San FranciscoNvidia partner Span installs home compute nodes that pay owners in free electricity.
Span says typical homes carry 80 unused amps; its nodes convert that capacity into AI cloud compute routed to providers on demand.
100 southwestern US homes host prototypes this fall; Span targets 1 gigawatt of capacity.
Sources: Aol
THE CONSTRAINT
AI compute is gated by power, not silicon. A modern H100 GPU draws 700W; a rack draws 40kW; a hyperscale cluster draws hundreds of megawatts. Utilities take 3-7 years to build new transmission. The bottleneck for AI in 2026 is the interconnect queue at the substation, not the fab in Taiwan.
THE 80-AMP CLAIM
A typical US home has a 200-amp service panel but draws only 30-50 amps on average. The remaining headroom is sized for simultaneous peak loads — AC, dryer, oven, EV charger — that almost never coincide. Multiplied across 80 million single-family homes, the unused residential capacity rivals the entire installed base of US data centers.
THE LATENCY TRADEOFF
AI training needs tightly-coupled GPUs with sub-microsecond interconnects — impossible across residential broadband. AI inference is different: a single forward pass is independent, tolerates 50-200ms latency, and can be sharded across thousands of nodes. Span's nodes can only ever serve inference, never training.
THE PRECEDENT
Distributed home compute is not new. SETI@home (1999) used screensaver cycles to search radio data for alien signals; Folding@home hit 2.4 exaFLOPS during COVID protein research, briefly the world's most powerful supercomputer. The novelty is monetization — paying hosts in kilowatt-hours rather than asking for volunteer cycles.
THE GRID ECONOMICS
Residential electricity in the US Southwest costs $0.10-0.15/kWh; hyperscale data center power runs $0.04-0.06/kWh on industrial contracts. Span's model only works if the marginal node displaces no new generation — i.e., it consumes power the homeowner already pays the fixed-cost portion of.
WHY SOUTHWESTERN US
The pilot region matters. Arizona, Nevada, and southern California combine high solar penetration, low residential electrical-code restrictions on auxiliary loads, and the same hyperscaler corridor (Phoenix, Las Vegas, Reno) where Microsoft, Meta, and Google are queueing for grid interconnects. The same scarcity that makes the home network valuable concentrates demand within a one-state radius.
THE GIGAWATT TARGET
1 GW is the size of a large nuclear reactor or roughly 1 million H100 GPUs at full draw. To hit it through residential nodes, Span needs ~500,000 homes at 2kW each — about the size of a single utility's residential customer base. The number is plausible at the unit-economics level but unprecedented as a coordination problem.