GPU infrastructure starts with the server configuration
AI infrastructure concentrates demand in ways that an ordinary rack count does not describe. The number and type of GPU servers, their cooling interfaces and the cluster interconnect determine the physical infrastructure. MMC UAE scopes AI data center work from the server schedule outward: rack fit, electrical demand, heat removal, network routes and service access. A design should distinguish the initial cluster from reserved future capacity.
Capture the exact server models, installed options, dimensions, weight and manufacturer operating requirements. Record both expected use and the maximum design load with the application team. A training cluster, inference deployment and mixed enterprise room can have different utilisation patterns and availability priorities. Avoid choosing a cooling method from a generic watts-per-server threshold.
Evaluate the complete cooling chain
High density does not automatically mean one liquid cooling technology. The server manufacturer must support the selected arrangement. Air cooling, rear-door heat exchange, direct-to-chip cooling and immersion each place different demands on the rack and facility. A liquid-cooled server may still reject some heat into the room, so residual air cooling belongs in the design.
For a water-connected system, define the facility loop, coolant distribution, operating temperatures, water quality, pressure limits and heat rejection. Establish ownership at every connection between the building, CDU, rack and server. Leak detection, isolation and maintenance access need to be specified alongside capacity. Immersion introduces separate questions about fluid compatibility, equipment preparation, lifting and component maintenance.
Power and physical design for dense racks
Electrical planning should cover rack feed ratings, UPS configuration, downstream distribution and the load pattern the installation must tolerate. Confirm connector types and usable capacity in the intended redundancy arrangement. Two nominal feeds do not justify allocating their combined full rating to a load that must survive loss of one. Coordinate these decisions with the electrical engineer and the actual server power supplies.
Heavy racks also affect the slab, transport route and service procedure. Review rolling loads, lift access and the space needed to remove a server or cooling component. High-bandwidth interconnect cabling may introduce bend-radius and reach constraints. Agree switch placement, overhead pathways and labelling so the cluster layout is practical to operate as well as electrically supportable.
ATTOM supply, installation and staged commissioning
AgileCore modular infrastructure, direct-to-chip cooling and rear-door heat exchange are ATTOM options to assess against the approved server requirement. MMC can scope equipment selection, a coordinated bill of materials, delivery interfaces, infrastructure installation and monitoring configuration. Product compatibility, selected model and any manufacturer engineering review should be confirmed before procurement.
Commissioning should follow a staged load plan. Begin with utility and cooling readiness, then validate the rack interfaces before compute is introduced. Agree the permitted test loads, sensor checks, alarm limits and shutdown coordination with the server team. Record coolant conditions, electrical measurements and the operating envelope at handover. An AMC can cover infrastructure inspection and servicing; GPU software, workload optimisation and server warranty work must be separately identified if required.
Plan for UAE retrofit constraints and future growth
In a Dubai or Abu Dhabi retrofit, the limiting factor may be heat rejection or available electrical capacity rather than floor area. Review the approved summer design conditions, plant location, pipe routes and ability to maintain a liquid loop in an occupied building. Identify the space and utility capacity needed for the next deployment batch before the initial cluster fills the room.
ATTOM’s PoliCloud manufacturer story provides an example of modular infrastructure associated with distributed AI computing. It is a reference for discussing architecture, not a performance promise or MMC delivery case. Bring your own GPU schedule, rack loads, cooling requirements, network topology and expansion dates to the enquiry. The resulting service scope should explain which infrastructure is being supplied, how it connects to the servers, what will be tested and who operates each part after handover.