Over the past 15 years, the number of data centers in the United States has quadrupled, with a similar surge seen globally. According to the Uptime Institute, 377 data center projects exceeding 100 megawatts were announced in just the last four years. While Big Tech races to expand compute capacity, a growing concern is emerging: the physical limitations of existing infrastructure. Can we simply upgrade old data centers to handle today’s AI workloads—or are we facing a wave of demolition and rebuilding? Experts say the answer is largely no. Retrofitting legacy facilities for modern AI is not feasible at scale. The core issue is weight. AI racks—metal enclosures housing servers packed with hundreds or even thousands of GPUs—are now far too heavy for traditional data center floors. Chris Brown, chief technical officer at Uptime Institute, explains that racks from three decades ago weighed between 400 and 600 pounds—comparable to a refrigerator or baby grand piano. Today’s standard racks weigh 1,250 to 2,500 pounds—equivalent to a grizzly bear or a compact car. AI-specific racks can exceed 5,000 pounds. This dramatic increase stems from dense packing of hardware. To maximize performance, AI racks cram in vast amounts of memory, GPUs, and cooling components. The goal is to minimize gaps between chips to avoid latency and boost training speed. But this density comes at a cost: massive weight and extreme power demands. Power consumption has surged too. While older racks used around 10 kilowatts, AI racks now consume up to 350 kilowatts—35 times more. This immense energy generates intense heat, requiring advanced cooling solutions like liquid-cooled plates filled with toxic coolant mixtures. Water alone weighs over 8 pounds per gallon, adding to the load. Cables and power delivery systems have also grown heavier. To handle the current, modern busways—copper plates that carry power—can weigh 37 pounds per linear foot. A single row of 10 to 35 racks can require dozens of these heavy components, compounding the structural burden. Most legacy data centers are built to support static loads of about 1,250 pounds per square foot, and even less for dynamic loads, such as moving a rack across the floor. The new AI racks far exceed this limit, risking floor collapse. Even if floors were reinforced, other issues remain. Racks have grown from 6 to 9 feet tall over the past 20 years, outgrowing many industrial doorways and standard freight elevators. The weight of the rack, its support frame, and the people moving it make it impossible to use existing elevators in many older buildings. As a result, companies like OpenAI, Microsoft, and others are building new data centers from the ground up. Colocation providers such as CoreWeave, Digital Realty, and Compass are following suit, constructing AI-optimized facilities to meet surging demand. Despite the AI boom, traditional data workloads remain essential. Universities, hospitals, local governments, and midsize businesses still rely on legacy data centers to store non-AI data. As McLean of Critical Facility Group notes, “All those people still need that legacy data center environment. It’s never going to go away.” In short, while the dream of repurposing old data centers is appealing, the physical reality of AI’s weight, power, and cooling demands means that for the most part, the future of AI infrastructure will be built anew—on solid ground, not on the bones of the past.