Science & Technology10 November 20264 min read
ADNOC and OpenAI sign oil-sector AI deal as Physical AI gains ground
Major technology and energy firms, including OpenAI and ADNOC (Abu Dhabi National Oil Company), signed an agreement at the Abu Dhabi International Petroleum Exhibition and Conference (ADIPEC) 2026 to use artificial intelligence for operational efficiency in the oil sector. The deal is being read as a shift towards "Physical AI", a trend Nvidia has framed as the "ChatGPT moment for robotics".
The shift speaks to the AI industry's core challenge, which is monetisation. Putting AI into real-world industrial systems gives companies a way to earn from it.
What Physical AI is, and how it differs
Physical AI refers to AI systems built to operate in and interact with the physical world, the realm of "atoms", rather than living only in digital settings such as chatbots or code assistants. Autonomous systems such as robots, camera-bots and self-driving cars can use it to perceive their surroundings, reason about them and carry out complex physical actions.
Technically, it brings together large language models, sensors such as cameras and LiDAR, control systems and actuators. These perceive, process, plan and execute actions. Computer vision, reinforcement learning, imitation learning and simulation-based training help the systems adapt to real environments.
The contrast with conventional robotics is sharp. Traditional factory robots follow rigid, pre-programmed paths and fail if an object shifts even slightly. An AI-powered robotic arm adjusts its gripping force and position to an object's shape and orientation on a conveyor belt. What it gains is autonomy, real-time adaptability and decision-making in dynamic settings.
Where it pays off
By moving into business and industrial use, Physical AI offers concrete returns to tech firms struggling to monetise consumer-facing tools. McKinsey analysts project that robotics and physical AI will create at least a trillion dollars in economic value by 2040, mainly through manufacturing and logistics efficiencies.
Other gains include:
- Flexible production lines instead of rigid automation. AI-assisted scheduling has helped manufacturers cut non-value-added time by up to 75%.
- Safer work, as machines handle dangerous tasks and robotic exoskeletons support assembly-line workers in demanding overhead jobs, reducing fatigue and injury.
- Models that log joint angles and sensor feeds and keep improving with use.
- Cheaper training through digital twins, virtual replicas of physical systems, and synthetic data, which lessen reliance on costly and hazardous real-world trials.
Several firms already use such tools. Chevron's in-house generative AI tool ApEX uses geospatial agents to interpret seismic data and assess prospective drill sites in the Gulf of Mexico. BP analyses real-time drilling data with AI, which has cut exploration costs by 7% and non-productive time by nearly 8%. Shell runs a predictive maintenance programme with C3.ai across more than 13,000 pieces of equipment, preventing failures and trimming emergency repair costs.
Foxconn pairs a scalable AI-powered robotic workforce with digital twin simulations, so robots can handle variable tasks such as screw tightening and cable insertion. The company reports deployment times down nearly 50% and error rates down 25%. In India, AI at Vizhinjam port helps monitor ship movements, flag navigational hazards and coordinate automated cranes.
Costs, rules and unresolved risks
The barriers are considerable. Developing and deploying high-fidelity robotic arms, tactile sensors and mobile platforms remains capital-intensive.
Safety is a larger worry once autonomous, learning machines share workspaces with people. Emerging frameworks such as the EU Machinery Regulation 2023, applicable from 2027, and the EU AI Act 2024 mandate strict safety and compliance standards for AI in physical workspaces.
Society may also resist. Scaling Physical AI, especially humanoid robots, runs into questions of public acceptance and a persistent fear of blue-collar job displacement.
Then there is the simulation-to-reality gap. Models trained in simulation can behave unpredictably under unfamiliar real-world conditions, and inadequate training data can leave rare or safety-critical situations unaddressed. That matters most in the settings where the technology is now being deployed: oilfields, factories and ports.