561,000 Wells, If Every One of Them Delivers: The Denominator Under a $230 Billion AI Forecast

A consulting firm's analysis put $230 billion of artificial intelligence value on oil, gas and oilfield services, and a trade title restated it on 2 September 2026 [1]. Our own weekly sweep rejected that item twice. Once because it reports a modelled projection and names no operator taking any action, which our contract calls commentary rather than a signal. Once because it is out of window: the underlying analysis was released on 25 August 2026 and only the trade restatement of 2 September falls inside the week we were reading. The sweep's own note on it is blunt, that the figure is a full-potential model output and must never be attributed to any operator, and nothing in what follows attributes it to one.

So this piece inverts a number our own process threw out, deliberately, with the rejection printed on the page. The reason is not the number. It is the grammar. A market-scale AI value total is arithmetically a count of assets multiplied by a value per asset, and not one of the totals in our pool publishes the count. Rystad's $500 billion opportunity sits in the same class, an analyst projection rather than an operator-reported figure, and our sweep could establish no date for it at all, so it is carried here undated and never as something we dated [2].

We happen to have the missing half. An illustrative scenario we published carries a per-well value that comes with its own denominator, and three of our own engagement records carry two independently measured fractions of wells that actually deliver usable data and a well-count ablation in which value per well stops rising and starts falling. Put those against a headline with no denominator and the headline implies a population.

What each of these sources actually is

The provenance comes before the arithmetic, because most of the external material here is weak and the weakness is part of the argument.

ExxonMobil, and the only operator-authored uplift figure this piece uses. SPE 219553, "Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence", was written by ExxonMobil authors and presented at the SPE Artificial Lift Conference and Exhibition Americas in The Woodlands, Texas, in August 2024 [3]. It describes 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 is the one hard, checkable number our sweep found in its week, and the only operator-authored uplift figure in this piece. It is also from August 2024, it is out of window, and it is never recent. It reached us through a survey article of 4 September 2026 cited here as a pointer and nothing more [5], and a restatement of the same result exists in the journal literature [4]. A separate claim on that same survey page, attached to a different operator and naming a well count an order of magnitude larger, could not be corroborated anywhere outside that one article, so it does not appear here in any form.

SLB, on the shape of the estate, with no AI on the page. A question-and-answer feature published on 8 September 2026 with Paul Sims, president of Production Systems at SLB, puts approximately 70% of global production on assets past peak, expects that to approach 80% by 2030, and puts conventional reserve replenishment close to 30% [6]. Those are SLB's figures, sourced on the page to recent analysis the page does not name, and no operator is named anywhere on it. It is a position piece, neither a deployment nor an intent, and our sweep records no AI or machine-learning content on it at all: the release says nothing about artificial intelligence. The question this piece puts to its 70% is ours and is attributed as ours throughout.

Baker Hughes and OGDC, the only in-window item here that carries AI wording. Baker Hughes announced on 2 September 2026 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 Pirkoh field [7]. Phase one is evaluation. Phase two is execution, and phase two is where the AI-enabled chemical injection programmes for flow assurance sit. Three things have to be said. This is a vendor's own account of its own contract and OGDC has not spoken, so the AI framing originates in the vendor's language rather than in anything the operator has said. Our own sweep rejected the release in the prior week partly because nothing AI-related has happened yet: the only AI mention on the page places those injections in a phase that begins after the evaluation, and the words "artificial intelligence" and "machine learning" appear on it zero times. And the 120 is a contract count, not a working set.

Five items that are here only as failure modes, each labelled. ADNOC's figures through AIQ and SLB's ENERGYai, 10 times seismic interpretation speed and a 70% precision increase, are dated 7 August 2025 in our trap table, are out of window, and are test-environment numbers on 15% of ADNOC data rather than production results: the vendors' figures about an operator's data, not ADNOC speaking. A separate $340 million agentic AI programme across 28 producing fields, again ADNOC through AIQ, carries two conflicting datings in our own record and both are carried here, our trap table dating it 10 March 2025 and 18 months out while our daily section of 5 September records the article carrying it as 1 May 2025 and 16 months old. Saudi Aramco's next-generation supercomputer with solutions by stc, $372.5 million stated as SAR 1.4 billion and 7 times the Upstream compute, is dated 6 May 2026 and out of window: a real figure with a real multiple and no stated base compute, which is the same shape of hole as a value total with no asset count. ONGC's Rs 125 crore three-year AI-led IT infrastructure contract with CIPL, 16 March 2026 and out of window, is the denominator problem in its most literal form, because trade copies disagree on whether it covers 47 or 450 locations and no first-party ONGC text was reachable to settle it. And a Petrobras executive's panel remark contrasting an agent count with a legacy-system count is named here only as the failure mode it is, a ratio with both halves stated by someone other than the operator.

Eight operators and service majors are named above. On substance, exactly one of those items is both in window and carries AI wording, and its AI sits in a phase that has not started. That is the state of the evidence.

A value total is a product, and only one half is ever published

Suppressing the second factor is not a rounding convenience. It is what makes a total unfalsifiable, because any total can be reconciled with any evidence by moving a count nobody has to state.

An illustrative scenario we published breaks that symmetry in one place. It is a composite case study, not a client engagement: a real-time advisory agent runs across a land rig fleet, and the scenario states both halves. Over 12 months and 94 wells, non-productive time falls from an average of 4.2 days per well to 2.6 days, a 38% reduction. Stuck-pipe incidents drop from 7 per 10 wells to 2 per 10 wells. The average cost saved per well is $410,000 once rig day-rate, logistics and lost production are factored in, and the aggregate is stated in the same file as $38 million saved across 94 wells [8]. A total, and the count it was computed over, in the same breath.

The mechanism decides where that figure can be carried. In the scenario, the agent learned a normal envelope per formation and lithology from 240 wells of historical drilling data, compared live one-second telemetry against it, and alerted when the deviation score crossed a threshold tuned over 6 weeks to catch 85% of incidents with fewer than 2 false positives per well. In operation it caught 82% of stuck-pipe events between 2 and 8 minutes before the driller would have seen the first console alarm [8]. So the saving is avoided drilling non-productive time: rig time that was going to be lost and was not.

Two delivery fractions, both ours and both measured

A count of wells reached is not a count of wells that deliver anything, and we have measured that gap twice on two different programmes.

On a three-phase carbonate borehole-imaging programme in Oman we ran the engagement off a two-column ledger, asked against delivered, updated monthly. Phase 1 recorded 5 asked and 0 delivered. Phase 2 recorded 10 asked and 8 delivered. Phase 3 recorded 25 asked and 8 delivered, and that row is described in the file itself as the single largest entry in the ledger [9]. Vug ground truth came from just 2 wells, and 5 horizontal wells arrived late as an unplanned add-on. The delivered column counts post-gate wells rather than shipments, which is why the gap stayed visible.

On a separate formation-evaluation programme with a mid-sized Middle East carbonate operator, more than 80 processed and interpreted high-resolution borehole image logs were already in hand to normalise, the Phase 2 scope was written explicitly as correlation across 2 to 80 wells, and the well-maturity ramp reported on the programme's own progress decks climbed 3 of 80, then 8 of 80, against a near-term target of roughly 20 to 25 of 80, with Phase 3 gated at a minimum on the order of 10 to 15 wells [10].

Two programmes, two gates, 8 wells at the sharp end of both, and the exhibit's two detents: 32.0%, which is 8 of the 25 asked, and 10.0%, which is 8 of the 80 in hand.

Value per well is not constant in the well count

Here the arithmetic stops being arithmetic. On a well-count ablation for a detection transformer picking fractures and beddings from borehole image logs, with architecture, augmentation and optimiser held fixed and only the training pool moving, validation class error fell from 93.115% at 3 wells to 18.370% at 6, to 1.055% at 9, to a floor of 0.817% at 11, and then rose to 2.536% at 14 [11].

The rise is the finding. The 3 wells added between the 11-well and the 14-well runs carried less consistent sinusoid picks than the core set, and in a set-prediction model trained with Hungarian matching the optimal assignment is computed against the labels, so a sinusoid one interpreter picked and another did not teaches the matcher a contradictory target rather than adding a hard example. The signature is in the same table: the L1 geometry loss kept falling through the reversal, from 0.083 to 0.059, while classification regressed. Geometry better and classification worse is a labelling problem, not a capacity one, and a parallel ablation in the same file put ResNet-34 at 26.759% class error against ResNet-10's 0.499%, so capacity was never the lever.

What that does to a market total is structural. If value per well is realised in proportion to how often the model is right, and the model gets less right as more inconsistently picked wells enter, the total is not linear in its own denominator. It rises, turns over, and falls. Every published total in this pool is silently linear in a denominator it does not state, and ours is measurably not. The same reversal also supports a partner-share argument about pooled exploration data; that argument, with its crossings, belongs to an earlier piece and is not restated here [12]. The reversal is used here for one thing only, the non-monotonicity of value per well.

The arithmetic, and it is ours

Take a value total $V$, the illustrative saving per well $s$, and the delivery fraction $\varphi$. The count the total implies if every delivering well earns that figure is a division, and the population that has to be reached for that many wells to arrive is a second one.

The exact figure is 560,976 well-years and the title rounds it to 561,000. At our measured 32.0% the population reached is 1,753,049 wells, at our measured 10.0% it is 5,609,756, and the analyst projection prices out at 1,219,512 well-years and 3,810,976 wells reached at the same 32.0%.

Now the ablation, read on a logarithmic well-count axis, the axis a learning curve is flat on, with the measured reversal carried forward on it.

That turnover sits at 4,468,141 delivering wells, where class error has reached 92.872% and value per well has fallen to $29,465. The most any pool can raise is $131.7bn. The headline is $230 billion. The gap is $98.3bn, at every radius, and there is no pool size anywhere on the curve that closes it.

The object

The exhibit is a pole and its shadow, because dividing by a delivery fraction is exactly what a low sun does to a pole. Stand a pole of height h in sun at an elevation whose tangent is the delivery fraction and its shadow is h divided by that fraction. The pole is the wells that deliver usable, consistently picked data and therefore earn the per-well figure. The shadow is the population that had to be reached to find them. Inside the pole stands an amber tower, 560,976 well-years, the headline priced at the illustrative $410,000, and it is the one body here that keeps its shape however low the sun sinks. The ray that grazes the top of the pole and lands on the tip of the shadow closes the right triangle, so the slope of that ray is the delivery fraction and the picture cannot disagree with the arithmetic.

Height and shadow length are linear in well count, not volumetric, and the plate says so. On a volumetric scale a factor of 10 in the population would read as a factor of about 2.15, which would flatter the forecast by hiding the exact inflation this piece is about. A tick on the shadow marks each forecast height, so the shadow is read in units of the forecast itself. A body's width is a fixed share of its height up to a cap and carries nothing. Linear scale costs the exhibit its ability to draw the small sets beside the large ones, which is what the dive is for.

The controls act on the sun and the pole and never on the forecast, which is the argument restated as an interaction. The delivery fraction slider runs from 100% down to 1% with 2 detents at our own measured values. Pick consistency is a switch and not a slider because we measured 2 states and not a continuum, one pool of consistently picked wells and one pool with 3 inconsistent wells in it. The dive slides the frame from the whole shadow down to the foot of the pole, and a counter prints how many times over the smaller of the two bodies the frame sits between fits into the larger.

What the object shows that the prose cannot

Start at the top detent. At a delivery fraction of 100% the sun stands at 45 degrees, the shadow is exactly one forecast long, and the population is 560,976, the tidy number a market total implies when nothing is lost. Drag to 32.0% and the sun sinks to 17.7 degrees while the shadow runs out to 1,753,049, three forecasts long on the ruler. Drag on to 10.0%, a sun of 5.7 degrees, and it reaches 5,609,756, ten forecasts long. Nothing about the forecast changes at any point in that drag, and watching its tower keep its proportions while the shadow of its denominator stretches to 10 times its length is the single perceptual fact this piece turns on. Prose can state the three numbers. It cannot make a reader watch the sun sink.

Switch pick consistency to as-delivered and the pole rises off the tower, because more delivering wells are now needed to raise the same money: the best pool sits at 4,468,141 wells, which is 7.96 times the core. Then the tower fails to fill. It is filled only to 57.2% of its height, the share the best pool can raise, and the top 42.8% stands as an empty red cage with $98.3bn short printed beside it. A value total that is not monotone in its own denominator does not merely bend. It cannot be filled, and a tower with its top left empty in the air is legible in a way that a level curve closing into a loop is not.

Then dive. The frame slides down the shadow to the foot of the pole, the tower rises out of the top of the plate, and in front of it stand the only countable well sets in this argument, at true scale and each labelled to its source: the 94 wells over 12 months from the illustrative drilling scenario, the 240 wells of historical data its envelope was learned from, the 80 interpreted borehole image logs with their 3 of 80 and 8 of 80 maturity steps, the 25 asked and 8 delivered, and the 11 and 14 wells of the ablation. Beside them, in outline and labelled a contract count rather than a working set, stands the more than 120 OGDC wells. And the largest hard, operator-authored well count anywhere in our pool, the more than 1,300 ExxonMobil Permian wells at about 2.2% uplift from August 2024, fits into the core 431.5 times over, into the population reached at our 32.0% fraction 1,348 times, and into the population at 10.0% 4,315 times. The illustrative 94 wells fit into that last one 59,678 times.

The far 70.0% of the shadow is drawn matte and striped, closing off exactly 70% of its area. That is the share of global production a vendor position piece of 8 September 2026 puts past peak [6], and on that stretch the reconstruction has no meaning at all. The illustrative $410,000 is avoided drilling non-productive time, and there is none of that on a well nobody is drilling, so the reconstruction is defined only on the fraction of the estate still being drilled. The reader ends up looking at a shadow that is both an order of magnitude too long and mostly undefined.

Where this model is weakest

The extension of the ablation past 14 wells is ours and it is the softest thing here. We measured 5 points between 3 and 14 wells and carried one slope forward across 5 decades. The choice of axis is the assumption: on a logarithmic well-count axis the extension puts class error at 16.550% by 100 wells, 49.376% by 10,000 and 82.202% by 1,000,000, which is severe but survivable, while the same measured slope carried linearly in well count would hit total error at 184.09 wells and make the whole reconstruction absurd. We chose the axis a learning curve is actually flat on, and the plate says which half is measured and which half is ours.

The second is the one the matte band draws. The only operator-authored AI uplift available to this piece is about 2.2% across more than 1,300 ExxonMobil Permian wells, and it cannot be converted into money without a barrel count and a realised price that no source in our pool carries, so the inversion cannot be completed on the mature estate at all.

The third is the size of the base. One illustrative drilling fleet of 94 wells over 12 months, a composite rather than a measured operator; two interpretation programmes of 25 and 80-plus wells; a pool of 14 wells at the widest point of the ablation. Those are the numbers a 5.6 million well population is reconstructed from, and nothing about a national estate is guaranteed to behave like any of them.

Two neighbouring arguments are deliberately not made here. An earlier piece takes a figure that already has a denominator and asks what its numerator is measured against, which is a counterfactual question rather than a population one [13]. Another widens a population and watches a rate degrade against a fixed downstream capacity, which is a funnel and not a rollover [14]. This piece takes a figure with no denominator at all and reconstructs the population it implies.

What a buyer should ask

What asset count was this value computed over, and is it wells, well-years or sites? Ask for the denominator before the total. A total with no count cannot be checked against anything you own, and a count that no first-party text can settle is no better than no count at all.

What fraction of the assets you reach will actually deliver usable data, and who measured it? On our two programmes the answer was 8 of 25 and 8 of 80. A business case that quietly assumes 100% is off by a factor of 3.13 on the first of those and 10.00 on the second, on the population it has to touch, and the value total will not move to tell you.

Does your value per asset stay constant as the asset count grows, and what happens when the picks come from more than one hand? This is the question that decides whether a programme compounds or quietly degrades, and our own answer was a floor at 11 wells and a reversal at 14. Its operational form is a label-consistency gate: score every new well's picks against the existing corpus before it enters the pool, and quarantine anything that would raise validation error rather than merging it. The harmonisation that moves the delivery fraction and flattens the rollover is a correlation problem, which is what ES W2W Correlation exists to do.

Is the value defined on the assets you actually have? Ours is not, on a mature estate that is not being drilled. If a forecast is aimed at a population the arithmetic is undefined on, the honest answer is that the number has no denominator you can supply, which is where reading reservoir, wells, lift, chemistry and facilities as one system with one denominator starts to matter, and where EarthScan 360 sits.

References

[1] Rigzone. AI Could Unlock $230B for Oil, Gas, OFSE Cos, McKinsey Says. 2 September 2026. Trade restatement of an analysis released 25 August 2026. Cited as the occasion only: it reports a modelled full-potential projection, names no operator taking any action, and the figure is attributed to no operator in this piece. https://www.rigzone.com/news/ai_could_unlock_230b_for_oil_gas_ofse_cos_mckinsey_says-02-sep-2026-184515-article/

[2] Rystad Energy. "$500 billion opportunity". Carried in EarthScan's own out-of-window trap table with no URL and no date the sweep could establish. An analyst projection, not an operator-reported figure, of the same class as reference 1. Source record: docs/signals/2026-W36-brief.md.

[3] Society of Petroleum Engineers. SPE 219553, Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence. ExxonMobil authors. Presented at the SPE Artificial Lift Conference and Exhibition Americas, The Woodlands, Texas, August 2024. Out of window; operator-authored. https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/550379

[4] Journal of Petroleum Technology. Restatement of the same gas lift optimisation result, volume 77, issue 10. https://onepetro.org/JPT/article-abstract/77/10/1/791497

[5] Journal of Petroleum Technology. Maximizing Every Barrel: Artificial Lift Embraces Automation. 4 September 2026. Cited as the pointer that surfaced reference 3 and for nothing else; the article is this week's, the work in it is one to two years old. https://jpt.spe.org/maximizing-every-barrel-artificial-lift-embraces-automation

[6] SLB. Q and A with Paul Sims, president of Production Systems. 8 September 2026. Partner release. These are SLB's figures, sourced on the page to recent analysis the page does not name; no operator is named on it, and the page says nothing about artificial intelligence. https://www.slb.com/newsroom/features/2026/2026-0908-paul-sims-qa

[7] Baker Hughes. Baker Hughes collaborates with OGDC to optimize existing well production, supporting energy security. 2 September 2026. A service major's own account of its own contract; OGDC has not spoken, and the AI-enabled chemical injection sits in an execution phase that begins after the evaluation phase. https://www.bakerhughes.com/company/news/baker-hughes-collaborates-ogdc-optimize-existing-well-production-supporting-energy

[8] EarthScan. Illustrative scenario: Cutting drilling non-productive time 38% with a real-time advisory agent. A composite case study, not a client engagement. Source file: content/case-studies/cutting-drilling-npt-real-time-advisory-agent.mdx. https://earthscan.io/case-studies/cutting-drilling-npt-real-time-advisory-agent

[9] EarthScan. 25 Wells Asked For, 8 Delivered: The Data-Scarcity Gap. Source file: 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

[10] EarthScan. The Interpretation Capacity Gap: 80 Wells, a Backlog and the ROI. Source file: content/case-studies/interpretation-capacity-gap-80-wells-backlog-roi.mdx. https://earthscan.io/case-studies/interpretation-capacity-gap-80-wells-backlog-roi

[11] EarthScan. How Many Wells Is Enough? A Well-Count Ablation for Fracture Detection. Source file: content/case-studies/geobfdt-well-count-ablation-how-many-wells-fracture-detection.mdx. https://earthscan.io/case-studies/geobfdt-well-count-ablation-how-many-wells-fracture-detection

[12] EarthScan. More Wells Is Not More Signal: The Partner Share Where Pooled Exploration Data Starts to Pay. 25 September 2026. Cited for the argument this piece deliberately does not restate. https://earthscan.io/insights/more-wells-is-not-more-signal-the-partner-share-where-pooled-exploration-data-starts-to-pay

[13] EarthScan. 12.2% Against What? The Missing Baseline Behind an AI Production Uplift. 6 October 2026. Cited for the separation: that piece asks what a numerator is measured against; this one asks what population a figure with no denominator implies. https://earthscan.io/insights/12-2-percent-against-what-the-missing-baseline-behind-an-ai-production-uplift

[14] EarthScan. Precision Is Not a Property of the Model: What Cheaper Prospect Screening Does to the Funnel. 14 October 2026. Cited for the separation: that piece moves a base rate against a fixed downstream capacity; this one moves a delivery fraction against a value total, and the finding here is the rollover. https://earthscan.io/insights/precision-is-not-a-property-of-the-model-what-cheaper-prospect-screening-does-to-the-funnel