SPAN's XFRA program, backed by Nvidia and homebuilder PulteGroup, will mount 16-GPU compute nodes on suburban homes starting Q3 2026, offering hosts $150-a-month utility and internet coverage in exchange for the wall space, as AI's appetite for power begins hitting household bills nationwide.
The pitch sounds almost too clean. Let a startup bolt a server to your house, and you stop paying for electricity and internet. SPAN, the smart electrical panel company, announced in May that it's doing exactly this, partnering with Nvidia and PulteGroup to deploy what it calls XFRA nodes on newly built homes in Arizona and Nevada, with a 100-home pilot scheduled for this summer. Each unit packs 16 Nvidia RTX Pro 6000 Blackwell GPUs and four AMD EPYC CPUs into a liquid-cooled enclosure mounted on the exterior wall. SPAN owns the hardware, sells the compute to hyperscalers and AI cloud providers, and hands the homeowner a fixed monthly fee covering both electricity and internet at roughly half what most Americans pay for those services combined. The homeowner provides the location. That's the whole deal.
It's a genuinely interesting model because the grid pressure that makes it necessary is already landing on your bill. According to Goldman Sachs research published in February, electricity prices jumped 6.9% in 2025, more than double that year's headline inflation rate of 2.9%. The capacity price that PJM, the grid operator covering a wide swath of the eastern US, charges to guarantee enough generating power has rocketed from $28.92 per megawatt-day two years ago to $329.17 for the year beginning June 2026. Goldman projects US data center power demand will climb from 31 gigawatts in 2025 to 66 gigawatts by 2027, and the bank is direct about what that means: higher bills, lower disposable income, a measurable drag on consumer spending. Data centers are expected to account for 40% of all electricity demand growth through the end of the decade, as CNBC noted in its coverage of the Goldman report.
The harder question is on SPAN's side of the ledger. Installation is free, maintenance is SPAN's problem, and the homeowner carries no capital risk. But SPAN plans to ramp to 80,000 homes by 2027, which it says would deliver more than one gigawatt of distributed compute capacity. That's a serious number. Selling that to hyperscalers at rates competitive with purpose-built data centres requires proving that fragmented residential infrastructure can match the reliability and density those buyers expect. The 100-home pilot this quarter is specifically designed to answer that. Whether it does will determine whether the economics work at all.
Not the First Distributed Bet
The model isn't entirely new. UK startup Heata has been running a quieter version of the same idea since 2022, installing server units directly onto household hot water cylinders so the waste heat from computation warms water that would otherwise need a gas boiler. British Gas signed on for a trial partnership to explore scaling the approach. Heata covers the homeowner's electricity and handles all maintenance. The thermal integration is clever, turning a data centre's most persistent problem - heat - into the product itself. Where SPAN is optimising for raw compute density, Heata is optimising for energy efficiency at the household level. Both are trying to move the argument about distributed compute from thought experiment to operational proof.
Frankly, the more interesting question for investors isn't whether homeowners will sign up. They will, especially as electricity bills keep climbing. The question is whether distributed residential compute can hold its position against hyperscalers pouring hundreds of billions into centralised capacity. Amazon, Microsoft, and Google are all deepening their data centre commitments for exactly the reasons XFRA was designed to address: permitting delays, grid interconnection queues, and the multi-year lead times for large facilities. SPAN's argument is that its nodes can be deployed in months rather than years, using existing grid connections at homes already built, and that aggregated capacity across tens of thousands of sites becomes a credible alternative for workloads that don't require one giant cluster.
That argument has traction because the bottleneck is real. What it hasn't proven is durability. A 100-home pilot in the desert southwest tells you whether the hardware survives and whether hosts follow through on their end. It doesn't tell you whether enterprise buyers will write serious contracts against residential infrastructure, or whether SPAN's operating margin holds when you're managing hardware across 80,000 separate households. Those are questions Q3 2026 won't answer. But given that the alternative is waiting years for grid-connected land in any major metro, the distributed model doesn't need to be perfect to win business. It just needs to be faster than what it's competing with. Right now, it is.
Also read: OpenAI launched ChatGPT Health and quietly moved your medical records outside HIPAA • An AI data center that promised to spare the Colorado River is now suing for 260 million gallons a year • Nearly 200 startups tell Washington that banning Chinese AI models would hand OpenAI and Anthropic a monopoly
