Each record states what the source contributes and what it cannot establish. Provider disclosures and independent research are useful for different reasons, and neither is treated as complete by default.
Selected studies that materially support the public synthesis.
Estimates global data-centre electricity use at 415 TWh in 2024 and projects about 945 TWh in 2030 in its base case.
The totals cover all data centres, not AI alone, and the projection is a scenario rather than a measurement.
Estimates 176 TWh of US data-centre electricity in 2023, with a 325 to 580 TWh scenario range for 2028. It also estimates direct and electricity-related water use.
The analysis is operational, US-specific and covers the full data-centre sector rather than AI workloads alone.
Provides a lifecycle view of France’s digital sector, including manufacturing, imported impacts and 117 million tonnes of resources mobilised in 2022.
The study concerns the whole digital sector. It does not isolate generative AI or provide an AI-specific material footprint.
Measures inference energy for 88 models on common hardware and shows that model and task choice can change operational impact by orders of magnitude.
The experiments concern specific open models and hardware. They do not measure current commercial AI services.
Estimates energy, water and carbon per query across 30 language models, with more than a 65-fold energy spread for a long prompt.
Provider hardware and environmental parameters are inferred, not observed. The paper is a preprint and its estimates have changed between versions.
Builds a wide 2025 range for AI-related operational carbon and water by combining estimated AI power demand with company-wide environmental intensities.
This is a perspective based on several approximations. Operators do not disclose the workload data needed to isolate AI directly.
Production measurements and omissions reported by AI providers themselves.
Reports 0.24 Wh, 0.03 gCO₂e and 0.26 mL of direct freshwater for a median Gemini Apps text prompt in May 2025.
The figures exclude training and use market-based carbon accounting. Prompt length, model mix, absolute totals and uncertainty are not disclosed.
Reports lifecycle indicators for Mistral Large 2 and marginal figures of 1.14 gCO₂e, 45 mL of water and 0.16 mg Sb eq for a 400-token Le Chat response.
The underlying model, infrastructure and full study package are not public. The response result excludes training allocation.
Cites an external estimate of about 0.3 Wh for a typical GPT-4o query and describes hardware and software efficiency efforts.
The query figure is not based on OpenAI measurements. The disclosure contains no absolute energy, water or carbon total for ChatGPT.
Names AWS and Google Cloud as training providers and states that company operations include cloud computing.
It provides no model-level electricity, water or carbon quantity and no reproducible environmental accounting boundary.