Racks for AI
& GPU infrastructure

AI infrastructure starts with the rack

Much is written about GPUs, network topologies and cooling concepts — and then installation fails on a question nobody asked: Can the rack actually carry it? A fully loaded GPU rack today weighs more than a tonne, draws double-digit kilowatts and needs airflow without any recirculation. ZPAS supplies the mechanical foundation: the Z-SERVER with up to 2,000 kg static load capacity (without levelling feet), reinforced 19″ levels and 86 % door perforation — individually or as a contained row.

The technical design step by step — weight, power, cooling, fire protection — is explained in the AI & GPU server rack guide.

Typical projects in AI infrastructure

Requirements for the AI rack — clearly classified

Standards for AI racks

For AI areas too, the DIN EN 50600 series of standards remains the planning framework for the data centre — availability, cooling concept and physical security are defined there, and the rack has to deliver what the concept requires. The 19-inch level follows IEC 60297 (482.6 mm standard), manufacturing is to ISO 9001/14001; ZPAS provides product-specific documentation for each order.

The practical difference from a conventional server rack comes down to three figures: load capacity (GPU racks reach 1,000–1,500 kg), airflow (86 % door perforation instead of standard perforation) and depth (1,200 mm for current GPU chassis including cable space). The AI rack guide shows in detail what planning for 30–100 kW per rack looks like.

Which rack for which task in the AI cluster?

The overview matches the series to the tasks — details on the series pages:

The racks are only the visible part: only the interplay of containment, cooling and monitoring turns a row of racks into an AI area that keeps its heat load under control for the long term. That is why ZPAS plans the row as a system — from the blanking panel range to the sensors in each zone.

Typical project: GPU cluster at a mid-sized company

Starting point: a company is setting up its own training cluster with six GPU racks — in a server room that was never planned for it. The critical questions: can the raised floor carry the point load? Where does enough cold air come from? And how do the loaded racks get into the room?

The solution: Z-SERVER 47 U placed directly on the floor without levelling feet (2,000 kg static), as a row with cold aisle containment against bypass air, with blanking panels and brush strips fitted at the factory. Redundant PDUs per rack, AKCP monitoring of temperature and power in each rack. The load distribution on the raised floor is clarified before delivery, and the racks arrive pre-assembled — the GPU systems roll straight in. After roll-out, the monitoring provides the data for fine-tuning cooling and utilisation; the next expansion stage starts as a repeat order to the same drawing.

Planning checklist: AI & GPU racks

These points should be clarified before ordering:

• Load capacity & floor: rack weight when fully loaded plus a reserve — and the point load rating of the raised floor at the installation site.
• Cooling concept: air cooling with cold aisle, rear door or in-row — the heat load per rack decides, not the average.
• Dimensions: 1,200 mm depth for GPU chassis, 800 mm width if power and data cables are to be routed separately.
• Power: redundant feed (A/B) and PDU fit-out per rack — fitting at the factory saves time during roll-out.
• Monitoring: temperature per zone, power consumption per rack, door contact — define alarm paths before commissioning.
• Expansion: plan the row and containment so that further racks can be added without conversion work.

Why ZPAS for AI projects

Short paths instead of corporate processes: at ZPAS, special requirements — reinforced levels, coolant feed-throughs, special depths — go straight into production, with lead times that fit project schedules. And the racks arrive ready to equip: PDUs fitted, cable routes prepared, cooling designed, monitoring integrated.