Heat Reuse & Waste-Heat Recovery (Engineering)

An AI hall is, thermodynamically, a furnace that happens to do matrix multiplication. A 100 MW IT load rejects very close to 100 MW of heat — essentially all the electrical energy that enters the building leaves it as low-grade warmth. That is an enormous, continuous, year-round heat source sitting next to cities that burn gas to stay warm. The temptation to sell it is obvious, and in 2026 it is increasingly not a temptation but a legal obligation. The engineering reality is harsher than the brochure: the quantity of waste heat is never the constraint — the grade is. Heat at 30 °C is worth almost nothing; the same joules delivered at 80 °C are a tradable commodity. And the grade you can deliver was largely decided one chapter upstream, when you set the facility-water supply temperature in Chapter 5.7.

This chapter is the canonical engineering home for waste-heat recovery: how the grades are defined and what each is good for; how a heat pump upgrades a sub-usable return temperature to a district-heating supply, and what that lift costs in COP and electricity; where the reuse heat exchanger ties into the loop and why warm-water DLC is the enabler that makes the whole exercise viable; and the offtake chicken-and-egg, the scheduling and counterparty problem that kills more heat-reuse projects than any thermodynamic limit. The economics, contract structures, carbon-accounting, and regulatory treatment that decide whether the project pencils are deferred to Chapter 15.5; here we engineer the heat.

Waste heat is graded by temperature because temperature, not energy content, is what an offtaker's distribution network and emitters require. A district-heating network has a supply-temperature spec set by its generation: legacy 2nd/3rd-generation networks run 70–90 °C (and some older systems higher), modern 4th-generation networks run 30–70 °C, and emerging 5th-generation (ambient-loop) networks run 10–30 °C and put a heat pump at every building. Your captured stream is only useful if it can meet the network it is feeding, directly or after a lift. That single matching condition governs everything that follows.

The grade you can capture is dictated by the cooling modality, which is why the cooling-architecture chapters are upstream of this one. Air cooling and rear-door heat exchangers (Chapter 5.2, Chapter 5.3) hand you a diffuse, low-grade stream — return air or RDHx water around 30–40 °C — spread over large flow rates and large surface areas, which is expensive to concentrate. Direct-to-chip liquid (Chapter 5.4) concentrates almost all the heat into a single small-flow, high-temperature stream: a warm-water loop returns coolant at roughly 45–55 °C off the cold plates, single-phase D2C systems commonly returning in the 35–50 °C band depending on inlet and delta-T. That concentration is the entire reason DLC is the enabler of credible heat reuse — not because it produces more heat, but because it produces usable heat in a form a heat exchanger can grab cheaply.

The table sorts halls by cooling modality. If your hall is air-cooled, you are in the top row — a large heat pump, a marginal COP, and an offtaker who must be either very close or very forgiving. If your hall runs warm-water DLC, you are in the second row — the only row where the numbers reliably work, because a modest lift from ~50 °C to a 4th-generation network's 60–70 °C supply runs at a COP of 4–6 and the captured heat is worth selling. This is the quiet reason that the cooling decision in Chapter 5.4 and the loop-temperature decision in Chapter 5.7 are the real heat-reuse decisions. By the time you reach this chapter, the grade is mostly already set.

A heat pump moves heat from a cooler source to a warmer sink and consumes electricity to do it. For waste-heat upgrade, the source is your DLC or RDHx return and the sink is the district-heating supply. The governing number is the coefficient of performance (COP): useful heat delivered per unit of electrical work. COP falls as the temperature lift (sink temperature minus source temperature) rises — this is Carnot's tax, and it is unavoidable. A heat pump lifting a 50 °C source to a 65 °C sink (a 15 °C lift) might run at a COP of 5–6; the same pump lifting a 30 °C air-side source to an 80 °C legacy network (a 50 °C lift) drops toward a COP of 2.5–3.5. Industrial high-temperature heat pumps for district heating typically span a COP of roughly 2.5–6 across this range.

The COP is the central engineering and economic variable, because it sets both the electricity you burn and the carbon you account for. At a COP of 5, you deliver 5 MW of heat for 1 MW of electricity; the captured 4 MW of waste heat plus 1 MW of pump work leave as sellable heat. At a COP of 3, you deliver 3 MW of heat for 1 MW of electricity — the same offtake now costs you 67% more parasitic load, and on a power-bound site that 1 MW is megawatts you could have spent on accelerators. This is the cleanest expression of the POWER-BOUND thread in the cooling stack: a low-COP heat-reuse scheme is a tax on the IT capacity of the building, and the warmer the source water, the smaller that tax. Designing the DLC loop warm (Chapter 5.7) is therefore the single biggest lever on heat-reuse economics, because it shrinks the lift before the pump ever runs.

Heat reuse adds a third thermal domain to the building. The technology-cooling loop (chip to CDU) and the facility-water loop (CDU to heat rejection) already exist from Chapter 5.6 and Chapter 5.7. Heat reuse interposes a fourth-quadrant heat-recovery loop that runs from the reuse HX, optionally through a heat pump, into a hot-water buffer, and across an isolation heat exchanger to the offtaker's network. The interface point — the boundary where your water meets the district utility's water — is a hydraulically and contractually critical plane, and it is almost always a plate heat exchanger that keeps the two networks chemically and pressure-isolated. Neither side wants the other's water, the other's chemistry, or the other's pressure transients; the HX is the demarcation.

The integration's hardest engineering problem is that the data center's heat output is constant and the city's heat demand is not. An AI hall rejects near-baseload heat 8,760 hours a year; district-heating demand swings seasonally by 5–10x between a January peak and a July trough, and the summer trough can fall below the heat the data center produces. The plant must therefore be designed to reject the heat anyway when there is no buyer — the dry coolers and towers of Chapter 5.8 never leave the building, because heat reuse is an opportunistic heat sink layered on top of a mandatory one, never a replacement for it. A common mistake is to size the rejection plant assuming the offtaker absorbs the load; the offtaker disappears every summer and on every demand dip, and the GPUs do not stop. Warm-water DLC is the enabler not only because it produces usable grade, but because it lets the same dry cooler that rejects heat in summer also pre-warm the reuse loop in winter — the warmer the loop, the more both jobs are done by simple, compressor-free equipment.

The reason most data-center waste heat is still dumped is not thermodynamics — it is coordination. Heat reuse requires two parties with very different planning horizons to arrive at the same place at the same time. The data center is built in 18–36 months against an interconnection slot and a depreciation clock. A district-heating network expansion, a new greenhouse, or an industrial-process tie-in is a multi-year municipal or industrial capital project with its own permitting, its own financing, and its own committee. Each side is reluctant to commit capital before the other has: the utility will not lay pipe to a heat source that may never deliver; the operator will not buy heat pumps and reuse HXs for an offtaker that may never connect. That is the offtake chicken-and-egg, and it is an engineering constraint, not just a commercial one, because the loop you build must survive the period when one side exists and the other does not.

The engineering responses are concrete. Plumb for reuse without committing to it: install the tie-in stubs, reserve the plant-room footprint and electrical headroom for a future heat pump, and route the pipe-rack so a reuse HX can drop in — because the irreversible substrate (loop taps, space, power) is cheap at construction and expensive to retrofit, exactly the reversible-vs-irreversible discipline from Chapter 1.1. Design the plant to run profitably with zero offtake on day one: the mandatory rejection plant must stand alone, so the reuse loop can sit dormant for a year or three until the offtaker materializes. Decouple the capital split contractually so neither party fronts the other's risk — the model that broke the deadlock at scale is Stockholm Exergi's Open District Heating, where the operator invests in the heat pumps and the utility invests in the network connection, and the heat trades as a standardized, temperature-indexed product. That platform now connects 30-plus data centers across 16 providers, and a single site can earn on the order of SEK 2M per MW per year for the heat it would otherwise have thrown away (Stockholm Exergi / Eurelectric, 2025).

For much of the industry's history, heat reuse was a discretionary sustainability gesture you did if the economics happened to align. In the EU as of 2026, that framing is obsolete. Under the recast Energy Efficiency Directive (EU) 2023/1791, Article 26, data centers with a total rated energy input above 1 MW must carry out a cost-benefit analysis of waste-heat recovery, and member states have gone further: Germany's Energy Efficiency Act (EnEfG) mandates a minimum energy-reuse factor (ERF) of 10% for new data centers from July 2026, escalating to 20% by July 2028, and France's Law 2025-391 requires heat valorization for facilities above 1 MW. An EU-level data-center energy-efficiency package was expected in Q2 2026 to build further requirements on the EED's monitoring data.

The engineering consequence is that, for a European greenfield over 1 MW, the heat-reuse tie-in is no longer an optional bonus loop — it is a permit and compliance condition that must be in the design basis from day one, and a hall plumbed for cold air with no reuse provision may simply not be permittable. This is the strongest argument yet for warm-water DLC and a pre-installed reuse HX: a mandate that requires you to deliver a measurable ERF is far cheaper to meet from a 50 °C concentrated stream than from a 35 °C diffuse one. The detailed regulatory, contractual, and carbon-accounting treatment — including how ERF is measured and how heat sales interact with PUE and emissions claims — is the subject of Chapter 15.5; the point here is that the regulation reaches back into the mechanical design and removes "do nothing" as an option.