# HEARTH

> A feasibility study for putting AI inference into ordinary houses instead of building new
> hyperscale data centers. All arithmetic is exact rational, not floating point.

The thesis: Berkeley Lab projects US data-center electricity use at 521-843 TWh in 2030. ERCOT
reported 410 GW of prospective large loads in April 2026, about 87% associated with data centers;
that queue is not a build forecast, but it shows the pressure on the connection process. Meanwhile
82.5 million US single-family detached homes already have meters and electrical services, though
eligibility, spare capacity and utility approval would need to be established site by site.

## Findings

- The model was run five times. Two passes concluded the home LOSES. Both corrections are findings.
- Comparing venues on facility cost is wrong: accelerators dominate and are identical in both, so
  the decision turns on how many kWh each venue spreads the same silicon over.
- Compare accelerators on cost per unit of delivered work at PRODUCTION BATCH, never on $/MW and
  never single-stream. Keep FP8 figures dense for every part.
- Model size decides the winner. At 8B a home fleet costs 0.49x a Blackwell data center per million
  tokens. At 32B it costs 1.16x. The 70B case is excluded by the study's single-GPU assumption;
  multi-GPU serving was not modeled.
- Three thresholds: duty cycle > 0.8433, residential retail rate < $0.2309/kWh, and (at 32B only)
  an interconnection wait > 13 months.
- One node per distribution transformer: 24/7 compute has no diversity, and residential
  distribution is engineered around diversity.
- The mesh's unique property is aggregate warm-weight capacity: 1,024 PB across 4M homes, which
  would take 76,417 GB200 NVL72 racks and a 9.17 GW grid connection to match.
- Pilot: 220,519 homes over 66 months behind four gates. Peak funding $1.51B (not the $4.44B of
  gross capex). Break-even token price $0.0246/Mtok against an assumed $0.030 — 18% headroom.
- Largest risk is consumer GPU licensing: the cost advantage rests on price segmentation, not physics.

## Pages

- [The vision](https://copyleftdev.github.io/hearth/): the case, the numbers, the pilot, the risks.
- [Prototype studies](https://copyleftdev.github.io/hearth/prototypes/): wall, tower and duct-integrated appearance and installation concepts, with validation questions.
- [The arithmetic](https://copyleftdev.github.io/hearth/report/): the full five-pass model with every threshold and stress case.
- [Source and data](https://github.com/copyleftdev/hearth): input datasets with citations, the model code, and every computed
  quantity as exact rationals plus decimals.

## Reproduce

    git clone https://github.com/copyleftdev/hearth
    cd hearth && for i in "" 2 3 4 5 6 7 8 9 10 11 12; do python3 calc/model$i.py; done

Requires `agent-calc`, an exact-computation CLI. Every number in the study comes from it.