AI’s Environmental Cost Depends on More Than the Model—and Companies’ Claims Are Hard to Verify

November 16, 2026
AI’s Environmental Cost Depends on More Than the Model—and Companies’ Claims Are Hard to Verify
  • AI’s environmental footprint depends on more than the model: processing location, electricity sources, chips, cooling systems and the full infrastructure lifecycle all matter, yet limited corporate disclosure makes reliable measurement difficult.

  • Data centers use electricity to power processors and cooling, which can account for up to 40% of a facility’s energy use; an EPRI report estimated that a prompt to an early ChatGPT model used about 10 times the power of a typical Google search, while complex tasks can consume much more than simple queries.

  • Energy use also varies by model and task, depending on processing architecture, chip types and numbers, and the tokens generated; smaller models are not automatically more efficient if they need more tokens to finish a task.

  • Where computing happens matters because electricity mixes differ: Scandinavian facilities can draw more on wind, solar and hydroelectric power, while rising US demand risks greater fossil-fuel use and China continues to rely significantly on coal despite expanding renewables.

  • As AI use expands to billions of requests, energy demand and emissions accumulate, with image and video generation and autonomous agents requiring more computing power; infrastructure construction and equipment disposal add impacts beyond daily operations.

  • Google said an average text request on Gemini apps uses 0.24 watt-hours and reported that it cut energy per request 33-fold and carbon emissions 44-fold between 2024 and 2025, but outside researchers say limited access to its chips and technical data makes those claims difficult to verify.

  • Efficiency gains can deliver sizable savings: Delta Electronics estimated that a 0.5% improvement in average power-supply efficiency could save data centers about 5 terawatt-hours of electricity and 2 million tons of carbon dioxide emissions a year.

  • John Bistline said businesses can act even without complete data by selecting efficient models and cleaner-power locations; choosing a different model alone can change the energy used for the same task by as much as 30 times.

  • Companies’ net-zero claims may count renewable-energy certificates even as data centers draw power from fossil-fuel-heavy local grids, increasing calls from consumers, investors, businesses and regulators for more detailed environmental data.

  • Disclosure requirements are taking shape through California rules for large businesses, the European Union’s CSRD framework, which affects companies in the region and may reach multinationals, and international frameworks such as ISSB.

  • In Vietnam, Resolution 57-NQ/TW identifies AI as a strategic technology, but investment in computing infrastructure and talent should also prioritize energy efficiency, resource-saving technology and environmental responsibility in data-center construction and operation.

Summary based on 1 source


Get a daily email with more AI stories

More Stories