US datacenters drew 176 TWh in 2023 — 4.4% of national electricity. Berkeley Lab projects 325–580 TWh by 2028, a gap of 149 to 404 TWh per year, much of it arriving as new concrete beside housing.
Meanwhile: 82.5 million single-family detached homes, about 60% passed by fiber, each idling at roughly 1.2 kW against a 24–48 kW service, and 46 GW of rooftop solar whose exports are increasingly paid at avoided cost instead of retail. Can the second pool absorb the first?
The model was run five times against quoted hardware prices. The first four answers were wrong, in ways worth showing rather than hiding — each correction is the actual finding.
A $10.7M/MW facility against a $3,000 circuit and enclosure.
Same accelerators, spread over 70% home duty vs 90% datacenter utilization.
Real prices — and the two venues do not run the same silicon at all.
192 GB of HBM buys concurrency 32 GB cannot hold. 32B model.
The home part is memory-limited where datacenter parts are compute-limited.
Pass 1 was seductive and wrong
A datacenter costs $10.7M per MW of critical IT capacity. Amortized over 15 years at 8% across a 90%-utilized megawatt, that is $0.1585 per IT-kWh of capital alone — more than the electricity it buys. A house's equivalent shell is a 240 V circuit, an enclosure and a PDU: about $3,000 for a 5 kW node, or $600,000 per MW, 17.8× cheaper. The number is correct. The denominator was not.
Pass 2 found the denominator
Accelerators at an assumed $25M/MW cost $0.9567 per IT-kWh and dominate everything else by an order of magnitude — and they were assumed identical in both venues, which makes the comparison turn on how many kilowatt-hours each venue spreads that silicon over. The home's shell advantage is a 4.8% sliver of $1.2267/kWh delivered, and a 22% duty-cycle penalty swallows it. This produced the 0.8433 break-even duty cycle that still governs the physical design.
Pass 3 broke the "identical silicon" premise
Quoted prices show the two venues would never run the same parts. The gap is not small.
| Accelerator anchor | Quoted | Per MW-IT | GB/s per $ | GB/s per W |
|---|
The assumed $25M/MW was about 7% low against real datacenter quotes. But consumer silicon is 3.6–10× cheaper per MW and 2.1–9.5× better per dollar of memory bandwidth, while sitting roughly level with Hopper per watt and 1.67× behind Blackwell. On a single-stream decode the home won by 2.4–4.6×.
That proxy is the home's best case, and it is not how anyone serves traffic.
Pass 4 batched, and the verdict flipped
Decode reads the full weights once per step and emits one token per sequence in flight, so throughput scales with batch until compute saturates — and batch is capped by KV cache, which is capped by GPU memory. This is where 192 GB of HBM stops being a spec-sheet line and becomes the whole argument.
| Part | Max batch | Single stream | Batched | Limited by |
|---|
At a 32B model the consumer card runs out of memory long before it runs out of compute: $0.028222/Mtok at home against $0.024277 in a datacenter — 16% dearer on MSRP hardware, 80% dearer at current street prices.
Pass 5 noticed the limiters differ
Bandwidth at home, compute in the datacenter. That asymmetry is a memory-capacity artifact, and it shrinks with the model.
| Served model | Best home | Best datacenter | Ratio | Home limiter |
|---|
HEARTH is a small-model inference fleet. That is not a hedge, it is the specification. At 8B the home fleet is 2.02× cheaper on MSRP hardware and 1.30× cheaper at inflated street prices, with no credit for the interconnection queue at all. At 32B it needs a 12-month queue to catch up. At 70B the weights exceed 32 GB, so the case is excluded by the study's single-GPU assumption; multi-GPU serving was not modeled.
Exact arithmetic will not narrow an uncertain input, but it does locate a boundary precisely. These three thresholds are the usable output of the study — the point estimates elsewhere are illustration.
Uptime below which idle silicon costs more at home than the whole house saves. The binding constraint on the architecture.
Retail tie point per kWh — $0.2811 in a resistance-heated home. The only input of eleven that flips the sign.
Delay beyond which a node energized today beats a facility energized later. Observed waits run 24 to 72 months.
Each input swept across its plausible range, one at a time, against the shell-level comparison. Bars extend from the low outcome to the high one; the centre line is zero advantage. Exactly one variable crosses it.
The second-largest lever is datacenter construction cost — and it is rising, not falling (6.8%/yr, $7.7M→$10.7M per MW since 2020, against 6.2%/yr on residential rates). Because the faster-growing term sits on the datacenter's side, the margin widens from $0.0590 today to $0.1348 by year 10. Only if datacenter capex flattens in real terms while retail rates keep climbing does the advantage go negative — in year 9.
Every choice below is derived from the 0.8433 duty threshold, not selected for elegance.
Solar is a garnish, not the meal
A solar-following node runs about 1,500 h/yr — a 17% duty — which drives accelerator cost above $5/kWh. Surplus solar therefore cannot be an operating mode; it can only be a price tier inside a node that runs anyway. At 85% duty, surplus hours are 20.1% of runtime, and buying them at 1.8× the NEM 3.0 export credit — leaving the homeowner strictly better off than exporting — blends power from $0.1834 down to $0.1655/IT-kWh. Real, but a rounding error beside duty cycle.
One node per transformer, and the reason is diversity
A 25 kVA transformer serving five homes has about 14.09 kW of continuous headroom after their mean draw — nominally room for 2.8 nodes. That arithmetic is a trap. Residential distribution is sized on diversity: the assumption that homes peak at different times. A 24/7 compute node has no diversity at all — it is permanently coincident with every other node on the circuit. Hence one node per distribution transformer, which caps the national fleet at 82.5 GW and buys cheap insurance against accelerated transformer aging.
Inference only, with pinned weights
Serving 200 concurrent streams at 30 tokens/s needs 0.19 Mbit/s of upload — token egress is free on any tier. A 40 GB checkpoint over a 300 Mbit/s downlink takes 17.8 minutes, allowing about 8 model swaps per day before loading eats a tenth of the node's life. Training is excluded outright: a 7B model in fp16 moves 28 GB per step in gradient all-reduce, which a residential uplink would need hours to carry. This is a fleet for latency-tolerant, interruptible, weight-stable work — not a substitute for a training cluster.
The thermal credit is narrower than it sounds
One kWh of compute heat displaces one kWh of resistance heat at full retail ($0.1747), but only 1/COP of a heat pump's electricity ($0.0624), and gas at the burner tip ($0.0482). Netted against a cooling-season penalty of $0.0499, the annual credit is $0.0562/IT-kWh in an electric-resistance home, $0.0094 with a heat pump, and $0.0035 — essentially nil — in the gas-heated majority. A cold-climate, resistance-heat phenomenon, not a general one.
Against August 2026 state rates, the $0.2309 tie point is a clean screen.
| State | Residential rate | Margin to tie point | Grid-power siting |
|---|
The complementarity is worth noticing: the states that fail this screen are the same states running punitive solar export tariffs — precisely where the surplus-solar price tier has the most to offer.
Restated per million tokens at production batch, an interconnection wait is worth $0.000331/Mtok per month. Against the 32B gap of $0.003945 that is a 11.9-month break-even on MSRP hardware — and 59.0 months at August-2026 street prices, which only clears at the extreme end of the observed range.
At 8B none of this is needed: the home fleet is already ahead, and every month of delay is additional margin.
Power delivery can become the datacenter schedule. In April 2026 ERCOT reported 410 GW of prospective large loads, about 87% associated with datacenters. That queue is not a build forecast, but it shows how much demand is converging on the same process.
Accelerators do not wait. On a four-year life, 1 MW of idle silicon burns $520,833 every month — $0.0165/IT-kWh per month of delay. A HEARTH node behind an existing meter may avoid a campus-scale connection, but site screening, panel capacity, permits and utility approval still decide whether it can connect.
| Datacenter wait | Effective DC cost | HEARTH advantage | Homeowner payment |
|---|
HEARTH sells cheap small-model inference. The interconnection queue is what extends that from small models to mid-size ones. If interconnection reform succeeds, the 32B case weakens and the 8B case does not move at all.
Solving the allocation exactly — minimize delivered cost to absorb the 149 TWh/yr low-growth scenario, subject to per-archetype eligibility at 20% opt-in and one node per transformer:
The binding constraint is the cold-climate resistance-heat pool, which saturates at 27.3 TWh; generic homes supply the remaining 121.7 TWh with 64% of their own pool still unused. Annual saving against building the same capacity and waiting 48 months for it: $123.4B/yr.
The high scenario is harder. Absorbing all 404 TWh would need 26.6% of fiber-passed single-family homes — 13.2 million households. Neither the transformer ceiling nor the eligible pools bind there; adoption does, and 27% of American homes opting into anything is not a plan. HEARTH covers the low scenario comfortably and takes roughly 75% of the high one before purpose-built capacity has to make up the balance.
Four million nodes is not four million computers if they are scheduled as one fabric. But the obvious hope — pool them to run models no single node can hold — is the one thing that does not work. What the fleet has instead is width.
Matching 1,024 PB of warm weights would take 76,417 GB200 NVL72 racks at 13.4 TB each — and a 9.17 GW interconnection. That is what lets the fleet keep an enormous catalogue resident and reduce weight swaps for served models: 50 distinct 8B checkpoints warm per house, and 200 warm replicas of every model in a million-model catalogue.
Quantified: a 5 GB checkpoint takes 2.5 s to reach VRAM, during which a GB200 forgoes 57,447 tokens of output. At 16.71 weight swaps per hour per GPU, a datacenter's cost per token rises to meet the mesh's — on cold starts alone, before any of the capital argument. Serve a long tail of fine-tunes and the datacenter swaps constantly. A stable mesh allocation can avoid many of those swaps.
No durable session affinity is the load-bearing invariant
One node still holds an in-flight request's KV cache; if that house disappears, the generation is lost. New requests can route around the dead house at an assumed 1.031× spare-capacity cost. Durable sessions would require a different replication and recovery model.
What the mesh still cannot do
Split a model across houses. A four-stage pipeline moves only 0.42 MB of activations per step — but three network hops per token at residential latency yields 5.67 tok/s on 100 Mbit upload and 11.74 tok/s on gigabit. Unusable for anything interactive. Train. Gradient all-reduce for a 7B model is 28 GB per step. Still disqualifying.
Wide, not deep: a long-tail model-serving fabric holding millions of small models warm, everywhere, at once. Not a bigger computer — a wider one.
An install is an EV-charger job and prices like one. Acquisition is the number everyone fears, and it is the wrong thing to fear.
| Cost line | Per home | Basis |
|---|
That reproduces the $3,000 assumed at the start, from first principles. Acquisition, anchored to instrumented residential-solar channels, runs $450 referral to $4,200 if HEARTH behaved exactly like solar per watt.
Unit economics per node, per year
| Line | Base | Street silicon | Bear |
|---|
A 70-day CAC payback and 43× LTV:CAC describes a business that should spend harder on acquisition. But buying silicon at street prices rather than MSRP is by itself the difference between +$6,000 and −$523 per node-year. Procurement is not a back-office function here; it is the business.
The frontier that actually decides it
Contribution per node per year, after paying the homeowner $3,000. Below $0.020/Mtok there is no amount of demand that makes a node pay.
| Wholesale price | u=0.30 | u=0.50 | u=0.70 | u=0.90 | Break-even util |
|---|
Neither supply nor acquisition. It is the realized price of small-model inference, in a market whose list prices keep falling. At $0.015/Mtok the break-even utilization exceeds 1.0.
Programme capital, against the alternative
| Line | HEARTH @ MSRP | HEARTH @ street | Datacenter, equal output |
|---|
Acquisition is 5.7% of programme capital. Silicon is 79%. The instinct that a four-million-home consumer programme lives or dies on marketing is simply wrong.
Field operations
The labour is not scarce — a five-year buildout needs 4,160 electricians, 1.48% of the 280,000-strong US solar workforce. The rate is the problem: out-installing the entire US residential solar industry by 60%, every year. The answer is to contract existing electrical firms rather than employ a field force — 2,080 crews across a few hundred of the ~70,000 US electrical contractors is four crews each. And it never stops: 12%/yr churn means 480,000 replacement installs per year forever, plus $216M/yr of uninstall cost.
| Buildout | Installs/yr | Crews | Electricians | vs US residential solar |
|---|
If the agreement conveys to the buyer with the house, a move stops being churn. Tenure rises from 8.33 to 15.38 years, LTV from $50,003 to $92,314, and the clause is worth $169.24B across 4M homes — twenty-nine times the entire acquisition budget. It costs nothing to write, and retrofitting it into signed contracts is impossible.
Do not build to four million
Global AI inference runs near 370 trillion tokens/day. The 4M-home mesh is 2.72% of all inference — but HEARTH serves only 8B-class work, so against an assumed 30% small-model share it needs 9.06% of the serviceable market. That is a market-share assumption, not an engineering one, and it is the least defensible number in this study.
| Share of serviceable small-model market | Homes | GW | Capital |
|---|
A 220,000-home entry at $4.44B proves or kills every assumption above inside two years. Four million homes is the ceiling, not the plan.
What the numbers say the strategy must be
- Door-to-door is structurally unavailable. One node per transformer means roughly one house in five on a street, and D2D economics depend on converting density by walking a block. The cheapest solar channel is the one HEARTH cannot use — acquisition runs on targeted digital ($1,150) and referral ($450).
- Targeting is unusually precise, which should push CAC to the low end. The qualifying home is knowable from public data before anyone is contacted: state rate below $0.2309/kWh, electric-resistance heat, fiber-passed, detached, one per transformer. Utility tariffs, county assessor records and FCC fiber maps give exactly that list.
- Buy silicon at MSRP or do not start. 79% of programme capital, and the sole difference between +$6,000 and −$523 per node-year.
- Sell the tokens before building the nodes. The frontier is unforgiving below $0.020/Mtok, and no field organisation rescues a price problem.
The transformer rule, not the screen, sets market depth
Screening for cheap power, electric-resistance heat, cold winters, fiber and detached stock points at the Pacific Northwest and the TVA footprint. Washington is nearly purpose-built: 59.68% of homes heat with electricity — largely resistance, a legacy of cheap BPA hydro — at the country's second-cheapest rate, 12.00¢.
| Market | Rate | Qualified | Transformer ceiling | Addressable | Binds on |
|---|
The one-node-per-transformer rule binds in every single market. Washington has 680,352 homes that qualify on heating and fiber and only 228,000 that can be reached — which makes the pilot 33.05% of the entire screened pool, a saturation level at which CAC climbs steeply.
The fix is to stop using the heating screen as a gate. A cold-climate resistance-heated home is worth $1,963/yr more — a powerful ranking signal and a poor eligibility test. Tier 2 (any fiber-passed detached home in a qualifying state, still one per transformer) is 9,901,200 homes, of which the pilot is 2.2%.
Four gates
| Gate | Months | Homes | Peak installs/mo | Crews | Staff | Capital through gate |
|---|
Node capex at install is $20,164; replacement on churn is $6,571 because hardware is recovered. The pilot requires 1,764,152 accelerators — not a retail purchase, and a buy at that scale is precisely what makes the GeForce licensing question live rather than theoretical.
Cash
Peak funding is $1.51B, not the $4.44B of gross capex — earlier cohorts fund later stages, and that difference is the single most useful fact for structuring the raise. First FCF-positive month 60.
IRR is undefined here and should not be quoted: the 48-month hardware refresh drives the series negative again after its first positive stretch, so it changes sign more than once. NPV at a stated discount rate is the meaningful figure.
The number the whole pilot turns on
Solving for the token price at which cumulative cash reaches exactly zero at month 84: $0.02561/Mtok on a fixed homeowner payment, $0.02459 on a revenue share. Against the assumed $0.030 that is 18.0% of headroom, and the break-even sits at 37.8% of 2026 list pricing for 8B output. Everything else in this programme has slack. This does not.
Stress: the base case works and every downside does not
| Scenario | Fixed $3,000 | 15% revenue share | Homeowner receives |
|---|
A fixed payment converts a price risk into a solvency risk — it does not shrink when the token price does. A 15% revenue share with a $600 floor recovers $0.6–0.7B across the price and demand scenarios and changes the modeled homeowner payment only slightly in the base case ($2,893 vs $3,000). But on street-price silicon it recovers only $0.09B: nothing in the contract hedges procurement. Only a supply agreement does.
The flaw in this staging, stated plainly
| Capital at risk | Question it answers |
|---|
$822M is committed before the question that actually kills the programme gets answered. The demand test is the cheapest one to run and it is sequenced third. It should be first: contract offtake, or resell rented datacenter capacity at the target price, and learn whether $0.030/Mtok clears — at essentially zero capital, before a single node is installed.
Pilot decision rules
- Do not pass G1 without signed offtake at ≥$0.026/Mtok. That is the programme break-even; below it, every later gate is a way of losing money faster.
- Do not pass G2 without a silicon supply agreement at or near MSRP. Street pricing alone turns +$1.78B into −$3.76B and no contract term recovers it.
- Write the conveyance clause and the revenue share into the first 250 contracts. Both are free at G0 and impossible to retrofit at G3.
- Rank on thermal value, do not gate on it — otherwise the pilot saturates a 667,200-home pool it needs a third of, instead of a 9.9M-home pool it needs 2.2% of.
Falsifiable, in the order they should be tested. The first two are now the whole ballgame.
- Duty cycle below 84%. Everything else is noise beside this. Any design that curtails frequently, follows solar, or tolerates residential downtime fails on contact.
- Interconnection reform. If datacenter waits fall below roughly 13 months, the surplus collapses to the 2% band and there is nothing left to pay homeowners with.
- Consumer-silicon licensing — now the largest risk. With quoted prices in, the whole result rests on consumer parts being 2.1–9.5× better per dollar of bandwidth. That gap is vendor price segmentation, not physics, and NVIDIA's GeForce EULA restricts datacenter deployment precisely to protect it. Every favourable number in passes 3 and 5 is legally contingent.
- GPU acquisition price. The RTX 5090 trades at $4,700 street against a $1,999 MSRP — a 2.35× premium on GDDR7 costs. At MSRP the 32B break-even is 11.9 months; at street price, 59.0. Viability is a procurement question as much as an engineering one.
- Batch efficiency at the edge. The batched figures credit a home node with datacenter MFU and occupancy. A node holding 28–108 sequences against residential arrival variance, with no cross-node batching, will do worse. The likeliest place these numbers erode.
- Transformer thermal aging. The one-node-per-transformer rule comes from a mean-load headroom calculation, not a thermal model of a real transformer population. It needs a utility partner and live telemetry before any dense deployment.
- Rate inflation outrunning capex inflation. Negative by year 9 in the adverse case.
Propagating all eleven uncertain inputs through the shell-level comparison as intervals gives an advantage spanning −$0.2305 to +$0.3965/IT-kWh. That straddles zero, and it should — it is the honest joint bound, pairing the cheapest conceivable datacenter with the most expensive conceivable home. Restricting the residential rate to the cheapest ~25 states narrows it to −$0.0655 to +$0.4271. Still straddling.
The intervals are wide because the inputs genuinely are, and no amount of exact arithmetic narrows them. What the exact arithmetic delivers instead is the boundary: duty > 0.8433, retail < $0.2309/kWh, wait > 13 months. Those are the decision rules. Everything else on this page is illustration.