Data Center Waste Heat Recovery Depends on Lift, Not Supply

Most of the waste heat in a data center sits in the 20 to 26°C range, according to a Siemens white paper written with Grundfos on connecting data centers to district heating. The networks that could take it run at flow temperatures of 60 to 80°C in the fourth generation and 80 to 100°C in the third. Liquid cooling is often said to close that gap by returning hotter water. This piece tests that claim against published heat pump, capture, and water temperature figures, and leaves out project finance beyond one worked case and all carbon accounting, because the temperature question can be settled without them.

Most waste heat arrives at 20 to 26°C

The Siemens and Grundfos paper says waste heat in that range has a limited number of applications, such as heating swimming pools or aquaculture farms. It adds that for most data centers the opportunity to export heat is limited by the higher temperature requirements of potential users, which is why the heat has to be upgraded with heat pumps.

The paper’s Table 1 sorts heat networks by generation. Third-generation networks, favored in Nordic countries, run flow temperatures of 80 to 100°C. Fourth-generation networks run 60 to 80°C, and the paper says they were developed to use low-grade sources such as solar and geothermal heat.

Only the fifth generation, which the paper calls an emerging concept and leaves out of its scope, runs at 10 to 25°C, a range that overlaps air-side exhaust. Between a 26°C exhaust stream and a fourth-generation flow of 60 to 80°C, the lift is 34 to 54 K.

Against a third-generation flow of 80 to 100°C the lift is 54 to 74 K.

The IEA’s 2019 commentary on data centres and energy names the same obstacle. It lists the technical challenge of getting sufficiently high temperatures among the barriers to waste heat use, next to taxation on the heat and contractual and legal questions. For third- and fourth-generation networks, then, exhaust-temperature heat is not a supply. The heat pump is part of the system, and its efficiency decides what the heat costs.

Heat pumps land near 60% of Carnot

The Siemens paper defines the coefficient of performance (COP) as useful heat delivered divided by the power consumed, and says it falls as the temperature difference between source and sink grows. For most heat reuse cases the paper puts the COP between 3.0 and 6.0. A reference value for any source and delivery temperature follows from the Carnot relation, COP = T_sink ÷ (T_sink − T_source), with both temperatures in kelvin.

The calculation takes the warmest temperature of the source stream and the warmest delivered temperature. That makes the result a reference point, not a strict bound, because a source stream that cools as it gives up heat has a lower average temperature than its hot end. For the Siemens “classic” connection, the source is the 26°C cooling return and the delivery is 80°C. The arithmetic is 353.15 K ÷ (353.15 K − 299.15 K) = 6.54. The paper’s assumed COP of 4.5 is 69% of that value.

Table 1. Realized heat pump COP against the Carnot reference value (COP in kWh heat per kWh electricity)

* Carnot reference COP = T_sink ÷ (T_sink − T_source), kelvin, from the warmest source and delivery temperatures; not a strict bound. Siemens COPs are business case assumptions for chillers only; the Finnish value is as reported, with delivery taken at the top of the reported 60 to 70°C.

Source: Siemens and Grundfos, “From Waste Heat to Usable Heat” (2023), Table 2; Lü, Lu and Zhang, “Optimizing Data Center Waste Heat Reuse” (2024); reference values calculated for this article

Table 1 repeats the calculation for the paper’s other two connection types and for a Finnish campus data center of about 1 MW that a 2024 conference paper used to validate its model.

That paper reports two heat pumps with a COP of about 3.6, water of 15°C falling to 10°C on the data center side, and delivery to a district heating return line at 60 to 70°C.

The realized COP lands between 55% and 69% of the reference value in all four cases. Four cases are too few to call that a rule, and the Siemens figures are assumptions in a business case, not measurements.

The pattern still shows where the lever is. A heat pump pays for every kelvin of lift, and the two cases that lift furthest, the booster and the Finnish delivery, sit at the bottom of the table.

A warmer source changes the arithmetic directly. Take the same 80°C delivery and a 45°C source, the upper facility water limit of ASHRAE’s W45 class. A return stream would be warmer still, so this understates the case.

The reference value becomes 353.15 K ÷ (353.15 K − 318.15 K) = 10.09, which is 54% above 6.54. In electricity terms, the reference value falls from 0.153 kWh to 0.099 kWh per kWh of heat delivered. If the 69% ratio carried over, the COP of 4.5 would move to about 6.9.

That is a real gain, and it is also the full size of it. A 45°C source still needs a heat pump to reach 80°C, and the pump still draws electricity.

Liquid cooling trades capture for temperature

A 2014 NREL report on Asetek’s RackCDU at the ESIF data center gives one measured pair of capture figures for a warm-water cold plate system, and its OSTI abstract is the only part used here. The abstract states that the system captured up to 64% of server heat into the liquid stream at an outlet temperature of 89°F.

At outlet temperatures approaching 100°F the capture was 48%. The system was designed to capture heat from the CPUs only, and the abstract says memory cooling could raise the capture.

In metric units those outlet temperatures are 31.7°C and 37.8°C. With an 80°C delivery, the reference COP is 353.15 K ÷ (353.15 K − 304.82 K) = 7.31 at the cooler outlet and 353.15 K ÷ (353.15 K − 310.93 K) = 8.36 at the warmer one, a gain of 14%.

The captured share fell from 64% to 48%, a loss of 25%. On a 10 MW IT load, the arithmetic gives 6.4 MW against 4.8 MW entering the liquid.

In that test, the 25% fall in capture is larger than the 14% rise in the reference COP. The abstract does not give the load, flow rate, or inlet temperature behind either figure, so the pair cannot be read as a pure temperature effect, and it is not a law of liquid cooling. It does show that outlet temperature and the share of heat entering the liquid are separate quantities, and that a higher temperature does not recover more heat by itself.

The second constraint is how long the high temperature lasts. ASHRAE TC9.9’s 2021 white paper renamed its water classes by upper limit: W17, W27, W32, W40, W45, and W+.

It says current liquid cooling products might operate on W45, but that newer, higher-powered products may be capable only of W32, and that cooling requirements could eventually push facility water temperatures into the W27 range. The paper says most cold plate solutions today run easily on W32 water with case temperatures of 80 to 90°C.

The paper’s case study shows the cost. For a dense 1U half-width node, class W27 supports up to 500 W per CPU and 12 W per DIMM with single-sided DIMM cooling.

Class W45 reduces the cooling capability to a point that may be comparable to an air-cooled system with a low cold aisle temperature. Case-to-coolant resistance, the quantity behind these limits, is the subject of a look at cold plate thermal resistance and a 28 K gap.

The water temperature that favors heat reuse is the one that rising socket power takes away. A heat sale priced on 45°C source water has to survive a hardware refresh that the same paper says could force the facility to W32 or W27.

The paper also describes a case where warm water was put to use. The SuperMUC-NG system at the Leibniz Supercomputing Center uses direct warm-water cooling at 40 to 45°C, and its waste heat is sufficient to reheat about 40,000 m² of office space. The heat also drives an adsorption chiller that produces 600 kW of chilled-water capacity for rear-door heat exchangers.

Network temperature and distance set the bill

Siemens’ 10 MW case compares three ways of joining the heat pump plant to the network. The “classic” type heats return water to flow parameters of 60 to 80°C, the “efficient” type heats return water to 60 to 65°C, and the “booster” type delivers 77 to 82°C at the far edge of the network.

The assumed COPs are 4.5, 5.5, and 3.5, and the marginal cost of heat is €23, €19, and €30 per MWh.

The paper sets that against €39 per MWh from natural gas, with gas at €25 per MWh and carbon at €60 per tonne.

The cheapest heat did not make the best project. Siemens reports that the booster type, the least efficient, earns the highest net cash flow, because it sells more heat at a higher price, €40 per MWh against €28 for the efficient type.

For the €6M heat pump system in that case, simple payback is about 9 years without subsidies. The buyer pays for delivered temperature, so a lower COP can still win when it is paired with a higher price.

Distance enters through capital cost, not through heat loss. The paper puts the connection share of total CAPEX at 3% for a 50 m run and 50% for 4,000 m.

It puts the heat loss of a well-insulated 70 to 80°C line at 0.5 to 1.5% per kilometer, which is 2 to 6% of the heat over 4 km. The pipe is the expensive part, and the loss is minor by comparison. Siemens adds that upgrading the heat at the network substation instead of at the data center can cut the loss by a factor of 3 or 4, since the long run then carries lower-temperature water.

Time matters as well. The paper observes that heat has most value when outside temperature is low, which is when many data centers use free cooling, so that supply and demand pull apart. It suggests that data centers exchange their waste heat for cooling services free of charge instead of using free cooling, and names Meta’s arrangement in Odense as an example. How delta T and flow set the capacity on the data center side decides how steady that supply can be.

The Odense plant is the clearest full-scale example. Ramboll, the owner’s engineer, describes electrically driven ammonia heat pumps that raise the recovered heat to 70 to 75°C, with about 45 MW of total heat production.

The plant was ready for operation in late 2019 and is designed to recover 215,000 MWh from the Tietgenbyen data center. Ramboll also notes that Denmark abolished its tax on using surplus heat in January 2022 for certified businesses. The IEA adds a contractual point: operators and suppliers have to agree on how heat is guaranteed to customers if the data center shuts down.

Whether warm return water can skip the heat pump

Can 40 to 45°C liquid-cooled return water feed a fourth-generation network directly? Not as a flow stream. Siemens lists fourth-generation flow at 60 to 80°C, well above 40 to 45°C. Fourth-generation return water runs at 40 to 70°C, so a stream at the top of the 40 to 45°C range overlaps only the low end and could at best preheat the return line. That is the logic of the paper’s “efficient” connection, and the heat pump remains in that design.

What share of the heat can a project actually sell? The two sources in this piece measure different things. The NREL test measured the share of server heat entering the liquid, for CPUs only. The 2024 Finnish-led paper modeled an 18 MW data center in design phase and found that 71% of its waste heat could be reused, assuming a heat pump COP of 4.5. The first is a measured capture and the second is a modeled reuse share, so neither number transfers to a different site without its conditions.

The claim holds as a COP gain

Liquid cooling does return warmer water than air cooling, and a warmer source raises the ideal heat pump COP by 54% in the calculation above. It does not remove the heat pump, because that water still sits below the flow temperature of a fourth-generation network. In the one NREL test reported, capture fell by a quarter between the two outlet temperatures, and ASHRAE’s own paper says rising socket power could pull facility water temperatures back down. What survives is a conditional improvement in efficiency that depends on a nearby low-temperature network, a demand profile that matches the supply, and a heat contract that outlasts a hardware generation. Qualified engineering review is required before any of these figures is used in a design.

References

• Siemens and Grundfos. From Waste Heat to Usable Heat: Connecting data centers and district heating networks. Siemens AG. 2023. https://assets.new.siemens.com/siemens/assets/api/uuid:ddbf25c8-95c5-4843-861b-6361c4289e4a/DataCenterWhitePaperCustomerReferenceGrundfos.pdf
• ASHRAE Technical Committee 9.9. Emergence and Expansion of Liquid Cooling in Mainstream Data Centers. ASHRAE. 2021. https://www.ashrae.org/File%20Library/Technical%20Resources/Bookstore/Emergence-and-Expansion-of-Liquid-Cooling-in-Mainstream-Data-Centers_WP.pdf
• Sickinger, D., Van Geet, O., Ravenscroft, C. Energy Performance Testing of Asetek’s RackCDU System at NREL’s High Performance Computing Data Center (abstract page). NREL via OSTI. November 2014. https://www.osti.gov/biblio/1164105
• Lü, X., Lu, T., Zhang, Q. Optimizing Data Center Waste Heat Reuse: Case Studies and Environmental Implications. Atlantis Highlights in Engineering 32, IRES 2023 proceedings. 2024. https://atlantis-press.com/article/126001376.pdf
• Ramboll. Meta: surplus heat to district heating. Ramboll (undated project page). https://www.ramboll.com/projects/energy/meta-surplus-heat-to-district-heating
• IEA. Data centres and energy: from global headlines to local headaches? IEA, Paris. 2019. https://www.iea.org/commentaries/data-centres-and-energy-from-global-headlines-to-local-headaches