The energy per AI query has fallen more than 30-fold in a year, and total AI emissions are still climbing. For CIOs and chief AI officers, the levers that matter most are model choice, workload design, and procurement.
In August 2025, Google reported that the median text prompt to its Gemini apps used 0.24 watt-hours of electricity, about as much as running a television for nine seconds. Energy per prompt had fallen 33-fold in twelve months.
Then, in May 2026, Google said it was processing more than 3.2 quadrillion tokens a month, up from about 480 trillion a year earlier. Its 2026 environmental report showed electricity use up 37% and total emissions up 18%.
That is the situation every enterprise AI program now faces. Each unit of AI keeps getting cheaper in energy, while the number of units grows faster. For the executives setting AI strategy, a sustainability plan that looks only at the provider's efficiency will miss where the emissions actually come from: how many tokens your own workloads use.
Published estimates have converged for simple chat prompts, and diverge sharply for everything else.
| Estimate | Workload | Energy per query | Notes |
|---|---|---|---|
| Google (Aug 2025) | Median Gemini text prompt | 0.24 Wh | Includes idle capacity and cooling; market-based emissions of 0.03 gCO2e |
| Epoch AI (Feb 2025) | Typical GPT-4o query | About 0.3 Wh | Rises to about 2.5 Wh with 10,000 input tokens |
| OpenAI's Sam Altman (Jun 2025) | Average ChatGPT query | 0.34 Wh | No methodology published |
| Microsoft researchers in Joule (Aug 2026) | Frontier models, median | 0.31 Wh | Earlier public estimates were 4 to 20 times too high |
| Microsoft researchers in Joule (Aug 2026) | Reasoning, about 5,000 output tokens | 3.91 Wh | 13 times a standard query |
| Zeke Hausfather (Aug 2026) | One agentic coding session | 150 to 1,200 Wh | One heavy user's logs; central estimate about 600 Wh |
The figures differ for three reasons, and anyone comparing them should know all three. Boundaries vary: some count only the active chips, others include idle servers and cooling. Emissions factors vary: Google's 0.03 grams is market-based, meaning it nets out clean energy purchases. MIT Technology Review noted it is about a third of what the local grid average would imply. And response lengths vary. Mistral's peer-reviewed lifecycle analysis of its Large 2 model, which includes training and hardware manufacturing, puts a 400-token response at 1.14 grams of CO2e, nearly 40 times Google's figure, largely because of these boundary choices.
The practical takeaway: a short chat prompt is a rounding error. A reasoning model, a long context window, or an agent is not.
Training a frontier model is a large, one-time cost. Serving it to millions of users is a continuous one. MIT Technology Review estimates that 80% to 90% of AI computing power now goes to inference. For an enterprise buyer this is good news, because inference is the part you control. You choose which model handles each task, how much context it sees, and how long it runs.
Reasoning and agents are the multipliers. Hugging Face's AI Energy Score found that turning on reasoning increased energy use about 30-fold on average, and 150 to 700 times for some models. It also found that 8 of 14 newer models used as much energy as their predecessors, or more. Newer models are not automatically more efficient, so procurement teams should test rather than assume.
For most enterprises, AI used through cloud APIs or software subscriptions falls under Scope 3, Category 1 (purchased goods and services). AI running on hardware you own or lease falls under Scope 2 for the electricity and Scope 3 for the equipment.
Three measurement approaches are in use, in order of accuracy:
The cloud providers' carbon dashboards help, but with limits:
Standards are catching up. The Green Software Foundation ratified its Software Carbon Intensity specification for AI in December 2025, with separate scores for providers and consumers and functional units such as tokens and inferences. Under the EU AI Act, providers of general-purpose AI models must document known or estimated energy consumption from August 2025, with enforcement powers active since August 2026. That documentation goes to regulators, not to customers, so buyers should ask for it directly.
| Lever | Typical reduction | Evidence |
|---|---|---|
| Route tasks to the smallest capable model | 15 to 50 times for well-defined tasks | UNESCO and UCL, 2025 |
| Make reasoning opt-in, not default | 5 times or more | Joule, 2026 |
| Use distilled models for high-volume tasks | 5 to 10 times | Joule, 2026 |
| Batch requests and use modern serving software | Up to 12.5 times per request | Hugging Face |
| Keep serving software current | 15% to 41% per token | ML.ENERGY, 2026 |
| Shorten outputs | About 54% when output length is halved | UNESCO and UCL, 2025 |
| Choose cleaner regions | Up to 13 times between best and worst cloud regions | Llopis, 2026 |
| Shift flexible jobs in time | Up to 80% for short jobs with a six-hour window | Dodge et al., 2022 |
| Newer hardware | 2.8 to 3.4 times throughput per GPU | Joule, 2026 |
Stacked together, the Joule authors estimate 8 to 20 times reductions are within reach. Two cautions apply. Quantization, running models at lower numerical precision, is often assumed to save energy, but Hugging Face researchers found it can increase energy use during text generation on some hardware, so measure before you rely on it. And prompt caching is priced at a fraction of fresh input on most APIs (Anthropic bills cache reads at 10% of the base input price), but its actual energy saving has not been independently measured.
Month one: get a baseline.
Month two: change the defaults.
Month three: put it in contracts.
For most companies, AI emissions are still small compared with the rest of their footprint. The same preprint estimated a 200-person firm's AI use at under one tonne of CO2e a year. That will not last. Agentic workloads can use hundreds of times the energy of a chat prompt, and enterprises are scaling them now.
The companies that build measurement and routing discipline while the numbers are small will be ready for disclosure rules, will keep their inference bills in check, and will be able to back their climate claims with evidence. Good carbon practice here is also good cost control: every token you do not need is energy and money you do not spend.
Chips and capital are no longer the binding constraint on the AI buildout. Firm power is. Here is how developers are closing the gap between now and 2032, and what each option costs in carbon.
Every hyperscaler matches 100% of its electricity with renewables, and every one of them reported higher emissions this year. The rules behind that paradox are being rewritten. Here is the standard we think the industry should hold itself to.