DataCentersX > Types

Data Center Types

A data center type is defined by the characteristic constraint set that drives its physical and operational design. An AI factory is a type because accelerator density and sustained training load shape everything about how it is built and run. A hyperscale facility is a type because aggregate scale economics and elastic multi-tenant workloads shape its engineering. An edge site is a type because geographic proximity to users imposes distributed deployment and ruggedized operation. The eight types below are organized around what each is optimized for, not around arbitrary size or branding differences, and each corresponds to a genuine clustering of facility engineering choices that appears in real deployments across operators.

Types and Workloads are separate pillars on DatacentersX because a single data center type typically hosts several workload families and a single workload family typically deploys across several types. An AI factory runs AI training workloads primarily but may also host some inference; a hyperscale facility hosts cloud, enterprise, analytics, AI training, and content delivery in separate zones; content delivery runs across edge, carrier, and hyperscaler origin infrastructure. Treating types and workloads as orthogonal pillars lets each be analyzed on its own terms while preserving cross-references where the fit between them is tight.

The eight types

Type	Primary Purpose	Design Characteristics	Dominant Workloads	Examples
AI Factory	Foundation model training and AI-native inference at scale	Dense accelerator fleets, direct-to-chip liquid cooling, low-latency lossless fabric, gigawatt-class power	AI Training, hyperscale AI inference	xAI Colossus, Stargate, Meta Hyperion, Tesla Cortex, Fermi Hypergrid
Hyperscaler DCs	Global-scale cloud and SaaS delivery	Standardized racks and rooms, multi-tenant elastic infrastructure, automation-first operations	Cloud and Enterprise, Content Delivery, mixed AI Training	AWS, Azure, Google Cloud, Meta, Oracle Cloud
HPC Clusters	Scientific and engineering simulation at national-lab scale	Hybrid CPU and GPU fleets, parallel file systems, batch queueing, specialized interconnect	HPC and Simulation, pharma research compute	Frontier (ORNL), Aurora (ANL), LUMI (EU), El Capitan
Enterprise DCs	Enterprise-owned and operated compute for business applications and regulated workloads	Custom-sized to organizational needs, on-premise or dedicated, integrated with enterprise IT	Enterprise Apps, regulated workloads retained on-premise	Banking data centers, hospital IT facilities, utility control centers
Colocation DCs	Multi-tenant facilities providing space, power, and cooling to third-party tenants	Carrier-neutral, cross-connect-dense, cage and suite granularity, service-level agreements	Cloud and Enterprise, Financial Services, Content Delivery, hybrid cloud	Equinix, Digital Realty, CyrusOne, QTS, Iron Mountain, CoreSite
Edge DCs	Latency-proximate compute close to users, devices, and network aggregation points	Smaller footprint per site, distributed fleet across many locations, carrier and CDN integration	Content Delivery, 5G MEC, distributed inference	AWS Wavelength, Cloudflare PoPs, carrier edge sites, EdgeConneX metro sites
Modular DCs	Prefabricated deployable data center capacity for rapid or unusual siting	ISO-container or purpose-built modules, factory-assembled, transportable, site-assembled	Edge compute, tactical deployments, disaster recovery, rapid capacity additions	Vertiv SmartMod, Schneider EcoStruxure Modular, defense mobile DCs
Orbital and Space DCs	LEO-based AI compute satellites powered by direct solar	Rad-tolerant silicon, radiative cooling, optical inter-satellite mesh, launch-survivable mechanical	AI Training and Inference at orbital scale	SpaceX Orbital Data Center constellation (AI Sat Mini, Tesla D3)

What distinguishes the types

Reading the table across its columns reveals the axes along which the types diverge. Understanding those axes makes the type taxonomy useful for analysis rather than just descriptive.

Density and thermal posture. AI Factories and HPC Clusters sit at the high-density end, with rack powers exceeding 100 kW and direct-to-chip liquid cooling as the default. Hyperscale, Colocation, and Enterprise facilities run at moderate densities (10 to 30 kW per rack) with air cooling dominant. Edge DCs accept air cooling because per-site density is low. Modular DCs cover a wide range depending on how the modules are specified. Orbital sits alone, with radiative cooling replacing fluid cooling entirely.

Scale and siting. Hyperscalers and AI Factories consolidate massive aggregate capacity at a small number of sites, with individual campuses reaching hundreds of megawatts and trending toward gigawatt scale. HPC Clusters, Enterprise, and Colocation facilities sit at moderate scale per site and are geographically tied to specific missions or regional demand. Edge DCs invert the scale relationship, distributing small capacity across thousands of sites. Modular and Orbital solve the siting problem differently, with modular going where terrestrial infrastructure has not yet reached and orbital leaving terrestrial siting entirely.

Tenancy and ownership. Enterprise DCs are single-tenant and owner-operated. Hyperscaler and AI Factory facilities are typically owned and operated by the companies whose workloads run there. Colocation is multi-tenant by definition. Edge and Modular span both models. Orbital is a new ownership model with the same operator typically handling launch, satellite bus, silicon, and workload.

Workload fit. Types and workloads are not in one-to-one correspondence, but certain fits are strong. AI Training has a strong fit with AI Factories and selected hyperscaler zones. HPC has a strong fit with HPC Clusters. Content Delivery has a strong fit with Edge and Colocation. Regulated workloads often prefer Enterprise or sovereign Hyperscaler regions. Understanding these fits is covered in the Workloads pillar, which discusses each workload's facility preferences from the workload side of the pairing.

The power scale separation

The eight types span roughly five orders of magnitude in per-site power. A single edge PoP may be 50 to 200 kW. A modular unit may be in the same range or smaller. An enterprise data center runs from single-megawatt to tens of megawatts. Colocation facilities span from tens of megawatts to several hundred. Hyperscaler campuses reach multiple gigawatts per campus when fully built out. HPC Clusters at exascale sit in the tens-to-hundreds-of-megawatts range. AI Factories are the most aggressive, with announced frontier facilities targeting 1 to 5 GW per campus and roadmaps pointing beyond. Orbital satellites currently target 100 kW per spacecraft with constellation-level aggregate capacity as the scaling unit rather than per-site capacity.

This power-scale separation is one of the reasons type boundaries matter. A 100 kW edge deployment and a 1 GW AI factory require fundamentally different grid interconnection strategies, different operational models, different security postures, and different supply chain relationships. Calling both "data centers" is correct but analytically insufficient; the type distinction is what makes the analysis actionable.

Where Types sits in the DatacentersX structure

The Types pillar answers the question "what kind of facility is this?" The Workloads pillar answers "what runs inside?" The Stack pillar answers "how is it engineered?" The Facility Operations and Compute Operations pillars answer "how is it run?" The AI Inference pillar covers inference specifically because it spans type boundaries (including on-device, which is not a facility at all). The Energy, Security, and GRC pillars cut across all types. Each type child page below can be read independently or cross-referenced against the other pillars that shape it.

Related coverage