ADNOC Already Had 70-Billion-Parameter AI. Then It Bought a Platform to Connect It.

On November 2, at the ADIPEC energy conference in Abu Dhabi, ADNOC announced a multi-year agreement with OpenAI and G42 to deploy OpenAI Frontier across its business. G42 signed through its subsidiary Inception42. No price was disclosed.
The framing in the press materials is that ADNOC becomes the first energy company to run Frontier, which OpenAI describes as a platform for connecting and coordinating AI agents across business operations. That sentence is doing a lot of work, and most coverage skipped past it.
ADNOC was not short of AI. It had a 70-billion-parameter model trained on more than fifty years of its own geological and operational data. It had a three-year, $340 million contract to roll that model out. What it did not have was a way to make those models talk to each other at the level where capital gets allocated.
The thing already in the ground
The existing system is called ENERGYai, built by Abu Dhabi AI company AIQ together with G42 and Microsoft. The model was trained on five decades of ADNOC data. In a 90-day trial completed in January 2025, it drew on data from more than 15% of ADNOC's onshore and offshore wells, and the company reported a 70% improvement in accuracy across major aspects of seismic interpretation.
That result bought a contract. AIQ won a three-year, $340 million deal to push ENERGYai across upstream operations, starting with five agents handling subsurface work. Those agents read seismic volumes, well logs and reservoir models. They are specialists, and they are good at their jobs within a narrow boundary.
A specialist agent that reads seismic data has no opinion about whether a particular field development plan should be sequenced differently to free up capital in the next fiscal year. That decision lives in spreadsheets, planning meetings, and the heads of people who have been doing it for twenty years.
What Frontier is actually for
OpenAI's platform is meant to sit above that layer. The company's description is coordinating agents across business operations. In ADNOC's case, the stated goal is connecting operational intelligence to capital and commercial decisions so those decisions get made together instead of in sequence.
Two applications are named as initial targets.
The first is field master planning. Today that work is a chain of separate studies: subsurface, facilities, economics. Each one takes weeks, and each one waits for the one before it. ADNOC says OpenAI's reasoning models will combine data from thousands of wells, multiple stacked reservoirs and hundreds of surface facilities with capital, cost and commercial factors, so the plan is optimised for an entire field rather than reservoir by reservoir. The claimed outcome is going from months to hours for complex development plans.
The second is more ambitious and vaguer. ADNOC wants to value each barrel while it is still underground. Agents connected across production, processing, refining, logistics, sales and trading would continuously forecast crude supply, demand, prices and margins, and weigh the commercial and operational effect of every decision.
Read those two together and the shape becomes clear. The first application compresses a planning cycle. The second tries to turn a quarterly exercise into a running calculation. Neither is a model capability. Both are integration problems.
The part that makes it deployable
Three details in the announcement matter more than the capabilities.
Sovereignty. The system runs on G42's AI infrastructure in the UAE. For a state oil company, where the inference happens is a board-level question, not an IT preference. A platform that connects agents across production, trading and commercial data cannot live on someone else's public cloud in another jurisdiction.
Human oversight. ADNOC says human experts keep oversight of consequential decisions. That is the right word choice. Consequential is doing the work: routine queries run, big calls get a person.
Adoption already in place. ADNOC rolled ChatGPT Enterprise out to more than 40,000 employees and says adoption passed 97%. That number is worth pausing on. Most enterprise AI programmes struggle to get past 30%. A workforce that already uses the vendor's chat product daily is a very different starting condition from one that has never seen it.
The agreement also covers cyber defence through OpenAI's Daybreak Blue initiative and workforce training. On the same day in Abu Dhabi, OpenAI signed with sovereign investor Mubadala and a five-year pact with Inception42 covering government, enterprise and education.
The comparison that matters
Oil majors have bought AI for a decade, and the results have clustered in two places: predictive maintenance on rotating equipment and seismic interpretation. Both are pattern recognition against historical data, and both produce a recommendation that a specialist can check in an afternoon.
Field planning is different because the answer is a plan, and a plan has no ground truth until years later. If the reasoning layer produces a development plan in four hours that a team of engineers would have produced in four months, there is no immediate test of whether it is a good plan. You find out when the wells come in.
That is why the human oversight line and the audit trail matter more here than in a chatbot. A platform that can reshape a production plan in an afternoon is a platform that can commit hundreds of millions of dollars of capital in an afternoon.
What is not in the announcement
There is no timeline for either application. There are no performance targets. There is no deal value, no seat count for the new platform, and no explanation of how the reasoning layer handles the fact that geological models are probabilistic while financial models need point estimates.
The months-to-hours claim is the one to watch. Compressing a planning cycle is measurable, and if ADNOC publishes before-and-after numbers in a year, that will be the most useful data point anyone has on enterprise agent platforms in heavy industry. Until then it is a target.
There is also a structural question. ENERGYai was built by AIQ, a joint venture involving G42 and Microsoft. Frontier is OpenAI's. Putting one on top of the other means two vendors, two roadmaps and two security postures inside the same decision path. ADNOC's chief executive Sultan Al Jaber framed it as moving beyond individual applications. The engineering version of that sentence is that integration cost just became a permanent line item.
Why energy first
Oil and gas is an odd choice for the first marquee Frontier deployment until you look at the data. A field has thousands of wells, decades of production history, hundreds of facilities, and a cash-flow model that everyone argues about. It is dense, structured, high value, and already instrumented. If a platform for coordinating agents can prove itself anywhere, it is here.
The industry also has a habit of buying software that promises to shorten planning cycles. What it has less experience with is software that changes the answer rather than the schedule. ADNOC's second application, the barrel-in-the-ground valuation, is that kind of software. It proposes a different way of deciding what to produce and when.
That is the bet. Not that reasoning models can read a seismic section, because they already can. That a connected layer above the specialists can produce a better field plan than the current chain of sequential studies, inside a sovereign infrastructure, with a human signing off.
ADNOC says it is demonstrating how frontier AI reshapes an entire industry. The verifiable part of that claim arrives when the first field plan comes out in hours, and someone outside the company checks whether it was a better plan.
Related articles
METR Spent 11 Weeks Inside Anthropic. The Number Is 1.9, and Most of the Evidence Is Blacked Out.
A policy threshold with a number in it, an auditor who says it may have been crossed, and evidence nobody outside can check.
One Percent of Firms Pay 80% of the AI Bill. That Cohort Just Stopped Growing.
A $600 million revenue cut, a 26.8% price decline, and a cohort that slipped below $10,000 per employee.
Claude Can Now Read Your Bank Statements. Anthropic Deliberately Gave It No Hands.
The read-only design is the product, and unlinking does not retract what is already in a chat.
Adobe Stopped Asking Marketers to Open Adobe
For a decade the strategy was own the interface. Adobe just bet the opposite.