A box the size of an air conditioning unit. Sitting in your side yard. Humming with AI tasks while you sleep.
It draws power from your home. It pays you in discounted electricity and internet. Or so the pitch goes.
This is XFRA. A distributed network of miniature computing nodes. Unveiled by Span, the smart-panel startup founded in San Francisco in 2008. They partnered with Nvidia. The idea? Stop building massive, grid-crushing data centers. Start using the grid you already have.
“Access to electricity has become one of the biggest constraints in the AI industry.”
Utilities are choked. Grids are full.
Want to plug in a 100-megahawatt facility? Wait four years. Seven, in some parts of the country. As of late 2025? More than 2,064 gigawatts of capacity are just… sitting there. Waiting. According to Lawrence Berkeley National Lab, the queues are absurdly long.
Span thinks it knows a workaround.
Don’t build one giant monster. Spread the bite. Across thousands of homes. Homes that are already connected. Homes that aren’t using all the juice they’re zoned for.
The Math Is Bothering Experts
It’s clever. It might work.
Jonathan Koomey, who studies data center energy, is skeptical. Not because the tech won’t fit. But because the economics are fuzzy. Big, purpose-built centers have scale. Massive scale. Can a backyard unit compete?
“We say it’s about speed,” Koomey notes. “But the benefits need to outweigh the scale.”
The hardware is serious stuff, though.
Each XFRA node packs 16 Nvidia GPUs, four CPUs, and three terabytes of memory. Mahadev Satyanarayanan from Carnegie Mellon calls it “pretty beefy.” A modest large language model runs right on there. No problem.
The energy hit is real.
One node draws 12.5 kilowats at full power.
Do the math.
Roughly 8,000 nodes equal the power hunger of a medium 100-MW data center.
Here’s the kicker: an XFRA unit running full tilt for three days eats as much energy as your average US household burns in a whole month.
Span argues that’s fine. Why?
Most modern houses are wired for 20 amps. They rarely use 80. Even if you set aside a safety buffer, there’s a chunk of capacity… just sitting there. Unused. Chris Lander, VFRA’s VP, sees that idle capacity as money left on the table.
But The Grid Hates This Idea
The extra load isn’t invisible to the system.
Rich Brown, another Berkeley Lab vet, worries. Grids rely on diversity. People don’t all turn on their AC at 5 pm. That creates a nice curve of peaks and valleys. Distributed data centers? They fill the valleys. They flatten the curve. Or worse, create new, sharp spikes.
Then there’s tomorrow’s load.
Solar panels. Heat pumps. Electric cars. All that stuff eats capacity. The “headroom” Span is targeting might not exist in five years. Koomey warns against ignoring that trajectory.
And let’s talk about speed. Or lack thereof.
AI needs chips to talk. Fast. Training a frontier model? Requires thousands of chips screaming data to each other in near real-time. You can’t scatter those tasks across suburban backyards. Latency will kill it.
Inference? Different story.
Inference is the Q&A part. Chat. Coding. Agentic tasks. Those don’t need a hive mind. They can happen independently.
“Proximity of the node matters a lot. The user sees the benefit.”
For voice assistants. Live translation. Augmented reality. Being close helps. Less travel time for the signal. Snappier responses. Satyanarayanan admits that performance win is real.
Testing The Waters
So who gets these boxes first?
New home builds. PulteGroup, one of America’s biggest builders, is rolling out XFRA units in fresh communities. They’re testing prototypes now. With paying customers.
Fall brings the next step: 100 units. Swelling the network to about 1.2 mega watts of compute. In the southwest. Hot, dry southwest. Where cooling becomes an immediate problem.
The units are liquid-cooled. Heat pumps pull heat out. No water. Quiet, Span claims. Quieter than your AC, anyway.
Homeowners pay nothing for hardware. Flat fee for power/WiFi. They earn credits based on usage. Span aims for 1 gigawatt of capacity eventually. That’s big. That’s enormous.
But is it practical?
A backup battery handles surges or outages. Span can throttle jobs or bounce them to another node if things get too tight. They promise the homeowner’s life won’t change. Not a flicker.
Satyanarayanan sees the catch.
Moving workloads around costs money. So do repairs. Span’s financial model assumes a smooth ride. Reality is bumpier.
“There are a lot of unknows on the business side.”
He’s sold on the tech. Completely convinced. The physics work. The engineering holds up.
The money part?
That remains an open question. We’ll see which side wins. The efficiency of distribution, or the brutal economies of scale.
Nobody really knows yet.
