# 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.

Source: https://greenai.guide/practice/edge-or-cloud/  
Updated: 2026-11-19

## In short

Neither edge nor cloud is greener by default. Central data centres usually run large models more efficiently, while on-device and edge AI can avoid network traffic, keep data local and work offline. Decide case by case by comparing energy per useful result across the whole system, including devices, network and hardware.

## Key points

- Run small, frequent or latency-sensitive tasks locally; send large or rare tasks to efficient central hardware.
- The IEA estimates data transmission networks used 260–360 TWh in 2022, about 1–1.5% of global electricity, with mobile networks around two-thirds.
- Network energy does not scale neatly with data volume, so per-gigabyte savings are often overstated.
- Billions of devices add up: count users' devices, their batteries and the hardware they may need to replace.
- Local AI keeps data on the device, which helps privacy, sovereignty and resilience.

## Is edge or cloud AI greener?

**It depends on the task, the model and the device, so compare them across the whole system rather than assuming.**
Central data centres run accelerators at high utilisation with specialised cooling, so they usually do a large
model's maths more efficiently. Edge AI, running on the device or a nearby server, avoids sending data across the
network and can switch off completely when idle.

A 2026 UK green AI sandpit of mathematicians, computer scientists and industry asked exactly this question. AI is
spreading into sensors, cameras, doorbells, cars and phone assistants. Multiplied by millions of devices, does local
AI cost more or less energy than sending the data to a data centre? Participants noted there are end-to-end energy
studies for some smaller vision models deployed at the edge, but few for large language models. The honest answer
today is: measure your case.

## What are the trade-offs?

**Each option wins on different things.**

| Factor | On-device / edge | Central data centre |
|---|---|---|
| Energy per computation | Usually higher on general-purpose chips; better on dedicated neural accelerators | Usually lower, especially at high utilisation |
| Model size | Small, often quantised models | Any size |
| Network energy | Little or none | Every request travels |
| Idle energy | Can sleep or switch off | Shared capacity kept ready |
| Hardware | Spread over billions of devices; may push upgrades | Concentrated, refreshed every few years |
| Latency | Fast, works offline | Depends on connection |
| Privacy | Data can stay on the device | Data leaves the user's control |
| Measurement | Hard: scattered, unreported | Possible, if providers report |
| Control | Distributed | Concentrated in a few providers |

## How much does the network matter?

**Networks use a lot of energy in total, but much of it is a near-constant baseline, so the effect of any one
request is smaller than per-byte calculators suggest.**

- The International Energy Agency estimates that data transmission networks used 260 to 360 TWh in 2022, around 1
  to 1.5% of global electricity, with mobile networks around two-thirds of that.
- Mytton, Lundén and Malmodin showed in 2024 that network energy is not directly proportional to data volume.
  Network equipment draws power whether busy or not, so converting gigabytes straight into kilowatt-hours
  overstates the saving from sending less.
- Cut data anyway: it saves energy on servers and devices, and avoided traffic can avoid future network capacity.
  But don't use per-gigabyte network savings as the main argument for edge AI.

Note that many published per-prompt figures leave the network and users' devices out. Google's 2025 Gemini figures,
for example, exclude both. Check the boundary before comparing edge and cloud numbers.

## What about the devices themselves?

**Count the energy, battery and hardware of the devices your AI runs on, because they are part of your footprint.**

- **Embodied emissions:** manufacturing devices produces carbon. The SCI specification includes a share of
  embodied emissions in a software system's footprint. If on-device AI pushes people to replace working phones or
  laptops sooner, that cost can outweigh any operational saving.
- **Battery drain** is a hidden energy cost the user pays. The sandpit flagged phone AI assistants as an example.
- **Adaptation:** the Green Software Foundation suggests switching to smaller models or lower frequency when
  battery is low or a device is hot.
- **Scale:** a tiny saving or cost per device becomes large across millions of devices running all the time.

## How do privacy and resilience come in?

**Local AI keeps data on the device, and keeps working when the connection or the central service fails.**

- **Privacy:** data processed locally does not need to be transmitted or stored by a provider. That supports the
  data-minimisation principle in UK and EU data-protection law, and suits sensitive data such as health or
  critical-infrastructure records. The sandpit noted that in critical-infrastructure sectors, sensitive
  operational data often must stay inside the organisation, which calls for efficient local models.
- **Sovereignty and control:** distributed inference spreads control away from a few large providers.
- **Resilience:** the sandpit also raised dependence: what happens when society relies on AI and the connection or
  power goes down? Local capability can keep essential functions running.
- **Accountability:** edge AI is harder to monitor. If you deploy it, plan how you will measure and report it.

## How do I decide, step by step?

**Work out the energy per useful result for each option, then weigh privacy, latency and resilience.**

1. **Define the task and its quality bar.** What counts as a successful result?
2. **Find the smallest model that meets the bar** ([choose the right model](/practice/choose-the-right-model/)).
   Quantised models need less memory: 8-bit inference halved memory needs in the LLM.int8() study.
3. **Estimate cloud energy per result**, including data-centre overhead and, where possible, network.
4. **Measure device energy per result** on representative devices, including older and low-power ones.
5. **Add hardware effects:** will this need new devices or dedicated edge servers?
6. **Consider frequency:** frequent, small tasks often suit the device; rare, heavy ones suit the cloud.
7. **Weigh privacy, latency, offline use and resilience.**
8. **Consider a split:** a small local model handles most requests and escalates hard ones to the cloud.
9. **Measure after deployment** and revisit as models and hardware change.

## Checklist

- [ ] Energy per useful result compared for edge, cloud and a split
- [ ] Boundary includes devices and, where possible, network
- [ ] Smallest adequate model chosen, quantised for devices where it holds up
- [ ] No device upgrades forced by the AI feature
- [ ] Battery and heat handled, with a lighter mode when low
- [ ] Privacy and data-protection benefits recorded
- [ ] Monitoring planned for edge deployments

## What to do next

1. Pick one AI feature and sketch its edge, cloud and split options.
2. Measure energy per result on a typical device and estimate the cloud equivalent; see
   [measure AI energy](/practice/measure-ai-energy/).
3. Test a quantised small model on your oldest supported device.
4. Decide what data must stay local for privacy or sovereignty reasons.
5. Read [AI at the edge](/articles/ai-at-the-edge/) for the wider picture, and
   [hosting and data centres](/practice/hosting-and-data-centres/) for the cloud side.
6. Find edge AI tools in the [products directory](/products/) and research groups in [research](/research/).

## Frequently asked questions

### Is on-device AI greener than cloud AI?

Sometimes. A small model on a phone avoids data-centre and network energy, but devices are less efficient per computation than data-centre accelerators, and buying new hardware to run AI has its own footprint. Compare the whole system.

### Does sending data to the cloud use a lot of energy?

Networks use substantial energy, but much of it is a near-constant baseline. Cutting data still helps devices and servers, but per-gigabyte network savings are usually overstated.

### What is edge computing?

Computing done close to where data is produced or used: on the device itself, or on a nearby gateway or small local server, rather than in a distant data centre.

### Why does privacy come into it?

Data processed on the device need not leave it. That reduces exposure and can make it easier to comply with data-protection rules such as data minimisation.

## Sources

- [Data centres and data transmission networks](https://www.iea.org/fuels-and-technologies/data-centres-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](https://doi.org/10.1111/jiec.13512), Mytton, Lundén and Malmodin, Journal of Industrial Ecology (2024)
- [Green Software Patterns: run AI models at the edge](https://github.com/Green-Software-Foundation/patterns/blob/main/docs/architecture/system-topology/run-ai-models-edge.md), Green Software Foundation (2025)
- [Green Software Patterns: optimise power consumption for edge AI inference](https://github.com/Green-Software-Foundation/patterns/blob/main/docs/operations/optimize-edge-ai-power-performance.md), Green Software Foundation (2025)
- [Measuring the environmental impact of delivering AI at Google scale](https://arxiv.org/abs/2508.15734), Google (arXiv 2508.15734) (2025)
- [Software Carbon Intensity (SCI) specification](https://sci.greensoftware.foundation/), Green Software Foundation (2024)
- [Principle (c): Data minimisation](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/a-guide-to-the-data-protection-principles/data-minimisation/), Information Commissioner's Office (2023)
- [LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale](https://arxiv.org/abs/2208.07339), Dettmers et al. (arXiv 2208.07339) (2022)
