Chips, compute and data centres
Edge AI
In short
Edge AI is artificial intelligence that runs close to where data is collected, on a device or a nearby server, instead of sending every task to a distant data centre. It is aimed at faster answers and less raw data sent away.
- Also called
- Edge intelligence, AI at the edge
- Topic
- Chips, compute and data centres
- Level
- Beginner
- First described
- Zhou and colleagues wrote in 2019 that "edge AI", or "edge intelligence", still had no formal definition. They defined it as using end devices, edge nodes and cloud data centres together to train and run a neural network, not only as running a model on the device2.
Edge AI, also called edge intelligence, means doing the AI near the data. Xu and colleagues define it as connected systems and devices that collect, cache, process and analyse data close to where it is captured. Set against a cloud setup that uploads everything to a remote data centre, local processing can protect privacy, shorten the wait and save bandwidth1.
It is an arrangement, not a particular chip. Zhou and colleagues say that equating it with a model that runs entirely on the device is too narrow, and that in 2019 there was still no agreed definition. Their scale runs from work in the cloud, through work shared with a nearby server, to fully on-device work where nothing is sent off. Moving up that scale sends less data off the device, which raises privacy and cuts network delay, while computation time and energy on the device go up. Phones, cars and surveillance cameras are the devices they have in mind2.
Xu and colleagues divide the job into caching, training, inference and offloading to a nearby server. The servers can be gateways, routers or micro data centres, including at mobile base stations and on vehicles. End devices still cannot run some deep learning models, for lack of processor, graphics chip, memory and network capacity, which is why the surveys review model compression. They write that cloud voice assistants such as Siri and Cortana stop if the network is down, and they report a face-recognition system, split between a phone and an edge server, whose latency fell from 900 milliseconds in the cloud version to 1691.
The IEA's 2026 glossary describes the same placement: processing on or near the device, cutting network delay and bandwidth, for real-time work such as autonomous driving, industrial robots and on-device inference3.
Example
A surveillance system that pulls high-definition video from cameras needs heavy computation, a lot of bandwidth and a short delay, and it handles sensitive footage. Zhou and colleagues describe keeping that analysis at the edge, near the cameras, instead of sending the stream to a cloud data centre2.
Why it matters
Doing the work near the data can cut the wait and the amount of raw data that leaves the device, which Xu and colleagues link to privacy. Zhou and colleagues add that pushing more of the job onto the device also raises its computation and energy use12.
Questions people ask
Is Edge AI the same as on-device AI?
Not always. Zhou and colleagues say the common habit of meaning "it runs entirely on the device" is too narrow. On their scale, on-device inference sends no data off the device, while other levels share the work with a nearby server or the cloud2.
Does Edge AI work without the internet?
Sometimes. Xu and colleagues note that cloud voice assistants such as Siri and Cortana stop working if the network is down. A fully on-device setup, in Zhou's scale, sends no data off the device. Shared setups still use a local server or the cloud for part of the task12.
Is Edge AI more private?
It can be, when less raw data leaves the device. Xu and colleagues say local processing protects privacy compared with uploading to the cloud. Zhou and colleagues say the privacy gain depends on how much data is offloaded, and that doing more locally costs energy and delay12.