Server racks for
AI & GPU data centres

AI training and high-performance computing have fundamentally changed the requirements for server racks. While a conventional server rack dissipates 5–10 kW, a fully populated GPU rack produces 30–100 kW — five to ten times as much. Ordinary racks and room cooling reach their limits here. Anyone planning AI infrastructure needs racks designed for extreme power density.

Why AI/GPU differs from conventional IT

GPU servers (e.g. with NVIDIA H100/H200) draw many times the power of conventional servers per rack unit. This affects the entire rack:

• More weight: dense GPU servers are heavy; a fully populated rack quickly exceeds one tonne.
• More heat in a confined space: 30–100 kW must be removed reliably, otherwise the GPUs throttle (thermal throttling) or shut down.
• More power: high power density requires powerful, redundant PDUs and well-planned supply cabling.
• Deeper designs: long GPU servers need deep racks (1,000–1,200 mm) with good cable management.

Requirement 1: load capacity

The stated load capacity must refer to the base frame, not just the rails. The ZPAS Z-SERVER carries a static load of 1,600 kg on levelling feet and 2,000 kg without levelling feet — designed for heavy GPU servers and battery systems. Also check the floor load capacity of the server room. The industry overview with projects and series is available under racks for AI & GPU infrastructure. Basics on rack units and load capacity: → What is a server rack?

Requirement 2: airflow and cooling

At high power density, airflow determines stability. Key points:

• High door perforation: Z-SERVER doors achieve 86 % perforation with an 80 % open cross-section — one of the highest values on the market. The more air flows through, the more efficient the cooling.
• Cold aisle/hot aisle containment: separates cold supply air from warm exhaust air, prevents mixing and reduces cooling energy demand by up to 40 %.
• High-performance cooling close to the rack: from around 15–20 kW per rack, room cooling is no longer sufficient. This is where rear-door heat exchangers (cooling directly at the rear door) or in-row coolers between the racks come in.
• Liquid cooling: at the highest densities (well above 50 kW), direct-to-chip or immersion cooling becomes relevant — the rack must be prepared for this.

Complete concepts with containment and precision cooling: → ZPAS data centre & server room solutions.

Requirement 3: power supply

• Redundant PDUs (A/B feed) with sufficient current-carrying capacity per rack
• Clean supply cabling and separation of power and data paths
• Monitoring of power, temperature and access for safe operation

Requirement 4: depth, cable management and security

Long GPU servers require rack depths of 1,000–1,200 mm. Structured cable management keeps the airflow paths clear — at high density a real performance factor. In colocation and multi-tenant environments, lockable cabinets or compartmentalised racks (individual lock per section) make sense.

The right ZPAS products

The ZPAS Z-SERVER is designed for high density: 42/45/47 U, 600/800 mm wide, depths up to 1,200 mm, up to 2,000 kg load capacity (without levelling feet) and 86 % door perforation. Combined with the Data Box cold aisle containment system and cooling close to the rack, this creates an infrastructure for AI and GPU loads. You can put together your rack quickly in the server rack configurator.