Span says it can install 8,000 exterior-mounted mini data centers six times faster and at one-fifth the cost of a comparable 100-megawatt facility. Nvidia and homebuilder PulteGroup are partnering with the startup to place compact, liquid-cooled server boxes, called XFRA units, on homes and small businesses. The boxes pair with Span's smart electrical panels to tap unused household grid capacity and link into a nationwide compute network for AI customers, potentially cutting demand for some large centralized data centers.
Span, a California startup known for smart electrical panels, unveiled a plan to install compact data-center modules on homes and small commercial buildings. The company calls the devices XFRA units. Span says the units mount on exterior walls like HVAC gear and tie into its home electrical system.
The system pairs three parts. Span's smart panel identifies unused local electrical capacity. An XFRA box houses compute hardware. A home backup battery helps manage power. Together the pieces let operators switch on compute when spare capacity is available.
How the boxes work
Span positions the network as a distributed alternative to building big, centralized data centers. The company says many local grids have unused capacity at certain times. Span's panel spots that slack. The XFRA boxes plug into that available power and link to other units across neighborhoods and regions.
Span has worked with Nvidia on the hardware and system design. The company said the units use liquid-cooled Nvidia RTX PRO 6000 Blackwell Server Edition GPUs. Those GPUs run without fans, so Span says the boxes produce no fan noise.
Nvidia framed the effort as a way to bring more AI compute online by using existing power access in many locations.
Arch Rao, founder and CEO of Span, described the idea as "an infrastructure play." He said the model aims to meet rising demand for compute more cheaply while also offering benefits to individual homeowners. Marc Spieler, senior managing director of global energy industry at Nvidia, said leveraging existing sites that already have power "makes a lot of sense."
Scale, cost and the data-center tradeoffs
Span provided a comparison for scale. The company said it can install 8,000 XFRA units about six times faster and at five times lower cost than constructing a centralized 100-megawatt data center of comparable compute capacity. Span also said hyperscalers and AI cloud providers would be able to tap the distributed network as they would a traditional data center.
The pitch addresses two pressures that have shaped recent data-center construction. One is public opposition.
Communities have pushed back against large facilities that consume lots of water and electricity. The other is grid stress. Rapid growth in AI and other compute needs has increased local power demand and, in places, pushed up household electric bills.
Span's approach tries to turn those trends into an asset. By placing compute where power already exists, the startup says it can reduce the need to site new, large facilities on greenfield land. By timing compute loads to match spare capacity, Span aims to avoid adding steady new draws on local grids.
PulteGroup, a major U.S. Homebuilder, is participating in the plan.
The company will work with Span to allow installation of XFRA units on houses in some developments. Span and PulteGroup say the arrangement will let builders offer an option to host the small data hubs as part of new or repaired home exteriors.
Span also highlights a potential homeowner benefit. The company's panels are designed to optimize energy use. Span says homeowners can earn compensation when their site's spare capacity is used for compute. That payment flow could offset higher bills that have been linked to increasing neighborhood loads in recent years.
Still, the model changes how compute is located and how utilities manage demand. The distributed network relies on many small power draws placed across municipal circuits. That could change local load profiles in ways utilities must monitor. Nvidia framed the move as accelerating access to AI solutions and helping affordability for compute customers.
Deployment and rollout questions
Span said the XFRA units are small and modular. The company argued the installation timeline is faster than the months or years required to permit and build a large centralized center. Span also emphasized hardware design choices, like liquid cooling, that limit noise and exposure to outdoor elements.
The plan requires cooperation across several actors. Homeowners must accept exterior-mounted boxes. Builders must include the option in new construction. Utilities must allow the additional, intermittent loads on local circuits. Cloud and AI providers must integrate the distributed nodes into their operations.
Span has pitched the network to hyperscalers and AI cloud customers as a way to augment existing data centers or to reduce the need for greenfield builds. The company says the distributed nodes can combine to match the capacity of a small to mid-sized traditional data center when aggregated.
Span's public descriptions stop short of detailed rollout calendars and specific financial terms for homeowners or builders. The company has given headline metrics for deployment speed and cost relative to a 100-megawatt centralized build. Those numbers anchor Span's claim that distributed, residential-hosted compute can be cheaper and faster to deploy than conventional centers.
Related Articles
- Gen Z splits: heavy AI use and a growing backlash
- MacBook Neo vs iPad Air: Which $599 Apple Buy?
- 5 new tools shaping how we build tech
"Fundamentally, it's an infrastructure play," said Arch Rao, founder and CEO of Span.
This article was created with AI assistance.
