The most interesting place to put a data center, it turns out, might be the garage. That is the bet from XFRA, a distributed data center solution from SPAN.io that is turning underutilized residential power capacity into AI inference compute. It is a clever, almost mischievous, answer to the grid's power crunch, swapping the three-year wait for a new hyperscale campus for a six-month sprint to install liquid-cooled GPUs in 100 new homes.
Arch Rao, SPAN's CEO and the former head of product at Tesla Energy, calls XFRA "the next logical step in what we've been building at SPAN" [LinkedIn, 2025]. The logic is straightforward. His company's core product is a smart electrical panel that manages home energy flows. XFRA adds a high-performance edge compute node, paired with that panel and a whole-home battery, creating a mini data center that can be orchestrated by software to run AI workloads when the grid has spare capacity [Latitude Media, 2025]. For homeowners, the pitch is a free installation of the SPAN system and a discounted, bundled rate for electricity and internet, reportedly around $150 per month [The AI Consulting Network, 2026]. For AI companies desperate for compute, the promise is gigawatts of capacity, unlocked not by building new power plants, but by quietly tapping the slack in the existing grid.
The hardware wedge
XFRA's strategy relies on a hardware wedge that few pure software plays could attempt. Each residential "XFRA Node" is a self-contained unit housing direct liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs [Business Wire, 2026]. It is physically installed alongside SPAN's smart panel and battery, creating a behind-the-meter power island that can draw from the grid or the battery as needed. This setup is not an afterthought; it is the product of years of SPAN's development in home energy management. The company has raised over $285 million to date, including a $163 million Series C in January 2026, giving XFRA a deep well of parent-company capital and hardware expertise to draw from [DataCenterRichness, 2026; Tracxn & CBInsights, 2026].
The initial pilot, rolling out in 2025, aims for 1.25 megawatts of capacity across 100 newly constructed homes, representing about 1,600 GPUs [Latitude Media, 2025]. Homebuilder PulteGroup is a partner in the early testing, deploying units in select communities [The MortgagePoint, 2026]. If that seems like a modest start, the ambition scales quickly. Rao claims XFRA can deploy 100 megawatts of compute in roughly six months at a capital cost of $3 million per megawatt. He contrasts this with a traditional data center, which he says takes three to five years and costs about $15 million per megawatt [LinkedIn Arch Rao, 2026]. The unit economics, in theory, flip the script on infrastructure build-out.
Why the grid says now
The timing is not an accident. U.S. data center demand is forecast to reach 74 gigawatts by 2028, with a projected power shortfall of 49 gigawatts [XFRA.ai, 2026]. That is a physical constraint that permitting and new transmission lines cannot solve quickly. XFRA's model sidesteps the bottleneck by using capacity that already exists but is often idle, particularly in newer homes with 200-amp service that rarely runs at full load. The company is targeting inference workloads, which are generally less latency-sensitive than training and can be batched and routed across a distributed network. It is a bet that the future of AI compute is not just bigger, centralized facilities, but also smarter, dispersed ones.
The counterfactual: latency and load
For all its elegant logic, XFRA faces two substantial technical questions. The first is network latency. Can a cluster of home-based nodes reliably serve enterprise inference workloads without introducing unpredictable lag? The company's software orchestration layer will need to be exceptionally sophisticated, dynamically routing jobs based on real-time network conditions and home power availability. The second is grid coordination. While the model uses "underutilized" power, a concentrated deployment in a neighborhood could still strain local transformers if many nodes activate simultaneously during a peak compute period. SPAN's grid-edge intelligence is meant to solve this, but it remains an unproven dynamic at scale.
The commercial model also presents a classic two-sided marketplace challenge. XFRA must attract enough homeowners to achieve meaningful density, while simultaneously selling that distributed capacity to compute buyers. The reported $150 monthly discount for homeowners is an attractive carrot, but scaling requires either deep partnerships with homebuilders like PulteGroup or a costly direct-install sales motion. A commercial node for businesses is planned for early 2027, which could simplify deployment but comes with its own set of power agreements and site complexities [Latitude Media, 2025].
The next twelve months
The coming year is about proving the pilot. Success is not just about keeping the 1,600 GPUs in 100 homes running, but about demonstrating the software can reliably monetize that capacity. Key signals to watch will be the announcement of a named compute offtaker,a cloud provider or large AI lab,and data on uptime and utilization from the initial nodes. Rovin Pulikken joined SPAN as Vice President of XFRA in 2026 to lead the initiative, indicating a dedicated operational push [LinkedIn Rovin Pulikken, 2026].
If the 1.25 MW pilot works, the back-of-the-envelope math is compelling. At Rao's claimed $3 million per MW, deploying 100 MW would cost about $300 million in capex. A traditional data center for the same capacity could run to $1.5 billion. The delta, $1.2 billion, is the theoretical economic moat. The real test is whether XFRA can achieve a comparable or better cost per inference at scale, factoring in all the software, networking, and maintenance overhead of a distributed fleet. To win, it must beat not the idea of a data center, but the incumbent efficiency of a well-run, hyperscale facility operated by the likes of Equinix or Digital Realty. That is a high bar, but also the only one that matters for the climate math: the lowest carbon cost per unit of compute.
Sources
- [Business Wire, April 2026] SPAN Announces XFRA, a Distributed Data Center Solution | https://www.businesswire.com/news/home/20260414372626/en/SPAN-Announces-XFRA-a-Distributed-Data-Center-Solution-to-Close-the-Speed-to-Power-Gap-for-AI-Compute-Demand
- [DataCenterRichness, 2026] SPAN Company Profile | https://datacenterrichness.com/company/span
- [Latitude Media, 2025] Span to launch mini AI data centers for distributed at-home compute | https://www.latitudemedia.com/news/span-to-launch-mini-ai-data-centers-for-distributed-at-home-compute/
- [LinkedIn, 2025] Arch Rao LinkedIn post on XFRA | https://www.linkedin.com/posts/arch-rao_xfra-is-the-next-logical-step-in-what-weve-activity-7449590066278211584-okTb
- [LinkedIn Arch Rao, 2026] Post on deployment speed and cost | https://www.linkedin.com/in/arch-rao
- [LinkedIn Rovin Pulikken, 2026] Profile indicating VP, XFRA role | https://www.linkedin.com/in/rovinpulikken
- [The AI Consulting Network, 2026] Article on XFRA homeowner economics | https://theaiconsultingnetwork.com
- [The MortgagePoint, 2026] Article on PulteGroup deployment | https://themortgagepoint.com
- [Tracxn & CBInsights, 2026] SPAN funding data | https://tracxn.com/d/companies/span.io
- [XFRA.ai, 2026] XFRA homepage with market sizing | https://www.xfra.ai