Nvidia wants to repurpose over 130 old telco buildings as tiny data centers, report claims — company in talks with Brightspeed after pitching idea to parent company Apollo Global Management - Tom's Hardware

Nvidia's real play isn't real estate—it's collapsing the two-to-three-year data center buildout into a quarterly deployment cycle. 130 telco buildings at roughly 100-300MW aggregate can't rival a single hyperscale campus, but the distributed topology is purpose-built for inference: latency-sensitive, geographically dispersed, modest per-node draw. The intent is to reposition GPUs from training factories into urban neural endpoints. Apollo's involvement signals a capital-structure shift. Private equity loading AI compute onto existing assets means compute is transitioning from a capex line item to a securitizable asset class. If Nvidia locks into the deployment layer, it slides from fabless chipmaker toward vertically integrated platform—mirroring Apple's in-house silicon trajectory, just inverted: Apple went upstream, Nvidia is going downstream. The hard constraint is power and thermal. Telco buildings were designed for 1-5MW loads; retrofitting liquid cooling is capital-intensive, and full-load PUE could breach 2.0, destroying unit economics. Deeper risk: Nvidia controlling both chip supply and deployment channel invites vertical-integration scrutiny under FTC or EU DMA frameworks, where compliance costs could erode scale advantages. On competition, AMD will likely replicate this model with alternative real-estate partners. But the structural threat is subtler: Google TPUs and AWS Trainium carry gentler power curves, making them natively suited to building-scale deployment. Within 18 months, GPU-as-urban-infrastructure will reprice REIT valuations, and telco tower assets will command their first AI premium.