For two decades, the dominant trend in the IT sector has been centralisation: applications, data and processing have migrated to data centres that are increasingly large and increasingly distant from end users—this is cloud computing. This centralisation offers considerable advantages in terms of economies of scale, availability and ease of management. But it faces an insurmountable physical constraint: the speed of light. The unavoidable latency between a device and a data centre located several hundred kilometres away—typically 20 to 100 ms round trip—is imperceptible for a web page or an email, but a deal-breaker for a remote surgical application, an autonomous vehicle that must brake in under 10 ms, or a robotic production line requiring millisecond-level synchronisation.

Edge computing (edge computing) provides the solution to this challenge by bringing data processing closer to its source. Rather than sending all data to the central cloud for processing and returning the results, everything that can be processed locally is processed locally, and only aggregated data or information that is genuinely useful for a comprehensive analysis is sent to the cloud. This approach reduces latency, consumes less network bandwidth, preserves the confidentiality of sensitive data (which does not leave the company’s premises) and allows for degraded operation in the event of a cloud connection failure.

1. Edge computing hierarchy

Edge computing does not refer to a single location but to a hierarchy of processing levels, each offering a different balance between proximity to the source and available computing power.

The EDGE DEVICE refers to processing carried out directly on the source device: a microcontroller embedded in an industrial sensor, a dedicated processor in a smart surveillance camera, or a signal processing chip in a portable medical device. The field of TinyML (embedded machine learning) enables AI inference models to be run directly on these ultra-low-power microcontrollers, for applications such as voice keyword detection, image classification or the detection of vibration anomalies in industrial machinery.

The NEAR EDGE refers to IoT gateways, local rack-mounted servers and industrial PCs located within a factory, building or 5G base station. This equipment can aggregate and pre-process data from tens to thousands of sensors, execute closed-loop control algorithms with latencies of a few milliseconds, and store a day’s worth of production data locally before sending it to the cloud during off-peak hours.

The FAR EDGE, or REGIONAL EDGE, refers to points of presence (PoPs) located close to users: edge data centres comprising a few dozen servers deployed by telecoms operators in medium-sized towns, co-location facilities near industrial estates, or MEC (Multi-Access Edge Computing) nodes integrated into 5G base stations. These nodes offer latencies of 5 to 15 ms and can host demanding applications such as 3D graphics rendering for augmented reality, real-time video analysis or local content caching.

Finally, the central cloud – the hyperscalers AWS, Azure and Google Cloud – remains essential for training AI models on large volumes of data, conducting comprehensive aggregated analyses, long-term storage, and applications that do not require low latency. In a modern architecture, the four tiers coexist and work in tandem, with each task being executed at the most appropriate tier.

2. MEC, Multi-Access Edge Computing

MEC (Multi-Access Edge Computing), standardised by ETSI since 2014, is an architecture that places computing and storage capabilities directly at the edge of the 5G access network, physically within or in the immediate vicinity of base stations (gNB). This proximity enables latencies of less than 5 ms between applications and connected devices, which is impossible to achieve by routing traffic through the network core and the cloud.

The most promising MEC applications are concentrated in three areas.

In Industry 4.0, machine vision cameras positioned on production lines send their video feeds to the factory’s MEC server, which performs real-time quality control using convolutional neural networks, detects defects within milliseconds and triggers the ejection of defective parts. The latency to the cloud would be too high for a production line operating at a rate of several dozen parts per second.

In the field of connected vehicles, the V2X (Vehicle-to-Everything) standard enables vehicles to communicate with road infrastructure, traffic lights and other vehicles. The MEC server located near a junction can aggregate information from all approaching vehicles, calculate speed or trajectory recommendations to avoid collisions in less than 10 ms, and relay them back to the vehicles. A latency of 50 ms via the central cloud would not allow for this vital responsiveness.

Enterprise augmented reality also benefits from MEC: lightweight AR glasses (Microsoft HoloLens, RealWear), which lack the computing power required for complex 3D rendering, offload this processing to the MEC server, which returns the enhanced video feed within a few milliseconds. A maintenance technician can thus view a machine’s wiring diagrams as an overlay without having to consult paper documentation.

3. CDN, Content Delivery Networks

CDNs (Content Delivery Networks) are global networks of distributed servers that cache static and dynamic content as close as possible to end users. First developed in the 1990s to distribute web pages and images, by 2025 they had become critical infrastructure carrying over 70% of global internet traffic, including video streaming, APIs, software updates and protection against DDoS attacks.

The basic principle of a CDN is simple: rather than having users from all over the world connect to an application’s origin server—which would result in high latency for remote users and a massive overload on the server—the CDN replicates the content across hundreds or thousands of geographically distributed points of presence (PoP). When a user requests a resource, they are redirected to the nearest PoP, which responds within a few milliseconds from its local cache.

Cloudflare exemplifies the recent shift towards comprehensive edge computing platforms.

With over 300 PoPs in more than 100 countries, Cloudflare combines content delivery with

· DDoS protection (capable of absorbing attacks exceeding 2 Tbps)

· The Web Application Firewall (WAF), Zero Trust Network Access (Cloudflare Access)

·  More recently, the ability to run code at the edge via Cloudflare Workers. This latest feature allows developers to run JavaScript or WebAssembly code directly within Cloudflare’s PoPs, within 50 ms of any user worldwide, without having to manage any servers.

Akamai, a pioneer founded in 1998, operates one of the most extensive networks, with over 4,000 PoPs and 340,000 distributed servers. This position enables it to offer security services that extend beyond the CDN, particularly following the acquisition of Guardicore for network microsegmentation. AWS CloudFront, integrated into the AWS ecosystem, enables developers to run code at the edge via Lambda@Edge and CloudFront Functions, with native integration with other AWS services. Fastly stands out for its near-instantaneous cache flushing capability (less than 150 ms to propagate a cache invalidation across the entire network), which is critical for news publishers who publish content continuously.

Modifié le: vendredi 9 octobre 2026, 10:02