No inventory shows AI cutting greenhouse gases as data-centre load heads toward 945 TWh
UN climate language in Antalya asked governments to align AI with emissions cuts, but IEA and hyperscaler filings show the opposite movement in electricity and reported greenhouse gases. Use-case pilots are not netted against data-centre load or Scope 3. A climate benefit exists only if a later inventory prints a location-based decline.
Mary Nichols, former chair of the California Air Resources Board, and Jeremy Tamanini of Dual Citizen wrote in MIT Technology Review on 9 November 2026 that the climate case for AI has not been made. The piece landed as governments met in Antalya and the host asked that AI be designed, powered, and deployed to support climate action. That is a diplomatic sentence. It names no avoided ton, no base year, and no inventory boundary.
The International Energy Agency’s Energy and AI report (10 April 2025) recorded data-centre electricity at about 415 TWh in 2024, roughly 1.5% of global supply, and projected about 945 TWh by 2030 in its base case, more than a doubling, with AI a principal driver of the increment. Google’s 2024 Environmental Report listed 14.3 million tCO2e in 2023, 48% above 2019. Microsoft’s 2024 Environmental Sustainability Report listed total emissions 29.1% above the 2020 baseline, led by data-centre supply chains. Those accounts rose. Claimed climate benefits do not appear as line items that reverse them.
What corporate climate pages and the Antalya ask leave unscored is the gap between a pilot and an inventory. Grid forecasts, materials search, and satellite monitoring can cut emissions inside a named system. None of those results is netted against the IEA load path or against hyperscaler Scope 3. Patterson et al. (arXiv:2204.05149, 2022) documented large drops in energy per training run and argued the training footprint would plateau, then shrink. Aggregate demand did not follow that path. Annual renewable certificates can zero a marketing footprint while the local hour stays fossil. Google’s move toward 24/7 carbon-free energy is evidence that annual matching overstated cleanliness. Efficiency is not abatement.
The operational test is a filing. A positive climate case requires a location-based Scope 1–3 decline after new data-centre electricity and embodied carbon, against a stated base year. The next meters are the 2026–2027 sustainability reports of Microsoft, Google, Amazon, and Meta, and the IEA’s next data-centre series. If terawatt-hours and reported tons keep rising while “AI for climate” stays qualitative, Antalya produced a commitment without a measurement. Deployment is the test. These deployments have increased the measured load.
Sources (4)
- [1]IEA Energy and AI(https://www.iea.org/reports/energy-and-ai)
- [2]Google 2024 Environmental Report(https://sustainability.google/reports/google-2024-environmental-report/)
- [3]Microsoft 2024 Environmental Sustainability Report(https://www.microsoft.com/en-us/corporate-responsibility/sustainability/report)
- [4]The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink(https://arxiv.org/abs/2204.05149)