Fast Company published a careful takedown of the Span-PulteGroup pilot that is supposed to let homebuilders put mini AI data center nodes alongside residential HVAC equipment. The pitch is straightforward. Span's smart utility box detects unused household electrical capacity, steers that capacity to GPU nodes mounted outside the house in HVAC-looking enclosures, and pays the homeowner a share of the resulting compute revenue. Nvidia is providing the GPUs. Pulte is providing the houses. Span is providing the box.
Reading the reporting carefully reveals the gap between the press release and the engineering reality. Span has not actually deployed at scale. Pulte told CNBC that exactly one prototype has been installed next to a single home. Span's VP Chris Lander declined to share technical studies showing the distributed compute architecture can actually deliver AI workloads at production quality. The company says the pilot will involve "upwards of 100" nodes "later this year." Neither timing nor location has been disclosed.
A GPU node sitting next to a residential HVAC unit has to reject heat somewhere. The most likely thermal architecture is a self-contained air-cooled enclosure with fans pushing waste heat into the air around the house. Two practical problems follow immediately.
First, the noise. Even a modest 5 kW GPU node running air cooling at full load will emit fan noise well above ambient residential levels. Residential HVAC compressors run around 50 to 65 dBA at the unit. A continuously running GPU enclosure at the same scale runs higher, especially during AI training workloads that hold the GPUs at sustained near-peak power for hours. The Span pitch assumes neighbors will not complain because the enclosure looks like an HVAC unit. The neighbors will hear the difference within a week.
Second, the thermal envelope. A residential lot in Arizona, Texas, or Florida will see ambient temperatures above 38°C for months at a time. Air cooling a 5 kW GPU at those ambient temperatures requires either oversized heat sinks, larger fans, or both. Power draw on the cooling system rises sharply with ambient temperature. The GPU's effective performance drops because it has to throttle to stay under thermal limits. The 100-node pilot has not been stress-tested through a Texas summer. The data on what the actual usable compute capacity looks like at 40°C ambient does not exist.
Span's marketing positions the distributed approach as a way to ease load on the regional grid by using residential transformer capacity that is sitting idle. The math is more complicated than the pitch suggests. Residential rate structures are the political third rail behind every hyperscale opposition campaign. Communities are angry because data center growth is correlating with rate hikes on their power bills. The Span model does not change that equation. It actually deepens it by drawing power through residential transformers and distribution lines that are billed at residential rates, while the compute revenue accrues to a third party.
If neighbors realize their house is hosting a GPU node and the homeowner is taking a share of the revenue, the political math becomes harder still. The pitch only works if the nodes are essentially invisible. The cooling noise, the visible enclosure, and the rate-structure questions make invisibility impossible.
The Span concept is not the first attempt to distribute AI compute into residential or edge environments. Edge data centers, micro DCs, and modular containers have all been tried in various form factors over the last decade. Most have not scaled because the thermal architecture trades efficiency for distribution, and the resulting cost per FLOP is hard to compete with what hyperscalers can deliver from a centralized 100 MW facility running optimized direct-to-chip cooling.
The cooling vendor base should treat the Pulte-Span pilot as a low-probability outcome that, if it works, would shift the addressable market for residential-grade thermal solutions in interesting directions. If it does not work, the political pressure that pushed hyperscalers to look for distributed alternatives will refocus on the hyperscale model and the existing cooling architectures that support it. The interim is going to be an experiment with maybe 100 prototype units, which is too small a sample to draw a credible thermal performance conclusion from. Wait for the second pilot before adjusting any product roadmap.