A vision for the compute crunch that does not require pouring any more concrete.

Read the vision → · Explore the prototypes → · Read the arithmetic →

The United States is in the middle of a data center buildout without precedent, and it is running into a wall that money cannot move.

US data centers drew 176 TWh in 2023. Lawrence Berkeley National Laboratory now projects 521–843 TWh in 2030, with a 649 TWh reference case. That capacity is being sited next to housing, drawing on local water and local grids, and increasingly meeting organised local opposition.

Power delivery is increasingly a schedule constraint. In April 2026, ERCOT reported 410 GW of prospective large loads, about 87% associated with data centers. That queue is not a forecast of what will be built, but it shows how much demand is converging on the same connection process.

So the industry is buying accelerators it cannot energize. A megawatt of idle silicon loses about $520,833 in value every month it waits.

There are 82.5 million single-family detached homes in the United States. Roughly 60% are passed by fiber. Each has a meter and an electrical service. That does not establish five kilowatts of spare capacity: panel condition, other loads, distribution equipment, permits and utility approval must be screened.

HEARTH asks what happens if you put a 5 kW compute node — the electrical size of an EV charger — behind an existing residential meter, pay the household for hosting it, and schedule a few million of them as one fabric.

The proposed hosting contract asks the household to buy nothing and take no loan. It assumes a dedicated 240-volt circuit, a revenue share and operator ownership. Installation time, heat reuse, insurance and transfer to a later owner are hypotheses for a field pilot—not promises.

This is a feasibility study, not a manifesto. It was run five times and two of those passes concluded the idea does not work. Both corrections turned out to be the most useful findings, and both are published here in full.

The answer is that it is a small-model inference fleet. That is the specification, not a hedge:

Three thresholds decide everything. The point estimates are illustration; these are the result:

• Duty cycle > 0.8433 — below it, idle silicon costs more at home than the whole house saves
• Residential retail < $0.2309/kWh — the only one of eleven inputs that flips the sign
• Interconnection wait > 13 months — needed at 32B, irrelevant at 8B

Four million nodes cannot be pooled to run one enormous model — a four-stage pipeline across houses yields 5.67 to 11.74 tokens per second, and training all-reduce moves 28 GB per step. The fleet is wide, not deep, and width is the product:

• 1,024 PB of aggregate warm model weights — matching that in racks takes 76,417 GB200 NVL72 racks and a 9.17 GW grid connection
• 200 warm replicas of every model in a million-model catalogue under the modeled allocation; node startup, failover and reassignment can still create cold starts
• A data center forfeits 57,447 tokens on every weight swap; at 16.71 swaps/hour its cost per token rises to meet the mesh's on cold starts alone
• No durable session affinity. One node still holds each in-flight KV cache; if it fails, that generation is lost. New work can route elsewhere at an assumed 1.031× spare-capacity cost

220,519 homes over 66 months, behind four gates, each buying the answer to one question.

• Peak funding $1.51B in month 59 — not the $4.44B of gross capital, because earlier cohorts fund later stages
• Free cash positive month 60; cash-whole month 74; NPV $0.86B at 8%
• Break-even token price $0.0246/Mtok against an assumed $0.030 — only 18% of headroom

That table also shows a flaw the study found in its own plan: $822M is committed before the question most likely to kill the programme is answered, and that question is the cheapest of all to test. It should be first.

Named and priced, because a study that hides its failure modes is just a slower way of finding them.

1. Consumer GPU licensing. The entire cost advantage rests on consumer parts being 2.1–9.5× better per dollar of memory bandwidth. That gap is vendor price segmentation, not physics, and the consumer licence restricts data center use precisely to protect it.
2. The token price. Break-even is $0.0246/Mtok in a market whose list prices keep falling. Below $0.020 no amount of demand makes a node pay.
3. GPU acquisition price. Street pricing rather than list turns +$1.78B into −$3.76B, and no contract term recovers it.
4. Batch efficiency at the edge. The model credits a home node with data-center-grade utilisation. It will do worse.
5. The transformer. One node per distribution transformer comes from a mean-load headroom calculation, not a thermal model of a real transformer population.

No figure here came from a spreadsheet. Every quantity is computed as an exact rational by agent-calc, then reported with both its exact fraction and a decimal.

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

Three generalisable lessons, each of which reversed a conclusion here:

1. A cost that is "identical either way" still sets the denominator. Accelerators cancel between venues — but different utilisation of that shared cost swamped a 4.8% facility advantage.
2. Compare hardware on delivered work at production batch, never on $/MW and never single-stream. A single-stream proxy flatters whichever part has less memory.
3. A first-principles performance model is a ratio machine, not a magnitude machine. The roofline here ran ~4× hot; one scalar calibrated to observed market pricing fixed the scale and left every ratio intact — then an independent benchmark confirmed it.

A feasibility study, published for scrutiny. Nothing has been built. The numbers are as good as their inputs, which are cited in data/, and the thresholds matter more than the point estimates.

Corrections are welcome and interesting. If a number here is wrong, the model is small enough to show exactly where, and an issue is the fastest way to say so.

MIT — see LICENSE. Use the model, the data and the arguments freely.