AI's enormous appetite for electricity is running into a physical problem: building giant new data centers takes years, requires massive amounts of power and water, and increasingly faces local opposition.
A startup called SPAN is testing a radically different approach. Instead of putting thousands of GPUs inside one enormous data center, its XFRA system would distribute smaller AI-compute nodes among individual homes. Think of something roughly the size of an outdoor AC unit containing 8-16 NVIDIA GPUs and potentially $150,000+ worth of computing equipment.
The clever part is how it uses the home. A typical 200-amp residential electrical service can theoretically provide around 48 kW, but homes use only a fraction of that capacity much of the time. SPAN's technology can manage household loads and tap some of that unused electrical headroom for AI computing.
But who pays the enormous electric bill from running all those GPUs? The idea isn't for the homeowner to simply absorb it. Instead, the home essentially becomes a host for part of a distributed data center. SPAN has discussed compensation packages that could include discounted or fixed-price household electricity and internet, battery backup and other benefits in exchange for hosting the equipment and providing access to the home's electrical infrastructure. The AI customers ultimately pay for the computing power.
SPAN is working with NVIDIA and homebuilder PulteGroup on an initial deployment of roughly 100 homes. If it works, SPAN wants to begin broader deployment in 2027 and eventually build more than 1 gigawatt of distributed AI computing capacity, potentially requiring tens of thousands of installations. Basically, turn thousands of homes into one enormous decentralized data center.
And investors are already looking at possible picks-and-shovels plays:
NVT - nVent Electric
Electrical enclosures and data-center infrastructure.
GNRC - Generac
Home energy management, backup power and growing data-center power exposure.
VST - Vistra
One of America's major electricity producers, with exposure to regions experiencing enormous data-center power demand.
None are a direct pure-play investment in SPAN, and XFRA itself is still in the proof-of-concept stage. But the underlying idea is fascinating. We've spent years asking where we're going to find enough land, electricity and infrastructure for the AI boom. Maybe part of the answer isn't another giant data center, maybe it's 80,000 tiny ones hiding in plain sight.
AI rewrite of this article?
Not from that article...I used AI to de-tease a promotional pitch for one of Angel Publishing's premium subscriptions, specifically from their Energy & Capital segment.
I get a lot of these pitch emails wanting $1000s to join their newsletter distros for their member picks. Typically there are enough details leaked in the promo that AI can figure them out, but not all of them.
I can do you one better: my next datacenter will be sitting in my living room!