buildings energy efficiency

All_posts

Cogeneration Server Rooms: No More Wasted Server Heat in Industry and in Our Homes

Data centre heat reuse: cogeneration server rooms for homes and industry

Large data centres have their place: banking, healthcare, public administration and the training of the largest AI models need physically protected, centralised facilities. The problem lies elsewhere: data centre heat reuse, or rather the lack of it. Most of them burn megawatt-hours of electricity on computing, then spend tens of percent more on cooling just to dump the heat into the air. Where evaporative cooling towers run, roughly another litre of water evaporates for every kilowatt-hour (an estimate from Nvidia). Meanwhile homes, apartment blocks, greenhouses, breweries, dairies or other industrial plants and municipal swimming pools pay for gas, electricity or district heat to warm water to exactly the temperatures that servers treat as waste.

The answer is modular, distributed server rooms: instead of one huge hall, thousands of small modules in homes, apartment blocks, plants and pools. Understood this way, a cogeneration server room turns one kilowatt-hour of electricity into two products: computing and heat. Distribution makes it possible to use waste heat widely, right where it is produced, to use the idle connection capacity of existing buildings instead of waiting for new grid connections, and to save the water that some large data centres evaporate in cooling towers.

The point is not to replace big data centres but to complement them where it makes sense: bulk, less sensitive computing can run inside buildings that need heat anyway.

The idea is simple: liquid-cooled compute modules are placed where heat is needed and connected to a heat buffer, a heat pump, energy storage and the building management system (BMS). The cost of cooling IT becomes a source of heating. Below is the concept through the lens of physics, hydraulics and automation, without the marketing gloss.

Overview diagram: liquid-cooled IT module, 25–40°C (up to 50°C) heat buffer, direct use, water-to-water heat pump, space heating and DHW, PV and BESS supply, backup boiler or electric heater feeding the buffer
General layout: the server releases heat into the buffer, and the heat pump raises the temperature while also cooling the processors.
Illustration: a liquid-cooled server module sends its heat through pipes to a greenhouse, a brewery and a swimming pool
Server heat going to a greenhouse, a plant and a pool instead of into the air.

💡 1. What is easier to send: information or heat?

What is easier to move over a distance: information or heat? Information. And not by a little. Over optical fibre, 1,000 km takes about 5 ms. Heat is another league: to carry 1 MW at a 10 K temperature difference you have to pump about 86 m³ of water an hour. That means a DN150 pipe, pumps, insulation and trenches (my own calculation). And a server turns almost all the electricity it draws into low-grade heat: 25–40°C from today’s liquid cooling, around 55°C from the newest servers (Nvidia, 2026). The scale? Data centres used about 415 TWh in 2024, roughly 1.5% of the world’s electricity, and are expected to reach about 945 TWh in 2030 (IEA, Base Case).

🏙️ Route one: a large data centre plus district heating

Heat from a big data centre can be piped into a city, but only where a heat network exists and someone builds the heat pumps. In Odense, Meta’s data centre delivers heat to the network only after heat pumps lift it to 70–75°C; the plant is about 45 MW (Ramboll). In Poznań, Beyond.pl and Veolia are planning about 30 MW (Beyond.pl).

The catch is that heat networks are not everywhere. District heat reaches about 40% of households in Poland, about 22% in Germany, about 2% in the UK and 68% in Denmark (Wikipedia’s overview, data from different years).

🔁 What about warming the return of an existing network?

Where a network exists, a simpler option is tempting: connect the data centre to the return pipe. It can be done, but it is a supporting role: the data centre assists the network, it does not replace its heat source. Figures from the Warsaw network’s control table (type I, from 1 December 2026): supply 68°C in summer (12°C or more outdoors), 85.5°C at 0°C and 117°C at −20°C; return 38°C, 44.5°C and 54.5°C respectively.

  • 🌡️ Temperature. 25–40°C coolant from today’s servers is colder than the return, so it cannot warm it without a heat pump. Around 55°C coolant from the newest servers can warm the return directly, but only by a few to a dozen or so degrees, and mostly outside frosty weather.
  • 🔥 The source stays. The network’s conventional source (or large heat pumps, as in Odense) still has to lift the water to supply temperature.
  • 📉 A warmer return has a cost. It reduces the temperature difference, and with it pipe capacity, and it worsens the operation of the CHP plant and flue-gas heat recovery. Customers are billed on return temperature, so a separate contract is needed: the data centre as a source, not a customer.
  • ☀️ Summer. When server heat is hardest to use, the network needs it least (hot water only).

Low-temperature (fourth-generation) networks fit better, but in Poland they are the exception. Where a network exists, this is a good complement. Where there is none, or the heat user is right next door, it is simpler to deliver the heat straight into the building.

🔀 Turning it around: computing travels to the heat

Since information travels more easily, let us not pipe heat into the city; let us send the computing to where the heat is needed. One kilowatt-hour, two products: computing and heat. Three benefits:

  • 📡 Information instead of heat: optical fibre instead of new pipes.
  • 💧 Efficiency and water: cooling the processor becomes heating the building (section 3).
  • 🧱 Scaling in small steps on existing grid connections (section 8).
Short (55 s): bits or heat?

🛠️ 2. IT reality: what such a server computes and why it can last 7–10 years

  • 🔒 What the module does not do. Critical systems (banks, healthcare, government), latency-sensitive transactions and training of giant language models, which requires thousands of GPUs linked by NVLink or InfiniBand, stay in large data centres.
  • ⚙️ What it does. Most cloud traffic is bulk, asynchronous “grunt work”: inference for everyday AI agents, image and video generation, 3D rendering, media transcoding, batch processing. These jobs do not need ultra-low latency between nodes, just cheap power and effective heat removal.
  • ⏳ Lifetime. Three to four years is mostly corporate depreciation logic: working hardware is replaced to pack more power into the rack. With 100% liquid cooling there are no thermal shocks, and power can be capped. Under such conditions a lifetime of seven to ten years is a reasonable design assumption, and older generations handle smaller models and batch jobs well.
  • 🧭 Computing follows the heat. The fleet sends jobs to wherever heat is currently needed or power is cheap, so seasonality is solved at fleet level, not building level. Rule of thumb: depreciated hardware goes to seasonal heat users (it can stand idle in summer), new hardware to year-round ones.
  • 🔌 Operator model (plug & play). The building owner provides the connection, internet and takes the heat. A remote operator manages the rack, service and replacement of compute cartridges.

🌡️ 3. Physics and hydraulics: heating the building is cooling the processor

Liquid cooling (direct-to-chip or immersion) removes the fans, so the module runs about as quietly as a heat pump. Almost all electricity drawn by the processors turns into heat. Its temperature depends on the hardware generation: today’s liquid cooling delivers 25–40°C, while in the newest servers coolant enters at 45°C and leaves at around 55°C (Nvidia, 2026). This wide range lets most of the heat be used directly, with a heat pump only where a warmer medium is needed. At every scale the core is a heat buffer and a water-to-water heat pump:

  1. 💧 Buffer and direct use. Server coolant releases heat into a 25–40°C buffer (warmer with the newest servers). Without a heat pump, directly through an exchanger, it can feed low-temperature loads: a 28°C pool basin, 30–35°C underfloor heating, greenhouse soil heating or low-temperature panel radiators (these usually need to be oversized because they run at a low temperature). It can also preheat mains water from 10°C to 30–35°C (higher with ~55°C coolant from the newest servers). A backup boiler or electric heater tops up the buffer, not the IT module itself.
  2. 🔥 Raising the temperature. Where a warmer medium is needed (domestic hot water at 55–60°C, older radiators, some process heat) a water-to-water heat pump uses the buffer as a stable, warm source (instead of −10°C winter air), so it does little work. A COP of 5.5–7.0 is realistic under specific conditions: a 30–35°C source and supply at around 55°C; the warmer the buffer and the lower the supply temperature, the higher it goes. With the newest servers, hot water can be made almost directly.
  3. ♻️ Symbiosis. By drawing heat from the buffer, the heat pump also cools the liquid returning to the processors. Heating the building and cooling the server are one thermodynamic process.

🦠 Water hygiene. Preheating drinking water to 30–45°C puts it in the Legionella growth range. So preheating is done instantaneously through an exchanger, the buffer stays on the heating-water side, and the water is finally raised to 55–60°C.

💧 Efficiency and water: what we are really comparing with

In a conventional data centre, cooling is a cost: historically up to 40% of the electricity (Nvidia). The average PUE in the 2025 Uptime Institute survey was 1.54 (as cited by Google), so about a third of the electricity never reaches the processors.

Water: where evaporative cooling towers run, roughly a litre evaporates per kilowatt-hour (Nvidia: about 1.1 l/kWh; the physics checks out, as evaporating a litre removes about 0.7 kWh of heat). To be fair, that is the top of the range. Microsoft reports an average of 0.27–0.30 l/kWh globally and, with air cooling, 0.03 l/kWh in Europe, the Middle East and Africa. Beyond.pl estimates a saving of 76,500 m³ of water a year at 30 MW, about 0.3 l/kWh. In Poland, water is therefore a secondary argument. The main one: the heat goes into the air.

The newest servers (45/55°C) also let large facilities move off water and chillers to dry coolers, but they still throw the heat away. A module in a building sells it. And the objection “it is just an electric heater”? We are not comparing it with a heat pump, but with the same server in a hall where the heat goes into the air while someone next door burns gas.

Short (40 s): cooling that heats

🏠 4. Micro scale: a single-family home, 0.5–5 kW module

Oversizing is a mistake at home, and an easy one to make. A 3–5 kW module running all year delivers 72–120 kWh of heat a day, or 26–44 MWh a year, while a new home needs something like 10–15 MWh. In summer most of it would go to the dry cooler. Honestly, there are two routes, both with a home buffer:

  • 🚿 All year, small module: 0.5–1 kW for hot water, with power modulation, because hot water alone is a few kWh a day. The home then becomes a year-round heat user.
  • ❄️ Seasonal, 3–5 kW module: in winter it can cover the base load of a modern home’s underfloor heating and hot water. This is where depreciated hardware fits, as it can stand idle in summer.
  • ☀️ Summer: hot water for a family of four takes about 5–6 kWh a day. Surplus heat goes to a garden pool or greenhouse, and beyond that the module simply throttles down.
  • 🌬️ Emergency dump: when nothing else can take heat, a simple dry cooler or AC unit rejects it outdoors.
Cross-section of a single-family home with a 3–5 kW module, buffer, underfloor heating, DHW, dry cooler and pool or greenhouse
Home with a 3–5 kW module: full heat use in winter, DHW, pool or greenhouse and a dry cooler in summer.

🏭 5. Macro scale: agriculture and food industry, heat 365 days a year

The sizing rule is the same as in cogeneration: IT power is matched to the year-round base heat demand, not to the peak, and the yardstick is the annual share of heat actually used. Germany already counts it: the Energy Efficiency Act (EnEfG) requires new data centres to reuse at least 10% of their energy if they start operating from July 2026, 15% from July 2027 and 20% from July 2028.

  • 🍅 Greenhouses: 25–40°C heat goes straight into root-zone heating, with no heat pump.
  • 🥛 Dairies, breweries, food processing and other industry: process water (pasteurisation, brewing, CIP cleaning) is heated all year. Modules provide free preheating from 10°C to 30–35°C (higher with ~55°C coolant from the newest servers); heat pumps top up to process temperatures.
  • 🛡️ Hygiene: heat passes through certified plate exchangers in the plant room, so servers are fully separated from food production zones.

🏢 Multi-family buildings: heat for the whole community

An apartment block or housing estate needs heat all year round, because domestic hot water (DHW) and its circulation run in summer too. The modules connect to the substation or the estate boiler room in two ways, often combined:

  • 💧 Directly: the 25–40°C buffer preheats mains water for DHW (from 10°C to 30–35°C, instantaneously through an exchanger; see water hygiene in section 3). This is the cheapest part of the energy and needs no heat pump. The same buffer can feed low-temperature systems such as underfloor heating in newer buildings.
  • 🔥 Boosted by a heat pump: a water-to-water heat pump with the buffer as its source lifts water to 55–60°C for DHW and to the parameters of the heating system (usually after lowering the supply temperature or a thermal upgrade). A cascade of several micro heat pumps adds redundancy.

The existing substation or boiler stays as peak top-up and backup, so residents never notice an outage. Modules placed next to the substation also use the building’s existing electrical connection, and their BMS-controlled power yields to induction hobs and EV chargers in the evening peak.

🏊 6. Swimming pools and water parks: the thermodynamic “holy grail”

Processor coolant (25–40°C) is as warm as pool water (27–32°C) or warmer, and pools need heat all year. Heat from the plant room flows in three streams:

  1. 🌊 Pool basin: the IT buffer heats 28°C pool water through a titanium exchanger. No heat pump, just a small circulation pump.
  2. 🚰 Make-up water: regulations require roughly 30 l of fresh water per visitor per day. 10°C mains water is preheated to 25–30°C for free.
  3. 🚿 Showers (55°C) and hall ventilation (around 32°C): handled by a water-to-water heat pump using the buffer as its source.

Scale: a 100 kW swarm running all year produces 876 MWh of heat. A pool can take it 365 days a year, so the annual share of heat used can exceed 90% here (the German minimum starts at 10%). Whether a particular pool has that much base load has to be calculated from its own data.

Liquid-cooled racks are sealed, so they do not draw in humid, chlorine-laden air, and titanium exchangers protect the system from pool chemistry.

Cross-section of a swimming pool plant room: IT cabinets, 25–40°C buffer, titanium exchanger, fresh water, heat pump, showers and ventilation
Plant room: three heat streams for the basin, fresh water, showers and ventilation.

🔄 7. Deployment: return-pipe integration, swarm architecture, backup

  • 📉 Return-pipe integration: the module heats water returning from the building before it reaches the boiler or substation. Most of the year the old source stays off; in frost it only tops up.
  • ⚠️ Condensing boilers: every degree of warmer return reduces the condensing gain, and condensation only stops above about 55°C (a typical value for gas boilers). That is why the module is best used for hot-water preheating, or the controls let the modules take the whole load.
  • 🐝 Swarm architecture: instead of one big server with one heat pump, a cascade of small modules (e.g. 20 × 5 kW), each with its own micro heat pump. Servicing two of twenty means 10% less output and no single point of failure.
  • 🔥 Backup: if compute jobs stop, BMS-controlled valves route water through the existing boiler, substation or an LPG peak boiler.
Diagram of data centre modules connected into the return pipe of an existing boiler room, with a swarm of 20 modules of 5 kW and bypass valves
Return-pipe integration: a swarm of modules ahead of the existing heat source, with a BMS bypass.

🔋 8. Protecting the grid connection: PV, BESS, Modbus and BACnet

A building’s power profile is uneven: induction hobs and EV chargers at home, machine start-ups in a workshop, pumps and attractions at a pool. A constant server load could trip main fuses or trigger contracted-capacity penalties. So the module works with PV, battery storage (BESS) and HEMS/BMS/EMS over Modbus TCP or BACnet IP. Energy storage is not a decoration for the PV here: its job is to squeeze more out of the capacity the building already has.

🗄️ Three stores in one building

  1. 🔋 The battery (BESS) stores electricity. It shaves building peaks (peak shaving) so grid draw stays below the contracted capacity and the modules do not have to shut down, and it shifts PV surplus to the evening and night. Its second role: backup power for the servers (UPS), which a data centre needs anyway.
  2. 🌡️ The heat buffer stores heat. Computing at noon on PV, hot water in the evening. 1 m³ of water at a 20 K difference holds about 23 kWh.
  3. 📋 The job queue stores computing. A batch job can be paused or shifted, so the module is the most flexible load in the building. The BMS sees it as a controllable load, a heat source and a buffer.

🎛️ Control order: never exceed the contracted capacity → PV goes to the modules first → surplus to the battery and the buffer → at building peaks the battery discharges and the modules throttle down last → with dynamic tariffs, more computing and buffer charging in cheap hours.

The result: more operating hours for the modules on the same connection, higher PV self-consumption and no penalties for exceeding contracted capacity. The servers use only the part of the connection capacity the building does not need at that moment.

⚡ An answer to the shortage of grid connection capacity

Grid operators increasingly refuse new connections or make investors wait for years. A data centre can be built in two to three years, but the energy infrastructure around it takes longer (IEA). Meanwhile, in existing buildings the contracted capacity sits idle for most of the day: homes, schools, plants and pools use their full capacity only in short peaks. Distributed cogeneration server rooms use and optimise this existing capacity instead of requiring new connections and new halls.

  • ☀️ Local PV surplus: solar energy that at sunny noon would be exported for next to nothing or curtailed by the inverter powers computing on site and charges the heat buffer.
  • 📈 Better use of the connection: a BMS-controlled module fills consumption valleys, and during building peaks it throttles down or the battery takes over.
  • 🧱 Small steps: 5 kW, 50 kW, 250 kW. Each step has its own heat user from day one. No new hall, no new power line.
Short (35 s): small steps on existing connections
Electrical supply diagram: PV, BESS energy storage, DSO grid, servers and BMS/EMS over Modbus TCP and BACnet IP
PV and BESS cut peaks, and the BMS sees the module as a load, a heat source and a buffer.

🧪 9. Who is already doing it, and what they learned

The idea is not new. Microsoft Research described it in 2011 in “The Data Furnace”: servers in homes and offices as a primary heat source. Since then it has been tested in practice:

  • 🔥 Qarnot (France) moved from heaters to digital boilers because, as its CEO admitted, you don’t need a heater all year round (TechCrunch). In July 2026 the company was acquired by Scaleway.
  • 🏊 Deep Green (UK): a pool in Exmouth; according to the BBC, server heat kept the water at about 30°C 60% of the time.
  • 🚿 Heata (UK): hot water in homes; according to a 2023 description, up to 4.8 kWh a day.
  • 🏢 Leafcloud (Netherlands): modules in apartment blocks (company description).
  • 📉 Nerdalize (Netherlands): servers in homes, declared bankrupt in December 2018.

The lesson from this list: year-round heat users (hot water, pools) win; seasonal space heating loses.

💰 10. Money and risks

The operator pays for electricity through a sub-meter and earns from computing. The building gets heat through a heat meter, cheaper or free. Rule of thumb: the heat is worth as much as the gas that was not burned, typically 25–40% of the module’s electricity bill. There are downsides: electricity on the building’s tariff costs more than at a large data centre’s connection, and service is spread over many addresses.

Risks to close in the design and the contract: 🔐 data on someone else’s premises (stateless jobs, encryption, seals, remote wipe), 🌐 a symmetrical connection, 🧯 fire safety and insurance, 🔊 pump noise and ⚖️ the legal status of the electricity customer.

📊 11. Summary: comparison of building types

BuildingTypical IT powerWinter heat useSummer heat useHeat transferViability
Single-family home0.5–1 kW all year (DHW) or 3–5 kW seasonallyapprox. 100% (underfloor, DHW)at 0.5–1 kW: DHW; at 3–5 kW: 10–30% (DHW, pool, greenhouse; rest via dry cooler or throttling)25–40°C buffer → underfloor directly; heat pump → DHW 55°CGood for a small DHW module; 3–5 kW only seasonally, on depreciated hardware
Greenhouses, agriculture10–100 kW100% (soil, hall)50–80% (nights)Directly from 25–40°C bufferVery high, no heat pump
Breweries, dairies, other industry20–250 kW (swarm)100% (process water, heating)100% (process and cleaning water)Preheat 10→30–35°C + heat pump on returnExcellent: flat 365-day profile
Multi-family building20–100 kW (swarm)100% (DHW, low-temp heating)100% (DHW and circulation)DHW preheat 10→30–35°C + heat pump to 55–60°CHigh: flat DHW profile, connection already there
Pools, water parks50–200 kW (swarm)100% (basins, DHW, ventilation)100% (water losses, evaporation)Direct 25–40°C → 28°C (titanium) + HP for DHWHighest: transfer without a heat pump

Power and heat-use figures are conceptual assumptions for analysing a specific site, not guaranteed results.

✅ Conclusions

Distributed, modular data centres do not replace large, secure facilities; they take over the bulk work that does not need them. Placing that compute where a grid connection already exists and sits underused, absorbing local PV surplus, and year-round heat demand, with liquid cooling, swarm architecture, return-pipe integration and open protocols, turns the cost of cooling processors into savings on heating and, where large data centres use evaporative cooling towers, into water that does not have to be evaporated.

Three theses. Information is easier to send than heat, so computing travels to the heat, not the other way round. Cooling the processor and heating the building are one process, so the electricity and water spent on throwing heat away are no longer needed. Growth comes in small steps on existing connections, each step with its own heat user. A complement to large data centres, not a replacement.

🔮 Next topic: working with dynamic tariffs and virtual power plants (VPP): ramping up compute and charging the heat buffer when energy prices are zero or negative.

Related reading

Leave a Reply