 Tables 11 and 12 — the spot and futures price tables — would be DISCONTINUED with their content moving to daily publication; Table 14 DISCONTINUED as derivative; and the Appendices DISCONTINUED, with Appendix A retired, Appendix B folded into the methodological documentation and Appendix C to be published as Table 4A in HTML format only. It also recorded that Table 4, which this row reads, would not be affected, and that the week ended October 9 was the last full week Tables 11 and 12 would carry. This print is the first read against the reduced table set: Table 4 served the week ended October 16 normally and this row is read off it, so the carve-out holds as stated. The consequences for the crack-spread rows, which did read Tables 11 and 12, are recorded in the crack basis note on this page rather than restated here, because a basis change belongs beside the figures it governs. A publisher notice recorded before the event and then borne out is worth leaving in place. The record should show that the change to this page's sources was anticipated and scoped in advance, not discovered afterwards when a path stopped resolving. 🔴 THE PRIOR SCHEDULE NOTICE IS RESOLVED, AND NO NEW ONE IS ASSERTED. The last entry recorded that the report would return to the usual cadence on WEDNESDAY, OCTOBER 21, 2026, after the federal closure on Monday, October 12 pushed the preceding release to Thursday, October 15. It appeared on Wednesday, October 21 as announced, and this row is read off it. That is the second consecutive schedule notice this row has carried forward and seen borne out. No schedule notice for the following release was captured in this read, so none is printed. The ordinary cadence would place it the following Wednesday, and this row will not say so, because the two notices that were worth carrying were both PUBLISHED notices and an assumed cadence is not one. The same distinction governs every withdrawal above: not in what the Institute read, and not published, are different claims, and only one of them is the Institute's to make. |
| US Strategic Petroleum Reserve | 282.7 million bbl | week ended October 16, 2026 | EIA Weekly Petroleum Status Report, Table 4, published October 21, 2026 | FELL 0.3 million barrels on the week to 282.7 million, from 283.0 million. Against the 384.6 million of two years ago it is 101.9 million, or 26.5%, below. What the weekly figure rests on, stated rather than assumed: the unrounded current level reads 282,683 thousand barrels against the prior week's 282,983, a decline of exactly 300 thousand barrels. Because the EIA rounds levels and differences independently — see the disclosure in the commercial crude row above — the Institute reads the change off the unrounded pair rather than off the rounded levels, which is the same discipline this row applied when the printed change was 0.0. 🔴 THE YEAR-AGO COMPARISON IS NOT CARRIED FORWARD, BECAUSE A YEAR-AGO BASE IS ITSELF A WEEKLY MEASUREMENT AND IT ADVANCED WITH THE WEEK. The prior entry ran this level against 407.7 million, which is the year-ago column for the week ended October 9. The year-ago column for the week ended October 16 is a different week and therefore a different number, and it was not captured in this week's read. Carrying 407.7 forward beside a figure dated October 16 would have produced a percentage computed across two different calendars — the Two-Clock failure this reference exists to avoid — and it would have looked exactly like an updated comparison. The two-years-ago base above WAS re-measured off this week's Table 4 and is printed; the year-on-year is withheld until it is read, and its absence here is a gap in the Institute's read rather than a gap in the release. The distinction matters: 'not in what we read' and 'not published' are different claims and only one of them is ours to make. 🔴 THE SEQUENCE NOW READS 1.2, 0.4, 0.4, 0.8, 0.8, 0.0, 0.3 — AND THE SEVENTH OBSERVATION HAS TO BE READ WITH THE SAME DISCIPLINE IN BOTH DIRECTIONS. This row refused, twice and on the record, to draw a line through repeated observations. The first refusal followed an error: two weeks of 0.4 million were called 'a rate, not a pause' and projected at about 21 million barrels a year, and the next print of 0.8 million refuted it. The second refusal came when 0.8 million repeated — the identical shape — and this row declined to annualise it at about 42 million barrels a year. The sixth print was ZERO, which is what vindicated both refusals. Both prior entries are retained and labelled rather than rewritten, because the instructive part is the error and the refusal, not the arithmetic, which was correct every time. The seventh print resumes the draw at 0.3 million, and the prior entry's closing sentence is what governs it: 'the same discipline that refused to extrapolate a draw refuses to extrapolate its absence.' That cuts both ways, and this is the week it has to be paid. The zero was NOT a floor, and the resumption is NOT a restart — a restart implies a programme stopped and began again, which is a mechanism, and seven weekly prints of a managed reserve cannot distinguish a mechanism from a sequence. What the seven observations support is narrower and unchanged since the first refusal: the draw is managed week to week, at a rate the operator sets rather than one a reader can project, and 0.3 million is now the smallest non-zero print in the series. 🔴 AND A FORWARD COMMITMENT THE WEEKLY SERIES STILL DOES NOT SHOW, AFTER FOUR PRINTS. The October 6, 2026 Short-Term Energy Outlook states that its forecast accounts for 40 million barrels of SPR EXCHANGES announced September 29, 2026. An exchange is not a sale — the barrels are contracted to return — and nothing in the level above reflects it yet. The four prints since that announcement have drawn 0.8 million, 0.8 million, nothing and 0.3 million: the series has moved 1.9 million in total against a 40 million commitment, which is under 5% of it. A reader should expect the weekly series and the forecast to diverge until the exchanges move, and should not read the announcement into the observed level, nor read the week at zero as evidence that the exchanges have been abandoned, nor read this week's resumption as evidence that they have begun. None of the three readings is supported. The only supported statement is that the weekly series has not yet shown them. |
| Global observed oil stocks | just under 7,900 million bbl | end-July 2026 | IEA Oil Market Report, published August 12, 2026 | The last LEVEL the Institute holds, and it is deliberately still on a July clock. Neither the September 2026 nor the October 2026 edition's public highlights restate the level; both state a stock CHANGE only. The gap is therefore now TWO months wide rather than one -- a July level above a September change -- and it widens by a month with every edition that states a change and not a level. Subtracting the changes from the level would produce a figure no publisher has printed, on a page whose entire claim is that it prints what was published. The first reading below 7.9 billion since April 2025. |
| Global observed oil stock change | fell 10 million bbl | September 2026 | IEA Oil Market Report, October 2026 edition, published October 14, 2026 (reporting month September 2026) | A small net number sitting on top of two very large offsetting ones, and the composition is the finding rather than the total. On-land inventories fell 113 million while oil on water ROSE 103 million — barrels in transit standing in for barrels in tanks, not a market in balance. Non-OECD crude stocks fell 76 million, 78% of it in China; OECD stocks fell 37 million, which is the opposite sign to August. The IEA states that on-land stocks are now at their lowest since at least 2016. Read the headline against the on-land leg rather than on its own: a ten-million net change understates the position by an order of magnitude, because the water leg is a float that can reverse in weeks and the on-land leg is what sets prompt scarcity. |
| Global observed oil stock change — the August reading | fell 95 million bbl | August 2026 | IEA Oil Market Report, September 2026 edition, published September 11, 2026 (reporting month August 2026) | 🔴 RETAINED AND LABELLED, not overwritten, per the Institute's Two-Clock rule. This was the current reading until the October 14, 2026 edition put September at a ten-million draw; the row above now carries the current month and this one keeps its own measurement date and its own citation. A draw of 3.1 mb/d across the month, and steeper than July's 69 million. The composition is the part worth reading: oil on water fell 65 million as tanker traffic out of the Middle East came under renewed attack, and non-OECD stocks drew 52 million led by China — while OECD stocks ROSE 23 million, commercial builds more than offsetting a 19 million draw in government stocks. Note what changed between the two months: in August the OECD was the one component building while the rest drew, and in September it drew with them. |
| Cumulative global stock draw | 509 million bbl | end-February to end-September 2026 | IEA Oil Market Report, October 2026 edition, published October 14, 2026 (reporting month September 2026) | An average draw of 2.4 mb/d across seven months. The rate is computed from the figure and the span — 509 across 214 days is 2.38 mb/d — and it reproduces the 2.4 mb/d the IEA itself prints, which is the check that the span named here is the span the publisher measured. 🔴 AND THIS IS A RESTATEMENT, NOT AN ADDITION. The September edition put the cumulative draw at 507 million through end-August. Adding the ten-million September draw to it gives 517, and the publisher prints 509 — so roughly eight million barrels were revised out of the earlier months in the same release that added a new one. The Institute carries both figures under their own citations rather than reconciling them: a publisher's prior-period column is a restatement, not a carry-forward, and arithmetic across two editions is the Institute's arithmetic rather than the agency's. |
| Cumulative global stock draw — the end-August reading | 507 million bbl | end-February to end-August 2026 | IEA Oil Market Report, September 2026 edition, published September 11, 2026 (reporting month August 2026) | 🔴 RETAINED AND LABELLED. Superseded as the current reading by the row above on October 14, 2026. An average draw of 2.8 mb/d across six months, against 410 million and 2.7 mb/d at end-July. Note that the EIA's independent read of the same quantity, in the balance table, is about 400 million barrels on an August clock — the two agencies agreed within ten million in July and were roughly a hundred million apart by August. |
Two Strategic Petroleum Reserve rows appear, on two different clocks, and the difference between them is the point. The 282.7 million barrel figure is the last official weekly reading the Energy Information Administration has published. The 282.7 million barrel figure is more recent and comes from press reporting rather than from the official weekly. The Institute prints both with their dates rather than substituting the newer for the older, because they are not the same kind of number: one is an official statistical release and one is a report of one. A reader who needs the authoritative figure should use the first; a reader who needs the current figure should use the second and know what it is.
The official weekly release for the week ended September 4, 2026 was pushed to Thursday, September 10 because of the federal closure on September 7. As of this page's publication date, therefore, the newest official reading is older than the newest reported one. That is a temporary condition and it will resolve on the next release, but while it holds it is exactly the situation in which a reader assumes the two numbers conflict when in fact they are consecutive.
Here is the reading a careful analyst gets from the standard indicator. United States commercial crude oil stocks stood at 412.5 million barrels in the week ended October 16, 2026, about one per cent above the five-year average for that week of the year. That is the number quoted on financial television, the number that moves the front-month contract on Wednesday mornings, and the number a screen reports when someone asks whether American crude supply is tight. It says supply is comfortable.
Here is what the same country's other tank shows. The Strategic Petroleum Reserve holds 282.7 million barrels, which is 38.9% of its 727 million barrel authorised storage capacity and the lowest level since November 1982 — before the majority of people now trading oil were born. It got there because 172 million barrels were authorised for release in March 2026, inside a 400 million barrel collective action by member countries of the International Energy Agency, the sixth and largest in that agency's history.
Those two facts are not in tension. They are causally linked, and the direction of the link is the finding: commercial crude looks normal because the strategic reserve was emptied into it. A government release does not vanish. It is sold to refiners, and it lands in commercial tanks, where it is counted as commercial inventory. The comfortable number is comfortable because the uncomfortable number paid for it.
This is a case of instrument blindness, and the Institute has documented the pattern elsewhere: a published series continues to read normally while the thing it was built to measure has changed underneath it. The commercial crude number is not wrong. It is answering the question it was designed to answer — how much crude is in commercial storage — and it is answering it accurately. It was never designed to answer the question a reader is actually asking, which is whether the country has cover.
Read the two together and the position is this. Total United States crude in storage, commercial and strategic combined, is 695.2 million barrels. The strategic share of it is 40.7%. When the authorised release completes, government stocks are expected to reach roughly 243 million barrels. At that point the release itself is no longer available as a tool, and the commercial number will have to stand on its own for the first time since March 2026.
Our reading is that the commercial-crude release should not be quoted as an indicator of American supply security for as long as a strategic release is running through it. During a release the two series are not independent, and the more visible one is the less informative one.
A chronology rather than a series, because the level is a consequence of decisions and the decisions have dates.
| Date | Event | Source |
|---|---|---|
| March 11, 2026 | Presidential directive authorises the release of 172 million barrels from the SPR over roughly 120 days, during the disruption of flows through the Strait of Hormuz. | Reported by Reuters and the Department of Energy |
| March 2026 | The release is coordinated inside a 400 million barrel collective action by IEA member countries — the sixth, and the largest, in the agency's history. | IEA |
| August 10, 2026 | SPR volumes fall below 300 million barrels, described at the time as the lowest since 1983. | CNBC, August 10, 2026 |
| September 8, 2026 | SPR reported at 285.4 million barrels, the lowest since November 1982. | Reuters / BOE Report |
| September 10, 2026 | EIA's weekly release confirms 285.4 million barrels for the week ended September 4 — the same figure the wires carried two days earlier, and 29.6% below the year-ago level of 405.2 million. | EIA Weekly Petroleum Status Report, published September 10, 2026 |
| September 16, 2026 | 285.0 million barrels for the week ended September 11, a draw of just 0.4 million against the prior week's 1.2 million. The level is 29.8% below the year-ago 405.7 million. The reserve is still falling and the rate at which it falls has slowed. | EIA Weekly Petroleum Status Report, published September 16, 2026 |
| September 23, 2026 | 284.6 million barrels for the week ended September 18, a second consecutive draw of 0.4 million. The level is 29.9% below the year-ago 406.0 million. A repeated rate is a rate: the 1.2-million weeks are not the current pace. | EIA Weekly Petroleum Status Report, Table 4, published September 23, 2026 |
| Expected at completion | US government stocks are expected to reach roughly 243 million barrels when the authorised release finishes. | Reported estimate |
The reserve's authorised storage capacity is 727 million barrels. Physical capacity and authorised capacity are not identical and the distinction matters when the reserve is refilled: drawing a salt-dome cavern down and filling it again degrades it, so a reserve that has been emptied cannot simply be refilled to its former level on the former schedule. The cost and the timetable of any refill are policy questions this page does not attempt to answer. The level, and how it got there, are matters of record and are above.
Where the two American series can be read against each other, the global series cannot be read two ways. Global observed oil stocks fell 10 million barrels in September 2026, a draw of 0.3 mb/d, and the cumulative fall end-February to end-September 2026 is 509 million barrels — an average of 2.4 mb/d sustained across seven months. The last stock level the Institute holds is just under 7,900 million barrels at end-July 2026, the first reading below 7.9 billion since April 2025.
That level is deliberately behind the change above it, and the gap is not an oversight — it is now two months wide rather than one. Neither the September 2026 edition nor the October 2026 edition restates the level; each states a change only, so the gap widens by a month with every edition that does so. Subtracting the changes from the level would produce a number no publisher has printed, on a page whose entire claim is that it prints what was published. The level will move when the IEA moves it.
The composition of the September draw is the part worth reading, and the headline is close to useless without it. Ten million barrels is a rounding error against world inventory of nearly 7.9 billion. It is the net of two movements an order of magnitude larger in opposite directions: on-land inventories fell 113 million barrels while oil on water rose 103 million. Underneath that, non-OECD crude stocks fell 76 million with roughly 78% of the fall in China, and OECD stocks fell 37 million. The IEA states that on-land stocks are now at their lowest level since at least 2016.
🔴 Note the sign change against the previous edition, because it is the substantive finding and a reader carrying August forward would have it backwards. In August the OECD was the one component building — it rose 23 million while oil on water and the non-OECD both drew. In September the OECD drew with them, and the component that rose was oil on water. The Institute retains the August reading on its own clock in the table above rather than overwriting it, so both can be read side by side.
Oil on water is the floating inventory in transit: cargoes at sea, in transit or waiting to discharge. It is the same barrels either way — what moves is where they are counted. A 103 million barrel rise on the water against a 113 million barrel fall on land is a transfer, not a build, and it says the barrels are in motion and have not arrived. The refining picture is consistent with that and is no longer uniform: global refinery utilisation runs at roughly 79%, and margins are now US margins at new highs; Europe and Asia crushed as of September 2026. Plants that can get crude are earning on it; plants that cannot are not. That is the signature of a constraint on crude arriving rather than on capacity to process it.
Weekly United States figures are measured on a Friday and published the following Wednesday. Global observed stocks are monthly, published with a lag of roughly six weeks, and revised. Comparing a weekly American reading against a monthly global one therefore compares a three-week-old measurement against a six-week-old one, and any inference about the direction of the last month rests on the older of the two.
The word observed is doing real work in the global figure. It counts stocks in locations that report, and a large share of world inventory sits in locations that do not — most consequentially China's strategic stockpile, whose level is not published. A fall in observed stocks is a fall in the observed part. It is not a census.
Prices, the supply and demand balance, refining, the two-clock rule and the plain-English glossary that serves all 13 ledgers.
Back to the hub →LedgerWhere every cent of a $4.48 gallon goes — crude, refining, distribution and tax — with each of the four components on its own clock and three of the four defined differently from how a reader assumes.
Open the ledger →LedgerSeventy-six years of crude prices deflated to constant September 2026 dollars. The 2008 records are still records in real terms and today's are not close.
Open the ledger →LedgerOne central bank, one date, one fixing time. A dollar barrel converted into 29 currencies without mixing thirty sources and thirty clocks.
Open the ledger →LedgerWhat the shape of the curve does to inventory and why it drives storage behaviour rather than describing it — then the full strip: 65 listed WTI months from the exchange's own Daily Bulletin, $97.97 front against $80.13 a year out, $17.84 of backwardation, open interest printed beside every settlement.
Open the ledger →LedgerTwo entirely different questions that share one word. The price a well needs and the price a treasury needs, kept rigorously apart — with the gaps in each named.
Open the ledger →LedgerAll 41 countries the OPEC bulletin itemises, ranked, with share of world total and reserve life computed two ways. Venezuela first at 303.701 billion barrels; the United States ninth with a reserve life of 8.4 years.
Open the ledger →LedgerEvery OPEC+ member's output for September 2026 against its implied target and its sustainable capacity. Effective spare capacity is 0.27 mb/d, or 0.26% of a 102.3 mb/d market.
Open the ledger →LedgerThe seven countries that hold the voluntary cuts — not eight, and not the full membership — their August and September 2026 decisions, and why quota discipline is currently the wrong lens.
Open the ledger →Ledger603 rigs turning in the United States, 460 drilling for oil, 1,919 worldwide. A forty per cent move in price has produced a 0.0 per cent move in the American count.
Open the ledger →LedgerCapacity measured on two clocks five months apart, crack spreads computed from same-week legs in both units the market quotes, and the one undated figure on this reference, labelled as such.
Open the ledger →LedgerDiesel at $6.276 a gallon, below a nominal record of $6.529 that was never a real one. Why distillate is tighter than gasoline, and why distillate fuel oil is not diesel.
Open the ledger →LedgerEight passages, six years of flows each, and bypass capacity that covers under a quarter of Hormuz. Plus what the industry itself expected in April, measured, and how wrong it was.
Open the ledger →Provenance. Reserve figures are transcribed from the OPEC Annual Statistical Bulletin; production, spare-capacity, stock and refining figures from the International Energy Agency Oil Market Report; United States production, refinery capacity and inventory figures from the Energy Information Administration; rig counts from the Baker Hughes North America and Worldwide rig counts; benchmark prices from published daily quotes. Each row names its source and the date the figure was measured. Where a figure was not published, or where a published series does not itemise a country, the cell reads n/d rather than carrying a fabricated estimate. Derived quantities — reserve life, shares of world total, capacity utilisation, spare capacity as a share of supply — are computed from the figures shown and are not separately sourced; they inherit the measurement dates of their inputs, and where the two inputs run on different clocks the page says so. The Baratelli Institute is a publisher operating under the Lowe v. SEC publisher exception. Nothing here is investment advice, a recommendation, or an offer to buy or sell any security or commodity. Figures change; verify against the primary source before relying on any of them.