HEARTH · Feasibility study · August 2026

America is pouring
concrete for compute.

There are eighty-two million houses already metered.

Data centers are landing in residential neighbourhoods because the compute has to go somewhere. This is the arithmetic for putting it in the houses instead — rebuilt five times in exact arithmetic, with two passes that said no.

0homes in the pilot
0GW of IT capacity
0peak funding, not $4.44B
01 — The problem

The compute has nowhere to plug in.

US data centers drew 176 TWh in 2023. Berkeley Lab now projects 521–843 TWh in 2030, with a 649 TWh reference case — a power-delivery problem as much as a construction problem.

The bottleneck is not just silicon or capital. It is a grid connection. ERCOT reported 410 GW of prospective large loads in April 2026, about 87% associated with data centers. That queue is not a forecast of what will be built; it is evidence of how much demand is converging on the same constrained process.

Meanwhile a megawatt of idle accelerators loses $520,833 every month it spends waiting.

0 GW
Prospective large loads in ERCOT, not a build forecast
00 mo
Reported data-center wait for power
$0
Value burned per MW, per month of waiting
0M
US detached homes, before eligibility screening

Don’t pour the concrete.
Light the windows.

A 5 kW node behind an existing house meter may avoid a campus-scale grid connection. Site screening, panel capacity, permits and utility approval still decide whether it can connect.

02 — The honest part

We tried to kill this five times.

Every figure here comes from exact rational arithmetic — no spreadsheet, no rounded intermediate. That discipline is why the answer flipped twice, and why we trust the one that survived.

Most decks show you only the pass that worked. Here is every pass, including the two where our own model said no.

Pass 1 compared buildings, and a house is genuinely 17.8× cheaper per megawatt than a datacenter shell. True — and the wrong denominator. Pass 2 put the accelerators back and the verdict inverted: they cost $0.9567 per IT-kWh and dominate everything else, so the contest is decided by how many kilowatt-hours each venue spreads the same silicon over, not by what the venue costs.

Pass 3 pulled real GPU quotes and broke the premise that both venues run the same parts. Pass 4 batched the workload the way anyone actually serves traffic, and 192 GB of memory beat 32 GB decisively. Pass 5 asked what changes with model size — and found the answer.

03 — What survived

It is a small-model fleet. That is the whole specification.

A consumer card runs out of memory where a datacenter card runs out of compute. Shrink the model and that asymmetry reverses completely.

Home node · RTX 5090 at list price
Datacenter · GB200 NVL72

At 8B the home fleet costs half what a Blackwell datacenter costs per million tokens — with no credit for the interconnection queue at all. At 32B the memory ceiling bites and we would need a twelve-month queue to catch up. At 70B the weights alone exceed 32 GB, so that case is excluded by this single-GPU model; multi-GPU serving was not modeled.

Small models are not a niche. Routing, classification, extraction, drafting, speculative decoding and every cheap tier all live there.

04 — The mesh

Not a bigger computer. A wider one.

Four million houses cannot be pooled to run one enormous model — the physics forbids it. What they can do is hold every small model in the world warm at once, which is precisely what a datacenter is worst at.

0 PB
Aggregate warm memory across the fleet
0 GW
Grid connection needed to match that in racks
0M
Warm model slots fleetwide
0
Tokens a datacenter forfeits on every weight swap

Matching 1,024 PB of warm weights takes 76,417 GB200 racks and a 9.17 GW interconnection. The memory is distributed across operator-purchased nodes rather than free. A million-model catalogue would get 200 warm replicas of every model.

At 16.71 weight swaps an hour, a datacenter’s cost per token climbs to meet ours on cold starts alone — before any of the capital argument. Serve a long tail of fine-tunes and it swaps constantly. A stable allocation can avoid many of those swaps.

One rule makes it work: no durable session affinity. The serving node still holds an in-flight request’s KV cache; if that house disappears, the generation is lost and must restart elsewhere. New requests can route around it at an assumed 1.031× spare-capacity cost.

05 — The household

What a family actually gets.

The proposed contract asks for no household purchase, loan or roof work. A dedicated 240-volt circuit and recurring payment are design assumptions still awaiting field validation.

Cost to the homeowner$0
Modelled install time~5 hours
Space requiredGarage wall
Winter heat creditConditional
Modelled annual payment$2,893

The heat is not a side effect. It is part of the offer.

Every watt of compute becomes a watt of heat. In a cold-climate, electric-resistance home that heat displaces electricity bought at full retail — worth $1,963 a year more than the same node in a gas-heated house.

The payment is a share, not a promise.

Households take 15% of what their own node earns, with a floor beneath it. A fixed payment would convert a market risk into a solvency risk for everyone; a share keeps the household and the operator on the same side of the trade.

It moves with the house.

The model assumes the agreement conveys to the buyer. That clause lifts average tenure from 8.3 to 15.4 years, and is worth $169B across a full fleet.

Three HEARTH concept appliances: a wall unit, a floor-standing thermal tower and a duct-integrated mechanical-room unit.
W1 · T1 · D1Concept status
05A — Prototype studies

Three ways to live with five kilowatts.

Wall, tower and duct-integrated form studies make the household appliance concrete enough to inspect — and to criticize.

These are appearance and installation concepts, not engineered products. Dimensions, acoustics, thermal performance, safety and utility approval remain validation work.

Explore the form factors
06 — The savings

Same tokens. A quarter of the capital.

Sized to produce identical output, a distributed fleet costs a fraction of a purpose-built campus — because the buildings, the land, the cooling and the grid site and meter already exist, while upgrades and utility work remain deployment costs.

HEARTH fleet · silicon, install, acquisition$80.65B
Datacenter · 5,959,111 GPUs across 9.93 GW$371.12B
0×
Capital advantage at equal token output
0%
Of programme capital that is silicon
0%
Of programme capital that is customer acquisition
$0B
Datacenter construction never built

That last pair is the counterintuitive one. Everyone assumes a four-million-home consumer programme lives or dies on marketing. Acquisition is 5.7% of the capital. This is a procurement business wearing a consumer business’s clothes.

07 — The pilot

220,519 homes, behind four gates.

Each gate buys the answer to exactly one question. If an answer comes back wrong, the programme stops there and the capital behind it was never committed.

The cash curve

Peak funding is $1.51B in month 59 — not the $4.44B of gross capital, because earlier cohorts fund later stages. Free cash turns positive in month 60 and the programme is cash-whole in month 74.

Cumulative cash Peak funding · month 59 Cash-whole · month 74
$0B
Net present value at 8%
$0B
Net present value at 20%
$0B
Cumulative cash at month 84
$0
Break-even token price, per million tokens

Eighteen percent.

That is all the headroom between the price we assume and the price at which this stops working. Everything else in the programme has slack. This does not — so it is the first thing we test, not the third.

08 — What would kill it

Named, priced and sequenced.

A pitch that hides its failure modes is just a slower way of finding them. Here are ours, in the order they should be tested.

    The last one is a flaw we found in our own plan. Staged by what is natural to build, $822M gets 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, and it can be: resell rented datacenter capacity at the target price before a single node is installed.

    09 — The ask

    Prove the price. Then light the windows.

    Nothing here needs four million homes to be true. It needs 250 homes, one signed offtake contract and one silicon supply agreement — in that order.

    $5M
    Gate 0. 250 homes in one metro, proving install, duty cycle and transformer tolerance inside nine months.
    ≥$0.026
    Per million tokens. Signed offtake before Gate 1 clears, tested on rented capacity at near-zero capital.
    List price
    A silicon supply agreement at or near list, before Gate 2. Street pricing alone turns +$1.78B into −$3.76B.
    $1.51B
    Total external capital to reach 220,519 homes, peaking in month 59 and repaid by month 74.

    The buildings are already there. The wiring is already there. The grid connection — the thing everyone else is waiting years for — is already there, behind eighty-two million meters.

    We are not asking to build a data center. We are asking to turn the lights on.

    10 — Questions

    Frequently asked questions.

    The ten questions this study gets asked most, answered from the model rather than from intuition.

    What is HEARTH?
    HEARTH is a feasibility study for putting AI inference into ordinary houses instead of building new hyperscale data centers. A 5 kW compute node sits behind an existing residential meter, the size of an EV charger, and the household is paid for hosting it. Every figure in the study is computed in exact rational arithmetic rather than floating point.
    Why not just build more data centers?
    Because power delivery is increasingly a schedule constraint. Berkeley Lab projects US data-center electricity use at 521 to 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 forecast of what will be built, but it shows the pressure on the connection process. A qualifying house already has a meter and service, though a HEARTH installation would still require site screening and utility approval.
    How much does a homeowner earn?
    About $2,893 a year in the modelled base case—not an offered or guaranteed payment—paid as 15% of what their own node earns with a floor beneath it. A revenue share rather than a fixed payment keeps the household and the operator on the same side of the trade if token prices fall. The proposed contract requires no household purchase or loan and assumes a dedicated 240-volt circuit; installation and contract terms still require field validation.
    Can a home node run large models like a 70B?
    Not under the study's single-GPU serving assumption. At 8B parameters a home node costs about half what a Blackwell data center costs per million tokens. At 32B the 32 GB memory ceiling caps the batch and the home loses. At 70B the weights alone exceed 32 GB, so that case is excluded from this single-GPU model; multi-GPU serving was not modeled. HEARTH is a small-model inference fleet, and that is the specification rather than a hedge.
    Can it train AI models?
    No. Gradient all-reduce for a 7B model moves about 28 GB per step, which a residential uplink would take hours to carry. Splitting one model across houses fails too: a four-stage pipeline yields 5.67 to 11.74 tokens per second because of three network hops per token. The fleet serves inference only, with weights pinned in place.
    How is this different from crypto mining or volunteer computing?
    Three ways. It is paid work with contracted commercial demand rather than a lottery or volunteering. It is capped at one node per distribution transformer, because a 24/7 load has no diversity and residential distribution is engineered around diversity. And the waste heat is part of the offer: in a cold-climate electric-resistance home it displaces electricity bought at full retail, worth about $1,963 a year.
    What does it do to the household electricity bill?
    The operator pays for the power the node draws, and in a resistance-heated home the node's waste heat displaces heating the household would otherwise buy. Nodes are curtailable and coordinated per transformer so they back off during local peaks. The study's tie point is a residential retail rate of $0.2309 per kWh; roughly forty states sit below it.
    What can four million homes do together that a data center cannot?
    Hold 1,024 petabytes of model weights warm at once. Matching that in racks would take 76,417 GB200 NVL72 racks and a 9.17 GW grid connection. It means a million-model catalogue can keep 200 warm replicas of every model in the modeled allocation. That can reduce weight swaps for the served catalogue, though node startup, failover and reassignment can still create cold starts.
    What is the biggest risk?
    Consumer GPU licensing. The entire cost advantage rests on consumer parts being 2.1 to 9.5 times better per dollar of memory bandwidth than data center parts, and that gap is vendor price segmentation rather than physics. The second risk is the token price: break-even is $0.0246 per million tokens against an assumed $0.030, which is only eighteen percent of headroom.
    How much capital does the pilot need?
    Peak funding is $1.51 billion in month 59, not the $4.44 billion of gross capital, because earlier cohorts fund later stages. The pilot reaches 220,519 homes across four gates over 66 months, turns free-cash positive in month 60 and is cash-whole in month 74.