How Much Energy Does an AI Query Actually Use?

Ask five people how much electricity one AI query uses and you'll get five answers, most of them off by an order of magnitude. The honest answer is that it depends on the model, the task, and where you draw the boundary around the measurement. But real published figures now exist, measured by vendors on their own production hardware and tracked by the International Energy Agency, and they're specific enough to plan around.

Here's what those figures actually say, and what they mean if you run a business in Burnaby, New Westminster, or anywhere else in Metro Vancouver.

The figure everyone quotes, and where it came from

In August 2025 Google published a technical report that metered energy per prompt inside its own production serving stack instead of estimating it from the outside. The headline finding: "the median Gemini Apps text prompt consumes 0.24 Wh of energy" (Elsworth et al., arXiv:2508.15734). The same work puts one median prompt at 0.03 gCO2e and 0.26 mL of water (Google Cloud).

Two details matter more than the headline.

The first is the boundary. Google ran the same measurement a second way, counting only active TPU and GPU draw, and got 0.10 Wh instead of 0.24 Wh. Google calls that narrower version "an optimistic scenario at best" that "substantially underestimates the real operational footprint of AI" (Google Cloud). Same prompt, same silicon, more than double the answer depending on whether you count idle capacity, CPU and memory, and data centre overhead. That one choice explains most of the disagreement you see in headlines about AI energy use.

The second is that 0.24 Wh is a median, for text, from a single vendor. Not a mean. Not an image, not a video, not a twenty step agent run.

Google also reported that energy for the median text prompt fell 33 times, and its carbon footprint 44 times, over the twelve months to May 2025. Efficiency is moving fast, which is another way of saying any figure you read here is perishable.

What that works out to on a Metro Vancouver power bill

BC Hydro's published Small General Service energy charge, the rate most small commercial customers in BC sit on, is 14.06 cents per kWh, plus a basic charge of 40.89 cents per day (BC Hydro business rates, read September 2026). Put the two numbers together:

• 1,000 median text prompts use about 0.24 kWh, roughly 3.4 cents of electricity.
• 10,000 prompts a month, which is a genuinely busy AI receptionist, comes to about 2.4 kWh, or roughly 34 cents a month.
• One desktop PC drawing an illustrative 100 watts across an eight hour day uses 0.8 kWh, more than 3,300 median prompts.
• Your daily basic charge alone, at 40.89 cents, outweighs a whole month of that AI workload.
• The API invoice for those queries will be hundreds of times larger than the electricity behind them.

So if you're weighing an AI deployment for your shop, electricity is not the line item to argue about. Vendor fees, integration work, and the staff time to supervise the thing are the real costs.

Why the aggregate numbers still sound alarming

Both things are true at once. The per query figure is tiny, and the total is large. The IEA put global data centre electricity use at "around 1.5% of the world's electricity consumption in 2024, or 415 terawatt-hours (TWh)" (IEA, Energy and AI, 2025).

The IEA's 2026 update has that consumption "roughly doubling from 485 TWh in 2025 to 950 TWh in 2030," reaching "around 3% of global electricity demand" (IEA, Key Questions on Energy and AI, 2026). Data centre demand "soared by 17% in 2025," against global electricity demand growth of 3% (IEA, 16 April 2026).

That same update is blunt about how small the text query slice of it is: "Simple text queries now typically consume less electricity than running a television over the same period of time," and if every conventional internet search were replaced with a simple AI text query, "it would consume less than 4 terawatt-hours (TWh) of electricity annually."

The growth is real and it is concentrated in a few regions. It is mostly not coming from you asking a chatbot to rewrite an email.

The number that should actually change your design

Here's the part with operational consequences. The IEA warns that newer, energy-intensive AI applications "can consume hundreds or thousands of times more energy per query than simple text generation" (IEA, 2026).

Peer reviewed work backs that up. Luccioni, Jernite and Strubell measured the energy and carbon needed for 1,000 inferences across model types and found that "multi-purpose, generative architectures are orders of magnitude more expensive than task-specific systems for a variety of tasks, even when controlling for the number of model parameters" (Power Hungry Processing, FAccT 2024).

A 2026 analysis goes finer, separating prefill, where the model reads your input, from autoregressive decoding, where it writes the answer, as distinct energy components (Vartziotis et al., arXiv:2607.26571). The practical read: tokens the model writes cost more than tokens it reads. Verbose answers and long reasoning chains are where energy, latency and money all go together.

Which gives you three levers that cost nothing to pull. Pick the smallest model that clears your quality bar. Cap output length. Don't reach for image or video generation when a template does the job.

Where this doesn't apply

Be careful about stretching any of this too far.

Google's 0.24 Wh is Google's own figure, self published rather than independently audited, measured on its own accelerators serving its own models. It is the most transparent production number publicly available, and it is not a rate card for every vendor.

A median also hides the tail. Long document summarisation and multi step agents sit well above it, and if that's most of your workload the headline figure understates you badly.

None of these per prompt numbers include model training or the embodied energy in the hardware. If you need auditable emissions figures for a tender or a sustainability disclosure, a vendor blog post won't survive scrutiny. You need the vendor's own reporting under a recognised standard, and most still can't hand you a per account figure.

Energy is also almost never the right reason to skip an automation. If your call volume is twenty a week, skip it because the setup and maintenance won't pay back, not because of the power draw.

The environmental case cuts both ways too. The IEA notes that proven AI applications "could help firms in energy-intensive industries reduce their energy costs by 3 to 10 percentage points" (IEA, 2026). That finding is about industrial energy users, not a dental clinic in New Westminster, so nobody should claim it as their own offset.

One last local wrinkle. BC's grid is unusually clean, but your queries almost certainly execute in a data centre outside the province, on that region's mix. Where the electron comes from isn't your choice unless you deliberately pick a vendor region.

The short version

Per query, AI energy use is small enough that it shouldn't appear in a small business case at all. In aggregate it's a genuine grid planning problem, which is a policy question rather than a purchasing one. And the design decisions that matter to you, model size and output length and task fit, happen to cut cost, latency and energy at the same time.

Sources

• Elsworth, C., Huang, K., Patterson, D., et al. "Measuring the environmental impact of delivering AI at Google Scale." Google, arXiv:2508.15734, 21 August 2025. https://arxiv.org/abs/2508.15734
• Google Cloud. "Measuring the environmental impact of AI inference." 21 August 2025. https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
• International Energy Agency. "Energy and AI: Executive summary." 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
• International Energy Agency. "Key Questions on Energy and AI: Executive summary." 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
• International Energy Agency. "Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions." News, 16 April 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
• Luccioni, A.S., Jernite, Y., Strubell, E. "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" ACM FAccT 2024, arXiv:2311.16863. https://arxiv.org/abs/2311.16863
• Vartziotis, T., Kosteli, R., Vartziotis, E., et al. "From Tokens to Watt-hours: Analytical Energy Estimation for LLM Inference on Modern GPUs." arXiv:2607.26571, 29 July 2026. https://arxiv.org/abs/2607.26571
• BC Hydro. "General Service Business Rates." Read 6 September 2026. https://app.bchydro.com/accounts-billing/rates-energy-use/electricity-rates/business-rates.html

If you're sizing up an AI deployment and want the real cost picture rather than the scary one, we're happy to walk through it with you. Autana Solutions builds and runs AI employees for small and mid-sized businesses around Burnaby, New Westminster and the rest of Metro Vancouver. Book a free call and we'll look at your actual volumes and tell you honestly whether automating is worth it yet.