Specify the class, not the setpoint
Procure IT against an ASHRAE W-class and you have bought a climate tolerance you can price. Procure against a nominal supply temperature and you have bought nothing enforceable.
INSIGHTThermal engineering · Analysis
Warm-water liquid cooling: why 45 °C is a feature
Every kelvin the IT concedes at the cold plate is a kelvin of climate the site no longer has to buy a compressor for. Here is the ladder that converts a coolant class into free-cooling hours — and into heat somebody will actually take.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
- 35 °C
- Ambient a 45 °C loop stays compressor-free to
- 55 °C
- 4GDH distribution supply a W45 return meets unaided
- 1.034
- NREL ESIF measured PUE, with no mechanical chillers
The short answer
Warm-water cooling is not a thermal compromise. It is a heat-rejection decision taken at the chip: ASHRAE names its liquid-cooling classes for the maximum supply temperature the IT equipment accepts, and every degree of that allowance converts almost one-for-one into ambient headroom for the rejection plant.[1] Under ordinary approach assumptions a 45 °C loop stays compressor-free up to roughly 35 °C ambient, where a 32 °C loop needs mechanical help above about 22 °C. The trap is that free-cooling hours and heat-reuse value are governed by two different temperatures — supply and return — and most designs optimise only one of them.
First principles
The temperature ladder, written out
Air-cooled rooms hide their thermal budget inside a single number — the supply air temperature. A direct-to-chip loop exposes it as a chain of approach temperatures, and that is what makes it tractable. ASHRAE's fifth-edition thermal guidelines renamed the liquid-cooling classes after the number that matters — W17 (formerly W1), W27 (W2), W32 (W3), W40 (new), W45 (W4) and an open-ended W+ (W5), all sharing a 2 °C lower limit.[1][2] The committee tightened the definition at the same time: operating within a class requires full performance across the entire range of that class under nonfailure conditions, so the label is literally the warmest supply water the equipment will take unthrottled.[2] The white paper frames the shift as a response to chip vendors raising thermal design power while lowering package case-temperature limits.[2] The hardware follows: NVIDIA describes the GB200 NVL72 as connecting 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design.[8]
Walk the chain outward from the die and the economics fall out of arithmetic. Coolant leaves the CDU at the class temperature. The CDU's plate heat exchanger sits a few kelvin above the facility water it is fed. A dry cooler can only pull that facility water down to some approach above ambient dry bulb. Chiller-free operation therefore requires nothing more than ambient dry bulb staying below class temperature minus the two approaches. Every kelvin the IT concedes at the cold plate is a kelvin of climate the site no longer has to buy a compressor for.
ASSUMPTIONS
CDU liquid-to-liquid approach 3 K · dry-cooler approach to ambient dry bulb 7 K · technology-loop ΔT across the rack 10 K · single-phase water/glycol · no adiabatic or evaporative assist · ≥ 95 % of rack heat captured directly to liquid, per federal-lab practice[3], with the residual air load rejected separately.
Ambient threshold = class temperature − 3 K − 7 K. Return = class temperature + 10 K.
Figure 1 · Original analysis
Climate headroom, bought one kelvin at a
Each bar is the ambient dry bulb below which the loop rejects heat with no compressor running, derived from the stated assumptions: class temperature minus a 3 K CDU approach minus a 7 K dry-cooler approach. The ladder is almost linear because both approaches are fixed — which is precisely why the coolant class, not the plant, is the variable worth negotiating.
7 °C
17 °C
22 °C
30 °C
35 °C
35+ °C
Table 1 · The numbers behind Figure 1
Coolant class, threshold, return, and what the heat is worth
Ambient threshold = class temperature − 3 K CDU approach − 7 K dry-cooler approach. Return = class temperature + the 10 K loop ΔT. The last two columns are where the framework stops being arithmetic and starts being a judgement about sinks.
| Ref | ASHRAE class (max supply) | Chiller-free up to | Return at 10 K ΔT | What the return heat can serve | Where it breaks |
|---|---|---|---|---|---|
| WW-01 | W17 — supply to 17 °C | 7 °C | 27 °C | Below almost every sink. Slab preheat or a heat pump, nothing else. | Needs mechanical cooling for most of the year almost everywhere. |
| WW-02 | W27 — supply to 27 °C | 17 °C | 37 °C | Low-temperature space heating. This is the NREL ESIF band in practice. | Dry-only rejection fails on summer afternoons; needs a wet stage or a trim chiller. |
| WW-03 | W32 — supply to 32 °C | 22 °C | 42 °C | Ultra-low-temperature networks and return-side injection; short of a 55 °C supply. | Marginal in hot-humid climates without adiabatic assist. |
| WW-04 | W40 — supply to 40 °C | 30 °C | 50 °C | Within a small lift of a fourth-generation district-heating supply. | Exceeded only on design-day afternoons in most temperate climates. |
| WW-05 | W45 — supply to 45 °C | 35 °C | 55 °C | Meets the 55 °C 4GDH distribution supply directly, with no heat pump. | Chip-side headroom, not the plant, becomes the binding constraint. |
| WW-06 | W+ — supply above 45 °C | 35 °C and above | 55 °C and above | Legacy network territory, still short of a conventional third-generation supply. | Cold-plate and interface-material qualification, and IT vendor support. |
Reading the ladder
Free-cooling hours are the wrong headline
The reuse column is calibrated against real sinks rather than optimism. Fourth-generation district heating is defined around a 55 °C distribution supply and a 20 °C return[6], which is exactly where a W45 loop lands without a heat pump — and exactly why a W32 loop does not. NREL's ESIF data center is the working proof of the lower band: 24 °C supply water, 35–40 °C return captured to heat offices and lab space, and no mechanical chillers at all.[3]
Once the ambient threshold is known, hours come straight out of published climatology. ASHRAE's climatic design conditions report cooling design dry bulb at annual cumulative frequencies of 0.4 %, 1.0 % and 2.0 % across 12,424 processed stations.[4]Those percentiles are exceedance fractions of the 8,760-hour year, so the conversion is one multiplication: find the percentile at which the site's design dry bulb equals your threshold, and free-cooling hours are 8,760 × (1 − p).
Table 2 · Climatology, converted
What a design percentile is actually telling you
One multiplication turns an ASHRAE exceedance fraction into an annual hour count. The fourth column is the part the hour count conceals.
| Design percentile met | Hours exceeded / yr | Free-cooling hours / yr | What it actually tells you |
|---|---|---|---|
| 0.4 % | ≈ 35 h | ≈ 8,725 h | Trim plant runs a handful of afternoons a year. |
| 1.0 % | ≈ 88 h | ≈ 8,672 h | Trim plant is a seasonal asset, still full capacity. |
| 2.0 % | ≈ 175 h | ≈ 8,585 h | Free cooling dominates OPEX; CAPEX barely moves. |
Free-cooling hours are an operating-expense metric; the trim plant is a capital one.
0.929
ESIF energy reuse effectiveness, first thermosyphon year
The decoupling
Supply and return are separate levers
Here is the part the hours number conceals. Ninety-nine percent free cooling sounds like a chiller you no longer need. It is not. Those 88 remaining hours still demand 100 % of the thermal capacity, so the trim plant is sized for the design day regardless of how rarely it runs. Free-cooling hours are an operating-expense metric; the trim plant is a capital one, and warm water moves the first continuously while moving the second not at all — until the ambient threshold clears the site's extreme annual maximum dry bulb, at which point the compressor disappears entirely. That is a step function, not a gradient, and it is the only transition in this whole analysis worth designing around. It is also why the marginal value of a warmer class is wildly site-specific: the same 5 K is worth almost nothing in a cool maritime climate and worth an entire chiller plant in a hot arid one.
The common failure in warm-water arguments is treating “coolant temperature” as one number. It is a pair. Free cooling is decided by the supply temperature, because that sets the ambient threshold. Heat reuse is decided by the return temperature, because that is what the off-taker consumes. They are joined by the loop ΔT, and ΔT is a design choice — a wider ΔT means lower flow, less pump energy, and a hotter return without touching the supply.
ESIF proves the decoupling empirically. Its supply water is 24 °C, which is a modest class by the table above, yet it produces 35–40 °C return heat useful enough to warm a building on the coldest day of the year, and it does so with no chillers.[3] It reaches that return from that supply by running a wide ΔT into 60–80 kW racks, and it protects the hottest silicon by cooling in series so that the most sensitive components see the coolest liquid.[3] The published outcome is a PUE of 1.034 and an energy reuse effectiveness of 0.929 in its first year of thermosyphon operation.[3]
So the honest version of “warmer is better” is narrower than the slogan. Raising the supply class buys climate tolerance. Widening ΔT buys reuse value and pump savings. Both spend from the same account — the thermal budget between the coolant and the chip's limit — and the binding constraint is the last device in a series path, not the average. A design that raises supply and widens ΔT at once has to prove the outlet device still meets its case temperature, which is precisely what the vendor-neutral cold-plate and CDU requirements coming out of the Open Compute Project's cooling work exist to make checkable across suppliers.[9]
The reuse-metric definition is ISO/IEC 30134-6[7]; the European duties are Articles 12 and 26(6) of the 2023 energy-efficiency recast[5]. Both are worth reading before a heat-reuse business case is written, because both measure quantity rather than value.
Practice
What this means for operators
Compute your ambient threshold before you shop for a site
Class minus the CDU approach minus the dry-cooler approach is a single number, and it converts any candidate location into a design percentile in one lookup.
Decide whether you are buying hours or eliminating a plant
If the threshold does not clear the site's extreme annual maximum, you are still building the compressor — so justify the warmer class on energy, not on capital.
Size the loop ΔT against the return your off-taker needs, then check the outlet device
Reuse value is set on the hot side; the reliability risk sits on the same side.
Confirm the sink exists before you value the heat
Energy reuse factor is standardised as reused energy over total energy consumed, and it counts only heat that leaves the data-centre boundary and is genuinely used.
In the EU, check the threshold that already applies to you
Member States must ensure data centres above 1 MW total rated energy input utilise waste heat unless infeasibility is shown, with public reporting from 500 kW of installed IT power.
Honest limits
What this does not prove
The ladder above is a framework built on stated assumptions, not a measurement, and it fails in identifiable ways.
- The approach temperatures are assumptions. A 3 K CDU approach and a 7 K dry-cooler approach are ordinary but not universal; they depend on coil sizing, glycol fraction, fouling, and altitude. Tightening either is a capital purchase, and every ambient threshold in the table moves with them.
- A W-class is an acceptance rating, not a performance guarantee. It certifies unthrottled operation within the class. It does not say the machine draws the same power there — semiconductor leakage rises with junction temperature, and the residual air-cooled fraction gets harder to cool as the enclosure warms. Nothing cited here quantifies that penalty, and it works against the warm-water case.
- Design percentiles are climatology, not a forecast. They give exceedance fractions, not consecutive-hour durations. Whether the exceedance arrives as 88 scattered hours or as four consecutive afternoons changes thermal-storage and load-shed design completely, and the percentile alone cannot tell you which.
- ESIF is one facility in one climate. Its 1.034 PUE and 0.929 ERE come from a dry, high-altitude Colorado site with an adjacent building that wanted the heat. Those figures are evidence that the architecture works; they are not transferable numbers for any other site.
- Heat reuse is not shown to be economic anywhere. ERF and ERE measure how much energy is reused, not what it is worth. The European duty carries an explicit technical and economic feasibility exemption. A 55 °C return with no off-taker inside pipe distance has a value of zero.
- Nothing here is a PODOS measurement. No figure in this article was produced by PODOS hardware, and the framework should be re-derived against a specific vendor's cold-plate data before it drives a purchase.
In the product
How PODOS treats the coolant budget
A factory-integrated unit turns this from a site negotiation into a specification. The PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, with the cold plates, CDU, loop ΔT, and enclosure specified together rather than reconciled on site. Because the whole thermal chain is fixed before shipment, the ambient threshold is a known property of the unit rather than an outcome of commissioning — which is part of why PODOS targets a 90-day window from order to commissioning for a standard unit. Note that a unit at that scale also sits at the level where the European waste-heat duty attaches, which makes the return temperature a compliance variable, not only an efficiency one.[5]
The downstream consequences are covered separately: what to do with the return heat in data-center heat recovery, the density that forces the loop in the first place in high-density GPU infrastructure, the site-side prerequisites in site power readiness, and the head-to-head in liquid cooling vs air cooling. Terms used above are defined in the AI infrastructure glossary, and the wider technical set sits under engineering.
Sources
- [1] Thermal Guidelines for Data Processing Environments, 5th ed. (TC 9.9, Datacom Series) — ASHRAE, 2021
- [2] Emergence and Expansion of Liquid Cooling in Mainstream Data Centers (white paper) — ASHRAE TC 9.9, May 2021
- [3] Liquid in the Rack: Liquid Cooling Your Data Center (O. Van Geet, NREL; NREL/PR-7A40-72046; ESIF operating data for 1 Sep 2016 – 31 Aug 2017) — LBNL Center of Expertise / NREL (DOE), 2018
- [4] Weather Data Viewer 2025 / Handbook—Fundamentals Ch. 14, climatic design conditions — ASHRAE, 2025 ed., accessed 2026-08-31
- [5] Directive (EU) 2023/1791 on energy efficiency (recast), Art. 12 and Art. 26(6) — European Union (EUR-Lex), Sep 2023
- [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] ISO/IEC 30134-6:2021 — Data centres key performance indicators, Part 6: Energy Reuse Factor (ERF) — ISO/IEC, 2021
- [8] GB200 NVL72 product page — NVIDIA, accessed 2026-08-31
- [9] Cooling Environments / Cold Plate sub-project (cold-plate-to-CDU technology cooling system requirements) — Open Compute Project, ongoing
Set the coolant budget before you pick the site
Bring your target W-class, the site's design dry bulb, and the return temperature an off-taker would actually take. Engineering will tell you whether the compressor leaves.