Houses Transformed into Mini Server Farms: The New Bet on AI with NVIDIA's Influence
The demand for computing related to artificial intelligence seems to have no end, and there are many ideas on how to cope with it. The latest, proposed by NVIDIA and the California startup SPAN, is the announcement of XFRA, a distributed computing platform designed to harness unused electric capacity directly in homes and small commercial buildings.
The idea behind it is to transform the so-called "grid edge", the point of junction between the distribution electrical network and end users, into a computational resource. XFRA consists of a network of locally installed nodes, referred to as XFRA units, that integrate enterprise-class GPUs, specifically the NVIDIA RTX PRO 6000 Blackwell Server Edition cooled by liquid, to manage AI workloads, with a particular focus on inference.
SPAN is a company founded in 2018 with the idea of marketing smart electrical panels designed to optimize home consumption. This technology now represents the core of the new proposal: the panels allow users to identify and free up unused electric capacity, which can be utilized to power the computing nodes. A typical installation includes a SPAN smart electrical panel, the XFRA unit, a battery storage system, and any solar panels. The node, mounted outside the home, integrates with existing infrastructures like HVAC and electrical systems.
Access is remote and allows cloud operators to utilize these distributed resources for AI services and cloud gaming. XFRA is not designed to replace centralized data centers but to complement them. The distributed model aims to alleviate certain critical issues, such as long construction times (over 10 years for new energy infrastructures), the difficulty of connecting to the electrical grid, and high energy consumption.
According to estimates, data centers in the United States consumed approximately 183 TWh in 2024, over 4% of the national total, with forecasts suggesting a possible exceedance of 9% by 2030. In this scenario, leveraging available capacity becomes a key element. SPAN claims that a network of XFRA nodes can achieve capacities comparable to small or medium data centers, with significant advantages in terms of cost and speed: up to six times faster to implement and costing about one-fifth compared to centralized infrastructure of 100 MW.
One of the central aspects of the project is the optimization for AI inference loads, which are expected to represent over half of all workloads by 2030. In this context, proximity to end users becomes crucial to reducing latency. According to NVIDIA, this approach allows meeting the power and latency requirements of modern AI models while maintaining high energy efficiency.
To accelerate adoption, SPAN has initiated collaborations with entities such as PulteGroup, one of the leading U.S. builders, with installations already underway in some new residential communities. The economic model offers holistic benefits: cloud providers have rapid access to distributed computing capacity, homeowners enjoy more predictable energy and connectivity rates, along with potential compensation, while utilities can benefit from better peak management and defer infrastructure investments.
SPAN aims to reach an overall capacity in the order of gigawatts by 2027, thanks to the highly scalable nature of the system. The stated goal is to reduce the gap between energy availability and computing demand ("speed-to-power gap"), one of the main limits to large-scale AI expansion. The real impact of this model in terms of adoption, sustainability, and acceptance by local communities remains to be seen. However, XFRA represents an attempt to rethink AI infrastructure, shifting part of the computational load from large centralized data centers to a dense distributed network.