Tiny AI boxes in yards aim to bypass power bottlenecks

XFRA distributed – A new system called XFRA proposes turning unused capacity in typical homes into distributed AI compute. Backed by Span and Nvidia, each compact liquid-cooled unit would draw power from the house, run dozens of GPUs, and compensate homeowners—offering a potenti

For now, the idea sounds almost too domestic to be serious: a unit about the size of an air conditioner, mounted in a side yard, quietly humming through artificial intelligence tasks.

In Span’s pitch. it would draw electricity from the home it sits beside—not from a new substation or a distant facility—while homeowners get discounted electricity and Internet in exchange. The system is called XFRA. a distributed network of miniature AI computing units that was recently unveiled by Span in partnership with Nvidia.

Span, which started in San Francisco in 2018, already sells hardware designed to help homes manage electrical loads. XFRA applies the same basic control logic to powering AI compute. The timing is sharp. Access to electricity has become one of the AI industry’s biggest constraints. with utilities struggling to connect power-hungry data centers to the grid fast enough.

In many parts of the U.S., substation upgrades to support a 100-megawatt data center now take four to seven years. At the end of 2025, more than 2,060 gigawatts of generation and storage capacity were sitting in interconnection queues, according to Lawrence Berkeley National Laboratory.

Span’s answer is to route around that delay. Instead of building one large data center that requires its own grid work—or, in some cases, on-site gas turbines—it spreads compute across thousands of homes that are already connected.

Each XFRA node is packed with hardware: 16 Nvidia graphics processing units (GPUs). four central processing units (CPUs). and three terabytes of RAM. “That’s pretty beefy. ” said Mahadev Satyanarayanan. a computer scientist at Carnegie Mellon University. known for his work on distributed and edge computing. “Even a modest-size large language model could run on a 16-GPU cluster.”.

At full power, an XFRA node draws about 12.5 kilowatts. Chris Lander, XFRA’s vice president, said that roughly 8,000 nodes would match the power demand of a medium-sized 100-megawatt data center.

Lander offered a way to visualize the scale: an XFRA node running at full power would consume as much energy in three days as the average U.S. household uses in a month.

He argues the capacity already exists in many neighborhoods. “Most newer single-family homes are wired for 200 amps of service but typically use closer to 80 at peak,” Lander said. Even setting aside a 40-amp buffer. he added. “that leaves roughly 80 amps of headroom” that he says is “just on the table. and it’s never used.”.

Utilities, in other words, size infrastructure for peak demand—and much of that capacity sits idle through much of the year. XFRA’s logic is to put that idle space to work, turning it into distributed computing power for AI cloud providers.

But turning peaks into compute isn’t purely a technical move. Rich Brown. who spent more than 30 years researching energy technologies at Berkeley Lab. cautioned that utilities and planners rely on diversity of loads to average out peaks and valleys. Distributed data centers “would eliminate some of the benefits of that diversity by filling in all the valleys and perhaps creating new peaks.”.

Jonathan Koomey. a longtime data center energy researcher who was formerly at Berkeley Lab. sounded a second kind of warning: even if headroom exists today. planners may lose it later. “Planning for these installations would need to account for growth in behind-the-meter solar. as well as electrification of heat. water heating and vehicles. ” he said.

Even if the grid can handle the idea, XFRA still has to match the way AI work actually runs. Many AI workloads depend on fast communication between chips connected by high-bandwidth networking. Spreading nodes across homes won’t fit every use case.

Satyanarayanan put it bluntly: “If you try to blindly take what a data center does and use a collection of XFRA nodes for it, it will not work very well.” The question becomes which workloads can be broken up, routed, and executed independently.

The dividing line, in Span’s framing, is training versus inference. Training frontier AI models requires thousands of chips to exchange huge amounts of data in near real time. and it still demands centralized. high-speed infrastructure. Inference—when trained models answer queries or generate content—requires far less coordination.

In Lander’s view, the fit is in inference tasks that can be dispatched across nearby nodes. “We know we can support the vast majority of inference compute for chat, for enterprise, for coding, for agentic AI,” he said.

That proximity is also the selling point for performance. For tasks that depend on fast back-and-forth—voice assistant functions. live translation. and augmented reality—placing compute closer to the user could ease congestion on long-distance networks and reduce response times. “The proximity of the node matters a lot,” Satyanarayanan said. “What the user sees are the benefits in performance.”.

For XFRA’s first commercial rollout, Span is working with PulteGroup, one of the largest U.S. home builders, to install XFRA units in newly built communities. Span said it has already tested prototype nodes with paying customers.

This fall, the company plans to deploy units in 100 homes totaling about 1.2 megawatts of compute capacity in the southwestern U.S., a region where thermal management will be stress-tested quickly.

Homeowners, under Span’s plan, would pay nothing for the hardware. They would pay a flat fee for power and Wi-Fi, and they would earn compensation based on how much compute and energy the network uses.

The cooling system is designed around the reality of being installed in residential spaces. Unlike the fan-driven air cooling typical of hyperscale data centers. XFRA units are liquid-cooled. with a heat pump pulling heat from a closed loop; no water is used. Lander said, “We expect them to be quieter than your standard HVAC [heating, ventilation and air-conditioning].”.

Span also built in safeguards. Each unit includes a backup battery in case a power outage occurs or a home’s power demand surges at the same time its XFRA node is running at full power. The company says it can throttle nonurgent workloads or transfer them to other nodes in the fleet. “We have a number of dials we can turn to make sure the customer experience for the home is untouched. ” Lander said.

Still, the business math may be the toughest part. Satyanarayanan said the cost of moving workloads around, along with expenses such as repairs, may be higher than Span is anticipating—factors that could determine whether XFRA scales or remains a clever concept.

“There are a lot of unknowns on the business side,” he said. On the technical side, he is more certain. “There are a lot of unknowns on the business side,” he said, but “on the technical side, though, he is completely convinced of the feasibility and the value.”

If XFRA succeeds. it would mark a striking shift in how AI infrastructure is imagined—less a remote. fenced-off power hog and more a distributed network bolted into everyday life. If it stumbles. the reasons may be just as familiar: the grid. the workload fit. and the question of whether thousands of homeowners can be turned into reliable compute partners without turning neighborhoods into new peak-demand battlegrounds.

AI data centers distributed computing Span Nvidia XFRA residential power grid constraints inference GPUs liquid cooling smart electrical panel PulteGroup interconnection queues