SPAN Plans to Install Mini AI Data Centers in 80,000 US Homes by 2027
May 13, 2026
SPAN Plans to Install Mini AI Data Centers in 80,000 US Homes by 2027
A San Francisco startup is betting that the future of AI computing does not lie in sprawling warehouse-sized facilities, but rather in the spare electrical capacity of American homes. SPAN has begun pilot testing a plan to install compact, liquid-cooled GPU nodes on the exterior of newly built houses, offering homeowners subsidized electricity, free internet, and a backup battery in return. The company aims to scale from a 100-home pilot in 2026 to 80,000 nodes nationwide by 2027, delivering over one gigawatt of distributed compute capacity.
Each XFRA node, as the company calls the system, sits beside a house and operates almost silently thanks to liquid cooling. The unit houses 16 Nvidia RTX Pro 6000 Blackwell Server Edition GPUs, 4 AMD EPYC Server CPUs, and 3 terabytes of memory. Alongside the node, each participating home receives a wall-mounted SPAN smart panel and a 16 kilowatt-hour battery. SPAN’s proprietary PowerUp software manages energy across the setup, and rooftop solar may be added where feasible. The company claims that 8,000 XFRA units can be installed for one-fifth the cost of building a typical 100-megawatt data center with the same compute output.
The financial arrangement for homeowners is designed to be attractive. SPAN takes over both electricity and internet bills, charging either a flat fee of $150 per month or nothing at all, depending on the agreement. The company notes that daily life inside the house would feel normal because the node uses only the slack in the home’s electrical service. Most US homes built in the last three decades have 200-amp utility service, and SPAN’s analysis shows about 80 amps are typically idle. “Virtually all homes with 200-amp utility services have 80 amps available at all times, so we set that as the maximum power consumption for a single XFRA node,” said Chris Lander, vice president of XFRA at SPAN. If household consumption spikes, the home battery kicks in first to keep the GPUs fed, and if that is not enough, the system temporarily throttles non-critical flexible loads like EV charging. Homeowners can use the PowerUp app to set their own priorities for which loads get curtailed. Full node shutdowns only happen during power outages, utility demand response events, or safety triggers, with affected workloads shifting elsewhere on SPAN’s network.
The broader industry context underscores the significance of this approach. Power companies across the US are grappling with AI-driven demand growth that often requires expensive grid upgrades, with costs eventually passed to customers. SPAN argues that distributed nodes use grid capacity that already exists. “Networks of XFRA nodes make electricity more affordable for the entire community because they increase sales over grid infrastructure that already exists, saving utilities from costly upgrades to support big data centers,” Lander said. However, experts have raised concerns. Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School, called the homeowner subsidy “fascinating” but questioned clustering effects. “If there’s a block that has several homes with these devices, maxing out compute and energy would force a lot of power to that local area,” he said, noting that local distribution networks were not designed for sustained heavy industrial loads in residential streets.
Benjamin Lee, a computer architect and engineer at the University of Pennsylvania, sees technical logic in the distributed model for certain workloads. “Computation for AI inference can and should be distributed at the ‘edge,’ deployed on smaller platforms closer to population centers and users,” he said. “The strategy could impose much smaller impacts on the grid because inference requires a few GPUs, unlike training which requires thousands of them working in concert.” SPAN’s nodes would handle downstream tasks such as cloud gaming, video streaming, and applying already-trained models to user queries, while training the largest AI models would remain at centralized hyperscale facilities. Lee also questioned whether shrinking compute to “the granularity of a few GPUs” is necessary, suggesting that building 20-megawatt data centers might solve similar problems with fewer logistical headaches.
Security and theft remain open questions. Distributing expensive hardware into thousands of suburban backyards creates risks that walled, guarded facilities do not face. Side-channel attacks, which extract data by measuring power draw or electromagnetic emissions, generally require close physical access. “Many side-channel attacks require physical proximity to the machine, which data centers can guard against,” Lee said. “Distributed GPUs in individual homes are much more difficult to protect.” Each Nvidia RTX Pro 6000 Blackwell sells for roughly $10,000, and a node contains sixteen of them, raising concerns about physical theft. “Of course, there is the risk of losing the actual hardware itself to theft,” Lee added.
The pilot, set for 2026, will produce the data needed to evaluate whether the economics, reliability, and security claims hold up in practice. While other parts of the AI industry chase ideas like orbital data centers and floating offshore facilities, SPAN’s suburban version may be the most grounded of the bunch—assuming homeowners, neighbors, and homeowners associations do not push back once the first nodes start humming next to someone’s fence.
Source: technology
