
A Verge report highlights an unexpectedly low‑tech constraint on the AI boom: many older data centers quite literally can’t handle the weight of dense GPU racks. As hyperscalers and AI companies cram more accelerators into each rack, floor loading, cooling and power distribution limits are forcing a wave of new construction rather than simple retrofits. Operators are having to design facilities with reinforced floors, specialized cooling and radically different power topologies to support modern AI clusters. The piece underlines that physical infrastructure is becoming just as much of a bottleneck as chips themselves in the race to scale models and agents. For investors, it’s a reminder that the AI capex cycle extends well beyond GPUs into construction, power and thermal engineering.