Data centre electricity consumption reached 415 terawatt-hours globally in 2024, roughly 1.5% of total global electricity use, and the IEA projects that figure will more than double to 945 TWh by 2030. AI-focused servers drove much of the 17% surge in data centre electricity demand recorded in 2025 alone. The energy cost of the AI build-out is real and measurable. Whether it is the right lens through which to evaluate AI’s relationship with the energy transition is a separate and more consequential question.
The dominant concern in policy and media circles frames AI as a net liability for decarbonisation, primarily because large language models consume substantial electricity to train and run while delivering outputs of uncertain climate relevance. That framing is not wrong, but it addresses the wrong category of AI application. The case for AI as a net positive for the green transition does not rest on what LLMs can do with text. It rests on what narrower, purpose-built machine learning systems can do with materials science.
What LLMs Can and Cannot Do With Energy Data
The distinction matters, and it is frequently collapsed in public discussion. LLMs are statistical pattern-matching systems trained to predict probable sequences of text. Their performance is directly correlated with the volume of text available for a given topic. On subjects where the training corpus is large, such as corporate sustainability commitments or general descriptions of renewable energy policy, they perform competently. On subjects where the corpus is sparse or where precise numerical values are required, such as a specific company’s Scope 1 greenhouse gas emissions for a given year, their performance degrades significantly.
The implication for emissions data is not trivial. When tested across six major LLMs on whether a British energy company has a net zero target, all returned the correct answer. When the same LLMs were asked for the company’s 2025 Scope 1 greenhouse gas emissions value, fewer than one in five returned the correct figure, and two-thirds returned an incorrect value. One declined to answer. The LLMs that produced wrong answers also consumed more energy per query than a human researcher would have needed to retrieve the correct value from a public disclosure. At the scale of millions of such queries, the energy cost of confident incorrectness is not negligible.
This performance profile defines the limits of LLM utility in emissions monitoring, corporate accountability, and data-driven climate analysis. These are not tasks where the most probable answer derived from a large corpus of internet text is reliable. They require precise retrieval of specific, often recently published figures, a task better suited to structured databases and conventional search tools than to neural network-based language models.
The Materials Discovery Problem and Why AI Changes It
The more promising application is in the laboratory, and the gap between traditional and AI-assisted materials discovery timescales is now well documented. Conventional battery materials research operates through trial and error, with researchers synthesising candidates, testing electrochemical properties, and iterating over timescales measured in months per candidate. The candidate space is enormous: tens of millions of inorganic materials are theoretically possible, most unexplored.
In 2023, Google DeepMind’s GNoME system identified 2.2 million potentially stable new inorganic crystal structures in 17 days, including 380,000 assessed as sufficiently stable for real-world applications. Among these were 528 potential lithium-ion conductors and 52,000 new layered compounds similar to graphene. The Materials Project, which had previously identified approximately 28,000 stable predicted materials over a decade of computational work, saw its database expanded by a factor of ten through a single AI-driven exercise. A parallel autonomous laboratory at Berkeley Lab subsequently synthesised and validated 736 of GNoME’s predicted structures experimentally, confirming that the computational predictions translate to physical reality.
The GNoME results have since attracted scrutiny. A 2025 analysis in Chemistry and Engineering News identified concerns about duplicate structures within the dataset, suggesting that some materials claimed as novel had already been recorded. The researchers affiliated with Google DeepMind at the time of the study dispute the extent of the problem. The methodological debate does not invalidate the broader finding that AI-assisted screening can explore candidate spaces orders of magnitude faster than experimental approaches alone, but it does reinforce the point that AI-generated material predictions require experimental validation before any practical utility can be claimed.
Argonne, NJIT, and the Next Layer of the Pipeline
Subsequent work has moved the AI materials pipeline further along the path from prediction to characterisation. In 2024, a team at Argonne National Laboratory trained one of the largest chemical foundation models built to date, focused on small molecules relevant to battery electrolyte design, using the ALCF’s Polaris supercomputer. The team is now developing a second foundation model for molecular crystals using the Aurora exascale system, targeting the structural building blocks of battery electrodes.
In research published in Cell Reports Physical Science in 2025, a team at the New Jersey Institute of Technology, led by Professor Dibakar Datta, applied generative AI to discover five previously unknown porous transition metal oxide structures with potential for multivalent-ion batteries. These batteries, which use ions of magnesium, calcium, aluminium, and zinc carrying two or three positive charges rather than the single charge of lithium ions, could in principle store significantly more energy per unit mass. The NJIT team validated the AI-generated structures using quantum mechanical simulations and stability tests before claiming experimental viability. The important caveat is the same as for GNoME: identifying a promising structure computationally is the first step in a pipeline that still includes synthesis, characterisation, integration into a full cell architecture, cycling stability testing, and eventually manufacturing scale-up.
Earlier work at Pacific Northwest National Laboratory demonstrated that the transition from computational identification to initial physical synthesis can be completed in under nine months for a single candidate, compared to the years that conventional research timelines typically require. The acceleration is real. But it applies to the early screening stages of development; the downstream stages of cell engineering, electrolyte optimisation, and manufacturing process development are not compressed by materials discovery AI and remain time and capital-intensive.
The Energy Balance and the Strategic Question
The IEA estimates that AI applications in energy-intensive industries could reduce energy costs by 3 to 10 percentage points for firms that adopt them effectively. Data centre electricity demand growth is running at 15% per year under the IEA’s base case through 2030, well above the 3% growth rate of total global electricity demand. Whether the energy consumed by AI infrastructure is outweighed by the energy savings it enables, and over what timeframe, is the calculation that the current debate rarely specifies with adequate precision.
For LLM-based applications deployed at scale for general-purpose tasks, the energy balance is ambiguous. The electricity consumed per query has been falling rapidly as model efficiency improves, with the IEA noting that power consumption per AI task has declined at a rate unprecedented in energy history. But the absolute consumption continues to rise because the volume of queries is growing faster than efficiency gains can offset.
For AI applied specifically to battery materials discovery, the relevant comparison is not the energy cost of the computation against the energy cost of the chatbot query it displaces. It is the energy cost of the computation against the value of accelerating the development of battery technology that the green transition demonstrably requires. Storage is the binding constraint on solar and wind deployment at scale. If AI compresses the timeline from laboratory discovery to commercially deployable battery chemistry by even a few years, the energy savings unlocked by those batteries at the grid scale will dwarf the electricity consumed in generating the AI models that enabled the discovery. The arithmetic is not close, even under conservative assumptions about deployment rates. The question is whether the laboratory AI pipeline can deliver at the required pace before the Paris Agreement’s 2050 timeline makes the point moot.
