Insights

ADNOC and AIQ signed a 340 million dollar three-year contract for ENERGYai, built on 70 years of proprietary data. Saudi Aramco announced Metabrain on 250 billion parameters, 7 trillion data points and more than 90 years of company history. Equinor has found more than 100 further AI use cases. ExxonMobil is running elastic full wavefield inversion on 4,032 Grace Hopper superchips. Every one of those is sold on the size of the archive, and the archive size is the one term in the problem that behaves benignly. Model retrieval as signal detection and the arithmetic is blunt: at a fixed retriever the pile you must read grows exactly linearly with the archive, 10 documents at 100,000 becoming 100,007 at a billion, while holding the pile at 10 instead needs the discrimination index to move from 3.719 to 5.612. Four decades of archive cost you either a factor of ten thousand in reading or 1.893 units of discrimination, and nothing in between. The term that decides which, the base rate of an answer-bearing document, is the one nobody measures before buying.

The blind horizontal azimuth metric broke: 65.4 degrees of mean error, close to reading a compass at random, against about 10 degrees on vertical wells. We measured that gap and published it in a companion research piece. This case-study is about the decision that number forced on the ground: instead of shipping one blended score, we and the operator's expert interpreter wrote a per-output hand-off protocol, agreed permissible-error tolerances output by output, and deployed the detector as a pre-processing aid that clears the vertical backlog and flags horizontal fractures for a human rather than pretending to resolve their azimuth.

Machine-learned velocity model building is scored the way every regression is scored, on the root-mean-square error of the residual. That scalar is a marginal statistic. What an undrilled trap depends on is not the depth error but the difference in depth error between the crest and the spill, and for a stationary field that difference has variance 2 sigma squared times one minus rho at the crest-to-spill separation. A long-correlated error is almost free and a short-correlated one costs the full square root of two, so an estimator that cuts sigma by 30% while shortening its correlation length from four closure widths to one improves its published score by 30% and makes the relief error 2.26 times worse. bp told the market in January 2019 that its full-waveform-inversion work had found a further billion barrels in place at Thunder Horse, Petrobras has published on heterogeneous salt velocity models feeding straight into gross rock volume in the Santos pre-salt, and Chevron and Shell are sanctioning multi-billion-dollar subsalt developments in the Gulf. The number to ask a vendor for is not the RMS. It is the semivariogram of the residual at the closure scale.

bp trenched over 100 km of cable and more than 10,000 sensors across roughly 45 square kilometres at Valhall in 2003 and has shot one to three surveys a year ever since. ConocoPhillips laid 200 km of permanent ocean bottom cable at Ekofisk from March 2010, in a system designed to be shot twice a year. Equinor put 380 km of fibre and more than 6,500 acoustic sensors over more than 120 square kilometres on Johan Sverdrup. Learned and regularised estimators are now pointed at those streams, and the published work scores itself in NRMS, the normalised RMS of the base minus monitor difference. Here is the arithmetic nobody prints beside the survey count. When every monitor is differenced against the same base, the base survey's own non-repeatable error lands in every difference with the same sign, the correlation between any two 4D differences is exactly one half, and stacking N monitors buys 3.01 dB in total no matter how large N gets. Two thirds of that is spent by the fourth survey. The lever that moves the floor is the depth of the reference, not the repeat count, and that is an acquisition decision rather than a modelling one.

Most accounts of an AI engagement are told through the model. The document that decides whether the operator ends up owning a capability or renting a dependency is the contract, and on a programme that runs across years it is not one contract but two: the founding research-service agreement and the year-two services agreement that follows. We read both, from a subsurface-AI programme built for a major operator in Oman on borehole-image logs from a fractured carbonate reservoir, as one design discipline whose object is a clean exit. Four levers carry that discipline. The obligation is written as effort, not a warranted result, on genuinely uncertain research. The foreground intellectual property is split 51 to the operator and 49 to the vendor, the smallest majority that still hands the operator control while leaving the vendor free to keep improving its own methods. A funded year-two agreement, USD 313,000 over twelve months, buys a handover rather than a subscription: six live training sessions, three online and three onsite, archived on a learning system, plus run-manage-operate and continuing engineering. And the money moves on a cadence that rewards transfer, five phase-gated tranches summing to USD 345,581.60 in year one, then a flat run-out in year two. A fifth structural move keeps the design clean: the academic partner's research is carved to a separate agreement so it does not entangle the operator's holdings. The two prior chapters of this programme, one reading the year-one paperwork as risk transfer and one reading the year-two agreement as a handover, are pointers here rather than re-derivations; the subject of this piece is the doctrine that spans both, and the argument that a vendor which designs its own replaceability into the contract is the one worth signing. Three interactive instruments make the case concrete.

Every methane monitoring procurement conversation opens on sensitivity, because sensitivity is the number a vendor prints on a datasheet. The source population says that is the wrong first question. Duren and colleagues surveyed more than 272,000 infrastructure elements in California and found 564 strong point sources, with 10% of them carrying roughly 60% of point-source emissions and a median persistence of 0.20, which means a single overpass sees a typical source one time in five. Fit a lognormal to two figures that paper prints and the arithmetic separates: moving from a 5 kg per hour airborne limit to the 100 kg per hour the same paper quotes for an equivalent instrument in low Earth orbit costs 7.7 points of detected mass and more than half the detected source count, and one extra pass buys the mass back. Twenty times the sensitivity is worth about one extra overpass.