Four of the largest operators in the world are buying the same capability at the same time, and the deals are public record.
In February 2025 Halliburton and Sekal delivered what they described as the world's first automated on-bottom drilling system, for Equinor on the Norwegian Continental Shelf. It is a closed loop: autonomous directional drilling, automated wellbore hydraulics and dynamic surface rig equipment control, orchestrated together and driven from a single button [1]. In March 2025 AIQ announced a 340 million dollar three-year contract with ADNOC to deploy ENERGYai across upstream operations, with five agents covering subsurface work in the first operational version, scaling out from a trial that reached over 15% of ADNOC's onshore and offshore wells [2][3]. Saudi Aramco's Metabrain, unveiled at LEAP in March 2024, reads drilling plans, geological data and historic drilling time and cost to recommend well selections [4]. PETRONAS runs TriCipta with Beicip-Franlab and AFED Digital, aimed at real-time data integration and continuous forecasting across subsurface evaluation and offshore operations [5].
Read those announcements and the emphasis falls on the model: agents, autonomy, generative reasoning over subsurface data [1][2][4][5]. What decides whether a downhole advisory is real time sits somewhere else entirely. It is the drill string.
The link is the budget
A downhole-to-surface link is a fixed-bandwidth pipe. Conventional mud-pulse telemetry runs at roughly 3 to 40 bits per second, commonly around 12 and closer to 3 on deep or difficult wells. Wired drill pipe runs at about 57,600 bits per second, over a thousand times mud-pulse bandwidth anywhere in that band [6][7]. Those figures are sustained rates, not burst rates, and that distinction is the whole argument, because hole is being made continuously. The link has to carry a fixed number of bits for every metre drilled.
Call that number the measurement cost per metre. The offered load is the measurement cost multiplied by the rate of penetration, and feasibility reduces to one division.
Below that ceiling the queue drains and the surface sees each metre a fixed lag after it is cut. Above it the queue does not drain while hole is being made. The backlog grows linearly in time on bottom and the advisory falls further behind with every metre. There is no middle state, and there is no graceful degradation into a slightly stale answer. What limits the backlog is not the link, it is the next connection or trip off bottom, when no new metres are being enqueued and the link finally gets ahead of what it is carrying. That is why the exhibit accumulates queue delay over one hour on bottom rather than letting it run: on a bandwidth-limited well the practical bound on the lag is the length of the stand.
What the advisor is actually looking at
On top of the link sits the sensor standoff, because the measurement is not taken at the bit. A directional-drilling study by Skoltech, IBM and Gazprom Neft researchers puts the logging-while-drilling sensors 15 to 40 metres behind the bit on a conventional assembly [8], and a trade account of the move to near-bit tools puts MWD survey sensors 50 to 75 feet or more back [9]. That gap is why near-bit and at-bit tools exist: Scientific Drilling's BitSub puts an azimuthal gamma measurement less than two feet from the bit [10]. The standoff is a depth lag someone has to remediate after the fact, and there is published machine-learning work doing precisely that for LWD gamma on Volve data [11]. The industry already treats it as a data problem worth a paper.
So the formation the advisory is reasoning about sits some distance behind the bit, and that distance has two parts: the standoff, and every millimetre the bit cuts during the round trip. The exhibit below opens its standoff control at 13 metres, just under the low end of those published figures, so its default case is the flattering one.
The transmission term carries the link. When the queue drains it is one metre's own serialisation. When the queue does not drain it is the whole accumulated backlog, and it keeps growing for as long as the bit stays on bottom.
Whole inference range, 0.05 s to 5 s, is worth 41 mm of hole at this rate of penetration, against a bit 216 mm across.
The band nobody prices
Leave the link on mud pulse and walk the measurement cost up from a decimated gamma curve toward image data. Somewhere between roughly a thousand and ten thousand bits per metre, the feasible rate of penetration falls through the rate the rig is actually drilling at. Below that band a well at 30 metres per hour is comfortably real time. Above it the same well is behind for as long as it stays on bottom, and the blind interval stops being a fixed sensor standoff and starts growing with every metre the bit makes.
The transition is a hyperbola, not a slope. The ceiling is the link rate divided by the measurement cost, so each doubling of the programme halves the ceiling, and the fall from comfortable to hopeless happens inside a single order of magnitude of data volume. In the procurement work we have sat in on, data volume gets treated as a scaling knob with a smooth cost curve behind it. On a bandwidth-limited well it is a cliff, and a demonstration run on wired pipe cannot show you where the edge is, because at 57,600 bits per second there is no edge anywhere in the range this exhibit covers.
Two consequences follow, and neither of them appears in the announcements above.
The first is that above the ceiling, the model running at surface on a mud-pulse well is not the model that was validated. It cannot be, or the link would not be feasible. Something upstream decimated the measurement, so the network is reading a summary of the data it was trained on. That is a distribution shift introduced by the transport layer, invisible to everyone above it, and it will not show up in a benchmark that was run against archived full-resolution logs.
The second is a geosteering consequence. Drag the measurement cost up in the exhibit and watch the amber sleeve run back past the ring where the well drops out of the target. The surface model has not been told the well left the zone. It is still confidently recommending a course correction for a formation the bit stopped being in some time ago.
The one term nobody should be optimising
Now move the inference-time control from one end to the other.
At 30 metres per hour the bit advances 8.3 millimetres per second, so a two second model run costs about 17 millimetres of hole. Across the exhibit's entire inference range, from 50 milliseconds to five seconds, the blind interval moves by 41 millimetres. Push the rate of penetration to 120 metres per hour, the fastest the exhibit allows, and the whole range is worth 165 millimetres against an eight and a half inch bit that is 216 millimetres across. The complete span of the inference budget is smaller than the hole the bit is cutting.
Everyone is optimising that term. Model latency is the number vendors quote, the number that gets benchmarked, and the number that justifies a GPU line item. It is not even the largest of the three delays in the loop: at the exhibit's default a metre of a standard LWD suite takes forty seconds to serialise onto a 12 bit per second link, against two seconds of inference and the two seconds of steering response the exhibit holds fixed. Set the two second model run against the blind interval it contributes to and it is one part in eight hundred. No amount of work on it changes the answer.
The terms that do move are the standoff and the link. On a draining link the standoff is nearly the whole blind interval, so halving it at the exhibit's default takes 48% of the interval with it. Switching from mud pulse to wired pipe multiplies the link rate by 4,800 against the 12 bit per second figure used here, and by at least 1,400 anywhere in the 3 to 40 bit per second band, which lifts the feasible ceiling above any rate of penetration a rig will reach across the whole measurement range this exhibit covers. The third option is to stop transmitting: run the inference downhole, next to the sensor, and send a decision rather than a measurement. That converts a bandwidth problem into a power, heat and qualification problem, which is a harder engineering job and the better trade.
Our own record here is a negative result
We should say plainly where this leaves our own delivery work, because it is directly on point and it does not flatter us.
Everything EarthScan has built on borehole image logs reads full-resolution image data after the fact. The pipeline that patches a processed image log for a detection transformer works on a raster about 690,000 pixels tall and 360 pixels wide, roughly 1.5 GB per image, sliced into 800-pixel patches that each span about 2.2 metres of hole [12]. Take that geometry at face value and one metre of section carries on the order of 130,000 azimuthal samples. At a single byte per sample that is about a megabit for every metre drilled. A 12 bit per second mud-pulse link moves 43,200 bits in an hour, so one metre of that data would take about a day to reach surface.
The exhibit's measurement-cost control stops at 300,000 bits per metre on purpose. Above that every case is the same case, and there is nothing left to learn from the surface.
None of our image-log work would survive being fed through a mud-pulse budget, and we are not going to claim otherwise. The per-metre QC dossiers, the depth-axis reconciliation, the patch-level audit trail: all of it assumes the full raster is sitting on disk and the well has already been drilled. That is a perfectly good product for post-well interpretation and a completely different product from a real-time advisor. Anyone selling one as the other has not done the division.
What to ask before signing
The useful question for an operator evaluating any of these systems is not how good the model is. It is what the measurement programme costs per metre, what the link carries, and what those two numbers make of the rate of penetration the field actually drills at. If the answer puts the operating point above the ceiling, the advisory is a post-well report with a live-looking dashboard on it, and no model procurement fixes that.
Key takeaways
- A downhole link is a sustained-rate pipe, so feasibility is one division: the rate of penetration a link can keep up with equals the link rate divided by the bits your measurement programme costs per metre.
- Between roughly 1,000 and 10,000 bits per metre a mud-pulse well crosses from comfortably real time to behind for the rest of the stand. The transition is a hyperbola rather than a slope, so it happens inside one order of magnitude of data volume.
- Once the queue stops draining the blind interval grows for as long as the bit stays on bottom, so what bounds it is the next connection rather than anything in the model, and the advisory is steering on a formation the well left some time ago.
- Inference time is not the term that binds. At the exhibit's default it is two seconds against forty seconds of serialising one metre onto a mud-pulse link, and across its whole range, 50 milliseconds to five seconds, it moves the blind interval by less than the diameter of the bit, so a faster model provably does nothing.
- The levers that work are sensor standoff, wired drill pipe at 4,800 times the rate of a 12 bit per second mud-pulse link, or moving the inference downhole so the measurement never crosses the link.
- EarthScan's own borehole image-log pipelines read full-resolution rasters after the well is drilled, on the order of a megabit per metre. That is a post-well product and it would not survive a mud-pulse budget.
References
[1] World Oil. Halliburton, Sekal deliver world's first automated on-bottom drilling system for Equinor in the North Sea. 26 February 2025. https://worldoil.com/news/2025/2/26/halliburton-sekal-deliver-world-s-first-automated-on-bottom-drilling-system-for-equinor-in-the-north-sea/
[2] G42 / AIQ. AIQ announces $340 million contract for large-scale deployment of agentic AI across ADNOC operations. 12 March 2025. https://www.g42.ai/resources/news/aiq-announces-340-million-contract-large-scale-deployment-agentic-ai-across-adnoc-operations
[3] ADNOC. ADNOC and AIQ successfully complete trial phase of agentic AI solution. 16 January 2025. https://www.adnoc.ae/en/news-and-media/press-releases/2024/adnoc-and-aiq-successfully-complete-trial-phase-of-agentic-ai-solution
[4] Offshore Technology. Saudi Aramco unveils industry-first generative AI model. March 2024. https://www.offshore-technology.com/news/saudi-aramco-unveils-industry-first-generative-ai-model/
[5] Future Digital Twin. PETRONAS builds AI capabilities with global tech firms. https://futuredigitaltwin.com/petronas-builds-ai-capabilities-with-global-tech-firms/
[6] Step-Change Improvements with Wired-Pipe Telemetry. SPE/IADC 119570-MS, SPE/IADC Drilling Conference and Exhibition. https://onepetro.org/SPEDC/proceedings/09DC/09DC/SPE-119570-MS/147850
[7] Oil & Gas Journal. New wired-pipe telemetry system tested in Arkoma basin. https://www.ogj.com/drilling-production/article/17224256/new-wired-pipe-telemetry-system-tested-in-arkoma-basin
[8] Klyuchnikov, N. et al. Data-driven model for the identification of the rock type at a drilling bit. arXiv:1806.03218v3, 25 March 2019. https://arxiv.org/abs/1806.03218
[9] Egypt Oil & Gas. Near Bit Sensors Create 'Formation Evaluation-While-Geosteering' Capability. 29 December 2014. https://egyptoil-gas.com/features/near-bit-sensors-create-formation-evaluation-while-geosteering-capability/
[10] Scientific Drilling. Staying in zone with at-bit gamma ray imaging from BitSub. https://scientificdrilling.com/resources/newsroom/staying-in-zone-with-at-bit-gamma-ray-imaging-from-bitsub/
[11] Applying transfer learning to address data scarcity: A case study on LWD gamma ray depth-lag remediation from Volve to another gasfield. Geoenergy Science and Engineering. https://www.sciencedirect.com/science/article/abs/pii/S2949891024006018
[12] EarthScan. From Binary Log to Dataset: Patching a 690,000x360-Pixel Image Log for a Transformer. https://earthscan.io/insights/patching-dlis-image-logs-for-transformers
