Compare eras
Four representative facilities, drawn from the same component library. Categories overlap and are not mutually exclusive; a hyperscaler can run neocloud-style pods, and many enterprises still run pre-cloud rooms.
| Attribute | Pre-cloudThe enterprise computer room. | CloudThe warehouse-scale computer. | Hyperscaler AIA power plant's worth of accelerators. | NeocloudThe AI factory as a service. |
|---|---|---|---|---|
| Period | 1990s to mid-2000s | Mid-2000s to late 2010s | 2020s onward | Current and emerging |
| Rack density | ≈1–5 kW/rack | ≈5–15 kW/rack | ≈30–150+ kW/rack | ≈40–150+ kW/rack |
| IT load per site | ≈0.1–2 MW | ≈5–60 MW | ≈100–1000 MW | ≈20–500 MW |
| Typical PUE | ≈1.9–2.5 | ≈1.1–1.6 | ≈1.1–1.4 | ≈1.1–1.35 |
| Default cooling | Traditional air | Chilled water | Direct-to-chip liquid | Direct-to-chip liquid |
| Representative rack | Air-cooled legacy rack | Enterprise accelerator rack | Advanced liquid-cooled AI rack (NVL72-class) | Advanced liquid-cooled AI rack (NVL72-class) |
| Building | Office-adjacent room or small purpose-built facility | Standardized warehouse shell, modular pods | Multi-hall campus, dedicated substation | Modular pods, retrofit shells, rapid expansion modules |
| Rack density | ≈1–5 kW/rack | ≈5–15 kW/rack | ≈30–150+ kW/rack in AI zones | ≈40–150+ kW/rack |
| Cooling | Perimeter CRAC units, underfloor plenum | Chilled water, containment, economizers | Direct-to-chip liquid + facility water, towers or dry coolers | Direct-to-chip liquid, dry coolers, immersion in some designs |
| Power | Utility feeder, pad-mount transformer, UPS room, diesel | Dedicated substation, modular UPS, busway | 115–230 kV interconnect, MV switchgear, 2N or distributed redundancy | Leased utility capacity, on-site gas turbines or bridging power |
| Network | Copper-heavy, three-tier, leased lines to carriers | Leaf-spine Clos fabric, fiber-heavy | Scale-up interconnect + scale-out AI fabric + front-end network | High-bandwidth low-latency GPU fabric, dark fiber to carriers |
| Operations | Manual, ticket-driven, physical media | Automated, fleet-managed, software-defined | Fleet-scale automation, AI-assisted operations | Utilization-driven, multi-tenant scheduling, GPU-as-a-service |
| Networking | Three-tier core / aggregation / access, mostly copper to the server, carrier T1/T3 and early metro Ethernet | Leaf-spine Clos fabric with oversubscription tuned for mixed north-south web traffic | Non-blocking or lightly oversubscribed rail-optimized fabrics (InfiniBand or Ethernet) beside a conventional front-end network | Dedicated per-tenant fabrics or partitioned shared fabrics; dark fiber and carrier diversity |
| Software model | Dedicated physical servers per application, manual provisioning | Virtualization, then containers and orchestration; servers are cattle, not pets | Cluster schedulers, collective-communication libraries, failure-tolerant training jobs | GPU cloud control plane: reservation, scheduling, health checks, tenant isolation |
| Business model | Cost center owned by the enterprise IT department | Public cloud (rent virtual machines, storage and services) and wholesale colocation | Internal AI products plus AI cloud services sold to customers | Reserved clusters and on-demand GPU-hours sold to model builders and enterprises |
| Primary bottleneck | Floor space and cooling distribution inside an existing building | Land, power and network reach in the right metro | Grid interconnection, transformers and switchgear, and accelerator supply | GPU allocation, energized power and the speed of capital |
| Geographic logic | Next to headquarters or the business it serves | Clusters near population and carrier hubs with cheap power and tax incentives | Wherever large blocks of power can be energized soonest | Anywhere with power available quickly: retrofits, colocation, build-to-suit and remote sites |
| Environmental trade-offs | Low absolute energy, poor efficiency: much of the power went to cooling and conversion losses | Large efficiency gains (PUE fell sharply) while total consumption grew | Hundreds of MW per campus; water use, grid impacts and embodied carbon become first-order questions | Speed-to-power can mean on-site fossil generation; choices vary widely by operator |
1990s to mid-2000s
An enterprise-owned room or small building on a raised floor. Individual 1U and 2U servers, SAN arrays and tape sit beside the occasional mainframe. Perimeter CRAC units blow cold air under the floor; a pad-mount transformer, a UPS room and a diesel generator keep it running. Operations are manual and physical.
Explore this era in 3DMid-2000s to late 2010s
A standardized warehouse building filled with identical racks of commodity servers, virtualized and orchestrated by software. Hot-aisle containment, chilled water, air-side economizers and modular UPS raise efficiency. A leaf-spine network replaces the three-tier tree so any server can reach any other at predictable latency.
Explore this era in 3D2020s onward
A campus-scale build with its own high-voltage substation, multiple data halls and purpose-built AI pods. Racks run tens to well above 100 kW, so direct-to-chip liquid cooling and CDUs join chillers, towers or dry coolers. A dense optical fabric links tens of thousands of accelerators that must behave like one machine during training.
Explore this era in 3DCurrent and emerging
A GPU-centric campus built for speed: prefabricated AI pods, leased or bridged power, liquid cooling from day one, and an orchestration layer that sells reserved clusters and on-demand GPU-hours. Utilization is the business; the control plane, tenant scheduling and job queue are as central as the substation.
DefinitionA neocloud is an AI-focused cloud infrastructure provider or GPU cloud operator designed primarily around accelerated computing rather than broad, general-purpose cloud services. It often emphasizes rapid deployment, high GPU utilization, dense clusters, flexible capacity access, specialized networking, liquid cooling, and close integration with AI developers, model-training customers, and GPU vendors.
Explore this era in 3DAll values are representative ranges for the category, not audited data for any facility.