In short
Running AI on phones, sensors and local servers instead of in distant data centres can cut the energy of moving data and protect privacy. But it spreads computation across billions of devices whose energy and manufacturing footprint is rarely measured. Whether edge AI is greener depends on the model, the device, how often it runs and what it replaces, and the evidence is still thin.
Key points
- Edge AI runs models on or near the device that collects the data, instead of in a central data centre.
- It avoids sending data across networks, which use energy, and keeps data private, but it gives up the efficiency of shared, fully used servers.
- Studies comparing edge and cloud inference disagree; results depend heavily on the model, hardware and baseline.
- Making devices has a large carbon cost, so AI features that shorten device lifetimes or demand new hardware can outweigh energy savings.
- Edge AI is hard to measure and regulate because there is no data-centre meter, no PUE and no reporting duty.
Most of the debate about AI's energy use is about data centres, because that is where the big meters are. But a growing share of AI runs elsewhere: on phones, laptops, cars, cameras, doorbells and factory sensors. This is "edge AI". It may be part of the answer to AI's footprint, or a way of moving the footprint somewhere nobody is counting.
What is edge AI?
Edge AI runs a model on or near the device that collects the data, instead of sending the data to a distant data centre. "On-device" AI is the extreme case, where the model runs on the phone or sensor itself.
Edge AI is already common:
- Apple's 2025 technical report describes an on-device language model of about three billion parameters, compressed with 2-bit quantisation-aware training so it fits and runs on phones and laptops.
- Google's Android AICore service lets apps use Gemini Nano, a model that runs on the device, on supported Pixel and Samsung phones.
- Industrial cameras, drones, cars and medical devices run vision and control models locally because they cannot wait for, or rely on, a network connection.
The number of devices is large and growing. IoT Analytics counted about 18.5 billion connected Internet of Things devices at the end of 2024, not counting phones, PCs and tablets, and projected about 39 billion by 2030.
Why might edge AI save energy?
Because moving data costs energy, and local processing can avoid it. It also keeps sensitive data where it is.
The International Energy Agency has estimated that data transmission networks use roughly 1–1.5% of the world's electricity, a similar share to data centres. Processing data where it is collected, and sending only results, can reduce that load.
There is a caveat. Mytton, Lundén and Malmodin showed in 2024 that network energy is not directly proportional to data volume: much of a network's power is drawn whether or not data is flowing. Sending less data does not save energy one-for-one, so simple "kWh per gigabyte" estimates overstate the saving.
Other potential benefits:
- Privacy and sovereignty. Data that never leaves a device, a hospital or a grid operator cannot leak in transit or be used by a third party.
- Resilience. A local model keeps working when the network or the cloud is down.
- Latency. Answers arrive without a round trip, which matters for vehicles and machinery.
Why might edge AI cost more?
Because small devices are usually less efficient than large, shared, fully used servers, and because making devices has a large carbon cost.
- Utilisation. A data-centre accelerator can batch many users' requests and stay busy. A phone or sensor runs one user's request, then sits idle.
- Manufacturing. Wu and colleagues at Meta reported in 2022 that embodied carbon, from making hardware, was becoming the dominant source of AI's overall footprint, and noted that computation shifting to the edge may run where access to renewable energy is limited. AI features that require new phones, or make older ones feel slow, add manufacturing emissions that can outweigh any energy saved.
- Multiplication. A small cost per device, times billions of devices, is a large cost. A doorbell or camera running a model constantly may use more energy in total than occasional cloud calls would have.
What does the evidence say?
It is mixed, and depends heavily on the comparison. Li, Islam and Ren (2025) compared generative AI inference in the cloud and on edge devices and found a reduced environmental footprint for edge deployment of several models, with energy savings of over 90% in their tests on a recent phone. Their cloud baseline was not a fully optimised production service, which matters: well-batched servers in an efficient data centre are a harder benchmark to beat.
End-to-end studies that follow a whole AI service, from device through network to data centre, are still rare, especially for large language models. A 2026 UK green AI sandpit of mathematicians, computer scientists and industry listed modelling the net energy effect of edge AI, across millions of devices, as an open question.
Why is edge AI hard to measure and regulate?
Because there is no single meter. Data centres report energy use and power usage effectiveness (PUE), and in the EU, larger ones must report to a common scheme. Phones, sensors and desk-side AI workstations have no such duty.
The energy shows up, if at all, as battery drain or a slightly higher electricity bill spread across millions of households and businesses. That makes it easy to claim savings by moving work to the edge without anyone being able to check. See measurement is the bottleneck.
What makes edge AI greener?
Small, efficient models, run only when needed, on hardware people already own. The same methods that cut data-centre energy matter more on a battery:
- Compressed models: pruning, quantisation and distillation. See the maths that makes AI more efficient.
- Event-driven processing: wake the model when something happens, rather than running it continuously.
- Sending summaries, not raw data, when the cloud is needed.
- Supporting older devices, so AI features do not force replacements.
GreenCode, a European research project on energy-efficient software, makes the same point: small tuned models, retrieval instead of brute-force context, routing to the cheapest capable component and scheduling are what make AI viable on battery power and constrained hardware, where there is no option to spend more energy.
What to do next
- Decide between edge and cloud case by case, using the questions in edge or cloud?.
- Count manufacturing as well as running energy, and avoid features that need new hardware.
- Measure device energy during testing, not only accuracy and speed.
- Run models on events, not continuously, and send results rather than raw data.
- If you buy edge AI products, ask suppliers for energy per inference and supported device lifetimes. See sustainable AI procurement.
Frequently asked questions
What is edge AI?
AI that runs on or near the device where data is collected, such as a phone, camera, car or factory sensor, rather than in a remote data centre.
Is on-device AI better for the environment?
Sometimes. It can save the energy of sending data to the cloud, but small devices are less efficient than shared servers, and new AI features can push people to buy new hardware. The net effect depends on the case.
Does on-device AI drain my phone battery?
Running a model on a phone uses battery like any other demanding app. Manufacturers use small, compressed models and dedicated chips to keep this down, but the energy is still used, and it does not appear in any data-centre statistics.
Sources
- Data centres and data transmission networks, International Energy Agency (2023)
- Network energy use not directly proportional to data volume: the power model approach for more reliable network energy consumption calculations, Mytton, Lundén and Malmodin, Journal of Industrial Ecology (2024)
- Number of connected IoT devices (State of IoT 2025), IoT Analytics (2025)
- Apple Intelligence Foundation Language Models: Tech Report 2025, Apple (2025)
- Gemini Nano with Android AICore, Google, Android Developers (2024)
- A case study of environmental footprints for generative AI inference: cloud versus edge, Li, Islam and Ren, ACM SIGMETRICS Performance Evaluation Review (2025)
- Sustainable AI: environmental implications, challenges and opportunities, Wu et al., MLSys (2022)
- Green AI and energy-aware computing, GreenCode (2026)
Related
Edge or cloud: where should AI run?
How to weigh on-device and edge AI against central data centres: energy, network, hardware, privacy and resilience, with a step-by-step way to decide.
Green AI hosting and data centres
How to choose where AI runs: grid carbon intensity, PUE and WUE, renewable energy claims, water use and siting, with questions to ask hosts.
Measurement is the bottleneck for green AI
Why AI energy figures are hard to trust or compare: self-reported numbers, estimates not meters, different boundaries, and what good looks like.
The maths that makes AI more efficient
Pruning, sparsity, sketching, random matrix theory and physics-informed models: how mathematics cuts the computation, memory and energy AI needs.