How to tame AI's voracious appetite for energy
AI data centers in the US consumed 224 terawatt-hours of electricity in 2025, representing over 5% of the country's total electricity use — a dramatic increase from just 1.9% in 2018 before the generative AI surge. A single median-length Gemini prompt consumes approximately 0.24 watt-hours, and with ChatGPT serving over 900 million weekly users and billions of daily queries, cumulative energy demand is escalating rapidly. New "hyperscale" AI data centers can consume a gigawatt or more of power —
Analysis
TL;DR
- AI data centers in the US consumed 224 terawatt-hours of electricity in 2025, representing over 5% of the country's total electricity use — a dramatic increase from just 1.9% in 2018 before the generative AI surge.
- A single median-length Gemini prompt consumes approximately 0.24 watt-hours, and with ChatGPT serving over 900 million weekly users and billions of daily queries, cumulative energy demand is escalating rapidly.
- New "hyperscale" AI data centers can consume a gigawatt or more of power — roughly a tenth of Los Angeles's electrical capacity — often relying on fossil fuel plants due to their placement in regions lacking abundant renewable energy.
- Without intervention, US data centers could emit 24 to 44 megatons of CO2 annually, with offset strategies falling short unless clean energy generation exceeds data center consumption.
- Researchers are actively developing energy-saving algorithms, more efficient processor designs, and rethinking data center siting and construction to reshape AI's energy trajectory.
Why It Matters
This article highlights an urgent and often overlooked dimension of the AI boom: its massive and accelerating environmental footprint. For AI practitioners and industry leaders, understanding and addressing energy consumption is no longer optional — it is a strategic, regulatory, and ethical imperative that will shape the sustainability and public acceptance of AI technologies for years to come.
Technical Details
- Transformer architecture (Google Brain, 2017) underpins most modern LLMs and generative AI systems, enabling parallel computation across word relationships but at high computational cost, primarily executed on NVIDIA GPUs originally designed for gaming graphics rendering.
- Per-query energy metrics: Google estimates a median Gemini text prompt uses ~0.24 watt-hours (equivalent to ~9 seconds of TV viewing), but aggregated across billions of daily queries across platforms like ChatGPT, this compounds into enormous total demand.
- Hyperscale data centers represent a new class of AI infrastructure, consuming 1+ gigawatts each compared to ~100 megawatts for pre-AI data centers, with major investments from Google, Meta, Amazon, OpenAI, Anthropic, Microsoft, and Oracle.
- Energy sourcing challenge: Many new data centers are built in regions without abundant renewable energy (hydropower, geothermal, solar, wind), leading to reliance on natural gas plants that release CO2; offset investments elsewhere do not reduce net emissions unless clean generation exceeds data center consumption.
- Expert-driven mitigation efforts include developing energy-efficient algorithms, next-generation processor chip designs, and strategic considerations around data center location, cooling systems, and water resource management.
Industry Insight
- Tech companies must treat energy efficiency as a first-class design constraint — from model architecture choices to hardware procurement — rather than an afterthought, as regulatory pressure and public scrutiny around AI's carbon footprint will only intensify.
- The strategic location of new data centers near renewable energy sources could become a competitive advantage, reducing both operational costs and emissions while improving corporate sustainability narratives.
- Investors and executives should monitor the emerging market for energy-efficient AI hardware and software solutions, as the companies that solve the AI energy problem will capture significant value in an industry under growing environmental pressure.
Disclaimer: The above content is generated by AI and is for reference only.