The next AI revolution will control and build

Large language models have defined AI for the past two years: systems that operate in the digital layer, generating text, code and workflows. The next phase matters for a different reason, as it begins to shape physical systems directly.

AI is moving from the operational dashboard into the design, control and composition of infrastructure. Driving that transition is the emergence of physical AI and more deeply, physics-infused AI, where machine learning is combined with physics, chemistry and materials science to change the physical assumptions on which infrastructure is designed.

Leaders across the technology and scientific landscape are already pointing in this direction. Jensen Huang, CEO of NVIDIA, has described the next frontier as AI that understands the laws of physics, while Demis Hassabis, CEO of Google DeepMind, has positioned AI as a driver of scientific discovery. The value of AI is no longer confined to what it can generate on a screen, but what it can alter in the world.

Energy developers are already using these approaches to design entirely new polymeric materials for hydrogen fuel cell membranes. Thousands of candidate structures are generated and screened computationally, tested for proton exchange and gas separation performance beyond what existing materials can achieve. Semiconductor manufacturers are doing something similar with machine learning–based interaction models, simulating atomic-level diffusion behaviour in next-generation lithography systems. That work has compressed materials design cycles from months to weeks and the atom-level insights it produces are guiding the creation of new chamber materials. These are early acts of invention and they put the frontier of AI in the molecular layer.

Progress in industrial chemistry and materials science has been constrained for decades by the pace of physical testing. New materials and chemical pathways had to be discovered through slow, iterative laboratory work. That is beginning to change. Hybrid approaches that combine AI with physics-based simulation now allow researchers to explore possible designs computationally at speeds and levels of accuracy previously out of reach.

Google DeepMind’s GNoME system, for instance, identified millions of potential new crystal structures and offers a glimpse of how computational discovery can compress decades of materials research into scalable, model-driven workflows. This does not eliminate the need for testing, but it sharply narrows the search space for next-generation batteries, semiconductors and other high-performance materials.

AI-driven digital twins are enabling the design of entirely new solid electrode materials for next-generation batteries, generating and evaluating candidate structures at a scale no conventional R&D process could match. Water treatment researchers are applying the same approach to design novel sorbent materials for PFAS capture and removal – simulating thousands of candidate molecular structures for binding affinity rather than testing them sequentially in a laboratory.

The starting point in each case is a performance requirement rather than an existing material. AI works backwards through chemistry and physics to find structures that have never been synthesised before.

The people who design, finance and operate physical assets are facing an economic story as much as a technology one.

Consider the energy sector. Industrial processes – from hydrogen production to sustainable fuels – remain constrained by the capital and operating costs of extreme heat, pressure and expensive catalysts. If AI can design catalysts that allow reactions to occur at lower temperatures with higher yields or under less extreme conditions, the effect is not marginal. It changes plant design, equipment requirements and in some cases, the commercial logic of the process itself. This marks a shift from optimising systems we inherited to engineering systems we can deliberately redesign.

Supply chains could be reshaped as well, as raw inputs are redesigned over time for efficiency, availability and local suitability rather than remaining tied to the legacy chemistry of the last century. The boundary of what is economically viable begins to move with the science.

None of this removes the hard part. Moving from computational discovery to physical deployment introduces a different order of complexity.

Engineering is and should remain risk-aware. It is shaped by operational realities, safety standards and the long test of durability in the field. It also operates within liability frameworks that require designs to be understood and that will not change simply because the tool generating them is more powerful.

Judgment remains the core skill, with a widening scope. Engineers have worked within relatively fixed constraints. They will increasingly have to interrogate a far larger set of computationally generated options and stand behind the ones they sign off.

The capability gap is real. Asset owners and engineering firms will need testing protocols designed for materials with no service history and the in-house skill to design industrial processes around them. Regulators are in the same position; current standards assume a material arrived through the laboratory, not the model.

There is also a question of timing. Because these advances scale first through computation and simulation before they scale through construction, they may arrive faster than traditional infrastructure planning cycles expect. For infrastructure leaders, the issue is whether today’s investment decisions reflect that trajectory.

The exposure sits with those committing to large-scale projects. Infrastructure has always been designed around a fixed set of material constraints and those constraints are loosening. A plant specified today for the chemistry of the last century may be locked in for the next thirty years. Organisations that grasp this early will help set the next cost curves of the physical economy.

AI has moved from a software layer sitting above infrastructure into the design and operation of the physical world itself. Under the pressure this century will place on energy and resources, that shift could prove as consequential as any digital revolution.