Insights

Three mature-asset announcements landed in five weeks and each published exactly one number. ADNOC and SLB, on 4 August 2026, said an AI-enabled Real-Time Operations Center now runs across more than 120 drilling rigs, which is a rig count and not a well count. Baker Hughes, on 2 September 2026, said it had signed a multi-year contract covering more than 120 OGDC wells, with the AI-enabled chemical injection in the second phase and no uplift figure anywhere. SLB, on 8 September 2026, published the denominator everybody now quotes, approximately 70% of global 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. The set a percentage uplift can actually reach is the roster times the past-peak share times the share of the archive a machine can read, and only the first of those three is ever printed. On the one archive we have counted end to end, 7,781 LAS files against 136,771 TIF rasters, the readable share is 5.38%, so a 1,300-well roster at SLB's 70% has 48.98 wells that can carry ExxonMobil's own 2.2%, which is 1.08 whole wells of annual flow against a headline reading of 28.60, an overstatement of 26.54 times. At that operating point a point of machine-readability is worth 13.00 points of past-peak share, and the entire move from 70% to 80% by 2030 is worth 7.00 uplift-bearing wells against 9.10 for a single point of readability. Triage cannot close an arbitrary gap either: lifting every raster at a 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, so a fully digitised 120-well programme matches a roster of 847.71 wells and never reaches 1,300.

On an illustrative two-train LNG plant with four 180,000 m3 tanks, the tanks reach their 95% working maximum 5.4 days after the last ship leaves if they were half full, and 1.8 days if they were 80% full. Turndown and train shut-ins all divide the same working room, so from 80% each is worth a third of what it is worth from half full, and the deadline for shutting a train shrinks with it: day 6.0 from half full with the ships back on day 12, 45 minutes from 80% with the ships back on day 6. Ras Laffan's published common storage, set against only the six mega trains that feed it, would hold at most 2.3 days from half full. Tank room is the cheapest insurance against a blocked export, and the loading schedule sets it weeks before the ships stop.

Turning receiver-function picks into a Moho map is the step between a study and a deliverable, and the usual interpolators are confident exactly where they know least. A graph neural field widens its intervals where the network thins. The harder lesson is about validation: under a random split the two approaches are indistinguishable, and only a regional hold-out shows the difference.

Two SPE panels three weeks apart, one in artificial lift and one in subsea well intervention, made the same argument about AI in production and carried no quantities at all. This is the arithmetic nobody attached to it, built on the per-class recall we recorded ourselves on eight real well-log scans: 0.97 on the class the archive represented best and 0.37 and 0.32 on the two it represented worst. At a worst class of 15% of the archive and an escape multiple of 10, the vendor account of the same workflow reads 61.08% of global production at a worst-class recall of 0.32 while the engineer account reads -10.32%. The engineer account is zeroed by a worst class of 13.02%, where the vendor still prints 91.28% of its stated ceiling.

A score-based prior makes the regulariser explicit and sample-able, and injecting a data-consistency gradient at each reverse step turns generation into posterior sampling. The practical catch is an ordering: coarse structure is decided early and the boundary depth late, so the image becomes convincing well before it becomes correct, and stopping when it looks right removes exactly the part that carries the answer.

ADNOC Gas said in August that scaling artificial intelligence and robotics across its assets has the potential to cut inspection costs by up to 75%. ADNOC said in May that a Taurob heavy-duty robot is conducting autonomous inspections at Taweelah. YPF opened a command room over more than 520 critical assets with 11 drones and used the words operational intelligence, not artificial intelligence. Suncor ran 19 autonomous trucks at Syncrude Mildred Lake and used no AI wording at all. Not one of the four names a detection rate for any fault class, and that is the number a saving of this size turns on. On the one review queue we have measured, the background class recalled at 0.97 and the two faint classes at 0.37 and 0.32, and 8 of 24 regions auto-accepted. Put those recalls behind a confidence gate and two things separate: the faults a 75% saving costs stay at 2.25% at every faint-class share where the saving is available, because the saving comes almost entirely from the strong class, while the ceiling falls with the faint share alone and takes 75% off the table above 25.7% of the population. There is no warning slope. The price does not rise; the option closes.