Developments elsewhere that bear on this problem: research, results, incidents and policy, each with its source. The UNSOLVED editors choose them every week, with AI research help. They are not verdicts: the review board rules only on contributions.
The authors compare models in agent workflows by energy, cost and time per successfully completed task. In a pilot with 15 open-weight models and measured GPU energy, which model used least energy per task depended on task length and type. The result suggests per-query figures can misrank models once agents retry and chain calls.
arXiv · When Is Smaller Greener? Right-Sizing Language Models for Agentic AI
Bloomberg reports that on-site cooling savings leave the larger draw untouched: water used by power plants supplying data centres. Meta attributes about 96% of its 2024 water use to that source. Google says it does not report the figure, Microsoft gave none, and Amazon says a standard method is needed first. US estimates range from about 211 billion to 4 trillion gallons a year.
Bloomberg (republished by Insurance Journal) · AI's Water Strain Includes Surging Demand From Electricity Use
The authors measured inference energy for 86 open-weight checkpoints on seven benchmarks and six hardware platforms. Turning on reasoning raised energy about 3.7 times on average, and ranking by generated tokens put 18–24% of model pairs in the wrong order. They propose that benchmark scores be reported with measured energy under a stated serving setup.
arXiv · Measuring the Inference Energy Efficiency of Open-Weight LLMs across Benchmarks, Reasoning Modes, and Hardware