13 Points of Past Peak for One Point of Readable Archive: The Denominator Behind 120 Rigs and 120 Wells

Three mature-asset announcements arrived inside five weeks, and between them they published one kind of number: a count of things.

On 4 August 2026 ADNOC said that, with SLB, it had deployed an artificial intelligence enabled Real-Time Operations Center across a fleet of more than 120 onshore and offshore drilling rigs, built on SLB DrillOps, developed in the United Arab Emirates and hosted in ADNOC's own sovereign cloud [1]. On 2 September 2026, at Oil and Gas Development Company headquarters in Islamabad, Baker Hughes announced a multi-year contract covering more than 120 wells across OGDC's Tando Alam Oil Complex and the Pirkoh field, evaluated in phase one and executed in phase two with AI-enabled chemical injection programmes for flow assurance, alongside workovers and high-graded interventions [2]. On 8 September 2026 SLB published a question and answer with Paul Sims, president of Production Systems, carrying the three figures every mature-asset pitch now rests on: approximately 70% of global production comes from assets past peak, that is expected to approach 80% by 2030, and the conventional reserve replenishment rate is close to 30% [3].

A rig count, a well count, and a share of world production. Not one of the three pages states the number that decides what any of those counts is worth to a model: the share of the archive behind them that a machine can read without a person first turning paper into numbers.

Three pages, and who is speaking on each

The provenance matters more than usual here, because only one of these three is an operator making a claim about its own production, and the one everybody quotes is a service company.

ADNOC's release is first-party and it is the strongest of the three. It is ADNOC's own newsroom, quoting Musabbeh Al Kaabi, its Upstream chief executive, and its figures are ADNOC's own: 30% to 40% less engineering effort, two to three times more rigs per engineer, reporting cycles cut from days to hours, incident response 4 to 12 hours faster and one to two days of rig downtime avoided [1]. What it counts is rigs. A rig is an activity, not an asset with a file, and the fleet it belongs to is in motion. Whether the span-of-control multiple in that release is real is a question a sibling piece owns, and it is not the question here. Here the release supplies one thing: a published count of 120, on the record, from a national oil company, and the fact that the count is of rigs rather than wells.

The OGDC item is Baker Hughes' account of Baker Hughes' contract. The release is dateline Islamabad and quotes Amerino Gatti, an executive vice president at Baker Hughes, and Ahmed Hayat Lak, managing director and chief executive of OGDC [2]; a fuller trade write-up carries the same terms [4]. It is first-party for the vendor and not for the operator. It states more than 120 wells and no split between the two fields, it states no production uplift figure, and the single AI sentence sits in the phase that begins only after the phase that does not. Our own sweep classified it as intent rather than deployment for exactly that reason. What has to be true inside phase one before phase two means anything is, again, a sibling's argument. What this piece takes from it is a second published count of 120, this time of wells.

SLB's Q and A is the frame, and it has to be read as one. SLB is a service company, not an operator, and it is naming no operator at all. Our sweep classified the page as neither a deployment nor an intent but a position piece. The three figures are SLB's own, and the page sources them to recent analysis that it does not name, which means a reader cannot go and check the population, the definition of peak or the weighting behind them [3]. Nothing on the page attaches an AI programme, a well count or an uplift to any of the three. So 70% past peak is not a finding here and it is not evidence of anything about AI. It is the opening position of an axis, and the question of what a percentage uplift is worth against it is ours, not SLB's. Whether past peak is even a measurable property rather than the output of a decision rule is a separate argument that a sibling piece makes; this piece takes SLB's number at face value and asks what it buys.

There is a fourth document, and it is the only one in the set with an uplift figure attached to an operator's own wells. SPE 219553, "Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence", with ExxonMobil authors, was presented at the SPE Artificial Lift Conference and Exhibition Americas in The Woodlands, Texas, in August 2024, so it is more than two years out of this window and is dated as such throughout [5]. It reports a closed-loop, well-by-well gas lift workflow deployed to more than 1,300 ExxonMobil Permian wells for a production uplift of about 2.2%, with no change to surface or downhole equipment. That 2.2% is the paper's own number for the paper's own wells. It is never attributed below to ADNOC, to OGDC or to anyone else.

The set the uplift can actually reach

A percentage uplift does not apply to a roster. It applies to the wells on the roster that the workflow can reach, and a well is reachable only if two things are true of it at once.

It has to be the kind of well the workflow is for. A production-uplift workflow on a mature asset is aimed at the declining end of the estate, which is the fraction SLB's 70% names. And its record has to be in a form the workflow can consume. A closed-loop optimisation needs the well's own history as numbers: rates, pressures, test results, completion geometry, the last intervention. If that history is a picture of ink on paper in a filing system, the loop does not close on that well, and no model choice changes it.

So the uplift-bearing set is

$$A \;=\; N \times s \times m$$

FormulaLaTeX

Copy LaTeX

with $N$ the announced roster, $s$ the share of production that is past peak, and $m$ the share of the archive a machine can read. Releases publish $N$. A vendor position piece publishes $s$. Nobody publishes $m$.

That plain product has two consequences, and neither depends on any of the numbers in it. The first is an exchange rate. Differentiate and the two axes trade at

$$\frac{\partial A/\partial m}{\partial A/\partial s} \;=\; \frac{s}{m}$$

FormulaLaTeX

Copy LaTeX

so whichever of the two shares is currently the smaller is the axis where a point is worth more. The second is a ceiling. Readability cannot exceed one, so raising it from $m$ multiplies the set by at most $1/m$. Triage can therefore overcome a roster ratio of $1/m$ and not a whisker more, whatever anyone spends on it. More wells is the only lever without a ceiling.

Our own archive, counted rather than assumed

We have counted one archive end to end. For a Texas onshore operator we built a pipeline on the public Texas Railroad Commission record, and the first deliverable was not a model but a census: every file opened, its true format confirmed, its dimensions recorded and a stable identifier assigned. The census found 136,771 scanned TIF rasters against 7,781 digital LAS files, a ratio the case study states as roughly 17.6 to 1 [6]. A TIF is a picture of a printed log. A LAS file is the same log already in depth-indexed numbers. So of the 144,552 files opened, the share a machine could read as it stood was 7,781 over 144,552, which is 5.38%.

That is one archive, in one jurisdiction, and it is public rather than an operator's own vault. It is also the only figure of this kind we have that was counted rather than estimated, and it sits well below the 10% to 90% band a reader would guess at, which is why the instrument's readability axis opens at 4% instead of 10%. A sibling piece asked how long an archive in that condition takes to become readable. This one asks a simpler question at time zero: what is the difference between readable and unreadable worth, in wells, today.

The second figure we bring is what happens when the digitiser is pointed at the rasters. On eight genuine field scans, with a single reviewer wired into the first production frontend, the background mask came back at recall 0.97 while the two faint curves came back at 0.37 and 0.32, and the delivered curves landed at a mean absolute error of 0.11 and 0.12 against the operator's own log [7]. Lifting a raster does not hand back the whole log. The yield used below, 0.345, is the mean of those two faint-curve recalls. That mean is a modelling choice made from two measured numbers, and the instrument says so on its own plate.

The bench

The exhibit below computes the set rather than drawing it. Left to right is the past-peak production share, 60% to 90%, with SLB's 70% and its 2030 figure of 80% both on the axis. Back to front is the machine-readable share of the archive, 4% to 90%. Height is the product of the two, which is the share of the roster that can carry the uplift, and the top of the height axis is 81%.

Three more objects sit on the block. The amber sheet is parity with the one programme on the record that published an uplift: ExxonMobil's more than 1,300 Permian wells at SLB's past-peak share and our measured readability, which is 48.98 uplift-bearing wells. Where the sheet cuts the block is the set of pairs at which your roster matches that programme. The aqua arc is what the raster digitiser buys, walking your point forward along the readability axis at our modelled yield. The grey rail in front of the block is the roster, 100 wells at the left to 10,000 at the right, carrying the three named counts. ADNOC's 120 and OGDC's 120 land on the same point of a logarithmic rail and share one label, which names the first as rigs and the second as wells, because a rig is not an asset with an archive and this instrument will not silently convert one into the other.

What is measured, what is quoted and what is modelled is printed on the plate. Measured, ours: the 5.38% and the curve recalls. Quoted, SLB's, and never computed with: the 70%. Measured, ExxonMobil's, and out of window: the 2.2%. Modelled, ours: the product itself, the digitiser yield of 0.345, and every readout that follows from them.

past-peak production share, 60% to 90%: 70.0%machine-readable as found, 4% to 90%: 5.38%roster, 100 to 10,000 wells: 1,300raster remainder put through the digitiser, 0% to 100%: 0%

The rail in front of the block is the roster, 100 wells at the left to 10,000 at the right. ADNOC and SLB sit at 120 rigs, which is an activity count and not a well count; Baker Hughes and OGDC at 120 wells; ExxonMobil at 1,300+ Permian wells, the only one of the three with an uplift figure of its own. Fully digitised at our recorded yield, a 120-well roster reaches 7.06 times its starting set and no further.

What the bench shows

At the opening settings, a 1,300-well roster at 70% past peak and 5.38% readable has an addressable share of 3.77%, which is 48.98 wells. Applying ExxonMobil's own 2.2% to those 48.98 wells gives 1.08 whole wells of annual flow. The reading a headline invites, 2.2% of 1,300 wells, gives 28.60. The gap between the two is 26.54 times, and every bit of it is the two factors the release does not print.

Now use the two shares against each other. Push the past-peak control up by one point and the set gains 0.70 wells. Push the readability control up by one point and it gains 9.10. The readout says 13.00 points of past peak per point of readability, and that is not a calibration: it is $s/m$ at an operating point where $m$ is very small. The consequence is worth stating slowly. SLB's whole 2026-to-2030 move, from 70% past peak to 80%, is worth 7.00 uplift-bearing wells on that roster. One point of machine-readability is worth 9.10. The denominator the entire sector is quoting will take four years to deliver less than a single point of the denominator nobody quotes.

Then drag the digitiser. At 25% of the raster remainder lifted, readability reads 13.54%. At 50%, 21.70%. At 75%, 29.87%. At 100%, with every raster in the archive put through, it reads 38.03% and stops there, because our own recorded curve recall does not give the whole log back. That ceiling is 7.06 times the starting set. The triage ceiling, which is what raising readability to a perfect 1 would buy, is 18.58 times, so the digitiser at our yield reaches 38.0% of what perfect triage would have been worth.

The crossover follows from the ceiling, and it is the finding this piece exists for. Drag the roster down to 120 wells, the OGDC count, and the parity readout turns amber and says 58.3%. That is the readability those 120 wells would need to match the 1,300-well programme, and it is on the block: 58.3% is a perfectly ordinary point on an axis that runs to 90%. The readout beside it is the one that closes the door. The digitiser ceiling from a 5.38% archive is 38.03%, and reaching 58.3% would need 1.62 of the raster remainder, which is more remainder than exists. The parity curve is visible and unreachable at the same time, which is what the amber is for. The largest roster a fully digitised 120-well programme can match is 847.71 wells. As found, those 120 wells carry 4.52 uplift-bearing wells and 0.10 whole wells of annual flow; fully digitised, 31.94 and 0.70. Set against the reference programme's 48.98, that is 65.2% of it and a shortfall of 17.04 wells, after spending everything the digitiser has.

Run the crossover the other way and it is a procurement rule. At 10% readable, 699.77 wells match the untriaged 1,300. At 20%, 349.88. At the 38.03% our digitiser reaches, 184.02. At 50%, 139.95. So a buyer choosing between a big roster and a readable one has a clean exchange: every halving of the roster has to be paid for with a doubling of readability, and readability runs out first.

Where the model is weakest, and what survives

Three objections are fair, and two of them are serious.

The first is that $s$ and $m$ are not independent. The argument treats past peak and machine-readable as two separate draws on the same well, and in reality the oldest assets are both the most likely to be past peak and the most likely to have a paper archive, so the two are negatively correlated and the true set is smaller than the product. That objection strengthens the conclusion rather than weakening it, which is why the product is kept: it is the generous case.

The second is the readability figure itself. Our 5.38% is a public state archive in Texas, and an operator's own vault on a field it has run for forty years may be in far better order, particularly for wells completed since digital acquisition became routine. That is exactly why readability is a control and not a constant, and why the axis reaches 90%. What does not move with the reader's number is the exchange rate: whatever $m$ is, a point of readability is worth $s/m$ points of past peak, so readability is the cheaper axis for any archive less readable than the estate is past peak. At SLB's 70% the two axes cost the same at 70.0% readable, and readability is still worth 3.50 points of past peak at 20.0% and 1.40 at 50.0%.

The third is the digitiser yield. Taking the mean of two faint-curve recalls from eight scans as the yield of a whole archive pass is the softest number in the piece, and it is ours. Move it to the background recall of 0.97 and the ceiling from our archive is 97.16%, worth 18.05 times; move it to the lower faint-curve recall of 0.32 and the ceiling is 35.66%, worth 6.62 times. The ceiling's existence does not depend on the value: any yield below 1 leaves a ceiling, and any ceiling makes the roster ratio triage can overcome finite. That is the part of the argument that survives every setting of the control.

One caveat on the instrument itself. The named counts on the rail are positions on an axis and not equivalent claims, and the two 120s share a single mark because they share a point. ADNOC's 120 is rigs. OGDC's 120 is wells. ExxonMobil's 1,300 is wells with an uplift figure attached to them. Converting a rig count into a well count needs a wells-per-rig-year figure that none of the three pages publishes, and the exhibit refuses to invent one.

Four questions before signing a mature-asset AI contract

What share of the wells on the roster have their production history in numbers rather than in images, counted rather than estimated? Not the share that exists somewhere, and not the share the data room says is available: the share a pipeline can read without a human intervening. On our own archive that number was 5.38%, and nobody knew it until the census was run.

What is the roster's past-peak share on your own definition, and who chose the definition? SLB's 70% is a global figure from a service company, sourced to analysis its page does not name [3]. Yours is a property of your estate and you can count it.

What does the vendor expect the uplift to be, and on which of the two sets? The only uplift figure on the record in this cluster is ExxonMobil's about 2.2%, in ExxonMobil's own paper, on ExxonMobil's own more than 1,300 Permian wells, in August 2024 [5]. If a proposal quotes a percentage against the full roster rather than against the readable past-peak subset, the difference at our measured readability is a factor of 26.54.

And what is the ceiling on triage? Ask the digitiser vendor for a recall figure per curve class, not an accuracy figure per scan. At the mean of our two recorded faint-curve recalls the ceiling from a 5.38% archive is 38.03%, which buys 7.06 times and no more, and that is what decides whether a small readable programme can ever match a large unreadable one. Reading reservoir, wells, lift, chemistry and facilities as one system is what makes an archive readable in the first place, which is the EarthScan 360° AI argument; the raster share that has to be lifted before the readability axis moves at all is where the ES Raster Digitizer sits.

Key takeaways

1. 1Three mature-asset announcements in five weeks published one count each and nothing else. ADNOC and SLB, 4 August 2026, more than 120 drilling rigs, which is an activity count and not a well count. Baker Hughes, 2 September 2026, more than 120 OGDC wells, with the AI-enabled chemical injection in phase two, no uplift figure and no split between the two fields. SLB, 8 September 2026, approximately 70% of production past peak and approaching 80% by 2030, sourced to recent analysis its own page does not name, with no AI programme attached to any of the three figures.
2. 2A percentage uplift reaches the roster times the past-peak share times the machine-readable share of the archive, and only the first of those three is ever printed. On the one archive we counted end to end, 7,781 LAS files against 136,771 TIF rasters in 144,552 files, the readable share is 5.38%.
3. 3At 1,300 wells, SLB's 70% and our 5.38%, the addressable share is 3.77%, which is 48.98 uplift-bearing wells. ExxonMobil's own 2.2% on those 48.98 wells is 1.08 whole wells of annual flow against a headline reading of 28.60, an overstatement of 26.54 times.
4. 4The exchange rate between the two axes is the past-peak share over the readable share, so at that operating point a point of machine-readability is worth 13.00 points of past peak. One point of readability buys 9.10 uplift-bearing wells; the whole move from 70% past peak to 80% by 2030 buys 7.00.
5. 5Triage has a ceiling and more wells does not. Lifting every raster in our archive at a modelled yield of 0.345, the mean of our two recorded faint-curve recalls, reaches 38.03% readable and stops, which is 7.06 times the starting set against the 18.58 times perfect triage would give. A fully digitised 120-well programme therefore matches a roster of 847.71 wells and can never reach 1,300, and 120 wells would need 58.3% readable to hold parity.
6. 6The crossover is a procurement rule: at 10% readable, 699.77 wells match 1,300 untriaged wells; at 20%, 349.88; at 38.03%, 184.02. Every halving of the roster has to be paid for with a doubling of readability, and readability runs out first.

References

[1] ADNOC. ADNOC and SLB Deploy AI Platform Across Over 120 Drilling Rigs to Strengthen Upstream Performance. 4 August 2026. https://www.adnoc.ae/en/news-and-media/press-releases/2026/adnoc-and-slb-deploy-ai-platform-across-over-120-drilling-rigs-to-strengthen-upstream-performance First-party, ADNOC's own newsroom. The count is of drilling rigs, not wells.

[2] Baker Hughes. Baker Hughes Collaborates with OGDC to Optimize Existing Well Production, Supporting Energy Security in Pakistan. 2 September 2026. https://www.bakerhughes.com/company/news/baker-hughes-collaborates-ogdc-optimize-existing-well-production-supporting-energy First-party for Baker Hughes' account of its own contract. OGDC is the operator being counted and is not the issuer.

[3] SLB. Q and A with Paul Sims, President, Production Systems. 8 September 2026. https://www.slb.com/newsroom/features/2026/2026-0908-paul-sims-qa A position piece by a service company, naming no operator. The three figures are SLB's own and are sourced to recent analysis the page does not name.

[4] Energy Connects. Baker Hughes Secures Multi-Year OGDC Deal to Boost Pakistan Oil and Gas Output. September 2026. https://www.energyconnects.com/news/oil/2026/september/baker-hughes-secures-multi-year-ogdc-deal-to-boost-pakistan-oil-and-gas-output Trade write-up of the same contract, carrying the same terms.

[5] SPE 219553. Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence. ExxonMobil authors. SPE Artificial Lift Conference and Exhibition Americas, The Woodlands, Texas, August 2024. https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/550379 Operator-authored and out of this window by more than two years. The restatement in the Journal of Petroleum Technology carries the same result: https://onepetro.org/JPT/article-abstract/77/10/1/791497

[6] EarthScan. Indexing a 136,771-Scan Raster Archive for ML Ingestion. https://earthscan.io/case-studies/indexing-136771-scan-raster-archive

[7] EarthScan. Eight Real Scans, One Reviewer: Standing Up Human-in-the-Loop Validation. https://earthscan.io/case-studies/human-in-the-loop-review-eight-real-scans