News: data centers and local AI

Reports on how much electricity data centers use, and research on running AI models on your own hardware. Running locally isn't free either: it pays off with small models, on hardware you already own, used when you need them.

  1. Small local models use a fraction of the energy per word

    A benchmark of 18 open models, from 0.5 to 7 billion parameters, on one consumer graphics card found the smallest used about 0.27 joules per token, and a 7-billion-parameter model up to 8.6 times more. Architecture and quantization mattered, not only size.

    Source: arXiv

  2. IEA: data center electricity use grew 17% in 2025

    Data centers' electricity demand rose 17% in 2025 and AI-focused data centers' rose 50%, while the world's demand grew 3%. The IEA expects data centers' use to double by 2030, to about 950 terawatt hours, about 3% of the world's electricity.

    Source: International Energy Agency, Key Questions on Energy and AI, reported by Enlit

  3. Compressing local models cut their energy use by up to 62%

    Researchers measured models running on a laptop's processor alone (a MacBook Pro M2) and on a Raspberry Pi 5. Compressing the models cut processor and memory use by up to 71.9% and energy use by up to 62%, with answers as good or better.

    Source: arXiv

  4. IEA: Energy and AI

    The IEA's first report on energy and AI put data centers' electricity use at about 415 terawatt hours in 2024, around 1.5% of the world's, and projected about 945 terawatt hours by 2030.

    Source: International Energy Agency