Topology
Devices plumbed in series see progressively warmer water, so the last position returns the most useful heat and keeps the least margin.
INSIGHTThermal engineering · Heat recovery
Data center heat recovery, explained
Data center heat recovery captures heat a facility would otherwise reject to ambient and delivers it to a useful load — building heat, a district network, an industrial process, or a power cycle. It succeeds or fails on one variable: the temperature the heat arrives at, because temperature, not quantity, decides what a load can accept. Below: heat grade, warm-water loops, ERE and ERF, district and campus reuse, ORC, and where reuse stops paying.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
- 55 °C
- 4th-generation district supply
- 1 MW
- EU reuse threshold, rated input
- ~12%
- Carnot ceiling, 60 °C over 20 °C
What you need to know
01
Grade, not quantity
Almost all the electricity a data center draws leaves as heat, but heat a few degrees above ambient does almost no work however much of it there is.
02
Warm loops are engineered backwards
A recovery-capable facility is designed from the off-taker's supply temperature, not from how cold the loop can be made.
03
PUE stops at the fence
ERE and ERF exist because a facility heating a district scores no better on PUE than one venting the same energy to the sky.
04
It needs a neighbour
Heat cannot be stored cheaply or shipped far. No off-taker, no recovery — and reuse never replaces the rejection plant.
First principles
Quantity is not the constraint. Grade is.
Essentially all electricity a data center draws leaves it as heat, and the quantity is no longer marginal: Lawrence Berkeley National Laboratory put US data-center electricity at 4.4% of national demand in 2023, on a path to 6.7–12% by 2028,[8] and the IEA puts data centres near 1.5% of global electricity demand in 2025, heading toward about 3% by 2030.[9] Almost none of it is useful as delivered, because energy has a quality as well as a quantity. Heat a few degrees above ambient does almost no work however much there is; the same watts captured in liquid at the die arrive far hotter. ASHRAE names liquid-cooling facility water classes by their maximum supply temperature precisely because that number, not the cooling method, is the design variable.[1]
Engineering the loop
Warm-water loops: how grade is engineered
A recovery-capable facility is designed backwards from the off-taker: not how cold the loop can be made, but how warm it may run while silicon stays inside its thermal envelope. Direct-to-chip capture is what makes that question answerable, collecting heat at the cold plate before it disperses into room air,[10] and federal-lab practice treats warm-water liquid cooling as the enabling step for free cooling and energy recovery, in that order.[3]
Where grade is lost
Three details decide how much grade survives
Approach temperature
Every exchanger between die and off-taker costs a few degrees.
Residual air load
Cold plates never capture everything, so size recovery on the liquid stream alone.
Reference
The heat-grade ladder
Five heat streams, ordered by the grade they arrive at, the loads each one can serve, and what reaching it costs.
| Code | Heat stream | Grade | Loads it can serve | What it costs to get there |
|---|---|---|---|---|
| HR-01 | Room return air | Lowest | Nothing off-site on its own | Free, and near worthless — diluted into a large air mass |
| HR-02 | Rear-door exchanger, air-coil water | Low | Building preheat, once a heat pump lifts it | Cheap to retrofit; the grade penalty is paid in heat-pump power |
| HR-03 | Cold-plate return, cool water class | Moderate | Space heating; a low-temperature secondary loop | Needs direct-to-chip capture, and leaves a chiller in place |
| HR-04 | Cold-plate return, warm water class | Good | Direct feed to a low-temperature district or campus loop | Needs IT rated for a warm class; costs thermal margin at the die |
| HR-05 | Heat-pump-boosted loop | High | Legacy high-temperature networks, hot water, process heat | Buys grade with electricity that must cost less than the heat |
Scoring reuse
ERE and ERF: the metrics that score reuse
PUE stops at the fence. It divides facility energy by IT energy and cannot credit heat that leaves the boundary usefully, so a facility heating a district scores no better than one venting the same energy to the sky. The Green Grid closed that gap with energy reuse effectiveness, ERE = (1 − ERF) × PUE, where the energy reuse factor runs from 0 to 1; with no reuse ERF is 0 and ERE collapses back to PUE.[4]
ERF is now standardised as ISO/IEC 30134-6: energy reused over total energy consumed by the data centre.[5]Two consequences are easy to miss. Reuse must be beneficial and outside the boundary, so internal recirculation does not count and an ERF target compels a real external off-taker. And ERF says nothing about how efficiently that off-taker uses the heat. NREL's Energy Systems Integration Facility is the federal reference — an energy-recovery loop spanning supercomputer, campus heating, and legacy IT, in a building reporting a PUE near 1.04[2] — and its lesson is architectural: the loop was designed in, not bolted on.
Reuse does not cut the data center's own consumption. The facility draws the same electricity; the saving lands on the off-taker's fuel bill.
0
Reduction in the facility's own draw
Destinations
District and campus reuse
District heating is the highest-value destination where a network exists, and it became viable for data centers because the networks got colder. Fourth-generation district heating describes low-temperature distribution around 55 °C supply and 20 °C return — temperatures a warm direct-to-chip loop can approach without a heat pump, where an older high-temperature network never could.[6] The test is not whether a city has district heating but which generation reaches the property line. Campus reuse is smaller and far easier to contract, because both sides of the meter share an owner: offices, labs, hot-water preheat, greenhouses. Its limit is that campus demand is a fraction of what an AI facility rejects.
Regulation is increasingly the forcing function. The recast EU Energy Efficiency Directive requires member states to ensure waste-heat utilisation at data centres above 1 MW of total rated energy input unless the operator shows it is not technically or economically feasible, with annual reporting from 500 kW of installed IT power.[7] The feasibility study becomes a compliance artefact — better done early enough to shape the design.
Feasibility screen
Does reuse pencil at this site?
Seven questions, in the order a feasibility review should ask them. A scheme that fails the first two rarely recovers on the strength of the rest.
| # | Question | Worth engineering when | Stops making sense when |
|---|---|---|---|
| 01 | Is there an off-taker, and how far? | A committed load within short pipe distance | Pipe is costly per metre and sheds grade en route |
| 02 | Do the loads coincide in time? | A year-round load — process heat, hot water, greenhouses | A heating-season off-taker idles the plant half the year |
| 03 | Does your loop meet their supply temperature? | A warm cold-plate return that feeds the network unboosted | A higher requirement forces a heat pump that can outcost the heat |
| 04 | Does the contract outlive the hardware? | Terms written around the site, not one GPU generation | Network operators plan in decades; fleets refresh far faster |
| 05 | Is the boundary metered and agreed? | A defined control volume, a meter, a rule on pumping cost | Without it the ERF figure is unauditable |
| 06 | Is there a regulatory driver? | A mandate that makes the study a cost you owe anyway | Absent one, reuse competes with every other use of capital |
| 07 | What else could that capital buy? | Capex small against the contract or permit it unlocks | The same money in compute or power usually returns more |
The stranded-heat fallback
The organic Rankine cycle, honestly
An organic Rankine cycle is a closed steam cycle that swaps water for an organic working fluid with a much lower boiling point, so it vaporises on heat too cool to raise steam: pump, evaporator against the hot loop, expander, condenser to a cold sink. The output is electricity, which needs no pipe and no neighbour.
Thermodynamics sets the expectation. Any heat engine is bounded by the Carnot efficiency 1 − T_cold / T_hot in absolute temperature, and data-center heat sits close to its own sink. At an illustrative 60 °C source against a 20 °C sink the ceiling is 1 − 293 K / 333 K, about 12%; at 45 °C against 25 °C, roughly 6%. A real machine returns a fraction of that, and its pumps come off the top. It also needs a genuinely cold sink, costs heavily per kW at small scale, and recovers only the liquid loop. It is a stranded-heat fallback — a design that leads with an ORC usually could not find a neighbour.
ηCarnot = 1 − Tcold / Thot
60 °C over 20 °C → 1 − 293 K / 333 K ≈ 12% · 45 °C over 25 °C → ≈ 6%
60 °C over 20 °C → 1 − 293 K / 333 K ≈ 12% · 45 °C over 25 °C → ≈ 6%
Honest limits
When heat recovery is not the right fit
Heat recovery is oversold more often than it is under-built. These limits decide most real projects.
- No off-taker means no recovery. Heat cannot be stored cheaply or shipped far, and a scheme built on a hypothetical future neighbour is a rejection plant with extra pipework.
- Reuse does not cut the data center's own consumption. The facility draws the same electricity; the saving lands on the off-taker's fuel bill.
- Full-capacity rejection stays mandatory. Reuse is never firm, so recovery is additive capital, never a substitute for the rejection plant.
- Summer inverts the economics. Heating demand collapses exactly when ambient temperatures make rejection hardest.
- Heat pumps relocate the problem. Boosting a cool loop burns electricity on site to displace fuel off site.
In the product
What this means for a modular unit
Recovery readiness is an architecture decision, made before the loop is plumbed. Each PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, with a closed direct-to-chip loop specified as part of the enclosure — heat stays at HR-03/HR-04 grade instead of dispersing into room air, and the reuse boundary lands on one meterable connection. Whether a site reuses that heat still depends on the off-taker, not the unit.
Modularity changes the siting question too: because the unit is factory-built and relocatable — PODOS targets a 90-day window from order to commissioning for a standard unit — capacity can be placed next to a thermal load rather than waiting for one to appear beside a finished building. Adjacent engineering: direct-to-chip liquid cooling and the power architecture that sets how much heat exists to recover; siting in deployment; the head-to-head in modular vs traditional AI data centers; vocabulary in the AI infrastructure glossary.
QUESTIONS
Frequently asked questions
What is data center heat recovery?
The capture of heat a data center would otherwise reject to ambient, and its delivery to a useful load outside the facility — building heating, a district network, an industrial process, or a power cycle.
What is the difference between ERE and ERF?
ERF, the energy reuse factor, is the share of energy entering the data center that is beneficially reused outside its boundary, on a scale of 0 to 1. ERE folds that into a PUE-style number: ERE = (1 − ERF) × PUE.
Does heat recovery lower a data center's own energy use?
No. The facility draws the same electricity; the saving lands on the off-taker's fuel bill. That is why reuse needs a metric outside PUE.
Is an ORC worth adding to a data center?
Mainly where no thermal off-taker exists and the loop runs warm. The Carnot ceiling at data-center temperatures sits in the single digits to low teens of percent, and a real machine returns a fraction of that.
Sources
- [1] Thermal Guidelines for Data Processing Environments, 5th ed. (TC 9.9) — ASHRAE, 2021
- [2] High-Performance Computing Data Center Waste Heat Reuse (ESIF) — NREL (DOE), accessed 2026-08-31
- [3] High-Performance Computing Data Center Warm-Water Liquid Cooling — NREL (DOE), accessed 2026-08-31
- [4] WP #29 — ERE: A Metric for Measuring the Benefit of Reuse Energy from a Data Center — The Green Grid, 2010
- [5] ISO/IEC 30134-6:2021 — Data centres key performance indicators, Part 6: Energy Reuse Factor (ERF) — ISO/IEC, 2021
- [6] 4th Generation District Heating (4GDH): Integrating smart thermal grids into future sustainable energy systems (Lund et al., Energy 68:1–11) — Elsevier, Apr 2014
- [7] Directive (EU) 2023/1791 on energy efficiency (recast), Art. 12 and Art. 26 — European Union (EUR-Lex), Sep 2023
- [8] 2024 United States Data Center Energy Usage Report (LBNL-2001637) — Lawrence Berkeley National Laboratory, Dec 2024
- [9] Energy and AI — Executive Summary — IEA, Apr 2025
- [10] Emergence and Expansion of Liquid Cooling in Mainstream Data Centers (white paper) — ASHRAE TC 9.9, c. 2021
Bring the off-taker into the design
Send the site, the loop temperature you can hold, and the load you would like to heat. Engineering will tell you what a pod-based loop can hand across the boundary.