Infrastructure
What a data centre in 100 backyards must prove.
Updated 13 July 2026. This analysis now reflects SPAN's June support material and Berkeley Lab's July queue update.
What would have to be true before you let a data centre share your home's electrical connection? That is no longer a theoretical question. SPAN is developing XFRA, a fleet of outdoor compute nodes installed with its smart electrical panel and a battery. The company's project page lists sixteen NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, four AMD EPYC CPUs, three terabytes of memory and liquid cooling in each node. NVIDIA is an initial launch partner, and SPAN says it is working with PulteGroup on the first rollout in new homes, according to SPAN's April 2026 announcement.
The first economic question needs a correction. This is not a published offer to rent any existing backyard, and SPAN has not confirmed a universal household payment, electricity bill or internet price. Its support page, updated June 23, 2026, describes a 100-home proof of concept with US new-construction builders. It says the XFRA equipment will be included in the price of those homes. It also says a monthly value proposition may include bundled services and host compensation, but the details vary by community and are still being developed. Included in a home price is not the same claim as free, and a possible benefit is not yet a household saving.
What the proof of concept is testing
XFRA's technical idea is to treat compute as a flexible load behind the household meter. SPAN's premise is that a home electrical service is designed for peaks while many circuits draw less than that limit much of the time. The company says its panel can identify unused headroom and operate the node without taking control away from the homeowner. Many nodes would then be coordinated as one inference fleet. Those are project claims that the proof of concept still has to demonstrate under real weather, household demand, network conditions and utility rules.
The demand behind the experiment is real, but it needs careful labels. The US Department of Energy's summary of the 2024 national data-centre energy report says data centres used about 4.4% of US electricity in 2023 and projects a range of 6.7% to 12% in 2028. That is a national scenario range, not a guarantee about one city or feeder. In a Federal Energy Regulatory Commission concurrence, Commissioner Judy Chang described the broader PJM challenge of interconnecting large loads and the generation needed to serve them, while warning against shifting costs to residential customers. Location and grid region matter.
One statistic does not prove the data-centre claim. Berkeley Lab's interconnection queue report tracks proposed large generation and storage projects seeking transmission access. It does not count data-centre demand waiting to connect. The report is relevant to the separate challenge of adding supply, but it should not be presented as a queue of AI facilities.
One home, one feeder, several decisions
Consider a household on a hot evening. The heat pump, oven and car charger may all need power while the compute node has work queued. Before signing, the buyer needs clear answers: Does household demand always take priority? When is compute curtailed? How much battery capacity remains for an outage? Who services the node, carries equipment risk and controls access to the property? What happens to the agreement when the home is sold? The current public material states that the homeowner retains control over home energy use, but it does not yet publish all of these contract and operating details.
The utility has a different decision. Spare capacity in one service panel does not prove spare capacity across the local transformer or feeder. It needs evidence about simultaneous peak demand, curtailment response, protection, metering and who pays if an upgrade becomes necessary. Claims that the fleet helps rather than merely uses the grid need evidence about how it responds when the local system is tight, as well as any off-peak use or deferred upgrades. Those questions are part of the product, not administrative fine print around it.
A compute node crosses three control boundaries
The proof of concept must work for the household, local grid and fleet at the same time.
- Household priority Heat pump, cooking, charging and outage reserve must retain defined precedence.
- Panel and battery control Curtail compute within electrical and homeowner limits.
- Transformer and feeder capacity Spare room behind one meter does not prove local network headroom.
- Fleet coordination Dispatch many nodes without creating a simultaneous new peak.
- Contract and accountability Name payment, maintenance, equipment risk, property access and sale terms.
The capability is credible only if hidden cost and risk do not move to households or the grid.
The signal, with its limits intact
My inference is that XFRA matters less as a verdict on the future of data centres than as a test of a new deployment boundary. The pattern resembles the move from a model to the harness around it: capability becomes more useful when the surrounding system can coordinate evidence, tools and limits. It also resembles the difference between renting model access and retaining an operating system, or building shared organisational memory. Here the surrounding system is physical: panels, batteries, fiber, cooling, contracts and grid controls.
Distributed nodes are not a general replacement for central data centres. SPAN itself describes them as an augmentation aimed at inference and other edge workloads. The proof of concept still has to establish reliability, security, maintenance economics, household acceptance, utility value and fair cost allocation. A hundred planned homes cannot answer the national grid question. It can answer a smaller and useful one: whether flexible compute can share existing residential infrastructure without transferring hidden cost or risk to the people and networks hosting it.
Sources & caveats. Project specifications and deployment claims come from XFRA and SPAN, including SPAN's support material updated June 23, 2026. National electricity figures come from the US Department of Energy. Load-connection context comes from FERC. Berkeley Lab's queue data are cited only to distinguish generation and storage interconnection from data-centre load connection. Product performance, host economics and grid benefits remain claims to be tested in the 100-home proof of concept.