AI's Energy Problem Isn't Your Chatbot Query

Hannah Ritchie's new Our World in Data breakdown of data-center energy use hit the Hacker News front page this week, and the headline numbers are genuinely calming: data centers consumed roughly 485 TWh in 2025, about 1.5% of global electricity, with AI-focused facilities accounting for around 155 TWh of that — a third of the total, half a percent of the world's power. A typical chatbot query costs 0.24–0.34 watt-hours, about a second of microwave time.

Those numbers are solid. They're also measuring the wrong era. The per-query figures describe the chatbot workloads of 2024–2025, while the industry — and every developer reading this — is busy shipping agents. That shift changes the math by two orders of magnitude, and it's the part of Ritchie's piece worth actually acting on.

The numbers hold up

The core figures trace back to the IEA's Energy and AI report, which projects global data-center consumption doubling from 415 TWh in 2024 to about 945 TWh by 2030 — just under 3% of global electricity, growing around 15% a year, four times faster than everything else on the grid. AI is the main driver: the IEA expects US data-center demand to grow by about 240 TWh over that window, and China's by around 175 TWh.

The per-query estimates check out too. Google's own measurement of the median Gemini text prompt — 0.24 Wh, 0.03 grams of CO₂, 0.26 ml of water, including cooling and power distribution — matches the low end of Ritchie's range, and OpenAI has cited 0.34 Wh for an average ChatGPT query. Worth noting: Google also reported the energy of that median prompt dropped 33x in a single year. Inference efficiency is improving absurdly fast, which is a big reason the doomier extrapolations from 2023 keep missing.

So the "your ChatGPT habit is boiling the oceans" framing is dead, and good riddance. Ritchie's arithmetic — an EU resident's daily electricity use equals thousands of long AI queries — settles it.

The 200x problem nobody's pricing in

Here's where the reassuring story gets shaky. Buried in the same analysis: a long-context query (~7,500 words of input) runs about 2.5 Wh. A maximum-length context costs around 40 Wh. An agent doing multi-step reasoning: roughly 50 Wh. That's a 200x spread between the median chat prompt everyone quotes and the workload every AI company is currently racing to make the default.

And agents don't just cost more per run — they multiply runs. A coding agent that plans, executes, reads tool output, and self-corrects might burn through dozens of model calls with fat contexts where a 2024-era chatbot made one. Nobody publishes a "median agent session" energy figure yet, and I'd bet that's not an accident. The honest per-interaction number for agentic products is plausibly three orders of magnitude above the 0.24 Wh figure doing PR duty.

This is why the IEA's base case has AI-focused capacity tripling to 465 TWh by 2030 even as per-token efficiency improves 30x+: efficiency gains are being converted directly into longer contexts, more reasoning tokens, and more autonomous loops. Jevons paradox with a GPU attached.

The grid is the real bottleneck — and it's local

The global 1.5% figure is a comfortable average hiding brutal concentration. Ireland's CSO reported data centers at 23% of national metered electricity in 2025, up from 5% a decade ago. Virginia is around 25%, per EIA data, with state studies projecting 36–51% by 2030. The US overall sits near 5%.

For developers this stopped being an abstract sustainability concern a while ago. Power availability — not chips, not capital — is now the binding constraint on where new capacity gets built, which is why hyperscalers are signing nuclear deals and why GPU capacity, spot pricing, and region availability increasingly track substation queues rather than silicon supply. If you've wondered why your preferred region has no H100 quota while some place you've never deployed to does, the answer is often measured in megawatts.

Your token budget is an energy budget

The practical upshot is cleaner than most sustainability advice, because for once the incentives align perfectly: API pricing is denominated in tokens, and tokens are a decent proxy for joules. The energy-efficient architecture and the cheap architecture are the same architecture.

Concretely, that means the optimization work you should be doing anyway: route requests to the smallest model that clears your quality bar instead of defaulting everything to the frontier tier; use prompt caching so a 50k-token system context isn't recomputed on every call; put a context budget on agent loops rather than letting them accumulate an entire repo into the window; batch offline workloads onto discounted (and better-scheduled) batch APIs. A team that cuts token spend 60% through routing and caching has cut its inference energy by roughly the same margin. No carbon dashboard required — your billing page is one.

Where this lands

Ritchie's conclusion — that how the electricity is generated matters more than how much AI uses — is right at the planetary scale, and the per-query guilt genre deserves the burial she gives it. But the developer-relevant reading is different: the comfortable numbers describe yesterday's workloads, the expensive tail (agents, long context, reasoning) is precisely where the industry is steering, and the grid constraints are already shaping what capacity you can rent and where. The good news is that in this market, frugality with tokens and frugality with watts are the same discipline — and it's one you're already paid to have.