Written by Ralph Sun
A new review argues that the energy crisis in artificial intelligence is not a single problem but seven interacting regimes — scaling, experimentation, hardware floors, data movement, centralised infrastructure, benchmark incentives and static inference — and that the brain’s 20-watt budget is evidence of what is possible, not a blueprint to copy.
Large language models have made fluency look cheap. The electricity is not. A review published in Discover Artificial Intelligence by Mahmoud Amiri, Azadeh Mokari and Thomas Bocklitz of the Leibniz Institute of Photonic Technology and Friedrich Schiller University Jena sets out to map the problem properly, and concludes that it cannot be solved at a single layer of the stack.
The scale figures are the starting point. Drawing on scenario-based 2025 estimates, the review reports that inference for a single model, GPT-4o, was modelled to consume between 391.5 and 463.3 gigawatt-hours of electricity annually, alongside 1,335 to 1,580 megalitres of freshwater and 138 to 163 kilotonnes of carbon-dioxide-equivalent emissions. Across 369 generative AI models, global training plus inference has been estimated at 25 to 41 terawatt-hours. Global data-centre electricity consumption reached roughly 415 terawatt-hours in 2024.
Against that, the human brain runs on about 20 watts, with roughly 0.1 watts attributed to cortical computation. The authors are careful here, and the care matters: they use the brain as contextual evidence that competent cognition can persist under tight constraints, not as a directly transferable efficiency benchmark.
Their method was a systematised narrative review that began with 6,284 identified records, screened 898 at full text and selected 150 focal studies. From that base they built a cross-layer framework of seven interacting regimes: diminishing returns from scaling, the cost of broad empirical experimentation, hardware energy floors, data-movement penalties, centralised infrastructure demands, benchmark-driven evaluation incentives and static inference allocation.
The first regime puts numbers on the scaling problem. Using fitted exponents, the authors calculate that a mere 10% reduction in fitted test loss would require roughly 3.03 times more training tokens, 4.00 times more non-embedding parameters, or 8.23 times more minimum training compute. Human learning generalises from a handful of examples by recombining structured prior knowledge rather than accumulating task-specific experience.
The second regime covers the cost of looking for answers. The authors’ analysis of 374 AI systems published between 1952 and 2022 found training compute doubling every 20.6 months before 2010, accelerating to every 7.3 months for regular-scale systems from 2010 to 2022 and every 6.6 months for higher-scale systems — and those figures exclude hyperparameter search and tuning overhead entirely.
Hardware is the third, and the trend is uncomfortable. Across 62 machine-learning accelerators from 2017 to 2025, thermal-design-power-normalised peak tensor throughput rose about 20.0% a year while reported thermal design power rose 13.1% a year. Between 2023 and 2025, median efficiency improved 49% while median power draw climbed 94%. Chips are getting better per operation and hungrier per device.
Data movement is the most technically striking. Accessing off-chip DRAM can cost roughly 200 times the energy of one multiply-accumulate operation, against about 6 times for an on-chip buffer and 1 time for a local register file. Between 2012 and 2022, transistor-switching energy per bit fell roughly tenfold, yet reported end-to-end energy per instruction remains between 2.3 × 10⁴ and 9.7 × 10⁶ times higher once fetching, buffering and control are counted.
The remaining regimes are institutional and deployment-level: frontier training concentrated in a handful of well-resourced labs, benchmark incentives that reward marginal accuracy while ignoring resource cost, and dense models that process every input through the full network regardless of difficulty.
The review’s most important caution is that efficiency claims rarely survive the full system boundary. Retrieval-augmented generation adds embedding, indexing and reranking costs. Mixture-of-experts routing reduces activated computation but can lose the savings to expert imbalance and inter-device communication. The prescription is cross-layer co-design evaluated with direct end-to-end energy measurements at matched capability — because, as the authors put it, a watt saved in one layer is all too easily spent, invisibly, in another.
