Case Study
Crafting a new model for urban AI inference capacity

Scroll
Client
Garbe DC
Location
Pan-European
Status
2025 - Ongoing
Capacity
6 to 12 MW
Partners
APL Data Centres, AKT II
Services
Blueprint, Vision
Overview
AI inference is creating demand for compute closer to the cities, institutions and businesses that generate it.
Conventional data centres are poorly adapted to this condition. Their familiar typology depends on large plots, perimeter security, extensive plant compounds and generous separation from neighbouring uses.
Green Cube starts from a different assumption: that urban AI infrastructure should be designed as a compact product from the outset. The result is a complete liquid-cooled AI facility contained within an 18-metre cube.
The challenge
Urban sites compress every part of the conventional data centre model.
There is less land for plant, less room for security standoff, less tolerance for servicing yards and less acceptance of industrial compounds that turn away from the street. At the same time, high-density AI infrastructure demands more from the building: heavier racks, liquid cooling, significant power, equipment replacement routes and strict security.
The task was therefore not simply to reduce the footprint of a standard data centre. It was to develop a different building logic.
Design response
The technical system is organised around a small number of repeatable relationships.
One structural bay, one rack row, one power skid and one cooling train are repeated four times. Structure, racks, distribution and replacement routes align to a common module within a steel moment-resisting frame based on a standardised grid.
The system is intended for prefabrication. Power and cooling assemblies can be manufactured, tested and coordinated away from the site before installation. The technical core remains stable while other elements can change around it. Rack mix, cooling strategy, grid interface and external treatment can respond to different customers and locations without altering the underlying product architecture.
Security is also reconsidered as part of the architecture. Rather than beginning with a fence and gatehouse, the secure perimeter moves into the envelope itself. The façade becomes windowless, anti-climb and security-rated, while doors, panels, louvres and service openings are designed according to the zones they protect.
Access is reduced to two controlled points: one for people and one for goods. This recovers the land normally occupied by perimeter standoff and makes the building materially easier to place on constrained urban plots.
The architectural expression remains deliberately restrained. A solid base and lighter acoustic crown establish the main order, while vertical fins absorb ventilation openings and rainwater routes. The underlying panel system remains constant, but the visible surface can change through metal, perforated screens, terracotta or glazed ceramic.
The product remains the same. Its appearance can belong to a particular street.
Outcome
Green Cube proposes a new model for urban inference infrastructure: compact, liquid-cooled, incrementally scalable and designed around factory assembly. Its security is substantial without becoming visually defensive. Its external character can change without redesigning the technical system behind it.
The project does not attempt to hide the data centre. It asks how a new category of AI infrastructure might become an ordinary and acceptable part of the city.