Heat Reuse & District Heating (Sustainability & Economics)

A gigawatt AI campus is, thermodynamically, a gigawatt heater. Essentially all the electricity it draws, minus a rounding error that leaves as light down fiber, becomes low-grade heat, and the building then spends more energy throwing that heat away. Chapter 5.9 covers the engineering of capturing it. This chapter covers the decision on top: is the waste stream a cost you minimize, or a product you sell?

You cannot freely choose the answer at commissioning. Three upstream facts set it: the temperature of the loop you plumb, whether a heat offtaker sits within trenching distance of the site you picked, and the regulatory regime of the country you build in. Get them right and waste heat becomes a modest revenue line and a strong social-license and compliance asset. Get them wrong and the building pays twice, once to make the heat and once to reject it, in a jurisdiction that is increasingly going to fine you for the privilege.

The engineering stays in Chapter 5.9. The subject here is economics and obligation: heat-grade pricing, the district-heating integration models (Stockholm's standardized network-access approach chief among them), the payback math and the EU regulatory drivers that turn reuse from a nice-to-have into a mandate, and designing for reuse as a constraint. Heat reuse is cheap if you design for it and expensive if you retrofit it, and the gap between those two worlds is decided before steel is cut.

Waste-heat economics start with the temperature of the heat you can deliver. Contracted volume, electricity price and trenching distance then decide whether that grade earns back the investment. Heat is not fungible: a megawatt of 30 °C heat and a megawatt of 60 °C heat are different products, because a district-heating network or an industrial process needs its input at a usable supply temperature. A conventional 3rd-generation network typically supplies water at 80–100 °C, less on weather-compensated shoulder-season curves, while 4th-generation low-temperature grids run about 50–70 °C. The IEA DHC STEP work uses 80/40 °C as its conventional reference pair, and 60/35 °C is the step at which server waste heat becomes directly usable. If your heat arrives below the supply temperature, someone has to spend electricity in a heat pump to lift it, and that lift dominates the economics of the deal.

The cooling decision in Chapter 5.4 reaches downstream into these economics. An air-cooled or chilled-water hall may return a stream around 30–45 °C. A network requiring 70–100 °C then needs a heat-pump lift, while a lower-temperature offtaker may not. Price electricity, capex, exchanger approaches and distribution losses against your actual source and sink pair, using Chapter 5.9.

A warm-water direct-to-chip loop can return facility water at 45–60 °C or higher. The supported facility-water class and the separate TCS operating point in Chapter 5.7 determine the grade you actually get. Most new high-density AI halls are liquid-cooled and vendors now permit warmer chip inlet temperatures, so the byproduct is getting hotter. Warmer supported water can widen economizer hours and shrink the heat-pump lift; for some low-temperature networks the lift disappears. It does not guarantee direct delivery, though. If the sink runs hotter than your return, you still pay a heat-pump electricity bill, just a smaller one than a cold loop would commit you to.

Heat reuse only works where there is a sink: an offtaker that accepts a contracted seasonal volume at a price, with ordinary rejection retained for any heat it does not accept. That is why the technology's center of gravity is northern Europe. Cold climates create substantial seasonal heating demand (domestic hot water gives a different, summer sink), and decades of public investment built dense district-heating (DH) networks that can absorb it. Stockholm Exergi’s network runs to roughly 3,000 km. That pre-existing pipe in the ground is the asset, and the data center is merely a new heat source plugged into it.

In a heat-sink-poor market, such as much of the US Sun Belt, the Gulf or Australia, there may be no nearby contracted heat sink. That is why US reuse remains marginal despite far larger compute, although industrial demand can change the result anywhere. The siting consequence is direct: heat reuse is a reason to value a Nordic, German or Dutch metro site, and its value in a desert needs a named industrial or other useful sink, not a presumed district-heating market. (The water-and-climate siting gate is in Chapter 3.7; this is its mirror image on the heat-rejection side.)

The decisive innovation is the commercial structure, not the heat exchanger. The naive model is bilateral: one data center negotiates a bespoke deal to dump heat into one utility, splitting capex and arguing over price annually. It scales one counterparty at a time, and it can finance when a creditworthy buyer commits volume, price and termination payments that repay the connection. Stockholm Exergi's Open District Heating rewrote this by making network access repeatable under standardized terms. Any heat producer can sell into the network under a published, temperature-indexed tariff, so the price you receive rises with the grade you deliver, settled like a spot market rather than negotiated one-off.

That turns waste heat from a charity donation into a product with a published price signal. Access is standardized at day-ahead prices, with delivery-temperature and capacity requirements set against forecast outdoor temperature; it is not a freely tradable commodity market. The model now connects 30-plus data centers across 16 providers. The volumes are still small. One press report puts Stockholm Exergi's 2025 recovered heat at about 113 GWh, over 100 GWh of it from data centers, which averages only about 13 MW across the year. Stockholm Data Parks aims for data centers to supply 10% of the city's heat by 2035. Today, according to Eurelectric (June 2026), Open District Heating supplies about 3.5% of Stockholm's heat. Those two figures do not square, because 3.5% of a whole city's heat is far more energy than 113 GWh. What they agree on is the scale: heat recovered from data centers still covers only a few percent of a large city's demand.

Which structure finances better is a counterparty question, not a foregone conclusion. An indexed tariff gives you a published access and pricing framework with temperature-dependent compensation, but leaves you the price and volume risk. A long-term bilateral contract with a creditworthy buyer, a minimum purchase obligation and a price floor is usually what a lender underwrites. Compare them on price exposure, volume commitment, connection ownership, curtailment rights and termination payments before you choose.

The table sorts by who you are selling to and why. If a real DH market exists, the indexed model scales across many producers without a bespoke negotiation each time: its price signal rewards grade and its standard terms are repeatable. With a single buyer, a long-term bilateral contract is the pragmatic choice. Meta's Odense facility works differently. Meta donates its surplus heat to the city for free, while the utility Fjernvarme Fyn owns and operates the ~45 MW heat-pump plant that lifts the heat into the Odense network. That is a social-license arrangement, not a revenue line.

In Germany, the primary obligation is to meet the applicable EnEfG ERF target; offering heat at generation cost and having the nearby network decline it is only one exemption route. France adds a tax lever. Since the 2021 REEN law (Loi n° 2021-1485), a data center gets the reduced electricity excise on its consumption above 1 GWh a year only if it feeds its waste heat into a heating or cooling network or meets a power-efficiency indicator set by decree. It must also keep its cooling-water use under a decree-set limit (Code des impositions sur les biens et services, art. L. 312-70). If you sited in a market with no network, your only reuse path is a captive on-site user (greenhouses, aquaculture, an adjacent industrial process, office heating), which is why the heat-reuse question is, at bottom, a siting question. Where no mandate applies, OCP's heat-reuse guidelines for local authorities (September 2024) ask councils to connect data centers with nearby heat users, offer grants or tax incentives, and secure heat-reuse agreements before approving permits. OCP's beta readiness self-assessment and its heat-reuse projects database let you test a site's case before that negotiation; the engineering is in Chapter 5.9.

The hyperscale Nordic example is Microsoft × Fortum in Finland. Fortum spent about €225M on heat-pump plants next to two Microsoft data-center sites in Espoo and Kirkkonummi, using air-source heat pumps and electric boilers. It said in May 2026 that it had started heat production there, with Microsoft's waste heat phased in from 2027. Once fully built out, the data-center heat is expected to cover about 40% of regional district-heat demand, which totals roughly 2 TWh/yr across ~250,000 users, a larger offtaker population than Meta's Odense project. There, the utility carries the heat-pump capex.

Trace. Receipts = (15,000 + 5,000) MWh × €40/MWh = €800,000/year. Incremental electricity = 4,000 MWh × €80/MWh = €320,000/year. Net annual cash = €800,000 − €320,000 − €100,000 = €380,000. The ten-year annuity factor is [1 − (1.08)⁻¹⁰]/0.08 = 6.7101. NPV = −€3.0M + €0.38M × 6.7101 ≈ −€0.45M. Simple payback is about 7.9 years but fails the 8% discounted hurdle. Reject the connection on these economics; preserve the easement only if its separately priced option is worthwhile.

Flip and rejection condition. Break-even delivered heat price = [€3.0M/6.7101 + €320,000 + €100,000]/20,000 MWh ≈ €43/MWh. At an assumed €45/MWh, annual cash is €480,000 and NPV is about +€0.22M. That price selects build only if the buyer’s minimum purchase, supply/return temperatures, meter location, curtailment rights, connection ownership and termination payment are enforceable. At that price, maximum supportable capex is €0.48M × 6.7101 ≈ €3.2M, so a longer trench taking capex to €4.0M rejects the deal again.

Stockholm Exergi's open tariff gives a sense of scale. It pays about SEK 2 million a year per MW delivered over most of the year, which is roughly €170,000, or €20–40/MWh depending on hours delivered. That sits under this case's €43 break-even, but the comparison is not like-for-like. Stockholm's network operator pays for the pipe, so the producer's capex is its heat pumps and production plant. Run your own deal on the buyer's real tariff and capex split rather than on this case's price.

An ERF obligation leaves a negative NPV negative. Either pick another way to comply, or establish the statutory exception that applies to you (Chapter 15.7). The heat balance itself is worked in Chapter 5.9, and the general cash-flow conventions behind this discounting are in Chapter 1.8.

Stripped of enthusiasm, the heat-reuse business case is modest and front-loaded with capex. Revenue is the heat price times the energy delivered, and DH heat is cheap energy: you are not selling electricity, you are selling lukewarm water. Against that you carry the heat-exchanger plant, the interconnection to the network, and, if your loop is cold, the heat pump and its lifetime electricity. Trenching and pipe are the silent budget-killer, because running a hot-water main several kilometers to reach a network can dwarf the in-building cost.

At a warm loop next to an existing network, delivered heat and the offtake tariff drive revenue; at a cold loop far from a sink, heat-pump electricity and trenching drive cost. Model recoverable MW, annual load duration, delivered temperature, heat-pump COP, power and heat prices, connection distance and cost per meter, plant capex/opex, and discount rate before claiming payback. The outcome is bimodal, cheap-and-good versus expensive-and-mandatory, and your design and siting choices determine which you get.

That is why Europe stopped leaving it to economics. The EU Energy Efficiency Directive, via Delegated Regulation (EU) 2024/1364, requires data centers at or above 500 kW IT load to report ERF to the European database, alongside PUE, WUE and REF. ERF is the energy-reuse factor, reused energy divided by total facility energy. Individual records are confidential, and specified Member-State and Union aggregates are public.

Reporting is only half of it. Article 26(6) of the same Directive goes further for large sites. Once total rated energy input is above 1 MW, a data center must put its waste heat to use, directly or through another waste-heat recovery application, unless reuse is shown to be technically or economically unfeasible. That showing is an Article 26(7) cost-benefit analysis, done when a site is newly planned or substantially refurbished. Member States enforce the duty through their own transposition, so the details differ.

Germany's EnEfG adds a numeric ERF floor. A site starting operation on or after 1 July 2026 needs an ERF of at least 10%, measured per DIN EN 50600-4-6. Sites that start from 1 July 2027 have to be designed for 15%, and those from 1 July 2028 for 20%. The start date fixes the floor, and it does not rise later for an existing site. Today a site gets two years to reach its floor as an annual average. Section 11(3) sets out exemption routes, which the callout below covers. One of them rests on a nearby network turning down an offer of heat at generation cost; the Act does not require you to sell heat at cost.

A government bill now before the Bundestag, Drs. 21/8027, would keep these percentages but relax how sites meet them. The committee heard it on 5 October 2026; it was not enacted as of 16 October 2026 and would take effect the day after promulgation. Sites would get four years to comply instead of two, and heat the operator uses in its own offices and plant rooms would count. Sites tied to a heat network would get a shortfall allowance. A new exemption would apply where a cost-benefit analysis finds that linking to an existing or planned heat network is not technically and economically reasonable. The bill would also set the size threshold at 500 kW of installed IT. Check where it stands before you lock a German design basis.

Disclosure is now turning into rules. The Commission adopted a delegated act creating a common EU rating scheme on 21 September 2026. Scrutiny normally runs two months, but Parliament asked for the two-month extension on 6 October 2026, so Parliament and the Council now have until 21 January 2027 to object. If neither does, the EU database will issue the first data-center labels by 15 August 2027, with PUE and WUE classes and an indication of waste-heat reuse readiness. On 21 September, the day it adopted the rating act, the Commission also launched a consultation on binding minimum performance standards. It closes on 14 December 2026, and a legislative proposal is planned for Q2 2027. The metric stack behind all of this (PUE, WUE, ERF, REF and the standards that define them) is built out in Chapter 15.1, and the broader disclosure regime in Chapter 15.7. For German sites starting operation from 1 July 2026, and under any Member State's strict reading of Article 26(6), heat reuse is a condition of operating; payback matters only through a documented exemption or the Article 26(7) analysis.

Some heat-reuse decisions are reversible and cheap to defer; others are irreversible and must be hedged at scoping. Irreversible (decide once): the site relative to a heat sink (you cannot move the campus to the network); the facility-water loop temperature architecture (a hall plumbed for a cold chilled-water loop cannot cheaply become a warm-water hall mid-life); and the physical provisions for a future connection, meaning pipe-rack space, knockouts, a heat-exchanger room and reserved easement for the trench to the network. Reversible (defer): whether you energize the connection now or later, the specific offtaker contract terms, and the heat-pump sizing, which can follow demand. Provisioning the substrate turns an irreversible decision into a cheap, deferrable one. Reserving a heat-exchanger room and a pipe easement costs little at design, while adding them to a finished, occupied, slab-poured hall is a major retrofit.

The emerging frame that ties this together is heat-as-a-product. The most sophisticated operators no longer treat waste heat as an externality to dispose of. They treat it as a co-product to sell, and they design the thermal plant (loop temperature, capture topology, metering) so that the heat leaving the building is merchantable. That reframing is what justifies the warm loop, the indexed contract and the reserved easement. It also turns a regulatory burden, the EnEfG ERF mandate, into a hedge: an operator whose metered seasonal deliveries meet the applicable ERF cohort has a compliance route, while a competitor who omitted capture, connection or a buyer may need a retrofit.

In a world where the binding constraint is power and every megawatt is scrutinized, showing that a fraction of your rejected heat warms a city is social license, and increasingly, license to build at all. The grid-and-energy-systems framing of that social-license argument continues in Chapter 15.8, and the embodied-carbon side of the sustainability account in Chapter 15.6.