On 8 September 2026 SLB published a question and answer with Paul Sims, its president of Production Systems, and it carries the three numbers that every mature-asset argument in this industry now leans on. Approximately 70% of global production comes from assets that are past peak. That share is expected to approach 80% by 2030. Conventional reserve replenishment is put at "close to 30%" [1].
Four things about that page matter before any of them is used. SLB is a service company, not an operator, so this is a supplier describing the market it sells into. No operator is named anywhere on it. The page attributes the figures to recent analysis that it does not name, so there is no document a reader can open to see what was counted. And no AI claim is attached to any of the three: the sweep brief that carries this item records no AI or machine learning content on it and classifies it as a position piece rather than a deployment or an intent. We have not re-read the page word by word, so this piece does not assert that the letters A and I never appear on it. The narrower claim is the one that matters: nothing on that page says these figures are an argument for artificial intelligence. The question below is ours and not SLB's.
Here is that question. What would have to be true for the number 70% to mean anything at all?
Three other items from the same fortnight, and who is actually speaking in each
The Journal of Petroleum Technology, the Society of Petroleum Engineers' own journal, published a survey on 4 September 2026 of how US majors are automating artificial lift [2]. It is JPT writing up what operators' engineers presented, not an operator publishing an account of itself. The hard figure in it comes from ExxonMobil's own SPE 219553, "Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence", presented in August 2024: a closed loop, well by well gas lift workflow running on more than 1,300 ExxonMobil Permian wells for a production uplift of about 2.2%, with no change to surface or downhole equipment [3]. ConocoPhillips is named in the same article for machine learning on gas lift injection rates, with no figures attached [2]. Nobody automated artificial lift this fortnight. The article is current; the work is one to two years old.
On 2 September 2026 Baker Hughes announced a multi-year contract with Oil and Gas Development Company, Pakistan's largest exploration and production company, covering more than 120 wells across OGDC's Tando Alam Oil Complex and the Pirkoh field. Phase one is evaluation and tailored redevelopment plans. Phase two is execution, and the AI-enabled chemical injection programmes sit in phase two only [4]. That item is worth carrying with its argument attached rather than laundered. Our week 36 sweep rejected it twice, once because OGDC is not on the stream's operator list and has not itself spoken, and once because the AI content sits in a phase that has not begun; the re-fetch that week confirmed that the words "artificial intelligence" and "machine learning" appear zero times on the release. Week 37 admitted it as the third leg of a mature-asset thread. Both readings are defensible and the reader should have both. It is a service company's account of a contract, and the AI in it is a plan.
On 27 August 2026 the SPE Artificial Lift Conference and Exhibition held a panel that JPT wrote up on 8 September [5]. Kreg Flowers, a production engineer at Occidental, warned that AI can create false expertise. Maggy Burns, an operations supervisor for produced water at ExxonMobil, argued for input data quality, domain knowledge and human oversight on safety-critical decisions. Araceli Rivera Mandujano of SLB warned about information overload and the difficulty of spotting a software error. The article carries no quantities at all [5]. Date it to the panel, 27 August, and read it as what it is: operator employees speaking in public, reported second hand.
Line those four up and something inverts. The biggest and most quoted number in the set, 70%, cannot be reconstructed by anyone outside SLB, because the analysis behind it is not named. The smallest number in the set, about 2.2%, can be argued with in detail, because the operator's own paper prints the population it ran on, the scope of the change and the condition that no equipment moved. The market rewards the figure whose cut is invisible. That is the subject of this piece.
Past peak is not a measurement
A peak is not observed. It is estimated, and the estimator is the argmax of a noisy series.
Take an asset producing on a long plateau. Month to month, its rate wobbles by some percentage of its own mean: wells come off for workover, a compressor trips, a cargo slips across a month end. Now suppose the underlying trend turns over and begins to decline slowly. The observed maximum of the series does not sit at the month the trend turned. It sits at whichever month the noise happened to be largest, which can be a year either side. The argmax of a flat signal plus noise is a biased, high variance estimator, and the flatter the plateau the worse it gets.
So the label "past peak" cannot be produced at all until the series has declined for long enough that the decline is separable from the wobble. Somebody, somewhere, chose how long that is. The classification is the output of a rule with a free parameter in it, and the value of the parameter is not published.
Two further facts follow, and they are what the exhibit below is built to show.
The first is a floor. A peak cannot be resolved more finely than the month to month scatter of the series containing it, in the same way that a fracture pick cannot be resolved more finely than the 3 cm of rock one pixel of a borehole image log covers. That parallel is an analogy and is treated as one below. The floor exists before any rule is applied, and no amount of confirmation discipline removes it.
The second is a lever. Every published mature-asset share is production weighted, not asset counted. A handful of very large assets carry more production than the entire long tail put together. If any of those very large assets are sitting near the rule's cut, then moving the cut by a few months moves the share further than a year of real announcements does, and it moves it in the direction that is commercially convenient.
The arithmetic
Write the asset's decline as per cent a year and its month to month scatter as per cent of its own mean. The trend gives up points of the mean each month, so the interval inside which the peak month cannot be resolved at all is the time the trend needs to give up one unit of scatter:
That is one quantity, not two, and it is the only thing about the asset that matters below. An asset with 3% scatter declining at 9% a year has months. The same asset quietened down to 1% scatter has months.
Now take the rule itself. A rule that says "declare an asset past peak once decline has been sustained for months" is running a trend test on monthly points, whether or not anyone writes it down that way. For an ordinary least squares slope over equally spaced observations the standard error is , so the test statistic collapses onto alone:
Ask for and the months below which no honest rule can classify the asset at all is the root of a cubic:
At that floor is 3.73 months at , 5.83 months at , 9.19 months at and 19.07 months at . So our 3% scatter asset on a 9% a year decline cannot be confirmed inside 9.19 months of observation no matter what window anybody publishes, and the quiet 1% version of it clears at 4.48 months.
The crest in the exhibit therefore stands at
with the confirmation window the reader sets. An asset is classified past peak when its observed months of sustained decline reach . Both halves of that are ours. No source states a window and no source states a floor, which is the whole point.
The exhibit
The object below is a population, not a surface. 240 producing assets, each one a separate solid cube standing either side of a sluice wall, each one countable. A body's volume is its production, so a handful of very large bodies carry more than the long tail: the largest single body is 7.9% of the modelled estate on its own, the largest 5 are 28.6% and the largest 10 are 42.7%, against 4.2% of the estate by count. Each body sits where a past-peak rule reads it. Up the wall runs the months of sustained decline observed since its provisional peak. Across the wall runs , the unresolvable interval, on a log scale, and into the depth runs its month to month scatter; the two together fix its decline rate, so all 3 of the model's coordinates are on screen. Through every body runs a thin glass needle that is the shell: its length is that asset's own months either side of the body. Quiet assets have almost no shell. Large assets on a long plateau are nearly all shell.
The reader drives one named control, the confirmation window, from 1 to 36 months. It does not slide along an axis of results. It drives timber boards dropped onto a fixed stone sill. The sill is each asset's own noise floor and rises to the right; the crest, the height a body must clear, is whichever is higher, the boards or the sill, which is the rule drawn as a wall. Nothing anywhere in this exhibit is read off the crest as a height. Bodies above the crest are classified past peak and counted in the glass bin to the left of the wall; the rest are in the bin to the right, whose sealed base is the share no board can reach. Two further controls sit underneath. The weighting switch turns every body between volume proportional to production and 1 identical unit cube per asset. The borderline control sets how much of the estate's production sits in the band the boards sweep, from 6 to 18 months, from 0 to 40%, and its opening setting of 30% is ours, because nobody publishes that either.
The population, the positions, the volumes, the shells and the wall are all ours and are labelled that way on the plate itself. The narrow shelf running behind the wall is not modelled. It carries measurements, drawn as solids whose physical thickness is their stated tolerance.
- Bodies in the left bin
- 133 of 240
- Shells the crest is inside
- 62, carrying 25.0%
- Carried by the last 6 months
- 30 bodies, 20.9%
- Largest of them
- 7.2%
- Undecidable at any window
- 29.2%
- Reachable at the loosest window
- 70.8%
What the exhibit shows that the prose cannot
Open it and the left bin reads 57.6%, production weighted, at a 12 month window, with 133 of the 240 bodies in it. Switch the weighting to asset counted and the same 133 bodies read 55.4%. Those two numbers are close, and that closeness is a trap, because what moves is not the level. It is the response.
Tighten the rule from 12 months to 18. Production weighted, the bin falls 20.9 points. Asset counted, it falls 12.5 points. The estate did not change. The same 30 bodies moved in both cases. What changed is that those 30 bodies are 12.5% of the population and 20.9% of its production, and the largest single one of them is 7.2% of the estate's production on its own against 0.4% of its count. That is a lever arm of 1.67 times, and it is the entire mechanism behind a production weighted share that nobody publishes a rule for.
The lever is not uniform, which is the second thing worth seeing. Loosen instead from 12 months to 6 and the production weighted bin gains 9.1 points while the asset counted bin gains 10.0 points, so on this stretch the weighting works slightly against the loosener. Whether a given nudge to the window is worth a lot or a little depends on where the very large bodies happen to be sitting relative to the cut, and that is not published either. A reader who is told only the answer cannot tell which stretch of the curve they are standing on.
Now drag the borderline control and watch what the lever is made of. At 0% the 18 month to 12 month tightening is worth exactly 0.0 points, because no production is sitting in the band the boards sweep. At 10% it is worth 7.0 points. At 30% it is worth 20.9 points. At 40% it is worth 27.8 points. The asset counted bin reads 55.4% at every one of those settings, because a unit cube is a unit cube whatever it produces. One unpublished number, the share of production sitting in the assets a rule is ambiguous about, sets how much a second unpublished number, the window, is worth.
The third thing is the shells, and it is the one that should stop a buyer. At the opening settings the crest is not cleanly past or short of every body. It is inside 62 of the 240 shells, and those 62 assets carry 25.0% of production and 25.8% of the estate. A quarter of the production in this population is labelled by the rule rather than read off the series. Those bodies are on screen, individually, at the same time, which is exactly what an averaged height field cannot do.
And there is a ceiling. 69 of the 240 bodies have observed less decline than their own noise floor requires, which means no window, however loose, can honestly classify them. They carry 29.2% of the estate's production. Run the window all the way down to 1 month and the bin stops at 70.8% production weighted and 71.2% asset counted. The remainder is not undecided. It is undecidable on the data the rule is reading.
One more body deserves naming. The single largest asset in the population carries 7.9% of production, declines at 5.6% a year, scatters by 6.56% of its own mean and has been declining for 6.4 months. Its unresolvable interval is 14.0 months and its own noise floor is 21.1 months. It cannot be classified at all, and it is the biggest thing in the estate. Any share that reports 2 buckets rather than 3 has put a body like that on one side of the line or the other, silently.
What is measured here, and the analogy it rests on
The shelf behind the box carries our own readings, and they are measurements rather than model output.
On a continuous, non-overlapping held-out strip of roughly 25 m of measured depth around 2.6 km down a vertical well the network had never trained on, our fracture model recovered roughly 90% of picks within a 10 cm depth offset, repeatably run to run [6]. The reason the window is 10 cm and not tighter is printed in the same file: at the native digital log resolution one pixel equals about 3 cm, so a plus or minus 3 cm depth uncertainty is baked in before the model makes a single decision, and asking for sub-centimetre agreement on a 3-cm-per-pixel image would be measuring noise [6]. The window is 3.33 times the floor beneath it, and it is drawn that way on the shelf: a slab with a solid core inside it that the slab can never be made thinner than.
The second reading is a tolerance ladder. On a blind July 2023 run at a 0.55 probability threshold and the 6.0 cm depth operating point, the detector scored F1 60.87% on horizontal wells, with depth mean absolute error of 2.11 cm and dip error of 11.12 degrees holding at workable levels [7]. The tolerances behind that number were agreed output by output with the operator's own expert interpreter: 2, 4 and 6 cm on depth, 1, 3 and 5 degrees on dip, 5, 10 and 15 degrees on azimuth [7]. The same file says the thing this piece is about, in its own limitations: the 6.0 cm depth operating point is one choice on a precision-recall curve, and a different tolerance moves the detection numbers [7]. The ladder is drawn as 3 nested shells with the 6.0 cm rung lit.
The transfer from a depth pick to a production classification is an analogy, and it should be stated as one rather than left for the reader to infer from a shared shelf. A fracture pick and a peak month are not the same object, the noise is not the same noise, and 3 cm of borehole image is not 3 months of production history. What carries across is the structure and only the structure: a cut whose width is stated, with a physical floor underneath it that no amount of method removes, against a cut whose width is not stated at all. The shells round the assets and the shells on the shelf are deliberately the same object drawn twice.
The contrast is the exhibit. Our readings are solids with a declared width and a floor. The market's denominator is a bin with a number on the front, filled by a crest whose position nobody published, through a population whose borderline members nobody itemised.
Where this model is weakest
The population is ours. Its absolute level means nothing, and the 57.6% on the plate is not a rival estimate of anything. Read it only as a response: what happens to a bin when a rule nobody published moves by a few months.
One coincidence needs saying out loud before somebody else says it. At the opening borderline setting the reachable ceiling is 70.8%, which is close to SLB's 70%. That is an accident of our own construction and nothing else. We did not fit anything to SLB's figure, the exhibit marks no setting as the one that reproduces it, and it never will, because no source states a rule that could be reproduced. Anyone who reads our 70.8% as a reconstruction of their 70% has made exactly the error this piece is about.
The trend test is the most arguable choice in the model. A 2 standard error threshold on an ordinary least squares slope assumes independent monthly errors, and real production series are autocorrelated, which makes the true floor worse rather than better. A real operator would also fit a decline curve rather than a straight line. Neither changes the shape of the result: there is a floor, it rises with scatter, and it is a cubic in months, so buying resolution with time is slow.
The borderline share is the softest input. It is on the slider precisely because it is unpublished, and the honest reading of the whole exhibit is the one at 0%, where tightening the rule costs 0.0 points. That is the world in which no very large asset is ambiguous. Nobody has shown that world exists.
Our own measured half is on borehole image fracture picks, not on production series. It is 1 confidential carbonate engagement, and the horizontal figures come from a blind evaluation on 5 wells [6][7].
Two things this piece deliberately does not do. It does not re-argue the ExxonMobil 2.2% as a re-test cadence question, which this stream published on 12 November, nor as a record-completeness question, which it published on 13 November. And it takes no position on the market total built on top of the 70%, which was the subject of the 5 November piece. This one is about the 70% itself.
One caution, and only one, on the OGDC contract. More than 120 wells in a contract is an ask, not a delivered and annotated corpus. Our own scarcity ledger on a carbonate programme was built around exactly that gap, and the gap, not the accuracy chart, was the artefact the programme was managed by [8]. A well count in a signed scope is a plan about data, not data.
What to ask before you use the number
What is the confirmation window, in months? Not "assets past peak" as a category, the rule that produced the category. If the answer is that there is no window because the data are annual, then the label is being produced from even fewer points than the exhibit assumes and the floor is worse.
What is the month to month scatter of the series the rule was applied to, and what share of production sits inside one scatter interval of the cut? That is the 25.0% readout in this exhibit, and it is the number that tells you how much of the share is a labelling decision rather than an observation.
Is the share production weighted or asset counted, and what do the two differ by? If a supplier cannot produce both, they have not counted anything twice, and the difference between them is where the entire lever arm lives.
How many assets are undecidable on the record available? An honest mature-asset denominator has 3 buckets, not 2. Ours has 29.2% of production in the third one.
And the test that separates a usable figure from an unusable one, which has nothing to do with how big the figure is: can you reconstruct it? ExxonMobil's about 2.2% across more than 1,300 Permian wells, reported second hand by JPT from the operator's own August 2024 paper and out of this window by more than a year, is checkable because the population, the scope and the condition that no surface or downhole equipment changed are all printed on it [2][3]. The 70% is not, because the analysis behind it is not named [1]. That asymmetry is not a vendor failing. The SPE panel of 27 August, where two operators' own engineers warned about false expertise and data quality, attached no quantity to anything either [5]. The discipline is missing across the board, which is why asking for it is worth something.
Key takeaways
- SLB's Paul Sims question and answer of 8 September 2026 carries approximately 70% of global production past peak, approaching 80% by 2030, and reserve replenishment close to 30%. SLB is a service company rather than an operator, no operator is named on the page, the figures are sourced to recent analysis the page does not name, and no AI claim is attached to any of them. The question asked in this piece is ours.
- Past peak is a label produced by a decision rule, not a reading taken off a series. A peak is the argmax of a declining trend plus operational noise, so the label only exists once somebody has chosen a confirmation window, and no published mature-asset share states one.
- There is a floor under the figure before any rule is applied. The unresolvable interval round a provisional peak is 12 times scatter over decline in months, and a 2 standard error trend test cannot confirm a peak inside the root of m^3 - m = 48 h^2. For a 3% scatter asset declining at 9% a year that is 4.00 months of unresolvable interval and a 9.19 month floor.
- The instrument is a population of 240 solid bodies held back or let through by a sluice wall, not a surface, and nothing in it is read off the crest as a height. At the opening settings the left bin reads 57.6% production weighted and 55.4% asset counted, with 133 of the 240 bodies in it, and the crest sits inside 62 shells whose assets carry 25.0% of production.
- The lever arm is the finding. Tightening the rule from 12 months to 18 costs 20.9 points production weighted and 12.5 points asset counted, from the same 30 bodies, because those bodies are 12.5% of the estate and 20.9% of its production and the largest one of them is 7.2% on its own. At a borderline share of 0% the same tightening is worth 0.0 points, so one unpublished number sets what the other is worth.
- 69 of the 240 bodies can never be classified at any window, because they have observed less decline than their own noise floor requires, and they carry 29.2% of production. An honest mature-asset denominator has 3 buckets and not 2.
- Our measured readings are on borehole image logs and state their width: roughly 90% of picks within a 10 cm depth offset on a plus or minus 3 cm per pixel floor, and F1 60.87% at the 0.55 probability threshold and the 6.0 cm depth operating point, against agreed depth rungs of 2, 4 and 6 cm. The transfer from a pick tolerance to a production classification is an analogy and is stated as one.
- The smallest figure in the fortnight is the only checkable one. ExxonMobil's about 2.2% uplift across more than 1,300 Permian wells with no change to surface or downhole equipment comes from the operator's own SPE 219553 of August 2024, reported second hand by JPT and out of window by more than a year, and it prints its population, its scope and its condition. The largest figure prints none of the three.
References
[1] SLB. Q and A with Paul Sims, President, Production Systems. 8 September 2026. First-party SLB, classified in our sweep as a partner release; no operator is named, and the 70%, the 80% by 2030 and the close to 30% are sourced on the page to recent analysis it does not name. https://www.slb.com/newsroom/features/2026/2026-0908-paul-sims-qa
[2] Journal of Petroleum Technology. Maximizing Every Barrel: Artificial Lift Embraces Automation. 4 September 2026. JPT is the Society of Petroleum Engineers' own journal reporting what operators' engineers presented; it is not an operator-published account. https://jpt.spe.org/maximizing-every-barrel-artificial-lift-embraces-automation
[3] ExxonMobil authors. SPE 219553, Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence. SPE Artificial Lift Conference and Exhibition Americas, The Woodlands, Texas, August 2024. Out of window by more than a year and used here only as the well-specified counterexample. https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/550379
[4] Baker Hughes. Baker Hughes Collaborates With OGDC To Optimize Existing Well Production, Supporting Energy Security in Pakistan. 2 September 2026. A service major's own account of its own contract; OGDC has not itself spoken, and the AI-enabled chemical injection sits in the execution phase that begins only after evaluation. https://www.bakerhughes.com/company/news/baker-hughes-collaborates-ogdc-optimize-existing-well-production-supporting-energy
[5] Journal of Petroleum Technology. New Tools, Old Lessons: Experts Stress the Importance of Artificial Lift Basics. Published 8 September 2026, reporting a panel held 27 August 2026 at the SPE Artificial Lift Conference and Exhibition. The article carries no quantities. https://jpt.spe.org/new-tools-old-lessons-experts-stress-the-importance-of-artificial-lift-basics
[6] EarthScan. Blind-Zone Validation and Repeatability on an Unseen Well. Our own case study, content/case-studies/blind-zone-validation-repeatability-unseen-well-milestone.mdx. https://earthscan.io/case-studies/blind-zone-validation-repeatability-unseen-well-milestone
[7] EarthScan. The Generalization Cliff: What Horizontal Wells Did to Our Fracture Model. Our own case study, content/case-studies/the-generalization-cliff-what-horizontal-wells-did-to-our-fracture-model.mdx. https://earthscan.io/case-studies/the-generalization-cliff-what-horizontal-wells-did-to-our-fracture-model
[8] EarthScan. 25 Wells Asked For, 8 Delivered: The Data Scarcity Gap. Our own case study, content/case-studies/25-wells-asked-for-8-delivered-the-data-scarcity-gap.mdx. https://earthscan.io/case-studies/25-wells-asked-for-8-delivered-the-data-scarcity-gap




