Three big things we still don't know about AI's energy burden (www.technologyreview.com)

After months of investigative efforts, recent disclosures from OpenAI and Google have provided the first concrete estimates of how much energy is consumed by AI models like ChatGPT and Gemini per query—around 0.24 to 0.34 watt-hours. This transparency addresses a long-standing gap in understanding AI’s environmental footprint, previously clouded by corporate secrecy. Yet, these figures represent only the energy used for simple chat interactions, excluding more demanding tasks like complex reasoning or multimodal outputs such as image and video generation, which are becoming increasingly common. The significance for the AI/ML community lies in acknowledging both progress and ongoing uncertainty. While individual AI queries may consume relatively little electricity—comparable to a microwave running for a few seconds—the aggregate growth of AI services and data centers is driving a rapid surge in global electricity demand, complicating tech companies’ sustainability goals. Despite promises that AI might eventually power climate solutions and improve energy efficiency, current evidence of such benefits remains anecdotal and insufficient to offset the industry’s growing carbon impact. Lastly, a critical unknown is whether AI’s projected exponential adoption will materialize or stall, a question that will shape the scale and longevity of its energy footprint. With AI prompting billions of queries daily and plans for more specialized data centers underway, the industry faces intense scrutiny over its expanding power needs. Whether this growth reflects a lasting transformation or a speculative “bubble” could determine AI’s true environmental cost—and the urgency with which the community must innovate toward greener AI infrastructure.