Racks of AI chips are too damn heavy (www.theverge.com)

The rapid expansion of AI data centers, driven by the increasing weight and power requirements of AI technologies, presents significant challenges for existing infrastructure. Over the past decade, the number of data centers in the U.S. has quadrupled, with new construction projects designed to accommodate the hefty racks of AI chips often reaching 5,000 pounds. Legacy data centers struggle to support these advancements due to their insufficient structural integrity and outdated designs, with most requiring complete demolition and reconstruction to meet modern demands. This trend is crucial for the AI/ML community as it highlights the urgent need for more robust data infrastructure to support the burgeoning AI sector. Advanced AI workloads can demand up to 350 kilowatts per rack, necessitating innovative cooling solutions and enhanced power delivery systems, which traditional setups cannot provide. Furthermore, while AI is driving much of the growth, non-AI workloads continue to increase, underscoring the ongoing relevance of legacy data centers. This dual demand for computational resources will shape the future of data center design, calling for a balance between new AI-focused facilities and the preservation of traditional data storage capabilities.