AI Compute Era: Data Center Power, Cooling, and Space Requirements

With the rise of AI, big data, and cloud computing, the demand for compute power has exploded, especially after the release of GPT‑3.5 and the subsequent AI‑generated content boom, prompting data centers to confront both opportunities and significant challenges.

Cabinet Power Requirements: Historically, most internet and cloud data centers operated cabinets at 4‑8 kW, but modern AI servers equipped with GPUs like the NVIDIA A800 (400 W per GPU) or H800 (700 W per GPU) can consume 6‑10 kW per cabinet, necessitating power upgrades from 6 kW to 12 kW or higher and often requiring a reduction in cabinet count or a complete redesign of power distribution.

Cooling Requirements: Traditional cooling designs for 4‑8 kW cabinets use air‑cooled units, but high‑density AI cabinets (10‑20 kW) demand more aggressive solutions. Options include enhanced air‑side cooling with row‑level units or adopting liquid‑cooling technologies such as cold‑plate or immersion cooling, each with its own integration challenges and economic considerations.

Space and Density Requirements: AI workloads drive the need for larger, higher‑density clusters; a single GPT‑4‑scale deployment may require 20‑30 MW of power and thousands of GPUs, implying data‑center floors of 40 MW or more. Continuous high‑density cabinets (12 kW+ per unit) are essential to reduce networking costs, especially when using expensive InfiniBand interconnects.

Construction Timelines: The fast‑moving AI market forces data‑center projects to be delivered within two months, creating tension between the need for rapid power and cooling upgrades and the extensive redesign work required for existing facilities.

Conclusion: AI compute will increasingly dictate data‑center design, with single‑cabinet power exceeding 30 kW and cooling thresholds around 50 kW prompting broader adoption of liquid‑cooling solutions. Embracing modular, high‑density designs and flexible power‑cooling architectures will be critical for meeting future AI workloads.