Homeowners Could Soon Host The Computing Power Behind AI
Home AI data centers are turning residential real estate into distributed compute infrastructure. Putting data centers inside people’s houses isn’t a hobbyist experiment anymore — it’s a serious commercial proposition.
Span, working with Nvidia and homebuilder PulteGroup, is installing compact compute hardware inside newly built homes. The setup taps the spare electrical capacity those properties already have and puts it to work processing AI workloads.
A cluster of startups is chasing the same distributed-compute model, driven by AI demand that’s overwhelming the centralized grid.
A separate report makes the case from the other direction.
One downed power line in Northern Virginia — home to the world’s largest concentration of data centers — was enough to show just how fragile centralized AI infrastructure has become.
Why Data Centers in People’s Homes Exist at All
The core problem is watts, not silicon. A conventional hyperscale data center draws between 100 and 500 megawatts of power.
The United States electrical grid was not designed to absorb dozens of these facilities appearing within a few years. Northern Virginia alone hosts roughly 35% of the world’s internet traffic, and grid operators there have warned of capacity shortfalls stretching into the 2030s.
The distributed-compute trend was detailed by Fortune on July 25 this year, while a report published the same day examined how the Northern Virginia power line failure exposed the fragility of centralized AI infrastructure.
The theory behind placing Data Centers in People‘s homes is that millions of residential properties already carry more electrical service than they use at any given moment. A typical American house is wired for 200 amps but rarely draws more than 50 during off-peak hours.
That unused capacity can host a small compute node. Aggregated across thousands or millions of homes, the total compute becomes meaningful.
Span makes smart electrical panels that can measure, route, and manage a home’s power draw in real time.
The panel already knows exactly how much spare capacity exists at any given second. Adding an Nvidia compute card to that system means the panel can spin workloads up when the home is quiet and throttle them when the oven or EV charger kicks on.
PulteGroup, one of the largest homebuilders in the United States, provides the distribution channel by wiring Span panels into new housing developments from the ground up.
From Grid Accident to Commercial Urgency
The Northern Virginia incident gave the Data Centers in People‘s homes thesis a harder edge. When a single downed transmission line hit in June this year, multiple data center campuses switched to backup diesel generators simultaneously.
Grid engineers watching the event said the coordinated surge nearly cascaded into a wider outage. The incident made national infrastructure planners take seriously what had previously seemed like a niche concern.
It is worth understanding what made Northern Virginia so vulnerable.
Ashburn, the county seat of the region’s data center corridor, sits at the end of a relatively thin spur of transmission infrastructure. The hyperscalers chose it for fiber density and tax incentives, not grid redundancy.
The more compute that concentrates there, the worse the single-point-of-failure risk becomes. Data Centers in People‘s residences, even if they handle only inference workloads rather than training, relieve that geographic chokepoint.
The distinction between training and inference matters here. Training a large AI model requires sustained, coordinated compute across thousands of chips running for weeks.
That cannot happen in a living room. Inference, the act of running a trained model to answer a query, is far lighter and can be split across many small nodes. Data Centers in People‘s homes would handle inference only.
That still represents a large and fast-growing slice of total AI compute demand.
Also Read: Anthropic Drone Benchmark: All Eight AI Models Fail Critical Test
The Business Model Behind Your Walls
Homeowners do not simply donate their electrical capacity. The model Span and its peers are pursuing is closer to a revenue-sharing arrangement.
The homeowner provides the panel space and accepts the compute hardware at low or no upfront cost. The startup routes AI inference jobs through the hardware, pays the electricity bill for those jobs, and shares a fraction of the compute revenue with the homeowner.
This structure echoes earlier distributed computing projects.
Folding@home, which launched in 2000, recruited volunteer home computers to simulate protein folding for medical research. The difference is that Data Centers in People‘s homes pay real money rather than relying on volunteers.
Inference as a service is a market that research firm Gartner projected would reach $100 billion annually by 2028, and it is growing faster than centralized capacity can match.
For Nvidia, the partnership with Span is both a supply-chain hedge and a market-expansion play. Every home compute node runs Nvidia silicon.
The total addressable market for residential AI compute hardware is enormous if even a fraction of new home builds include it. PulteGroup built roughly 31,000 homes in 2025.
If each included a compute node, that alone would add hundreds of megawatts of distributed inference capacity per year.
The regulatory picture is still forming. Utilities in several states have begun asking whether Data Centers in People‘s homes constitute commercial operations under existing tariff structures, which could expose homeowners to higher electricity rates.
That question is unresolved, and it represents the most immediate policy risk to the model’s scale-up.
Read Next: Can Bitcoin ETFs Reclaim The Flow Crown? Ethereum Has Won Three Weeks Running
