Nvidia’s Workaround For Data Center Backlash: Retrofitting Old Telephone Buildings - The Information

Nvidia's pivot to retrofitting legacy telecom buildings is not a siting compromise—it is a confession that the AI compute race has hit a physical-infrastructure wall no chip roadmap can solve. The technical consequences are severe. Those structures were engineered for 2-5 kW per rack of switching gear; GB200 clusters demand 40 kW+ per cabinet. Retrofitting means rebuilding liquid-cooling manifolds, busbar distribution, and floor-load reinforcement from scratch. More critically, spatial constraints compress the physical radius of NVLink domains, potentially forcing trade-offs in scale-up interconnect design that erode Nvidia's cluster-efficiency edge over AMD's MI300X. Regulatorily, this is textbook arbitrage. New-build data centers in the US face 18-to-36-month approval cycles—environmental reviews, zoning fights, NIMBY litigation. Converting existing structures sidesteps most new-construction permits, but at the cost of fragmented, capacity-capped sites with stricter structural-safety scrutiny. Strategically, hyperscalers watching delivery timelines slip will accelerate in-house silicon (Maia, MTIA) to decouple from a turnkey-GPU-cluster model. AMD will weaponize "fits your existing facility" as a MI300X differentiator. Within 18 months, the binding constraint shifts from "can we make the chip" to "can we find a building to put it in." Containerized DCs, subterranean facilities, and offshore platforms enter the mainstream. This is not a workaround. It is the early signal of a paradigm shift in how compute gets deployed.