What Long Periods of Low IT Load Do to a Liquid Cooling System
Liquid cooling is specified at the design point and operated far below it. What happens to a cold-plate loop at 15 percent load: pump power under the cube law, heat-exchanger effectiveness driving supply temperature toward the dew point, CDU control instability below the stable band, laminar-flow fouling, and the part-load efficiency inversion. With measured utilisation data and the specification items that prevent each failure.
Liquid cooling is sold on its full-load numbers. The cold plate carries 130 kW at a fraction of the volume an air stream would need, the CDU holds a tight supply temperature, and the facility water loop runs warm enough that mechanical cooling is optional for much of the year. Those numbers are real, and they are also the operating point that many AI halls reach for a small share of their lives.
The failure mode this article is about is not a design error. It is a design that is correct at rated load and mis-specified everywhere else. A liquid cooling system sized for peak density, installed in a facility that fills in phases or runs training jobs in bursts, spends most of its hours far from the point it was optimised for — and the physics at that point are not a gentler version of full load. They are different, and several of them are worse.
The number to keep in your head: a CDU whose pump is held at full speed while the rack idles is running a motor whose power scales with the cube of flow. Drop flow to 15 percent and pump power falls to about 0.3 percent — a factor of roughly 300. A fixed-speed CDU foregoes that entirely, and in a facility that lives at low load it is the single largest avoidable energy item in the liquid loop [1][2].
Why low load is the normal case, not the edge case
Start with how much of the time a liquid-cooled AI hall actually runs near nameplate. The honest answer, from multiple independent measurements, is: less than the planning documents assume.
| Measurement | Figure | Source |
|---|---|---|
| China smart-computing supply vs actual use, 2025 | ~3.8 billion card-hours supplied, ~1.4 billion used — ~36.8% utilisation | CAICT data, reported September 2026 [3] |
| National average GPU utilisation, smart-computing centres | ~30% | Inspur AI Research, Q1 2026 monitoring [3] |
| Enterprise self-built clusters | Average below 30%, some only in the low teens | Operator survey, industry press [4] |
| Enterprise general-purpose compute centres | 10–15% | IDC, reported 2025 [5] |
| Large GPU clusters (international panel) | 30–50% | Clockwork / SemiAnalysis panel, November 2025 [6] |
| A single 1,000-GPU regional site | Rack fill <50%; of installed servers, utilisation <40% | Industry survey [4] |
Table 1 — Measured utilisation figures. The spread is wide, but the direction is consistent: a liquid-cooled hall is likely to spend its first years — and in many facilities its whole life — at a fraction of the density it was sized for.
Two structural reasons keep it that way. First, phased build-out: a hall designed for 130 kW racks is commissioned with the racks it can fill, and the rest arrives in tranches. Second, workload shape: training runs are bursty and inferencing is diurnal, so even a fully populated hall oscillates between high and low power draw on a timescale of minutes to hours. Neither is a planning failure. Both mean the part-load operating point deserves the same engineering attention as the design point, and it usually does not get it. It also means the monitoring scope has to catch a part-load fault, which presents as a slow drift rather than an alarm.
The CDU is sized at the design point and operated off it
A CDU does four jobs: isolate the secondary loop, regulate supply temperature, control flow and differential pressure to the racks, and instrument the loop. Every one of those four behaves differently at 15 percent load than at 100 percent, and the second and third are where the trouble concentrates.
Liquid-cooled halls are not 100 percent liquid-cooled, which adds a second part-load variable. A contemporary direct-to-chip server rejects something like 70 to 80 percent of its heat to the coolant, with the remainder going to chassis airflow — power supplies, drives, NICs, DIMMs beyond what the cold plates reach [7]. When the compute load drops, both paths drop, but not in lockstep: the air-side share is dominated by fixed losses in the power conversion chain, so the rack's heat profile migrates toward the air side as load falls. The CDU sees a diminishing thermal load while the room sees a rising relative one.
Failure one: the pump runs at design speed to deliver part-load heat
The most common and most expensive symptom. CDU pumps in a shared row loop are typically set by the worst-case branch: the pump speed that guarantees adequate flow to the hottest, most flow-hungry cold plate in the loop. Every other branch then receives more flow than it needs, and that excess is pure overhead — the pump does not know which chips are hot and which are idle, and it pushes the same total volume regardless of workload distribution [1].
The energy consequence is governed by the affinity laws, which are unforgiving in the operator's favour if the pump is allowed to slow down, and equally unforgiving against them if it is not.
Pump affinity arithmetic. The cube law is why variable-speed control matters more at part load than any other single measure; the saving is not a percentage of something small, it is essentially the whole pump bill [1][2].
Two caveats keep that number honest. First, the cube law assumes the system curve is dominated by friction rather than static lift, which holds for a closed loop but wears off as valves throttle. Second, and more important operationally: a pump cannot be slowed below the point where the loop's own hydraulics still work. There is a minimum controllable flow, set by the least-favoured branch, the cold plates' own pressure-drop curves, and the pump's minimum-speed and cooling requirements. The hydraulic behaviour here is the same problem as coolant loop sizing, run backwards: instead of choosing a diameter for a design flow, you are asking what flow the installed pipework can still distribute evenly when the demand falls away. Case-study data from rack-mounted CDU deployments shows the flow-matching problem directly — an 8 to 15 kW rack demanding a nominal 30 to 50 L/min can see flow fluctuation beyond ±15 percent and efficiency loss beyond 25 percent with a control scheme that does not track the load, with reliability falling away sharply above 30 kW [8].What to specify rather than assume: the CDU datasheet should state the minimum controllable flow, the turndown ratio, and the pump curve — not just a nominal flow figure. A unit that cannot hold a stable differential pressure at 15 percent of design flow will be operated at 100 percent instead, and the cube law then works against the operator for the life of the install [8][9].
Failure two: the heat exchanger becomes too effective, and the supply temperature walks down
This is the counter-intuitive one, and it is the mechanism behind most part-load condensation events. Engineers are taught that a heat exchanger degrades with fouling and that low flow means poor heat transfer. In a CDU's primary-to-secondary heat exchanger, low flow on the secondary side does the opposite to effectiveness: the coolant spends longer in the exchanger, the number of transfer units rises, and the approach temperature collapses.
Heat-exchanger effectiveness arithmetic. The supply temperature is not held by the heat exchanger — it is held by the mixing valve. The exchanger's job is only to make that control possible, and at low load it hands the control loop a much narrower working range.
The mechanical consequence is a nearly-saturated heat exchanger with almost no thermal headroom left. The control consequence is worse: the mixing valve is now trying to trim a supply temperature whose un-mixed floor has moved down to within half a degree of the facility water temperature. Small valve movements produce large temperature excursions, and the loop begins to hunt.
Published work on exactly this regime is unusually specific. A University of Texas at Arlington study of cold-plate CDUs found that typical CDUs are designed to operate at 20 to 30 percent of rated heat load to achieve a stable secondary supply temperature, and that when operated at the extreme low end — 1 to 10 percent of rated capacity — large fluctuations in secondary supply temperature appear, and those fluctuations can cause the three-way valve on the primary side to fail. The study's proposed remedy was a flow control valve on the primary side combined with PID tuning, holding the secondary supply temperature within ±0.5 °C at those very low loads [10].
Read that as a procurement requirement, not a research note. The stable operating band of a CDU is a datasheet parameter that is rarely printed. A unit specified only by its rated capacity may be a perfectly good machine with a stable band of 20–100 percent and an unstable one below that — which is exactly the region a phased AI hall occupies in its first year. Ask for the stability band and the turndown ratio, and treat "±0.5 °C from 1 percent load" as a claim that should come with a test report [10].
Failure three: the supply temperature falls toward the dew point
Condensation is the part-load failure that destroys hardware rather than efficiency, and the mechanism connects directly to the previous section. The safe operating rule is simple and absolute: the cold plate supply temperature must stay above the hall dew point by a margin that covers sensor uncertainty and transient humidity swings — industry practice puts the margin at 2 to 3 °C, with the supply held at the greater of the manufacturer's minimum or the measured dew point plus margin [11][12][13].
Warm-water operation makes that easy at design. A loop supplied at 35 to 45 °C against a hall dew point that typically sits in the 10 to 18 °C range has 20 °C or more of margin, and the whole point of warm-water design is that the margin is structural rather than controlled [7][12][13]. The margin exists because the mechanical cooling plant would have to fail catastrophically for the supply to drop that far.
Part load removes that structural protection in a way that is easy to miss. The supply temperature is no longer pinned by the balance between a large heat load and a controlled facility water temperature; it is drifting down toward the facility water temperature, because the heat exchanger has become too effective to hold it up. If the facility water loop is itself running cold — which happens in shoulder seasons, in facilities that kept a chilled-water primary loop for legacy reasons, or when a chiller is staged on for another load — then the secondary supply can follow it down while the hall humidity stays where it was.
Margin arithmetic. At low load the design margin is preserved only if the control logic actively holds the supply temperature up. Nothing in the physics holds it up for you.
The field failure mode that operators actually report is subtler than a single set-point error. A humidity sensor drifts and under-reports dew point, so the CDU supplies water below the real condensation threshold for weeks until moisture becomes visible on cold plates. Or a firmware update raises a GPU's power draw, an operator lowers the CDU set point to chase the resulting delta-T, and the water temperature crosses the dew point before the control loop can react. Both are control-logic problems, and both are more likely at part load, because at part load the supply temperature is being actively managed rather than passively fixed by the load [13].
The interlock that prevents this: a hard limit that refuses any set-point change below the current dew-point floor, computed continuously from dry-bulb and relative-humidity sensors placed at rack inlet height and at the CDU inlet, taking the higher of the cold-aisle and hot-aisle readings. Log supply temperature, dew point, and valve position together so an excursion is visible as a trend rather than discovered as a puddle [12][13].
Failure four: low velocity turns the cold plate into a settling tank
Direct-to-chip cold plates have channels spaced hundreds of microns apart and the flow through them is intentionally laminar to keep distribution even across all the plates [14]. That is a deliberate design choice for thermal uniformity, and it removes the mechanism that most industrial cooling systems rely on to stay clean. High velocity keeps fouling down; laminar flow in narrow channels is the most fouling-sensitive geometry in common industrial use [14].
At full load, the resulting velocities are still high enough that deposits struggle to settle. At low load they are not. A loop that sits at 10 to 20 percent of design flow for months is a loop whose cold plates and heat exchanger surfaces are seeing near-stagnant conditions, and the fouling mechanisms that follow are well quantified:
| Mechanism | Measured effect |
|---|---|
| Heat-exchanger fouling, general | Heat transfer degraded up to 14%; pressure drop up 45%; exchanger efficiency down up to 30% |
| Biofilm, 20 µm thickness | Heat transfer efficiency reduced by ~30% — a layer thinner than a human hair |
| Biofilm versus mineral scale | 10–100% more thermally resistant than calcium carbonate, and attaches readily to carbonate crystals |
| Glycol concentration for biofilm control | Below roughly 17% propylene glycol, biofilm formation accelerates |
| Cost of 25% PG blend (the recommended mix) | ~8% loss of heat capacity plus a measurable pumping-power penalty from viscosity |
Table 2 — Fouling effects in direct-to-chip loops. The pressure-drop figure is the one that compounds: higher pressure drop pushes the pump back up the curve, which spends the part-load energy saving the control scheme was supposed to deliver [14][15][16].
Two aspects of this are specific to low-load operation. First, biological growth is a function of residence time and temperature, and warm low-flow loops are close to ideal incubator conditions. Industry guidance is explicit that biofilm can develop undetected because most facilities do not routinely monitor water quality, and that by the time cooling performance visibly degrades, remediation of the affected cold plates is difficult [14][15][16]. Second, and more subtly, the failure is self-reinforcing: deposits raise pressure drop, pressure drop forces the pump to a higher speed to hold flow, and if the control scheme is set to hold constant differential pressure rather than constant flow, the loop's total flow stays roughly constant while individual branches starve. The temperature alarms fire on the starved branch, the operator raises pump speed again, and the loop settles into a slightly-worse equilibrium that nobody chose.
Maintenance intervals should scale with load, not with calendar. A loop running at design load is self-cleaning by velocity and needs the standard programme. A loop running at 15 percent load for a year has accumulated the same biological exposure as a far longer period at design flow. Coolant testing at commissioning, annually in normal operation, and semiannually for high-density AI GPU loops is the baseline guidance [15] — in a low-utilisation hall, the argument for the higher frequency applies to the whole fleet, not just the busy fraction of it.
Stagnation in the loop also depletes the chemistry. Glycol degrades under thermal and oxidative stress into organic acids, which drops pH and accelerates corrosion of mixed-metal components; inhibitor packages deplete over time and have to be re-dosed against measured levels rather than by schedule [17]. A partial wet lay-up — the condition a low-utilisation hall drifts into when a row is powered down for months — is documented to support acid-producing bacteria that disperse through the system on start-up, and to present a Legionella risk at restart unless circulation and biocide residuals are maintained [18]. The mitigations are known and cheap: circulate at least weekly for at least four hours, hold zero sulphate-reducing bacteria and aerobic plate counts below 100,000 cfu/ml, and monitor corrosion coupons monthly [18]. None of them happen by themselves in a hall that is quiet. This is the same fouling pathway that fouling prevention and remediation addresses, arriving earlier than usual because the velocity that suppresses it is missing.
Failure five: the efficiency claim inverts
The headline economic argument for liquid cooling is efficiency, and at part load that argument needs re-examination rather than repetition. Liquid cooling's advantage over optimised air is partly a function of the load at which it runs.
The auxiliary power of a liquid loop is not negligible. Chinese Refrigeration Society figures put the combined pump and fan auxiliary load of a cold-plate liquid-cooled facility at roughly 8 to 12 percent of IT load, rising to 15 to 20 percent for immersion systems, where higher fluid viscosity increases flow resistance [19]. The same analysis notes that the part-load efficiency advantage of liquid cooling is not automatic, and that at low load a liquid system can be worse than an optimised air-cooled design — because the auxiliary load is proportionally larger and the fixed losses do not scale down [19].
Part-load efficiency arithmetic. PUE is a ratio, and its numerator does not shrink when its denominator does. A design that hits 1.2 at full load can sit well above it at 30 percent load without anything being broken [19][20].
For a facility that plans to fill in phases, that inverts the commissioning story: the PUE measured in the first year is not the PUE the design promised, and the gap is not a defect to be chased — it is the arithmetic of a fixed overhead against a partial denominator. The honest response is to report PUE against measured IT load and to size the first phase of the cooling plant to the load actually present, adding capacity as the racks arrive.
What a low-load-thoughtful specification asks for
None of the failures above require a different technology. They require a specification that names the operating range and a control scheme that is tested at both ends of it.
| Requirement | What to ask for explicitly | Failure it prevents |
|---|---|---|
| Pump turndown | Minimum controllable flow, turndown ratio, and pressure-control mode (constant ΔP vs constant flow) — as tested values, not nominal | Cube-law energy loss; branch starvation |
| Temperature stability band | The load range over which secondary supply holds ±0.5 °C, with a test report at the low end | Valve hunting; three-way valve failure |
| Dew-point interlock | Hard set-point floor computed from live dew point at two elevations; refusal, not warning | Cold-plate condensation |
| Flow measurement | Per-branch flow or differential-pressure evidence, not just CDU display totals | Uneven distribution masked by rack-level ΔT |
| Fluid management | Baseline at commissioning, inhibitor and pH monitoring, biofilm control, low-load circulation schedule | Fouling, inhibitor depletion, stagnation |
| Commissioning load test | Dummy-load testing at 10–20% as well as at design | Discovering part-load behaviour in production |
Table 3 — Specification items that exist because of part-load behaviour. Each maps to one of the five failures above.
The commissioning row deserves emphasis because it is the cheapest of the six and the most often skipped. Thermal response changes with actual server power, and baseline records should capture supply and return temperatures, flow or differential pressure, processor temperatures, pump speed, and room conditions — because without a baseline, a high return temperature months later cannot be attributed between workload growth, fouling, reduced flow, and a controls problem [21]. A dummy-load unit capable of simulating electrical load across a working flow range lets the hydraulic and thermal design be validated before live equipment is fully deployed [9], and running that test at 15 percent load as well as at design is the only way to find out whether the stability band claimed on the datasheet is the band the installed system actually has.
Frequently asked questions
Is running a liquid cooling system at low load actually harmful, or just inefficient?
Both, and the harmful part comes from the same mechanisms as the inefficient part. Low flow costs money through the cube law, and it also removes the velocity that keeps cold-plate channels clean. Low load raises heat-exchanger effectiveness, which makes the supply temperature harder to control and creates the condensation window. And low load pushes the CDU below the stability band its control was designed around, which is where three-way valve failures occur. None of these damage anything at 40 percent load. They matter when a system sits in the low teens for months.
Why does the supply temperature drop at low load instead of just staying where it was set?
Because the heat exchanger gets more effective as secondary-side flow falls, so the unmixed floor that the supply can reach moves down toward the facility water temperature. At full flow in our worked example the supply sits about 2.6 °C above the facility water; at 10 percent flow it sits about 0.5 °C above it. The mixing valve is then trimming a very small band, which both makes control unstable and removes the margin that protected against condensation. Warm-water loops have large margins at design — the loss of margin at part load is the thing to watch, not the design value.
How low can a CDU be turned down safely?
That is a datasheet question with a datasheet answer, and it is usually not printed. Published research describes typical CDU design intent as 20 to 30 percent of rated heat load for stable secondary supply temperature, with the 1 to 10 percent region showing large supply fluctuation and three-way valve failure unless a primary-side flow control valve and tuned PID parameters are used — which then achieved ±0.5 °C stability [10]. Treat 20 percent as a starting assumption, ask the vendor for the real figure, and verify it with a dummy-load test rather than accepting a table value.
Should the pump run at constant flow or constant differential pressure at low load?
Constant differential pressure is more forgiving of branch imbalance; constant flow is more defensible for cold-plate heat removal but starves branches when resistance diverges. The important point for low-load operation is that neither mode should be left unexamined: a pressure-drop-based control strategy exists specifically to hold the same per-server flow when some cooling loop modules are decommissioned from the rack one at a time, which is exactly the transition a phased hall goes through [10]. Whatever mode is chosen, per-branch flow or differential-pressure evidence is what distinguishes a genuinely balanced loop from one whose imbalance is hidden behind an acceptable rack-level delta-T [21].
What is the fluid-maintenance implication of running at low load?
The intervals should shorten. Fouling effects in direct-to-chip loops are severe — up to 14 percent heat-transfer degradation, up to 45 percent pressure-drop increase, and biofilm 10 to 100 percent more thermally resistant than mineral scale, with a 20 µm layer costing around 30 percent of heat transfer [14][15][16]. Because biofilm is driven by residence time and temperature, a low-utilisation loop accumulates exposure at a rate closer to a much busier one. Coolant testing at commissioning, annually in normal operation, and semiannually for high-density GPU loops is the baseline guidance; a hall running at 15 percent load should be treating the whole fleet at the higher frequency [15].
Does this problem go away with immersion cooling?
No. Immersion has a larger auxiliary load fraction — 15 to 20 percent of IT power per the CAR analysis, against 8 to 12 percent for cold plates — because the fluid's higher viscosity increases flow resistance, and that overhead is similarly insensitive to load [19]. Immersion also brings its own low-load concerns around fluid circulation and filtration, and its fluid-loss costs are a documented operational burden. The part-load problem is a property of pumping a loop, not of the cold plate.
What is the cheapest single intervention for a hall that will run at low load?
Variable-speed pump control with a real turndown range, verified at low load. It is the item with the largest energy consequence (a cube-law effect), it is the item that also determines whether the loop can hold stable temperatures at part load, and it is the item most often compromised by a specification that names only rated capacity. A dummy-load commissioning test at 10 to 20 percent load then tells you whether the system you bought behaves the way the datasheet said it would.
If you are specifying a liquid cooling loop for a hall that will fill in phases, or retrofitting one that is running at a fraction of its density, send us the project scope with the rack design power, the expected initial fill, and the facility water temperature band, and we will work the part-load operating point with you.
References
- The Cooling Report, "30% of Data Center Electricity Goes to Cooling — The Savings Are Hiding at the Component Level" — pump over-drive in shared loops, the pump power cube law, flow imbalance costing energy twice, and precision per-cold-plate flow control. thecoolingreport.com
- ProphetStor, "Predictive Liquid Cooling for AI Data Centers" — fixed-flow versus variable-flow CDU operation, 15–40% pump energy wasted at low load under fixed flow, cubic affinity law, and OCP design-flow guidance of 1.25–2.0 L/min/kW. prophetstor.com
- Securities Times (证券时报), "Structural imbalance in computing capacity utilisation" (September 2026) — CAICT 2025 smart-computing supply of ~3.8 billion card-hours against ~1.4 billion used (36.8% utilisation) and Inspur AI Research Q1 2026 national GPU utilisation of ~30%. epaper.stcn.com (Chinese, PDF)
- Cailian Press (财联社), "High-end compute shortage versus under-utilised smart computing centres" — a 1,000-GPU regional centre below 50% rack fill with under 40% utilisation of installed servers, and operator experience that enterprise self-built clusters average below 30% utilisation, some in the low teens. cls.cn (Chinese)
- China Financial Information Network (中国金融信息网), "Idle compute behind the smart-computing boom" (July 2025) — IDC data on enterprise general-purpose compute centre utilisation of 10–15% and GPU utilisation below 30% at some smart-computing centres. cnfin.com (Chinese)
- NextFin, "Enterprise AI at Scale: Why Infrastructure, Observability and Fault-Tolerance Drive ROI" — Clockwork / SemiAnalysis panel, November 2025: large-scale GPU cluster utilisation running at 30–50%, with the three named causes. nextfin.ai
- Mislier, "What is direct-to-chip liquid cooling?" — CDU functions and what breaks without each, direct-to-chip heat capture ratio of 70–80% in current servers, warm-water loop economics, and the retrofit questions that determine whether a design ages well. mislier.com
- TOPSFLO, "How to Select a Liquid Cooling Pump for a Data Center" — centralized versus distributed CDU behaviour at partial load, rack cooling load versus required minimum flow bands, flow fluctuation and efficiency loss figures, and pump selection and cavitation criteria. micro-dc-pump.com
- Liangdi Energy Saving Technology, "How to Choose a Rack-Mounted CDU Based on Heat Load, Flow Rate, and Redundancy" — oversizing reducing controllability at low load, minimum controllable flow and pressure-control mode as selection criteria, and dummy-load testing across a working flow range to validate hydraulic design. liangditech.com
- Pardeep Shahi, "Cold Plate Based Dynamic Liquid Cooling at Data Center and Chip Level" (University of Texas at Arlington, 2022) — typical CDU design operation at 20–30% of rated heat load for stable secondary supply temperature; large secondary supply fluctuations and primary-side three-way valve failure at 1–10% load; stabilisation to ±0.5 °C with a primary-side flow control valve and PID tuning; pressure-drop-based control to hold flow when loop modules are decommissioned one by one. mavmatrix.uta.edu
- LCT Expo, "The biggest misconception in liquid cooling O&M: lower CDU secondary supply temperature is not better" — ASHRAE-aligned optimal secondary supply band of 35–45 °C, COP loss of 2–3% per degree of unnecessary sub-cooling, hall dew points of 10–18 °C, and the safe floor of dew point plus 2–3 °C. lctexpo.com (Chinese)
- Leviathan Systems, "Condensation and Dew-Point Control in Liquid-Cooled GPU Halls" — continuous dew-point computation from sensors at rack inlet height and CDU inlet, taking the higher of cold-aisle and hot-aisle readings as the safety floor, and the requirement for a hard interlock refusing set-point changes below that floor. leviathansystems.co
- PatSnap Eureka, "How to Manage Condensation Risks in Liquid-Cooled Servers" — condensation as distinct from leakage, hidden-area formation and detection difficulty, dew-point management and preheating strategies, and the operating-envelope argument linking coolant temperature to room dew point. eureka.patsnap.com
- Heat Exchange Institute (HEI) Horizons, "Data centers: Water quality and fouling risks in direct-to-chip cooling systems" — D2C channel spacing of hundreds of microns with intentionally laminar flow, the resulting sensitivity to deposits, biofilm thermal resistance relative to calcium carbonate, and the 17% glycol threshold below which biofilm accelerates. horizons.htri.net
- Schneider Electric, "Liquid cooling fluid management: How to protect AI infrastructure" — measured fouling effects (up to 14% heat-transfer degradation, 45% pressure-drop increase, up to 30% exchanger efficiency loss), ASHRAE coolant cleanliness guidance, and the commissioning / annual / semiannual testing intervals for high-density GPU loops. blog.se.com
- ALVIM, "White papers: biofilm monitoring in data centers" — a 20 µm biofilm layer reducing heat transfer efficiency by 30%, biofilm clogging risk in microchannel cold plates, and the limitation of liquid-phase water-quality parameters versus surface-specific biofilm monitoring. alvimcleantech.com
- Advantage Europe, "Water Treatment in Glycol-Filled Cooling Systems for Data Centers" — glycol degradation to organic acids, pH-driven corrosion in mixed-metal systems, buffer and inhibitor depletion, and microbiological control including bacteria that feed on glycol compounds. advantage-eu.com
- Ecolab / Nalco Water, "Data Center Condenser Protection" (technical bulletin) — wet lay-up procedures, the effect of stagnant conditions in supporting acid-producing bacteria that disperse on start-up, Legionella risk at restart, and the maintenance targets of zero SRB bacteria, aerobic plate counts below 100,000 cfu/ml, monthly coupon inspection, and circulation at least weekly for at least four hours. ecolab.com (PDF)
- China Refrigeration Society (CAR), "Data Center Liquid Cooling Technology White Paper" (2023), via the 2026 commercialisation bottleneck study — auxiliary pump and fan power at 8–12% of IT load for cold-plate liquid cooling and 15–20% for immersion, arising from higher fluid viscosity and flow resistance, and the finding that part-load efficiency advantage is not automatic and can be worse than optimised air at low load. renrendoc.com (Chinese)
- The Cooling Report, "30% of Data Center Electricity Goes to Cooling" — cooling as roughly 30% of total data center electricity, rising toward 40% in older air-cooled facilities above PUE 1.6 and falling to 15–20% in modern liquid-cooled facilities at PUE 1.2. thecoolingreport.com
- Market Logic Desk, "Are Cold Plate Liquid Cooling Systems Reliable for Long-Term AI Server Use?" — loss-of-flow and uneven distribution diagnostics, condensation from poor dew-point control, and the commissioning baseline records (supply and return temperature, flow or differential pressure, processor temperature, pump speed, room conditions) required to attribute later faults. marketlogicdesk.com