Could the next AI data center be attached to your house?

Smart-panel startup Span, based in San Francisco, has introduced XFRA, a system that transforms spare household electricity into AI computing power. Developed in partnership with Nvidia, XFRA units are designed to be installed outside homes, drawing power from the existing electrical supply to run AI tasks. The company aims to address the significant electricity constraints currently facing the AI industry, where utilities struggle to connect power-hungry data centers to the grid quickly.

Span's approach is to distribute computing across thousands of homes, bypassing the need for large-scale substation upgrades typically required for 100-megawatt data centers, which can take four to seven years. Each XFRA node contains 16 Nvidia GPUs, four CPUs, and three terabytes of RAM, consuming about 12.5 kilowatts at full power. While experts like Jonathan Koomey question if the benefits outweigh the economies of scale of traditional data centers, Span argues that most newer homes have significant unused electrical headroom. Utilities, however, rely on load diversity, and filling these valleys could create new peaks, according to Rich Brown, a former Berkeley Lab researcher.

For its initial commercial rollout, Span is collaborating with PulteGroup, a major U.S. home builder, to install XFRA units in new communities. A pilot program deploying units in 100 homes with 1.2 megawatts of compute capacity is planned for the southwestern U.S. Homeowners receive discounted electricity and internet in exchange for hosting the units, which are liquid-cooled and quieter than standard HVAC. The system also includes a backup battery and mechanisms to throttle non-urgent workloads to ensure the home's power experience is unaffected. The feasibility of scaling, business-side costs, and implications for network communication between distributed chips remain key unknowns.