The honest climate case for AI

The snapshot: small today, noisy tomorrow

It has been reported that global data centers used roughly 415 TWh in 2024, about 1.5% of worldwide electricity, and that AI-specific servers consumed around 93 TWh in 2025 — roughly 0.3% of global power. Those are tidy-sounding numbers. A short query to a standard model allegedly uses about 0.3 Wh — three seconds of microwave time. So, are you saving the planet by deleting your ChatGPT account? Not really. Quitting a chatbot while still eating steak and driving an old sedan is, the author argues, climate theater.

Why this matters more than you think

But it has been reported that the “average query” is a moving target. Newer reasoning models and agentic workflows — think o3, GPT‑4.5, Claude with extended thinking, and production agents that plan, search and verify — can be 10–100× more energy-hungry per user task. Benchmarks allegedly put some of these at tens of Wh per query, turning a single user task into dozens or hundreds of inference calls. That flips the unit of cost: not prompt, but task.

And efficiency gains? Real, yes. NVIDIA’s Blackwell and algorithmic tricks deliver huge per-token improvements. Yet demand appears to be compounding faster than those savings. It has been reported that prompts rose from about 1 billion per day in late 2024 to 2.5 billion by mid‑2025, and that the IEA’s scenarios put data-center demand far higher by 2030–2035. This is Jevons paradox in action: as AI gets cheaper and better, people—and companies—use a lot more of it.

So what now? The honest climate argument for AI isn’t a moral finger-wag at individual users. It’s a call to focus on aggregate demand, the power mix that feeds data centers, and policy that steers efficiency gains toward emissions reductions rather than unchecked uptake. We can celebrate engineering wins and still be worried — because efficiency alone doesn’t guarantee a greener future.

Sources: dev.to/dcc, Lobsters