The honest answer to "how much energy does AI use" depends on what you're asking. A single chat reply is small — about as much as a bright LED bulb running for a minute. A short generated video is two orders of magnitude more. Multiply by the volume of queries served in 2026 and the cumulative number is large enough that grid operators are now planning around it. This guide gives the per-query figures, the per-model figures, and the trend lines that matter.
What changed in 2026
- Frontier-lab transparency improved. OpenAI, Google, and Anthropic now publish quarterly inference-energy disclosures under EU AI Act reporting requirements.
- Per-query energy dropped roughly 40% in two years thanks to mixture-of-experts routing, speculative decoding, and better hardware (Blackwell, MI400, TPU v6).
- Total demand grew faster than efficiency. AI inference is now ~3–4% of US electricity, up from ~1% in 2023. Microsoft and Google are signing nuclear PPAs to lock in capacity.
Per-query numbers (2026)
Real measured ranges from public methodology papers and lab disclosures:
| Workload |
Energy per request |
| Short text reply (frontier model) |
1.5–4 Wh |
| Long-context text reply (1M tokens) |
8–25 Wh |
| Image generation (1024×1024) |
5–15 Wh |
| Short video (5 sec, 720p) |
80–250 Wh |
| Voice agent (1 min interaction) |
6–12 Wh |
For comparison, a typical Google web search is around 0.3 Wh. An AI chat reply is 5–15× a search; a generated 5-second video is closer to running a microwave for a minute.
Where the energy actually goes
About 60% of inference energy is the GPU itself, 25% is cooling, 10% is power conversion, and the remaining ~5% is networking and storage. Cooling is where geography matters: a hyperscale facility in Phoenix uses 3–5× the water of one in Dublin or Stockholm because evaporative cooling does more work in dry heat.
Training is a different conversation. Training a frontier model in 2026 takes roughly 50–150 GWh — comparable to a small city's annual consumption. But training is amortized over billions of queries; the per-query training-amortized figure is small. The pressure on the grid is primarily inference at scale.
The buildout reality
US hyperscalers are bringing roughly 25–35 GW of new data-center capacity online over 2026–2027. That's the equivalent of 25 large nuclear reactors. The supply mix is the live question:
- Renewables-only is a marketing claim, not a 24/7 reality — even Google admits round-the-clock carbon-free is closer to 70% than 100%.
- Nuclear PPAs (Microsoft–Three Mile Island, Amazon–Talen) are real but won't add net capacity for 18–36 months.
- Natural gas peakers are filling the gap in regions where load grew faster than transmission. That's where the carbon is hiding.
What you can actually do
If you're a builder, the cheapest emissions reduction is using the smallest model that solves the job. Routing an order-status query to GPT-5 instead of a 4B model wastes 95% of the energy. Set up routing rules and measure them.
If you're a consumer, the marginal impact of one chat is tiny — focus on aggregate behavior (don't generate 4K video for fun) rather than guilt over a quick lookup.
FAQ
Is AI making climate change worse?
On current trends, yes — at the margin. AI demand is real new electricity load, much of it served by gas. The counterargument that AI accelerates clean-energy R&D is plausible but unverified at scale.
How much water does ChatGPT use per query?
Roughly 0.05–0.5 liters per long reply, depending on facility. Cumulative numbers get big at billions of queries per day.
Are smaller models meaningfully greener?
Yes — usually 10–50× less energy per query than frontier models for the workloads they can handle.
What's the trajectory?
Capacity grows faster than efficiency through at least 2027. After that, depends on whether scaling laws keep paying off and whether grid buildout keeps pace.
Where to go next
For more on running models efficiently see Local LLM setup guide for 2026, Small language models on the edge in 2026, and AI observability tools in 2026.