What is an edge data center?

An edge data center is a small facility — typically 100 kW to 5 MW — located close to end users or devices rather than in a major hub, cutting round-trip latency to single-digit milliseconds. It complements, rather than replaces, regional and hyperscale facilities by handling latency-sensitive and bandwidth-heavy processing locally.

How big is the edge data center market?

Roughly $16.9 billion in 2026, projected to reach $71.9 billion by 2035 at a 17.5% CAGR. Micro data centers are the largest segment at 36.3% share, and the modular edge sub-segment (sub-5 MW deployable compute) was valued at $7.29B in 2025, growing 19.8% annually toward $25.5B by 2034.

How much does an edge data center cost?

Containerized micro sites of 100-500 kW typically cost $1-5M deployed — often quoted at $7-12M per MW equivalent, above hyperscale unit costs because fixed engineering is amortized over less capacity. The compensation is speed (3-9 months to live versus 24-48 for greenfield) and placement value that a distant facility cannot deliver.

What is the difference between an edge data center and a regional data center?

Scale, distance, and mission. Regional facilities run 10-100+ MW, serve a whole metro or country at 10-40 ms latency, and optimize for cost per kW. Edge sites run 0.1-5 MW at under 5-10 ms from users, are usually unmanned (lights-out), and optimize for proximity. Workloads split accordingly: training and bulk storage centrally; inference, caching, and real-time processing at the edge.

What are the main use cases for edge data centers?

AI inference close to users, CDN caching and video delivery, telco 5G/MEC functions, industrial IoT and real-time control, retail and healthcare on-site processing, and low-latency finance and gaming. AI inference is the current growth driver — market analyses point to inference overtaking training in workload volume around 2027, redistributing capacity outward.

Are edge data centers replacing hyperscale data centers?

No. They grow together: centralized facilities keep winning on cost per unit of compute for training and batch workloads, while edge absorbs the latency-sensitive slice. The edge market's 17.5% CAGR comes on a $16.9B base — a fraction of total data center spend — and most edge architectures depend on hyperscale cores behind them.

Who builds and operates edge data centers?

Four groups: specialist edge operators and modular vendors; telcos converting exchange sites and cell aggregation points; hyperscalers extending on-ramps (Local Zones, edge nodes); and enterprises deploying on-premises micro data centers. Deployment models range from owned containers to colocation in regional edge facilities to fully managed edge-as-a-service.

Edge Data Centers Explained: Economics, Use Cases, and Deployment Models

Edge data centers are small facilities (0.1-5 MW) placed near end users to cut latency. The market hits $16.9B in 2026, growing 17.5% annually to 2035.

An edge data center is a small facility — typically 100 kW to 5 MW — deployed close to end users to bring compute within single-digit-millisecond reach, instead of the 20-100 ms round trips to centralized hubs. The segment is real money in 2026: roughly $16.9 billion globally, growing at 17.5% annually toward $71.9 billion by 2035, with micro data centers (36.3% share) the largest format and AI inference the demand engine.

Key takeaways

Market size: ~$16.9B in 2026 → $71.9B by 2035 (17.5% CAGR), per InsightAce Analytic.
Micro data centers lead with 36.3% market share in 2026 on fast, plug-and-play deployment (Coherent Market Insights).
Modular edge (sub-5 MW deployable): $7.29B in 2025, 19.8% CAGR toward $25.5B by 2034 (Moduledge).
Economics trade unit cost for placement: ~$7-12M/MW-equivalent deployed and higher per-kW opex than hyperscale, paid back through latency, bandwidth savings, and 3-9 month deployment.
The structural driver is inference: analyses converge on inference overtaking training in workload volume around 2027, pushing capacity outward from central clusters.
Edge complements, never replaces, core capacity — architectures pair edge sites with regional and hyperscale backends.

Compare edge against regional colocation economics in our price index and SEA facility catalog.

Definitions: what counts as “edge”

The term spans several distinct formats:

Format	IT load	Location	Typical operator
On-premises micro DC	5-100 kW	Inside enterprise/retail/factory sites	Enterprise, managed-service vendor
Containerized/modular edge	100-500 kW	Parking pads, cell aggregation, substations	Edge specialists, telcos
Metro edge facility	0.5-5 MW	Tier 2/3 cities, network aggregation points	Regional colo, edge operators
Regional DC (for contrast)	10-100+ MW	Major metros/hubs	Colo majors, hyperscalers

Working definition: a facility is “edge” when its placement is the product — it exists because being under ~5-10 ms from a specific user population, machine cluster, or radio network creates value a bigger, cheaper, farther facility cannot.

Two engineering consequences follow. Edge sites are usually lights-out (unmanned, remotely operated, DCIM-heavy) because staffing hundreds of small sites is uneconomic. And they standardize hard: identical modules, identical spares, fleet-level monitoring — closer to telecom network practice than to campus data center operations.

Micro and modular economics: 0.1-0.5 MW

The dominant deployment quantum is the containerized 100-500 kW module — integrated power, cooling, racks, fire suppression, and security in a factory-built unit.

Metric	Micro/modular edge (0.1-0.5 MW)	Hyperscale reference
Capex, deployed	~~$1-5M per site (~~$7-12M/MW-equiv.)	~$10-11M/MW (air-cooled)
Time to live	3-9 months	24-48 months greenfield
Staffing	0 on-site (remote NOC)	15-25 FTE per 10 MW
PUE	1.2-1.5 (free-air/DX; some liquid)	1.2-1.4
Expansion	Add modules	Add halls/buildings

The per-MW premium versus hyperscale reflects amortizing engineering, site works, and network over small capacity. The offset is threefold: speed (a module ships while a greenfield is still in permitting), granularity (deploy exactly where and how much demand exists, avoiding stranded capacity), and avoided backhaul (processing video or sensor data locally can cut WAN bandwidth requirements by an order of magnitude — often the single largest line in the business case).

Failure economics differ too: a 250 kW site can justify N+1 on cooling but rarely 2N across the board. Fleet designs instead lean on software-level redundancy — traffic fails over to the next-nearest site — which is acceptable precisely because edge workloads are stateless or replicated.

Use cases in 2026

AI inference. The growth engine. Serving models near users cuts token latency and egress cost; region-wide analyses point to the 2027 inflection where inference overtakes training in volume, redistributing capacity from central clusters outward. Practical today: recommendation, vision QC, speech, and compact-LLM serving on 1-8 GPU nodes per site (GPU cost context: /gpu/).
CDN and video. The original edge workload — caching, live-event distribution, game downloads — still the bandwidth majority.
Telco 5G/MEC. Core network functions and multi-access edge computing at aggregation sites; telcos are converting exchanges into revenue-bearing edge colocation.
Industrial IoT. Real-time control loops, predictive maintenance, machine vision at plants, ports, and mines — where round trips to a distant cloud violate control deadlines and data-sovereignty rules keep footage on-site.
Retail, healthcare, finance. POS analytics, imaging pre-processing, and latency-sensitive trading/gaming nodes in second-tier metros.

Edge vs. regional data center: choosing

Decision factor	Choose edge	Choose regional colo
Latency requirement	<10 ms to a defined population	10-40 ms acceptable
Workload	Inference, caching, real-time control	Training, batch, storage, general compute
Cost priority	Bandwidth/backhaul savings, placement	Lowest $/kW (see index)
Geography	Dispersed users, Tier 2/3 cities	Concentrated in major metros
Ops model	Lights-out fleet	Staffed facility, richer on-site services

Most real architectures are hybrid: train and store centrally, serve and cache at the edge. In Southeast Asia the pattern is pronounced — hyperscale campuses concentrate in Johor, Bangkok, and Batam while archipelago and secondary-city coverage (Vietnam’s provinces, the Philippines, Indonesia beyond Jakarta) increasingly falls to sub-5 MW edge and modular builds; regional supply is mapped in our data center catalog.

Deployment models

Own and operate: buy modules ($1-5M each), secure sites and power, run the fleet. Maximum control; requires critical-facility ops capability.
Edge colocation: rent racks in someone else’s metro-edge facility — pricing resembles retail colo with a placement premium (get comparable numbers via /quote/).
Edge-as-a-service / managed: vendor owns and operates the hardware to an SLA; you consume compute. Fastest, least capex, least control.
Hyperscaler edge: AWS Local Zones/Outposts, Azure Edge Zones and equivalents put cloud APIs in metro sites — simplest for cloud-native stacks, at hyperscaler pricing and lock-in.
Telco partnerships: compute hosted at operator aggregation sites, often the only structured option deep in emerging-market geographies.

Worked example: a 250 kW inference edge site

To make the economics concrete, consider a containerized 250 kW site in a Tier 2 Southeast Asian city serving regional AI inference and content caching:

Line	Value
Capex (module, site works, grid connection, network)	~$2.5M
IT fit-out (32 inference GPUs + storage/network)	$1.5-2.5M depending on GPU choice
Power @ $0.09/kWh, PUE 1.35	~$265k/year
Connectivity (2 diverse 10G+ routes)	$40-80k/year
Remote ops allocation (share of NOC + field visits)	$50-80k/year
Cash opex	~$360-425k/year

Against revenue: 32 GPUs sold as managed inference capacity at even $1.50/GPU-hour equivalent utilization-adjusted (see current GPU market rates) generates ~$420k/year at 100% notional — meaning realistic 50-70% utilization requires either premium local pricing or the bandwidth-savings side of the ledger to close the case. This is the honest state of edge economics in 2026: single-site P&Ls are tight; the returns are portfolio effects — dozens of standardized sites amortizing one NOC, one spares pool, one engineering team, plus enterprise contracts that pay for latency and sovereignty rather than raw compute. It is also why the operators scaling fastest are those attached to an existing asset base (telco sites, regional colo footprints) rather than pure greenfield builders.

Outlook

Three things to watch through 2027: whether inference demand at the edge materializes at the projected scale (the 17.5% CAGR depends on it); how far GPU-dense modules push edge power envelopes — liquid-cooled 50-100 kW racks are entering formats that were designed around 5-15 kW; and consolidation among subscale edge operators, whose site portfolios are more valuable as fleets than as companies. The market data — $16.9B now, ~$72B by 2035 — describes steady structural growth rather than a hype cycle: the physics of latency and the economics of bandwidth do not go away. Current pricing across formats is tracked in stats.