Why ‘physical AI’ marks the next frontier in industrial applications

At a time when renewed unease over artificial intelligence (AI) spending is weighing heavily on the technology sector, there are some answers to the related concern over how to tangibly monetise AI tools. Abu Dhabi’s state-owned oil company ADNOC, for instance, has signed a new deal earlier this week with AI firms G42 and OpenAI to deploy advanced AI capabilities across the energy major’s operations. The mandate: leverage AI to help improve operational efficiency and address complex engineering challenges in the oil sector.

“For a long time, the AI industry has talked about what it needs from the energy industry. And now I think we can repay the favour,” OpenAI CEO Sam Altman said Monday (November 2) after the signing of the agreement at the ADIPEC conference in Abu Dhabi. The agreement is a harbinger of what is to come next in the AI deployment strategy — “physical AI”, or AI systems that operate in and interact with the physical world, rather than existing only in digital environments or as pure software products.

What is physical AI?

Physical AI refers to the use of AI to enable autonomous systems such as robots, camera-bots, and self-driving cars to discern, comprehend, and perform complex actions in the physical world — effectively, AI systems that operate in the physical world. This involves the combination of AI models with sensors, actuators, and other control systems that allow models to act upon real-world environments. 

Most of the AI that humans interact with on a day-to-day basis today exists purely digitally: Things like talking with a chatbot, using AI for image generation, or using it as a code assistant – these all live in the world of bits and bytes. Physical AI, on the other hand, is the kind that crosses over into the world of atoms, things in actual physical space. These are AI systems that can perceive their environment, they can reason about it, and can take action. Take robotics, for instance. Before AI, robot behavior was typically scripted, and robots could only perform a predefined range of tasks within specifically engineered environments, like a robotic arm that welds the same seam a thousand times a day on an automotive production line.

In contrast, robotic AI agents equipped with general understanding from large language models (LLMs) can be imparted with a powerful “common sense” about the world, according to IBM Think’s Cole Stryker. These models can be paired with reinforcement learning techniques in high-performance hybrid architectures so that robots can possess both general knowledge and a specialised understanding of a specific use case. So, these AI-infused robots get better with use. According to Stryker, physical AI “goes far beyond individual robots to entire AI-powered factories, energy-efficient smart grids or fleets of automated vehicles”.

Nvidia CEO Jensen Huang is widely credited with flagging the term “physical AI” and framing it as the next big wave of AI-driven innovation. In keeping with his reputation for “seeing the future” and an obsessive pursuit of mapping out long-term technological trajectories, Huang, in a January 2026 podcast interview, predicted a future with “a billion robots” and a new global economy around developing and maintaining all these new robots — which he projected would be nothing short of a second industrial revolution.

“The ChatGPT moment for robotics is here,” Huang said, referring to the potential of robotics and physical AI. That is proving to be clairvoyant indeed, with AI progressively moving beyond digital applications into the physical world through physical AI.

Physical AI deployment

Story continues below this ad

Generative AI models such as ChatGPT or Meta’s Llama are trained on huge amounts of text and image data, largely scraped off from the internet. While these AI models have significant capabilities in generating human language, visuals, and abstract concepts, they are limited in their grasp of the physical world. Physical AI extends current generative AI with an understanding of spatial relationships and the physical behaviour of the world by taking in multimodal inputs such as images, videos, text, speech, or real-world sensor data and converting them into insights or actions that an autonomous machine can then execute.

Research papers have demonstrated that synthetic data alone is not useful. The environment where the deployment would happen plays a vital role in the training of the robots. Data is going to be equally important, because there is no existing data — one needs to really capture the real-world data, fuse it into multiple sensory signals, and give it to the labs. That would mean AI tools deployed in machines and real-world systems can progressively get better with time, a senior executive with a logistics-linked robotics company said.

Real-world applications

The ADNOC deal, though, is not by any means an isolated instance. There are more such instances.

Earlier this year, oil and gas major Shell had implemented a predictive maintenance system with California-based C3.ai, using machine learning to deploy and operate an enterprise-scale predictive maintenance programme. This monitors more than 13,000 pieces of equipment to prevent failures and cut emergency repair costs across Shell’s asset operations.

Story continues below this ad

Rival BP now applies AI to analyse real-time drilling data, optimising parameters to reduce non-productive time and improve drilling precision. With this, the petroleum major claims to have cut evaluation times from months to days, lowering exploration and drilling costs by 7%, and reducing non-productive time by nearly 8% to make well placement more precise.

Chevron is now using AI-driven seismic interpretation to improve subsurface imaging and increase exploration success rates. Launched in August 2024, ApEX, Chevron’s in-house genAI tool, was first used in deepwater exploration in the Gulf of America (traditionally known as the Gulf of Mexico) and employs multiple search agents to run tasks. The system’s geospatial agent owns tasks related to maps, while an exploration review team agent aids in assessing a prospective drill site’s chance of success.

Foxconn, a Taiwanese electronics contract manufacturer, has started to use a “scalable AI-powered robotic workforce” to respond to rising labour costs and local manufacturing trends. It employs AI and digital twin technology to simulate and automate precise tasks such as screw tightening and cable insertion, which traditional rule-based robots previously found challenging. Digital twin simulations, Foxconn says, have already helped prune deployment times for new systems by nearly 50%, while AI-powered robots have improved cycle times by 20-30% and lowered error rates by 25%.

The monetisation challenge

The physical AI deployment push is also an attempt to tackle what has been the biggest bugbear for the AI sector: monetisation.

Story continues below this ad

“I think they are thrashing around for revenue ideas. You can’t vibe code a phone. It is a deep ecosystem. Others such as Amazon have tried,” a US-based technology analyst told The Indian Express.

The path to monetisation is a struggle, despite companies such as Anthropic doing better on the B2B side. “That’s the story that Anthropic is doing better in enterprise. They do distribute more widely through all the CSPs (cloud service providers). OpenAI was hobbled by their weird contract with Microsoft. Apparently that’s coming to an end soon.”

The OpenAI-G42 deal with ADNOC, and other deals struck with AI service providers by Shell, Chevron and Foxconn are all trying to plug the monetisation gap. At the opening ceremony of ADIPEC 2026 in Abu Dhabi on Monday, OpenAI’s Altman described the relationship between AI and energy as a major opportunity for both industries, arguing that increasingly capable AI models could help address some of the sector’s most difficult challenges. AI’s potential, he said, extends beyond improving workplace productivity to accelerating scientific discoveries and solving problems that have historically proved difficult for researchers.

According to Ani Kelkar, a partner based in McKinsey’s Greater Boston office, between now and 2040, robotics and physical AI is expected to create at least a trillion dollars in economic value, most of that in manufacturing and logistics.

Story continues below this ad

“That will happen not through worker replacement and productivity but through our ability to innovate what products we make, how we make them, and what tasks humans perform versus which ones are automated, elevating the nature of work for humans to more problem-solving and to sort of a catcher role, rather than just relying on our physical dexterity,” Kelkar said.