June 3, 2026 · Tags: ai, infrastructure, data-centers, energy, nimby, span, nvidia
There's a headline circulating that Nvidia will pay a homeowner $22,000 a year to put a mini data center in their backyard. It is, as far as viral AI infrastructure claims go, exactly the kind of number that makes you stop scrolling. It is also, almost certainly, not a real figure tied to any actual program. The $22,000 came from an X post last year about the NVIDIA DGX Spark — a $2,999 desktop AI computer — where the author calculated the cloud GPU bill you'd avoid by running inference locally. Roughly $1,900 a month in cloud time, times twelve. That figure was never a payment from anyone. It got copied, pasted, and stapled onto a different story.
The different story is real, though, and it's more interesting than the headline.
What's actually happening #
On April 13, 2026, a San Francisco startup called Span announced XFRA — a "distributed data center" product that puts a small, white, AC-unit-sized box outside a home. Inside the box, sealed and liquid-cooled, are 16 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, four AMD EPYC server CPUs, and 3 TB of RAM. It draws about 12.5 kilowatts at full load — the kind of headroom most modern homes already have on a 200-amp service. It is, for all practical purposes, a data center. Just a small one. Sitting next to the HVAC condenser.
Homebuilder PulteGroup is the launch partner. The arrangement is a new-construction pilot: 100 Pulte homes in the southwestern U.S. will get XFRA nodes installed during the build, beginning in Q3 2026. CNBC reported in early May that exactly one home has the unit installed so far. The first commercial deployments of this size usually start this small.
The economics, for real this time #
The homeowner does not get a $22,000 check. They don't get a check at all. They get a $150 monthly bill that covers electricity and internet — positioned by Span as a discount on what a typical household would pay for those services combined. They also get a free Span smart panel and a roughly 15 kWh home battery installed, which Span owns (think solar-lease model — the homeowner gets the use, the company keeps the title).
Span, in turn, owns the GPU fleet. Hyperscalers, AI companies, and neocloud providers rent compute from Span's distributed network exactly the way they'd rent it from a centralized data center. The pitch is that inference workloads — the part of AI that responds to a prompt in real time — don't need to be in one giant building. They can be scattered across thousands of homes, on existing electrical service, on existing internet, with no new transmission lines and no new cooling-water permits.
Span claims it can deploy 8,000 XFRA nodes in roughly a sixth of the time and at one-fifth the cost-per-megawatt of a traditional 100 MW data center. The target is gigawatt-scale annual deployment by 2027.
Why this exists at all #
The reason anyone is experimenting with backyard compute is the simpler story most people have heard but few have quantified: a lot of big data centers are not getting built.
A research group called Data Center Watch — backed by an AI-security outfit called 10a Labs — published a report in mid-2025 claiming that $64 billion worth of U.S. data center projects had been blocked or delayed by local opposition between May 2024 and March 2025. That's 16 projects, 28 states, mostly over noise, water consumption, and the kind of grid strain that ends up on a homeowner's electric bill. By their own Q2 2025 update, the figure had grown to $98 billion in a single three-month period.
The reasons aren't subtle. A 100 MW data center's worth of cooling can pull something like five million gallons of water a day. The diesel and natural-gas backup generators emit the kind of nitrogen oxides that make the EPA write letters. The transformer hum from the substation that has to be built next door is the kind of 24/7 noise that gets people showing up at planning commission meetings. A property that used to be corn becomes a place where you can't sleep, and the tax breaks that attracted the developer in the first place are the same tax breaks the school district would rather have had.
Baird, the investment bank, has its own count: 25 data center projects were canceled in 2025, up from 6 in 2024 and 2 in 2023. The cancellation rate has roughly doubled every year. Most of the opposition is bipartisan and locally organized. In Warrenton, Virginia, every town council member who voted for an Amazon data center lost re-election. In Cascade Locks, Oregon, voters recalled two port authority officials who supported a project; the new board killed it on day one. In Prince William County, Virginia, a judge voided the rezoning of a 2,100-acre, 37-building data center corridor because the public notice was insufficient.
The viral "$22,000 a year" framing is wrong, but the underlying problem it gestures at — the sense that someone is trying to put industrial infrastructure where it doesn't belong, with little explanation of who benefits — is exactly the anxiety XFRA is being designed to dodge.
The interesting bet #
The XFRA thesis is that if the problem is visibility, the answer is invisibility. A 12.5 kW box next to the air conditioner is much easier to ignore than a 100 MW warehouse. The homeowner's electric bill goes down. Their backup battery works for them during outages. They don't see the GPU humming. Their HOA might, but that's a different fight.
The trade-offs are real, though, and the experts cited in the press coverage so far are not all convinced.
Grid management. A single XFRA node fits on a normal home service. Eight thousand of them on a single suburban feeder line is a different problem. Rich Brown at Lawrence Berkeley National Lab told Scientific American that distributed nodes could "fill in all the valleys" of grid load, eliminating the diversity benefit of having different homes draw power at different times. The local peaking problem is solvable, but it requires the utility to know in real time which houses are running inference at full tilt — which requires Span to share data, which raises its own questions.
Theft and security. Each RTX PRO 6000 Blackwell GPU retails for around $10,000. Sixteen of them in a box outside a house is, mechanically, a $160,000 target behind a fence. Ars Technica quoted UPenn's Benjamin Lee on the security problem: a node is far easier to physically access than a centralized data center, and side-channel attacks that need hardware proximity become more practical.
Resale and insurance. A house with a Span hosting agreement is, in the eyes of a mortgage appraiser, an unusual property. Does the agreement transfer to the next buyer? Does the homeowner's insurance cover a server farm in the yard? Does the GPU count as personal property for tax purposes? Span has answers to some of these. None of the answers are tested yet in court.
Future headroom. This is the one that quietly makes the economics look shakier. The 80-amp headroom XFRA depends on, in a typical 200-amp home, is fine for the average household in 2026. It is not fine for a 2035 household with a heat pump, two EV chargers, an induction range, and a hot tub. Jonathan Koomey at Stanford told Scientific American that the headroom XFRA relies on may shrink significantly as homes electrify. Span's business model assumes houses stay roughly as power-hungry as they are now.
Why this matters beyond the box #
Span is not the only company thinking along these lines. Vast.ai and RunPod already run distributed GPU marketplaces that crowdsource consumer hardware. Akash Network does something similar on the blockchain-flavored side. What XFRA is doing differently is controlling the entire stack: it designs the node, owns the hardware, orchestrates the workloads, owns the smart panel, and now has a homebuilder partner putting them in new construction. It is, in a real sense, a vertically integrated distributed cloud.
The interesting question isn't whether 100 homes in Nevada will get XFRA nodes in Q3 2026. They will. The interesting question is whether this scales. The numbers Span is targeting — gigawatt-scale by 2027, multi-gigawatt after that — would represent a fundamentally different shape for AI infrastructure. Instead of a few hundred enormous buildings, mostly clustered in Northern Virginia, you get millions of small nodes, distributed wherever there's a Pulte development with fiber and a 200-amp panel.
The bet is that, ten years from now, the average suburban backyard will be running about 12 kilowatts of compute in addition to the usual appliances. Whether that future is desirable, and who exactly would benefit from it, are the questions the "$22,000 a year" headline gestured at without actually answering. The real version of the story is more complicated, less viral, and worth taking seriously.
Sources
Span.io — "SPAN Announces XFRA, a Distributed Data Center Solution" (April 13, 2026): https://www.span.io/blog/span-announces-xfra-a-distributed-data-center-solution-to-close-the-speed-to-power-gap-for-ai-compute-demand
XFRA White Paper (April 2026): https://ap.span.io/XFRA_White_Paper.pdf
CNBC — "Nvidia, PulteGroup partner with Span to put mini data centers on homes" (May 5, 2026): https://www.cnbc.com/2026/05/05/nvidia-pulte-span-mini-data-centers-on-homes.html
Ars Technica — "The newest AI boom pitch: host a mini data center at your home" (May 2026): https://arstechnica.com/ai/2026/05/the-newest-ai-boom-pitch-host-a-mini-data-center-at-your-home/
Scientific American — "Span Wants to Turn Homes into Mini Data Centers" (May 2026): https://www.scientificamerican.com/article/span-wants-to-turn-homes-into-mini-data-centers/
Latitude Media — "Span to launch mini AI data centers for distributed at-home compute": https://www.latitudemedia.com/news/span-to-launch-mini-ai-data-centers-for-distributed-at-home-compute/
Realtor.com — "Nvidia, PulteGroup, Span: Date Center in the Backyard" (May 6, 2026): https://www.realtor.com/news/trends/nvidia-pultegroup-span-date-center-backyard/
PV Magazine — "Span and Nvidia to develop AI data centers in your backyard" (April 15, 2026): https://pv-magazine-usa.com/2026/04/15/span-and-nvidia-to-develop-ai-data-centers-in-your-backyard-lowering-electric-bills/
The Hill — "16 data center projects worth $64B blocked or delayed" (Nov 2025): https://thehill.com/policy/technology/5605667-data-center-criticism-study/
Data Center Watch — Q2 2025 Update ($98B in a single quarter): https://www.datacenterwatch.org/q22025
Construction Dive — "Data center project cancellations quadrupled in 2025" (Baird analyst data, April 22, 2026): https://www.constructiondive.com/news/data-center-project-cancellations-power-public-pushback/818157/