Your house as a data centre: what Span's XFRA node actually means

Span, Nvidia, and homebuilder PulteGroup are running a 100-home pilot that bolts 16-GPU compute enclosures onto newly built suburban houses in the US southwest. The pitch is simple: homes waste electricity, AI companies need compute, let's connect those two facts. What makes this technically interesting — and technically unresolved — is everything the press coverage doesn't explain.

Why this exists at all

The data-centre queue problem is real. Getting a large AI data centre connected to the US grid currently takes four to seven years in many markets. Utilities operate interconnection queues — formal waiting lists for new industrial-scale power connections — and those queues are backed up badly. Community opposition to hyperscale facilities adds more delay. Neither of those problems is going away quickly.

Span's answer is to sidestep the queue entirely by using capacity that is already connected: your home's electrical panel.

A standard US residential panel carries a 200-amp service rating. At 240 volts, that is roughly 48 kilowatts of potential draw. The average American home actually draws about 1.2 to 1.3 kilowatts continuously, spiking to 3 to 5 kilowatts during peak use. The gap between what the panel is rated for and what the household actually uses sits idle most of the time.

Span's existing product, the Span Panel, is a connected replacement for a conventional electrical panel. It monitors load per circuit in real time and can shift or shed load automatically. XFRA extends that idea: the panel that already knows your electrical budget now hosts a compute node within that budget.

What an XFRA node actually is

Each XFRA enclosure contains 16 Nvidia GPUs (RTX PRO 6000-class), 4 AMD EPYC CPUs, and 3 terabytes of RAM. This is not a consumer gaming rig. This is a dense inference server, the kind of hardware you would normally find in a rack inside a temperature-controlled colocation facility.

It mounts on a newly built PulteGroup home. Homeowners pay $150 per month covering electricity draw and wi-fi uplink, and receive a revenue share on GPU utilisation. They do not buy the hardware.

The 100-home pilot is planned for Q3 2026 in Nevada or Arizona. Span's stated ambition is 8,000 units.

The mechanism the coverage skips: distributed inference

Here is the part worth pausing on. A single 16-GPU node on one house is not, by itself, interesting. What Span is proposing is that 8,000 of these nodes, spread across suburban developments, collectively act as a data centre.

That requires solving a hard coordination problem.

Inference — running a trained model to produce output — is not infinitely divisible. Some AI workloads parallelise cleanly across many machines. Others have tight dependencies: step two cannot start until step one completes, and the result of step one needs to travel over a network before step two can begin. Every millisecond of network transit adds latency. In a centralised data centre, the GPUs are connected by high-bandwidth, low-latency interconnects (NVLink, InfiniBand) that move data at hundreds of gigabytes per second. Across 8,000 homes connected by wi-fi, you get something considerably slower.

The honest answer is that not all workloads do. Batch inference — processing large volumes of requests where individual latency doesn't matter — is a plausible fit. Real-time inference for a chatbot or a voice assistant is harder. Stateful inference, where a model needs to maintain context across many steps, is harder still.

Projects like Gensyn and io.net have explored distributed GPU networks before, but with single consumer GPUs, not 16-GPU enclosures. The scheduling, fault-tolerance, and networking stack required for XFRA's configuration has not been publicly described as of the pilot announcement.

The electrical headroom problem, precisely

Span's pitch relies on a statistical reality: most homes draw far less than their panel's rated capacity, most of the time. That is true. But "most of the time" hides a load-bearing caveat.

An XFRA node with 16 GPUs generates substantial heat. A single Nvidia H100 — a comparable high-end GPU — has a thermal design power of around 700 watts. Sixteen units in one enclosure produce roughly 11 kilowatts of heat that has to go somewhere. That number has not been publicly confirmed for the RTX PRO 6000 variant, and Span has not released a thermal management specification for XFRA.

Here is why that matters in Nevada and Arizona specifically. Southwest US summers produce peak HVAC load on the same afternoons you most need AI compute capacity. A home running air conditioning at full tilt is already drawing 3 to 5 kilowatts for cooling. Add 11-plus kilowatts for a GPU node, and you are well into the range where the statistical headroom in the panel rating gets used up by the home's own needs.

The Span Panel's per-circuit monitoring is designed to manage exactly this kind of conflict. The system can prioritise loads and throttle others. But the resolution — what happens to an in-progress inference job when the panel decides the air conditioner takes priority — is not addressed in any public material.

The grid headroom that makes XFRA possible in February is not the same headroom available in August, at 4pm, in Phoenix.

The reliability gap

Data centres are engineered for high availability. A well-run facility targets 99.9% uptime or better — meaning less than nine hours of downtime per year. Achieving that requires redundant power feeds, generator backup, uninterruptible power supplies, and professional operations staff.

A residential node has a homeowner. The homeowner might trip a breaker. The wi-fi router might reboot during a firmware update. The power might flicker during a summer storm. None of these events are catastrophic for the homeowner, but they are meaningful interruptions for an inference workload that was mid-run.

The design response to this is fault tolerance in the scheduling layer: the system that assigns work to nodes needs to treat each node as potentially unreliable, checkpoint work frequently, and reroute in-flight jobs when a node drops. That is solvable engineering. It is also non-trivial engineering, and the public description of XFRA does not include it.

What Span gets right

The deployment speed claim is the most interesting part of the pitch. Not the technology — the logistics. Getting a 16-GPU server installed in a newly built house during construction is, in principle, much simpler than getting a new grid interconnection approved for a 100 MW facility. PulteGroup builds tens of thousands of homes per year. If XFRA becomes a standard option during construction, the rollout path looks genuinely different from anything a hyperscaler can execute.

The $150 flat fee also makes the homeowner economics legible without requiring the homeowner to understand GPU utilisation rates. The revenue share upside is real but speculative; the flat fee is the thing that makes the pitch concrete.

What to watch

The 100-home pilot will tell us things the announcement cannot. Specifically: what thermal management Span actually deploys (active liquid cooling at this density would be significant infrastructure, not a quiet backyard appliance), what the noise envelope looks like in practice, and whether the wi-fi uplink is genuinely sufficient for the workloads being run or whether fibre backhaul becomes necessary.

The other thing worth watching is liability architecture. Hosting a persistent, network-connected compute node introduces data-residency questions, power contract terms, and potential obligations that the $150/month framing does not address publicly. Those questions will surface as the pilot scales.

I am not sure XFRA works at the reliability and performance levels AI inference companies require. I am fairly sure the deployment-speed-and-cost argument is real. Whether those two things resolve in Span's favour is what the pilot is actually testing.

Footnotes and links

Further reading

• [Span Panel product overview]: https://www.span.io/panel — the existing product XFRA extends
• [Gensyn distributed training network]: https://www.gensyn.ai — prior art on distributed GPU coordination
• [io.net distributed GPU marketplace]: https://io.net — consumer-GPU distributed inference, a less dense predecessor to XFRA's approach
• [US grid interconnection queue data]: https://www.energy.gov/sites/default/files/2024-12/egrid2023_technical_guide.pdf — Lawrence Berkeley Lab annual report on interconnection queue lengths

¶