The House That Pays Its Own Bills: GPU Boilers, Closed Loops, and the Decentralized Home

The House That Pays Its Own Bills: GPU Boilers, Closed Loops, and the Decentralized Home

There is a boiler in Belgium that heats a family's water by running other people's computations. Cold water flows in, 65°C water flows out, and the "fuel" is a rack of servers whose heat — roughly 95% of every watt they draw — is captured instead of being vented into the sky. This is not a concept render. Qarnot has been shipping compute-heaters for years, and its boiler-integrated modules already run in Belgian houses, a hospital, and a municipal swimming pool, replacing part of a gas heating system.

This post is about what happens when you take that one device seriously and follow the logic to its end: a house that generates its own electricity, heats itself with computation, drinks from its own roof and its own air, grows food in its basement, and sells its surplus — kilowatt-hours and inference alike — to decentralized networks. In 2026, every component in that sentence exists as a shipping product. What has changed is not any single technology. It is that they have quietly converged, and the sum flips the sign on the oldest line in a family's balance sheet: the house stops being a box that consumes utilities and taxes, and becomes a node that earns.

The reason these components compose into something larger than a gadget list is, I will argue, an accounting question: where you draw the boundary decides what counts as waste. That is an ODTOE question, and we will get there. Engineering first.

One watt, two jobs

Start with the physics, because it is embarrassingly simple. Essentially 100% of the electricity a server consumes exits as heat. A GPU does not "use up" energy computing; it converts electrical energy into thermal energy while, incidentally, arranging some bits. A datacenter therefore pays twice: once for the watts, and again for the chillers, cooling towers, and water needed to throw the resulting heat away. A home, meanwhile, pays for heat directly — in most cold climates it is the single largest utility line.

Co-locate compute and dwelling, and one watt does two jobs. The same kilowatt-hour that a datacenter buys as an input and then pays to reject becomes, in a house, both a unit of sellable computation and a unit of needed heating. Nothing is invented; a redundancy is deleted.

Companies already run this in production. Qarnot's Q.rad puts servers inside home radiators; its QBx module is a compute boiler — cold water in, 65°C out, about 95% of server heat recovered. Heata in the UK mounts server-boilers directly on domestic hot-water cylinders. Deep Green heats British swimming pools with immersion-cooled compute. These are operating businesses with customers, not white papers.

Now the maximal home version — the one worth designing around in 2026. A liquid-cooled 8-GPU node: eight H200 SXM cards at up to 700 W apiece is 5.6 kW for the silicon alone, and the whole server — CPUs, memory, fans, power conversion — lands around 7–10.2 kW under sustained load (a DGX-class box peaks near the top of that range). Wire the panel for 10–12 kW so there is headroom. Plumb it into a heat exchanger and you have a boiler that thinks. Run it at 10 kW around the clock and it draws 7,200 kWh a month — and, because essentially every one of those watt-hours exits as heat, it is also a 10 kW heater that never sleeps: roughly 240 kWh of heat a day, 7,200 kWh a month.

The systemic consequence deserves stating plainly. If compute migrates into buildings that need the heat, the pressure to build ever more centralized datacenters — with their heat-rejection infrastructure, their water draw, their strain on transmission grids — shrinks. And reliability moves in the same direction: a distributed fleet of ten thousand home nodes has no single point of failure, no one cooling plant whose outage takes down the whole region's inference capacity.

Ten kilowatts against a floor plan

Heating engineers size homes at roughly 30–150 W of heat per square meter, depending on climate and insulation. For a 100 m² house that works out to:

Put the 8-GPU node's 8–10 kW against those numbers and the match is genuine, not rhetorical: it can carry 100–150 m² in a mild climate, 70–100 m² in a cold one, and 150–250 m² of well-insulated southern house.

One honest asymmetry has to be stated before it flatters anyone into a purchase. A GPU is, thermodynamically, a resistive heater that happens to think: 1 kWh of electricity in, 1 kWh of heat out. A heat pump delivers 2.5–5 kWh of heat per kWh of electricity. On heat alone, the GPU boiler loses to a heat pump, and not narrowly. It wins only because the same kilowatt-hour also earns compute revenue on its way to becoming heat. If the node is not earning, you own the world's most expensive space heater.

Then summer. The tempting move — converting the excess heat back into electricity — is a thermodynamic dead end and should be named as one: server coolant runs at 40–70°C, and 10 kW of heat at those temperatures yields perhaps 100–500 W of electricity through the best available machinery. Spend a fortune, recover a nightlight. The engineering answer is a priority ladder, not a generator: first, domestic hot water — 10 kW raises about 200 liters by 40°C in roughly an hour, so showers and laundry run essentially free; second, the pool or hot tub, a thermal buffer of glorious capacity; third, drying and technical heat; and whatever remains goes out through an outdoor dry cooler, which costs little and complains less. Seasonal storage — banking summer heat in the ground for winter — is theoretically pretty, but a node emitting 7,200 kWh of heat a month is not a small-tank problem. Winter: space heating. Summer: water, pool, dump the rest. The GPU never stops earning; only the destination of its heat changes with the season — and a few neighbors on a mini district-heating loop widen the sink in both.

The energy stack

A GPU boiler needs watts, so the next layer down is generation. The 2026 menu for a single plot is broader than most people assume:

Solar PV on roof or ground remains the workhorse. Micro wind earns its place on open, gusty sites. Micro-hydro, where a stream crosses the property, is the quiet aristocrat — small, steady, day and night. Ground-source heat-pump loops do not generate electricity but multiply it: each kWh driving the compressor moves 3–5 kWh of heat into or out of the ground, which makes geothermal the highest-leverage line in the whole stack for heating and cooling.

And then the unglamorous one: biogas. A household septic tank upgraded to an anaerobic digester turns organic waste — kitchen scraps, garden matter, the septic stream itself — into methane for cooking and backup heating. Nobody photographs it for architecture magazines, but it is the most literally closed of all the loops: the house's own biological output returns as fuel.

Storage is where 2026 differs most from even three years ago. LiFePO4 wall batteries handle daily cycling — safe chemistry, thousands of cycles, boring in the best sense. The headline, though, is that vehicle-to-home bidirectional charging is now real and shipping: most systems deliver 9.6–11.5 kW of continuous power, enough to run a whole house including heavy loads. A roughly 100 kWh EV can carry an average home for two to three days; a Ford F-150 Lightning with its 131 kWh pack manages about three days of normal use and up to ten days rationed. The family car quietly became the family's largest battery — one that also drives to the shops.

Surplus closes the layer. Net metering and feed-in tariffs let the house export what it does not use. But notice what the GPU node adds here: it is a dispatchable load. A midday solar surplus that would otherwise be exported at a mediocre tariff can instead be converted, on the spot, into compute revenue and stored heat. The house gains a choice of what to turn sunlight into.

Water down, food up

Water follows a three-tier logic. The well remains the classic bulk source where geology allows. Rainwater capture handles irrigation, the grow room, and greywater duties. The genuinely new tier is atmospheric water: SOURCE Hydropanels are solar-powered panels whose hygroscopic material extracts drinking water from air down to about 10% humidity — roughly 3–5 liters per panel per day, at $2,500–3,000 installed, with a January 2026 expansion into 15 new countries under a "Community Water Farm" model. The honest framing: hydropanels will not run your shower. Well and rain do bulk; panels are potable-water insurance — a drinking supply that keeps working when the pump fails or the aquifer disappoints. Resilience, not volume.

Food moves in the opposite direction — downward, into the basement, which turns out to be ideal farmland: stable temperature year-round, and the absence of windows is irrelevant when spectrum-adjustable LEDs light the crops anyway. The 2026 generation of stackable automated hydroponic units is consumer-real: about 90% less water than soil agriculture, LED spectra tuned per crop cutting lighting energy 40–60%, AI-controlled nutrient dosing and climate with failure alerts to your phone. Robotics increasingly handles the seeding and harvest cycles; 3D food printing is emerging at the frontier for texture and protein shaping.

And here the loops interlock with a satisfying click. GPU waste heat keeps the grow room at growing temperature through winter. The grow lights run on the midday solar surplus that had nowhere better to go. Plant waste feeds the biogas digester, whose gas backs up the kitchen. Each subsystem's "waste" is the next subsystem's input — hold that thought, because it is the entire philosophical payload of this post.

Selling thought

What actually runs on the node? Two workloads, and the order matters.

First, the family's own AI. A local model on local hardware means private inference at home: the household's conversations, documents, camera feeds, and medical questions never leave the building. The same node runs the house's nervous system — a local model dispatching energy (when to charge the EV, when to export, when to compute harder because the pool is cooling), tending the hydroponics, watching security. The house acquires an on-premises brain that answers to no cloud.

Second, idle capacity gets sold. This is the piece that was missing until very recently, and 2025–2026 supplied it twice over. Cocoon — Pavel Durov's Confidential Compute Open Network on TON, announced at Blockchain Life 2025 and live since November 30, 2025 — already has GPU owners earning TON for serving private AI inference, with encryption that keeps jobs confidential even from the GPU owner. Gonka AI takes a different route to the same destination: a purpose-built blockchain where consensus and rewards are tied to verifiable AI computation — its proof of work is useful inference and training rather than burned hashes. Two live networks, two trust models, one consequence: a home GPU now has a market to sell into without asking anyone's permission.

Then the final layer, the one that turns houses into infrastructure: build homes as constructor kits with fiber interconnects laid in from day one. A neighborhood mesh where energy (microgrid), heat (district loop), water, and compute (distributed cluster) are four flows over one topology. A street of such homes is a datacenter, a power plant, and a farm — with residents.

The hardware ladder

What does the node itself cost? An 8×H200 server realistically budgets at $350k–500k, with DGX-class builds running to $600k and beyond. The generation ladder above it climbs fast: B200 (about 1 kW per GPU; 8-GPU systems around $380k–550k), then B300 / Blackwell Ultra at roughly 1.4 kW per card, and at the top the rack-scale GB200 NVL72 at $3M+ — a datacenter row, not a basement. AMD's Instinct MI325X and MI355X compete across the same span. But the ladder also has a pragmatic lower rung: 8× RTX PRO 6000 Blackwell Server Edition at roughly $80k–180k — the same cogeneration logic at about a fifth of the capital cost, and for inference workloads that is often exactly enough. And the concept scales all the way down: Heata-style 1–2 kW hot-water nodes already exist at appliance prices. The boundary-redrawing works at any wattage; only the ledger changes size.

The honest ledger

Here is where flagship posts usually reach for a spreadsheet with suspiciously confident numbers. The numbers now exist, so let us reach for it — with the confidence intervals left showing.

Revenue is one formula: 8 GPUs × price per GPU-hour × 24 × 30 × utilization. On today's decentralized and rental markets that spans an uncomfortably wide range:

Against it: electricity, 7,200 kWh a month — $720 at $0.10/kWh, $2,160 at $0.30 — plus cooling, connectivity, network fees, idle time, and the depreciation of hardware that ages fast. Payback on a $350k–500k box runs from about 18 months in the excellent case to 70+ months at weak utilization — at which point the silicon risks being obsolete before it is paid off.

A worked local example, because abstractions hide tariffs. In Kazan, Tatarstan, private-home electricity from January 1, 2026 costs 4.26–10.88 ₽/kWh depending on the home's category and monthly volume tier; the node's 7,200 kWh month comes to roughly 36,000–50,000 ₽ (about $450–630). Against even the weak-scenario gross of ~$2,600 that is a workable spread — if the utilization materializes.

That "if" is the honest center of the whole ledger. Utilization on young DePIN networks — Cocoon, Gonka, Akash, io.net, Render — is not guaranteed, and the research behind this post reaches a conclusion worth quoting in spirit: do not buy an 8×H200 box on DePIN hopes alone. Secure a compute sales channel first — your own workloads, B2B rental contracts — and treat DePIN as supplemental load that fills the gaps. An option, not a salary.

The wider household ledger keeps its structure:

Directionally: the heating bill gets offset, compute becomes income, energy surplus becomes income — and the house can plausibly cover its own running costs and taxes. That is the asset flip. Plausibly, not certainly: the word is chosen with care.

And the whole economics compresses into a three-conditions rule. The node makes sense only when all three hold at once: cheap and stable electricity (ideally self-generated), guaranteed or near-guaranteed GPU utilization, and heat that gets usefully absorbed at least part of the year. The rational stack is a compute business with heat as a valuable byproduct and DePIN as supplemental utilization — never the reverse. Hold that shape; the next section shows it is not just prudence but arithmetic.

The ODTOE reading

Now the frame, because without it this is just a well-researched list of gadgets, and with it the list becomes a single move performed six times.

ODTOE's central claim is that what actualizes for an observer is R = Ô(Ψ) — the result depends on the observation operator, and the operator includes where the observer draws the boundary of the configuration. Read "waste heat" through that lens and it dissolves: waste is not a property of the joules. It is a property of the boundary. A datacenter's heat is waste only because the configuration was drawn to exclude the buildings next door. Redraw the boundary — put the compute inside the dwelling — and the identical joules become the product. Nothing physical changed; the observer's operator did. The autonomous home, seen this way, is boundary-redrawing as engineering practice: heat, greywater, organic waste, idle GPU cycles, midday solar surplus — every stream that some narrower configuration labeled "waste" gets re-included in a loop where it is an input.

Second move: the home becomes a low-dimensional observer in its own right. ODTOE parametrizes an observer by (B, A, H) — coherence, attention structure, memory. Fit a house with sensors and a local AI and it becomes a self-observing configuration maintaining its own coherence: homeostasis of energy, water, heat, and food. Its autonomy is its coherence, and the multiplicative weak-link logic of B = F·E·(1−σ)·Λ applies without mercy: autonomy fails at the weakest subsystem — power or water or heat or connectivity — no matter how strong the others are. That is not poetry; it is the design rule that says every flow needs its own closed loop and its own reserve. And notice what the ledger's three-conditions rule turned out to be: the same algebra. Cheap watts, guaranteed utilization, absorbed heat — three multiplicative factors, not three additive bonuses; any one at zero zeroes the venture, exactly as any factor at zero zeroes B. The economics of the node and the coherence of the observer have the same weak-link shape — not a metaphor imported for effect, but one structure surfacing in two ledgers. And Λ, the human factor, here means something concrete: the skill to maintain what you own. A house full of systems nobody can service is a coherence with a hidden hole in it.

Third, decentralization — and this is where the collective-observer machinery earns its keep. ODTOE's collective-observation postulate reads P_coll(E) = 1 − ∏(1 − Bᵢᵏ): a network of observers actualizes an event if any of them does. A network of autonomous homes is disjunctively robust — any healthy node keeps the collective alive. A centralized grid or datacenter is the opposite shape: conjunctively fragile, a ∏ of dependencies where one substation or one cooling failure takes down everything downstream. Decentralization, in this reading, is not an ideology and not an aesthetic. It is arithmetic: moving the collective from ∏-fragility toward 1−∏(1−·) robustness. Cocoon and Gonka are that arithmetic applied to compute; microgrids apply it to energy; the fiber-mesh street applies it to all four flows at once. As the multi-agent coherence work argues from the AI side, and coherence-as-business-variable from the organizational side, the collective's strength is a function of the topology, not just the nodes.

And the asset/liability flip itself is configuration-relative identity. The same house — same walls, same roof — is a liability in one network topology (a terminal consumer node, an endpoint that only draws) and an asset in another (a contributing node that heats, computes, feeds, and exports). ODTOE reads identity as role-in-configuration rather than intrinsic property — the same move the corpus makes when it replaces profession-as-identity with trajectory. Nothing about the house changed. Its role in the configuration did, and in an observer-dependent ontology, that is what identity is.

Honest boundary of the frame, stated as plainly as the token-volatility warning: ODTOE supplies the accounting lens — boundaries, coherence, collective robustness. It does not derive Qarnot's balance sheet, and it does not make hydroponics cheaper. The engineering facts above stand entirely on their own. What the frame explains is why they compose into something bigger than a list — six technologies, one boundary operation — and where the fragilities remain: connectivity as the weak link in the multiplicative chain, token economies as the volatile term, and maintenance skill as the Λ nobody budgets for.

A street with residents

None of this requires a breakthrough. That is the strange, quiet fact of 2026: the house that pays its own bills by thinking is available as a bill of materials — panels, a heat-pump loop, a battery, a bidirectional charger, a liquid-cooled node at whichever rung of the hardware ladder your capital reaches, a rack of grow trays, a digester where the septic tank was, and a fiber run to the neighbors. What it requires is assembly, skill, and a decision about boundaries: whether your house is the end of someone else's network or a node in your own.

The deeper claim — that networks of small coherent observers can out-cohere one large fragile one, and that this is arithmetic rather than sentiment — is developed formally in the collective observer paper at odtoe.org. The street-sized version of that argument is now being built, one boiler at a time.