Why Climate Tech Needs ‘Physical AI’
Large Language Models (LLMs) have dominated discussions about artificial intelligence. They are powerful tools for generating text, summarising research, and managing knowledge. But the climate crisis is not a problem of words. It is a physical challenge, rooted in energy flows, industrial processes, and environmental systems.
Last year, the global decarbonisation rate was just 2.5 percent, according to PwC’s Net Zero Economy Index. To stay on track for 1.5°C, the required rate is 17.2 percent. Closing that gap will not be achieved by incremental improvements in digital tools alone. It will require AI systems that can sense, reason, and act in the real world — what leading researchers call physical AI or embodied AI.
These systems go beyond predicting the next word in a sentence. They integrate data from sensors, cameras, industrial machinery, and IoT networks to perceive their environment, make decisions in real time, and optimise complex systems. For climate tech, that means moving from analysis to direct action — detecting faults, improving efficiency, and enabling faster deployment of low-carbon solutions.
The limits of text-based AI for climate impact
LLMs are optimised for token prediction, not for modelling continuous, high-dimensional physical systems. As Meta’s AI Chief Scientist Yann LeCun explains, “The physical world is high-dimensional and continuous,” making it profoundly different from the discrete patterns language models are built to process.
In climate tech, LLMs can help accelerate literature reviews, support compliance reporting, and improve technical communication. But they cannot directly manage a live energy grid, design a next-generation catalyst, or predict equipment failure based on sensor input.
Attempts to adapt language-based models to predict video or physical events have faced fundamental barriers. The architectures behind LLMs are not designed for world modelling, which requires understanding physical laws, spatial relationships, and causality. This is why industrial and environmental applications increasingly demand models built from the ground up for physical interaction.
How physical AI works
Physical AI systems combine perception, reasoning, and action. They may be robotic, software-based, or hybrid. Typical applications include:
Industrial optimisation: AI controllers that continuously adjust manufacturing parameters to reduce energy waste.
Predictive maintenance: Robotics-assisted inspections of wind turbines and solar farms, anticipating failures before they occur.
Materials discovery: Machine learning models that identify new low-carbon materials, shortening R&D timelines.
An emerging approach is Joint Embedding Predictive Architectures (JAPA), which enables models to learn abstract representations of real-world states. This allows them to anticipate failures, plan multi-step actions, and adapt to changing conditions — critical capabilities for industrial decarbonisation.
For example, DeepMind’s work on AI-optimised fusion control shows how predictive modelling can stabilise highly dynamic physical systems, offering insights for renewable energy integration.
Where physical AI is already delivering results
Energy grid optimisation
AI is increasingly used to balance supply and demand in real time, integrating intermittent renewables and reducing emissions. In the UK, National Grid ESO’s AI-enabled control systems have helped absorb more wind power by forecasting and responding to fluctuations more effectively.
Industrial decarbonisation
Predictive maintenance and process control have reduced downtime and energy use in heavy industry. One European cement producer achieved a 5 percent energy saving using AI-driven kiln optimisation, cutting both costs and CO₂ emissions.
Climate resilience
AI-enabled early warning systems are improving disaster response in climate-vulnerable regions. The TECH 4 Resilience Challenge has supported start-ups developing drone-based and sensor-based monitoring for floods, droughts, and wildfires.
Materials innovation
Machine learning has accelerated the discovery of low-carbon cement formulations and next-generation battery chemistries, as shown in recent MIT research. This shortens the path from lab discovery to market-ready product.
Overcoming the “pilot trap”
The reality is that 92 percent of AI projects never scale beyond pilots. In climate tech, the blockers are rarely the algorithms themselves. More often, they are:
Weak data governance.
Poor integration with existing operations.
Lack of leadership alignment.
Gaps in ecosystem support and funding.
Physical AI adds extra complexity — integration with hardware, real-world testing, and sector-specific compliance requirements. Without a clear deployment plan, even the most promising system risks remaining a one-off demonstration.
Nexus Climate works with founders to design for scale from the outset, bridging technical, operational, and market considerations. This includes building cross-disciplinary teams, establishing robust data frameworks, and identifying the right industrial partners.
Environmental considerations for AI in climate tech
While AI can accelerate decarbonisation, its own environmental footprint must be managed. Data centres consume significant electricity and water, and chip manufacturing has a material carbon cost.
Researchers such as Dr. Sasha Luccioni have called for carbon-aware AI development, where energy use is minimised and emissions are transparently reported. This may involve:
Using energy-efficient model architectures.
Locating data centres in regions with renewable-heavy grids.
Extending hardware lifespans and recycling components.
The goal is to ensure AI is a net positive for climate outcomes, not a hidden source of emissions.
Building the ecosystem for physical AI
No single company or institution can deliver this transition alone. Open collaboration between researchers, industry, and government is essential. As LeCun has observed, “Good ideas come from the interaction of a lot of people and the exchange of ideas.”
This means:
Shared research infrastructure.
Cross-sector working groups on data standards.
Policy frameworks that encourage responsible AI deployment in critical infrastructure.
Q&A: Physical AI in Climate Tech
What is embodied or physical AI, and how does it differ from LLMs?
Embodied or physical AI refers to systems that can sense, reason, and act in the physical world. They use inputs from hardware such as sensors, robots, or industrial controllers. Unlike LLMs, which process and generate text, physical AI directly interacts with real-world systems.
How is AI accelerating materials discovery for climate solutions?
Machine learning can predict the properties of new materials and optimise synthesis routes, reducing the need for slow trial-and-error experiments. This has enabled breakthroughs in battery technology, solar cells, and carbon capture materials, as seen in work by the Materials Project among others.
What are the main barriers to scaling AI in real-world climate tech?
The “pilot trap” is the biggest barrier — many projects work in demonstration but fail in operational environments. Causes include poor integration with existing systems, data governance issues, and insufficient investment in deployment.
How does Nexus Climate support climate tech start-ups with physical AI?
Nexus Climate provides practitioner-led advisory, market access, and opportunity mapping. This helps founders move from pilot to commercial scale, with a focus on industrial and environmental applications.
What should innovators consider when developing sustainable AI for climate?
Energy use, hardware impacts, and transparency are key. Carbon-aware design, collaboration across disciplines, and lifecycle thinking are critical for ensuring AI’s benefits outweigh its costs.
From ambition to action
Physical AI represents a shift from information processing to real-world transformation. In MENA, Europe, and beyond, it is already enabling more efficient grids, cleaner industrial processes, and faster innovation cycles in materials science.
The next step is scale. That will take leadership, robust data practices, and partnerships that bridge the gap between lab success and market impact.
If you are developing AI that can see, sense, and act in the physical world — and want to bring it to market faster — contact Nexus Climate.
