Large
AI cluster
High-density AI cluster
Edge
AI cluster
Small
AI cluster
Power & resilience
Storage
Connectivity
Growth
Primary AI archetypes
Rack density (kW/rack)
Typical block size
Location
Cooling pattern
Real-time control and autonomous systems, computer vision inferencing
On the enterprise premise (e.g. on the factory floor)
5-30
1-8 racks per cell
Air
Single or N+1 UPS
Limited local storage, long-term retention offloaded
Basic resilience only
Scale by replicating edge cells at additional sites
N+1 UPS and distribution
High storage capacity to meet long-term retention requirements
Carrier diversity (two fibre circuits)
Scale by adding pods (2-12 racks each) with repeatable designs
Pattern recognition and anomaly detection, predictive
Enterprise-owned data centre OR colocation deployment
10-45
2-12 racks per pod
Air with hot/cold aisle
N+1 or 2N power train, dedicated UPS blocks
High storage capacity to meet long-term retention requirements
Carrier diversity (two fibre circuits) and path diversity with particular focus on availability and latency
Scale by repeating pods
(each 12-40 racks),
each with a standard design
Generative and interactive, enterprise-wide AI inference
Enterprise-owned data centre OR colocation deployment
30-80
12-40 racks per pod
Hybrid, liquid cooling for GPU racks likely
2N power train,
high capacity busway
High throughput,
low latency storage optimised for parallel I/O
Carrier diversity
(two fibre circuits) and path diversity with dynamic routing (BGP) and a focus on throughput and reliability
Scale by adding AI blocks/halls in a standard pattern
Model training,
scaled private
LLM inference
High performance colocation deployment
80-120+
Any, typically organised into blocks of 16-32 GPUs
Liquid, usually D2C or immersion
