TL;DR
- Edge data centres are small facilities — typically 50 kW to a few MW — sited close to end users to minimise round-trip latency and reduce backhaul to central regions.
- Forms range from telco central-office racks and street-cabinet micro-DCs to containerised 'micro modular' facilities and Tier III-equivalent edge regions.
- Driven by latency-sensitive workloads (real-time AI inference, computer vision, AR/VR, autonomous systems, 5G MEC) and by data-sovereignty requirements that mandate in-country processing.
- Operational challenge: lights-out automation, remote hands, and standardised modular builds replace the on-site engineering teams of hyperscale facilities.
Definition#
Edge data centres are facilities deployed close to where data is generated or consumed — at the network edge — rather than centralised in a small number of large regions. 'Close' is workload-relative: for autonomous-vehicle inference, it may mean within 5 ms of the road; for a regional inference cache, within 20-30 ms of a metropolitan area; for sovereign data residency, simply within a country boundary.
Compared with hyperscale regions (typically 20-300 MW), edge sites are small: 50 kW to perhaps 2-5 MW. They trade economies of scale for proximity, and operational density for distribution.
Typical Forms#
| Form | Capacity | Typical siting |
|---|---|---|
| Street cabinet / pole-top | 1-10 kW | 5G base stations, smart-city installs |
| Telco central office (CO) | 20-200 kW | Existing telco real estate |
| Micro modular DC (container) | 50-500 kW | Industrial sites, ports, hospitals |
| Regional edge / metro edge | 0.5-5 MW | Tier-2 cities, county-level sites |
| Sovereign in-country edge | 1-20 MW | Capital city, primary metro |
Driving Workloads#
- Real-time AI inference where round-trip latency must stay below 20 ms — content recommendation, vision pipelines, voice assistants.
- 5G multi-access edge compute (MEC) — RAN-adjacent compute for slice-specific workloads.
- Industrial IoT and computer vision — factory-floor inference, port logistics, retail analytics where uplink bandwidth is constrained.
- Sovereign / data-residency — in-country processing required by regulation (UK, EU, India, GCC).
- Content delivery and caching — CDN PoPs, regional model caches.
Engineering Constraints#
Edge sites are not just small data centres. The operating model is materially different from a central region:
- Unstaffed by default: most edge sites are 'lights-out'. Remote-hands contracts replace permanent staff.
- Standardised modules: variability in design across hundreds of edge sites would be unmanageable. Most operators settle on 1-3 reference designs.
- Environmental tolerance: edge sites land in industrial parks, telco shelters, basements, rooftops — ASHRAE A3 or A4 tolerance is often required.
- Power constraints: utility supply is often single-source; battery and generator runtime become more important.
- Connectivity: dual diverse fibre paths are the exception, not the norm. SD-WAN over multiple bearers (fibre + 5G + satellite) is common.
- Tier classification: most edge sites are Tier II-equivalent; redundancy is achieved across sites rather than within one site.
Where Edge Sits Relative to Hyperscale#
- Hyperscale: training, large-batch processing, long-term storage. 20-300 MW per region; 2-4 regions per market.
- Regional / sovereign: production inference, regional sovereign workloads. 2-20 MW per site; one per metro.
- Metro edge: latency-sensitive inference, CDN, caching. 0.5-5 MW per site; several per metro.
- Far edge: real-time control, industrial vision, 5G MEC. 1-200 kW per site; thousands per market.
Operational Pitfalls#
- Site sprawl: managing hundreds of sites with consistent security posture, patch state, and configuration is harder than running one big site.
- Cooling at small scale: a 50 kW micro-DC has no economies of scale on cooling. PUE is structurally worse than hyperscale (typically 1.4-1.7).
- Physical security: small unstaffed sites need camera, intrusion, and tamper detection plus a documented response model.
- Spares logistics: getting parts to a remote edge site within an SLA window is a real problem; pre-positioned spares are common.
- Connectivity outages: edge sites are at the end of the network. Multi-bearer connectivity (fibre + 5G + satellite) protects against backhaul outages.