# Energy-efficient AI: how to make AI use less

> Energy-efficient AI in practice: the levers that reduce AI energy most, ranked, from smaller models and quantisation to routing, caching and location.

Source: https://greenai.guide/articles/energy-efficient-ai/  
Updated: 2026-11-20

## In short

Energy-efficient AI means getting the same useful result for less electricity. The biggest savings usually come from choosing a smaller or task-specific model, switching off long reasoning when it isn't needed, and running in an efficient data centre on a clean grid; techniques such as quantisation, distillation, routing, caching and batching then reduce AI energy further. Published studies report savings from tens of per cent up to 90% or more for individual levers, and far larger when they are combined.

## Key points

- UNESCO and UCL found that small task-specific models can cut energy use by up to 90%, and that shorter prompts and responses can cut it by over 50%.
- Reasoning models used on average 30 times more energy than standard models in Hugging Face's AI Energy Score tests; switching reasoning on raised one model's energy 150 to 700 times.
- In a 2024 study, generating 1,000 images used about 2.9 kWh, against 0.002 kWh for 1,000 text classifications.
- Google researchers estimated that the choice of model, data centre and processor together can reduce the carbon footprint of training by up to around 100 to 1,000 times.
- Efficient serving matters: vLLM's memory management gave 2 to 4 times the throughput at the same latency, and speculative decoding made one model 2 to 3 times faster with identical outputs.
- Starting short training jobs at low-carbon times cut emissions by up to 80% in one study, but by less than 1.5% for a long job.

Energy-efficient AI is AI that gets you the result you need for less electricity. This page sets out the main ways
to reduce AI energy, roughly in order of how much they usually save, with the published figures behind each one.

## What is the quickest way to make AI use less energy?

**Use the smallest suitable model, and don't switch on heavy modes you don't need.** Model choice usually outweighs
every other lever because energy scales with how much computation each answer takes. A 2025 report by UNESCO and
University College London (UCL) found that small task-specific models can cut energy use by up to 90%, and that
shorter prompts and responses can cut it by over 50%. Used together, the report says, such measures can reduce
energy consumption by up to 90%.

The second big lever is location. Google researchers estimated in 2021 that the choice of model, data centre and
processor together can reduce the carbon footprint of training by up to around 100 to 1,000 times. Efficient AI
models, run in efficient places, are where most of the gain lies.

## Which levers save the most? A ranked list

**The order below reflects the typical size of the saving in published studies, largest first.** Your own results
will depend on the task, the quality you need and where the work runs, so measure before and after (see
[how to measure the energy and carbon of AI](/practice/measure-ai-energy/)).

### 1. Don't use generative AI for jobs it isn't needed for

**Generation is far more expensive than simpler tasks.** In a 2024 study, Sasha Luccioni, Yacine Jernite and Emma
Strubell measured image generation at a mean of about 2.9 kWh per 1,000 images, against 0.002 kWh per 1,000 text
classifications: well over a thousand times as much per request. They also found that general-purpose generative
models are orders of magnitude more expensive than task-specific systems for the same task.

Video is heavier still. A 2025 study of open text-to-video models measured nearly 90 Wh for a short clip from one
model and over 415 Wh from a larger one, with energy growing quadratically with the size and length of the video.
See [how to decide if you need AI for a task](/how-do-i/decide-if-i-need-ai-for-a-task/) and
[how to choose classical machine learning instead of an LLM](/how-do-i/choose-classical-ml-instead-of-an-llm/).

### 2. Right-size the model

**Pick the smallest model that meets your quality bar.** Beyond the UNESCO–UCL finding, research on cascades
shows how much headroom there is: FrugalGPT, which tries cheaper models first and escalates only when needed,
matched GPT-4's performance with up to 98% lower cost. Cost is not energy, but for hosted models it is a reasonable
rough guide to computation. See [choose the right model for the job](/practice/choose-the-right-model/),
[how to choose a smaller AI model](/how-do-i/choose-a-smaller-ai-model/) and
[how to find a task-specific model](/how-do-i/find-a-task-specific-model/).

### 3. Use reasoning modes only when needed

**"Thinking" modes multiply energy.** Hugging Face's AI Energy Score tests in late 2025 found that reasoning
models used, on average, 30 times more energy than standard ones, and that the same model used between 150 and 700
times more energy with reasoning on than off. Agents, which make many calls in a loop, can add up the same way. See
[how to use reasoning modes only when needed](/how-do-i/use-reasoning-modes-only-when-needed/) and
[agentic AI and token waste](/articles/agentic-ai-and-token-waste/).

### 4. Choose efficient hardware, data centres and low-carbon locations

**Where and on what the work runs can change its carbon by several times.** Jesse Dodge and colleagues found that
training the same BERT model emitted about 7 kg of CO₂ in the most efficient cloud region and about 26 kg in the
least efficient, almost four times as much. Hardware keeps improving too: Epoch AI estimates that leading machine
learning chips become about 40% more energy-efficient each year. See
[green AI hosting and data centres](/practice/hosting-and-data-centres/),
[how to pick a low-carbon cloud region](/how-do-i/pick-a-low-carbon-cloud-region/) and
[how to choose energy-efficient hardware for AI](/how-do-i/choose-energy-efficient-hardware-for-ai/).

### 5. Shorten prompts, context and outputs

**Energy rises with how much the model reads and writes.** Epoch AI estimates a typical ChatGPT query at about
0.3 Wh, rising to about 2.5 Wh with a 10,000-token input and almost 40 Wh with a 100,000-token input. Resending
long conversations and asking for long answers are the everyday causes. See
[how to write prompts that use less energy](/how-do-i/write-prompts-that-use-less-energy/),
[how to limit how much an LLM writes](/how-do-i/limit-how-much-an-llm-writes/) and
[how to stop resending the whole conversation](/how-do-i/stop-resending-the-whole-conversation/).

### 6. Route requests to the right model

**Send easy requests to a small model and hard ones to a large one.** Routing research reports large savings:
Hybrid LLM made up to 40% fewer calls to the large model with no drop in response quality, and RouteLLM reduced
costs by over two times in some cases without compromising quality. See
[how to set up model routing](/how-do-i/set-up-model-routing/) and
[orchestration over generation](/articles/orchestration-over-generation/).

### 7. Quantise the model

**Store and compute with fewer bits per number.** [Quantisation](/glossary/#quantisation) reduces memory and data
movement, which dominate energy in large models. LLM.int8() cut the memory needed for inference by half while
keeping full-precision performance, and the UNESCO–UCL report found compression such as quantisation saved up to
44% of energy. See [how to quantise a model](/how-do-i/quantise-a-model/).

### 8. Distil a large model into a small one

**Train a small "student" model to copy a large "teacher".** [Distillation](/glossary/#distillation) is how many
efficient AI models are made. DistilBERT was 40% smaller and 60% faster than BERT while keeping 97% of its
language-understanding ability. See [how to distil a large model into a small one](/how-do-i/distil-a-large-model-into-a-small-one/).

### 9. Prune and exploit sparsity

**Remove weights the model doesn't need.** SparseGPT showed that very large language models can be
[pruned](/glossary/#pruning) to at least 50% [sparsity](/glossary/#sparsity) in one shot, without retraining, at
minimal loss of accuracy. Energy savings depend on whether the software and hardware can skip the zeros, so pruning
saves less in practice than the headline sparsity suggests. See [how to prune a neural network](/how-do-i/prune-a-neural-network/).

### 10. Cache, batch and serve efficiently

**Avoid repeating work and keep the hardware busy.** Serving software matters: vLLM's paged
[KV cache](/glossary/#kv-cache) management improved throughput 2 to 4 times at the same latency, and speculative
decoding made T5-XXL 2 to 3 times faster with identical outputs. Providers price reuse accordingly: OpenAI's Batch
API costs 50% less than standard requests, and Anthropic charges 10% of the normal input price for cached prompt
reads. See [how to cache LLM calls](/how-do-i/cache-llm-calls/), [how to batch AI requests](/how-do-i/batch-ai-requests/),
[how to use speculative decoding](/how-do-i/use-speculative-decoding/) and
[how to cut the energy of serving a model](/how-do-i/cut-the-energy-of-serving-a-model/).

### 11. Schedule flexible work for cleaner times

**Carbon-aware scheduling cuts emissions, not energy, and works best for short, flexible jobs.** In Dodge and
colleagues' study, delaying the start of a short training job to a low-carbon time cut emissions by up to 80% in
one US region, but by less than 1.5% for a long job; pausing and resuming a long job saved up to about 25%. See
[how to schedule training when the grid is greener](/how-do-i/schedule-training-when-the-grid-is-greener/) and
[carbon-aware computing](/glossary/#carbon-aware-computing).

## Energy-efficient AI levers at a glance

**This table summarises the levers, who can use them and the savings reported in the sources.** Figures are from
different studies with different baselines, so read them as indications of scale, not as additive totals.

| Lever | Who can use it | Typical saving reported | Source |
|---|---|---|---|
| Use a task-specific model instead of a generative one | Developers, organisations | Orders of magnitude for tasks such as classification | Luccioni, Jernite and Strubell, 2024 |
| Smaller, task-specific model | Users, developers | Up to 90% energy | UNESCO and UCL, 2025 |
| Reasoning off unless needed | Users, developers | 150–700 times less for one model; 30 times on average across models | Hugging Face AI Energy Score, 2025 |
| Model, data centre and processor together | Developers, organisations | Up to about 100–1,000 times less carbon (training) | Patterson et al., Google, 2021 |
| Low-carbon cloud region | Developers, organisations | About 3.7 times between best and worst region | Dodge et al., 2022 |
| Shorter prompts and responses | Users, developers | Over 50% energy | UNESCO and UCL, 2025 |
| Routing | Developers | Up to 40% fewer large-model calls | Ding et al., 2024 |
| Quantisation | Developers | Up to 44% energy; half the memory | UNESCO and UCL, 2025; Dettmers et al., 2022 |
| Distillation | Developers | 40% smaller, 60% faster, 97% of ability kept | Sanh et al., 2019 |
| Pruning | Developers | At least 50% of weights removed | Frantar and Alistarh, 2023 |
| Efficient serving and batching | Developers | 2–4 times the throughput | Kwon et al., 2023 |
| Speculative decoding | Developers | 2–3 times faster, same outputs | Leviathan et al., 2022 |
| Carbon-aware start time | Developers, researchers | Up to 80% (short jobs); under 1.5% (long jobs) | Dodge et al., 2022 |
| Parameter-efficient fine-tuning (LoRA) | Developers, researchers | 10,000 times fewer trainable parameters, 3 times less GPU memory | Hu et al., 2021 |

## How can users reduce AI energy?

**Choose lighter options and ask for less.** Most people can't change the data centre, but they can change what
they ask for:

- Use a fast or "mini" model for everyday tasks, and save the largest models and reasoning modes for hard problems.
- Keep requests clear and ask for short answers; start a new chat instead of continuing a very long one.
- Avoid regenerating again and again: refine the request instead. See [how to stop regenerating AI answers](/how-do-i/stop-regenerating-ai-answers/).
- Generate images and video only when you need them. See [how to cut the impact of AI image and video generation](/how-do-i/cut-impact-of-ai-image-and-video-generation/).
- Turn off AI features you don't use. See [use AI efficiently: everyday habits](/practice/use-ai-efficiently/).

For scale, Google reported in 2025 that a median Gemini text prompt used 0.24 Wh. Individual text prompts are
small; the totals come from very high volumes and from the heavier tasks above. See the
[guide for the public](/guides/public/).

## How can developers build efficient AI models and systems?

**Right-size first, then optimise serving, then measure.** A practical order:

1. Set a quality bar on your own data and test the smallest candidate models against it.
2. Add routing or a cascade so the large model handles only what the small one can't.
3. Quantise, and consider distillation if volumes are high.
4. Serve with an efficient runtime: batching, paged KV cache, prompt caching and speculative decoding.
5. Cap output length and agent loops; see [how to stop AI agents wasting tokens](/how-do-i/stop-ai-agents-wasting-tokens/).
6. Fine-tune with LoRA rather than training from scratch; see [how to avoid training a model from scratch](/how-do-i/avoid-training-a-model-from-scratch/).
7. Measure energy per task before and after; see [how to track AI energy per task](/how-do-i/track-ai-energy-per-task/).

More detail is in [efficient AI training and inference](/practice/efficient-training-and-inference/) and
[maths for efficient AI](/articles/maths-for-efficient-ai/).

## How can organisations reduce AI energy at scale?

**Set defaults, buy well and measure.** The biggest organisational levers are which model is the default, which
features are on by default, and where workloads run:

- Make a small, fast model the default in internal tools, with larger models available on request.
- Ask suppliers for energy per task and data-centre figures; see [sustainable AI procurement](/practice/sustainable-ai-procurement/).
- Choose providers and regions with clean power; see [hosting and data centres](/practice/hosting-and-data-centres/).
- Set a policy and train staff; see [how to write a green AI policy](/practice/green-ai-policy/) and
  [how to train staff to use AI efficiently](/how-do-i/train-staff-to-use-ai-efficiently/).
- Set a baseline and targets; see [how to set a baseline for AI energy savings](/how-do-i/set-a-baseline-for-ai-energy-savings/).

## Does energy-efficient AI reduce total energy use?

**Only if total use doesn't grow faster than efficiency improves.** Google reported that the energy of its median
Gemini prompt fell 33-fold in a year, yet data-centre electricity use keeps rising worldwide because use is growing
faster. This is the [rebound effect](/glossary/#rebound-effect). Efficiency lowers the cost of each task;
sufficiency asks whether the task needs doing. Both are needed, as
[efficiency is not sufficiency](/articles/efficiency-is-not-sufficiency/) explains, and
[how to stop efficiency gains being cancelled out](/how-do-i/stop-efficiency-gains-being-cancelled-out/) gives
practical steps. For the wider picture, see [AI and data centre electricity demand](/articles/ai-data-centre-electricity/)
and [AI energy statistics](/articles/ai-energy-statistics/).

## What to do next

- Test whether a smaller model meets your quality bar, and switch to it if it does.
- Turn off reasoning, image and video modes by default where they aren't needed.
- Add caching, batching and routing to any AI feature with real volume.
- Choose a low-carbon region and an efficient provider for new workloads.
- Measure energy per task before and after each change, and report the result.
- Find help in the [consultants directory](/consultants/) or learn more through [courses](/courses/).

## Frequently asked questions

### What is energy-efficient AI?

It is AI that delivers the result you need for less electricity, through smaller or more suitable models, leaner software, efficient hardware and data centres, and using AI only where it adds value. It is part of green AI, which also asks whether the total use is sensible.

### How can I reduce AI energy use?

As a user, choose a smaller or faster model, keep prompts and requested answers short, avoid regenerating and use reasoning or image and video modes only when needed. As a developer, right-size the model, then add quantisation, routing, caching and batching, and run in a low-carbon region.

### Are smaller AI models more energy efficient?

Usually, for the tasks they can do well. UNESCO and UCL report that small task-specific models can cut energy use by up to 90% on suitable tasks. Check quality on your own work: a small model that has to be re-run several times may save nothing.

### Does quantisation save energy?

Usually yes, because the model moves less data and needs less memory. UNESCO and UCL report savings of up to 44% from model compression such as quantisation, and LLM.int8() halved the memory needed for inference. The saving depends on whether the hardware supports the lower precision.

### Which AI uses the most energy?

Video generation, then image generation and long reasoning or agent tasks. In a 2025 study, one open model used nearly 90 Wh for a short video clip and a larger one over 415 Wh, and reasoning models used about 30 times more energy than standard ones on average.

### Does efficiency actually reduce AI's total energy use?

Not automatically. Cheaper, faster AI tends to be used more, the rebound effect. Efficiency reduces energy per task; keeping totals down also needs sufficiency, meaning using AI where it is worth it.

### Do AI providers save energy with prompt caching and batch processing?

They avoid repeated work, which is why they charge less for it: OpenAI discounts batch jobs by 50%, and cached input on several providers costs around a tenth of the normal rate. Prices are a rough guide to compute saved, not a measurement of energy.

## Sources

- [AI large language models: new report shows small changes can reduce energy use by 90%](https://www.unesco.org/en/articles/ai-large-language-models-new-report-shows-small-changes-can-reduce-energy-use-90), UNESCO and University College London (2025)
- [AI Energy Score v2: refreshed leaderboard, now with reasoning](https://huggingface.co/blog/sasha/ai-energy-score-v2), Hugging Face (Luccioni and Gamazaychikov) (2025)
- [Power hungry processing: watts driving the cost of AI deployment?](https://arxiv.org/abs/2311.16863), Luccioni, Jernite and Strubell (FAccT 2024) (2024)
- [Video killed the energy budget: characterizing the latency and power regimes of open text-to-video models](https://arxiv.org/abs/2509.19222), Delavande, Pierrard and Luccioni (2025)
- [How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), Epoch AI (2025)
- [Carbon emissions and large neural network training](https://arxiv.org/abs/2104.10350), Patterson et al., Google (2021)
- [Good news about the carbon footprint of machine learning training](https://research.google/blog/good-news-about-the-carbon-footprint-of-machine-learning-training/), Google Research (2022)
- [Measuring the carbon intensity of AI in cloud instances](https://arxiv.org/abs/2206.05229), Dodge et al. (FAccT 2022) (2022)
- [Measuring the environmental impact of delivering AI at Google scale](https://arxiv.org/abs/2508.15734), Elsworth et al., Google (2025)
- [DistilBERT, a distilled version of BERT](https://arxiv.org/abs/1910.01108), Sanh et al., Hugging Face (2019)
- [LLM.int8(): 8-bit matrix multiplication for transformers at scale](https://arxiv.org/abs/2208.07339), Dettmers et al. (2022)
- [SparseGPT: massive language models can be accurately pruned in one-shot](https://arxiv.org/abs/2301.00774), Frantar and Alistarh (2023)
- [FrugalGPT: how to use large language models while reducing cost and improving performance](https://arxiv.org/abs/2305.05176), Chen, Zaharia and Zou (2023)
- [Hybrid LLM: cost-efficient and quality-aware query routing](https://arxiv.org/abs/2404.14618), Ding et al. (2024)
- [RouteLLM: learning to route LLMs with preference data](https://arxiv.org/abs/2406.18665), Ong et al. (2024)
- [Efficient memory management for large language model serving with PagedAttention](https://arxiv.org/abs/2309.06180), Kwon et al. (2023)
- [Fast inference from transformers via speculative decoding](https://arxiv.org/abs/2211.17192), Leviathan, Kalman and Matias (2022)
- [LoRA: low-rank adaptation of large language models](https://arxiv.org/abs/2106.09685), Hu et al. (2021)
- [Batch API](https://developers.openai.com/api/docs/guides/batch), OpenAI (2026)
- [Prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching), Anthropic (2026)
- [Leading ML hardware becomes 40% more energy-efficient each year](https://epoch.ai/data-insights/ml-hardware-energy-efficiency), Epoch AI (2024)
