Heat balance
Q̇ = PIT
For a steady rack boundary, 40 kW electrical draw means approximately 40 kW thermal rejection. Use metered input or the vendor thermal report.
01 / Quick reference
The density ladder
Bands overlap deliberately. “Viable” means technically defensible under the listed assumptions; “default shortlist” is the first architecture to price and model, not a mandate.
| Regime | Screening band | Viable when… | Default / optimal when… | PUE evidence, not promise | Relative capex | Ops load |
|---|---|---|---|---|---|---|
| Room air | ~5–10 kW | Inlet temperatures are uniform; bypass/recirculation is controlled; CRAH/CRAC and floor/duct paths close at N+1. | Mixed legacy racks, low peaks, no facility-water loop at rack. | 2024 fleet average: 1.56; not an air-system rating. | $ | Familiar; airflow balancing and filter maintenance. |
| Containment | ~10–20 kW | Sealing, blanking panels, fan control, and enough return path eliminate mixing. | The room has capacity but distribution—not heat rejection—is the constraint. | DOE Jefferson Lab reached 1.27 after a multi-measure retrofit including hot-aisle containment. | $–$$ | Door/roof discipline; pressure and inlet sensors. |
| In-row air | ~15–30 kW | Short airflow path, contained aisle, sufficient coil and fan capacity, and a maintainable condensate/chilled-water path. | A local hot row must be added without rebuilding the whole room. | No transferable PUE band; include unit, pumps/chiller, and partial-load controls in the facility model. | $$ | More distributed fans, valves, drains, and controls. |
| Rear-door HX | 20–80 kW products | Server fans can overcome door pressure drop; water and condensate risk are engineered; discharge remains room-neutral. | Standard air-cooled IT must stay intact, or DTC readiness is staged. | LBNL projected a facility change from ~1.4 to ~1.1 in a broader liquid-cooling project. | $$–$$$ | Water balancing, hinges/hoses, leak detection, coil cleaning. |
| Single-phase DTC | ~30–120+ kW | Cold plates cover dominant heat sources; a CDU and secondary loop reject heat; residual air is budgeted. | OEM-supported AI/HPC trays, warm-water opportunity, conventional service workflow desired. | Do not infer PUE from heat-capture ratio; model pumps, dry cooler/chiller, and remaining air separately. | $$$ | CDUs, chemistry, quick disconnects, firmware, leak response. |
| Two-phase DTC | 20–120 kW current rack units; higher claims | Qualified dielectric/refrigerant loop and condenser can be supported; component and fluid lifecycle are accepted. | Very high chip heat flux, low flow, or waterless white-space requirement outweighs ecosystem novelty. | Vendor efficiency claims are not comparable to whole-facility metered PUE. | $$$–$$$$ | Specialized fluid, pressure, recovery, and supplier skills. |
| Immersion | 100–361 kW product examples | Hardware, materials, cabling, fluid, lifting, fire, and warranty workflows are qualified as one system. | Maximum floor density, harsh/dusty environment, or fanless server design justifies a new service model. | One vendor claims <1.03; require boundary definition and meter it yourself. | $$$$ | Tank service, lifting/drip control, fluid analysis, hardware adaptation. |
02 / Fundamentals
Turn watts into flow before choosing hardware
Nearly every watt delivered to IT becomes heat in the data hall. Cooling capacity and transport are separate problems: a chiller may have spare tons while a rack starves for airflow.
Airflow
CFM ≈ 3,160 × kW ÷ ΔT°F
From Q = 1.08 × CFM × ΔT. At 20°F rise: ~158 CFM/kW. Altitude, humidity, leakage, and fan curves change the result.
Water flow
L/min ≈ 14.3 × kW ÷ ΔT°C
For water near room temperature. At a 10°C loop rise: ~1.43 L/min per kW before glycol/property and margin corrections.
Worked example — a sustained 40 kW rack at 20°F air rise
1 / Heat40 kW = 136,480 Btu/h
2 / Airflow136,480 ÷ (1.08 × 20)
3 / Result≈ 6,319 CFM
4 / Meaning~149 air changes/min through a 42U rack face*
*Illustrative velocity framing, not a fan selection. Using a nominal 24 × 79 in rack face (~13.2 ft²), 6,319 CFM implies ~479 ft/min face velocity. Real open area is smaller and pressure drop matters.
~3,000×
Volumetric heat-capacity advantageLBNL training material states it takes roughly 2,000–3,000 times the volume of air to cool what water can; the point is transport work, not that every liquid system is automatically efficient.
18–27°C
Recommended inlet bandASHRAE’s recommended air temperature envelope is 18–27°C (64.4–80.6°F). Measure server inlets—room thermostat temperature can hide recirculation.
RCI = 100%
No recommended-range excursionsDOE’s Rack Cooling Index framing: 100% means no intakes outside the recommended range; below 90% signals poor thermal conditions. Trend top/middle/bottom inlet probes.
120 kW
A real rack anchor, not a forecastNVIDIA documents approximately 120 kW for an NVL72 rack. The OCP MGX spec also uses 120 kW rack power and an 85% liquid heat-capture requirement, leaving residual air cooling.
03 / Claim reconciliation
Why “the air limit” moves
These figures describe different boundaries: installed fleet, room distribution, passive door rating, active door rating, or server thermal architecture. They do not contradict once the boundary is named.
| Quoted figure | Source / date | What it actually describes | Assumption producing the number | Decision use |
|---|---|---|---|---|
| 1–2 kW/rack | Schneider Electric, 2011 | Order-of-magnitude design density for legacy room-oriented systems. | Long shared air path; room average; older cooling distribution. | Explains why a legacy room can fail far below a modern air-cooled hall. |
| 8–12 kW/rack | LiquidStack article, 2021 | Vendor statement about what “most air cooling systems” can handle. | Unstated typical system; no containment or acoustic boundary disclosed. | Treat as sales framing; require measurements before using it. |
| over 10 kW/rack | Schneider Electric, 2011 | Operator-experience threshold in then-typical facilities. | Blade-era room cooling and limited high-density operating history. | A warning that distribution and redundancy become first-class design work. |
| ~20 kW/rack | Vertiv white paper / 2025 article | Practical conventional-air transition point and start of RDHx consideration. | Rising fan/distribution work; liquid heat exchanger relocates heat transport. | Good price-both trigger—not a physical law. |
| 20 kW/door | Vertiv Liebert XDR brochure, 2024-era publication | Passive, fanless rear-door module rating. | Server fans drive the door; specified approach and refrigerant system. | Demonstrates that “RDHx” spans passive products with modest ratings. |
| 50 kW/door | Vertiv Liebert DCD manual, 2023 | Chilled-water rear-door product family design point. | Door geometry, headroom, water conditions, and server airflow apply. | Validate site clearance, hose path, pressure drop, and N+1—not just kW. |
| 80 kW/door | Vertiv CoolLoop RDHx launch, 2025 | Maximum capacity claim for an active chilled-water door. | Product test conditions; active fans; facility water available. | Shows why “air-cooled server” can survive at 80 kW while heat transport is already liquid. |
| 100–361 kW/tank | Submer product pages, 2023–2026 | Immersion tank heat-dissipation ratings. | Tank, CDU, fluid, operating condition, and compatible IT treated as a system. | Use for footprint screening only; do not compare one tank directly with one rack. |
04 / Working knowledge
Seven regimes, seven different constraints
A regime earns its place when it changes the limiting mechanism or the operating model—not merely because a vendor renamed a coil.
CRAC / CRAH room air
Purpose: reject whole-room heat through DX refrigerant (CRAC) or chilled-water coil (CRAH).
Use: low/mixed density with measured inlet control. Gotcha: installed tons do not prove air reaches the rack; verify fan curve and return path at failure state.
Hot/cold aisle containment
Purpose: stop supply and exhaust mixing so each CFM performs useful transport.
Use: before buying exotic cooling. Gotcha: open U spaces, cable cutouts, fire-system interfaces, and doors left open defeat the pressure boundary.
In-row air
Purpose: shorten the source-to-coil path and add capacity beside a dense row.
Use: localized retrofit. Gotcha: unit failure can create a sharp local transient; distributed valves, drains, fans, and controls multiply maintenance points.
Rear-door heat exchanger
Purpose: transfer server exhaust to liquid without modifying server internals.
Use: 20–80 kW product range or staged liquid transition. Gotcha: passive doors add pressure drop to server fans; concentrated hot spots can underuse coil area.
Single-phase DTC
Purpose: circulate water/glycol or dielectric liquid through cold plates while it remains liquid.
Use: OEM-qualified CPUs/GPUs and warm-water loops. Gotcha: memory, drives, PSUs, NICs, and switches may remain air-cooled; capture ratio is never assumed to be 100%.
Two-phase DTC
Purpose: boil dielectric fluid at the cold plate and condense it in a closed loop.
Use: high heat flux or waterless white space. Current ZutaCore rack units span 20–120 kW (as of Jul 2026). Gotcha: fluid availability, recovery, seals, and supplier dependence.
Immersion
Purpose: submerge compatible IT in dielectric fluid, removing most server air paths.
Use: dense standardized fleets with redesigned service. Gotcha: warranty/material compatibility, lifting, optics/cabling, drip control, fluid aging, and tank-level fault domains.
Hybrid is the normal end state
Purpose: liquid-cool dominant silicon while air serves residual components and conventional racks.
Example: NVIDIA documents liquid-cooled CPU/GPU/NVSwitch devices and air-cooled remaining components. Gotcha: “liquid cooled” never licenses removal of the air budget.
Heat rejection is still required
Purpose: move heat from CDU/tank loop to dry cooler, cooling tower, chiller, district heat, or another sink.
Gotcha: closed-loop liquid can have zero direct water consumption and still reject heat through a water-using facility plant—or vice versa.
| Regime | Published capacity anchor | White-space fluid | Retrofit fit | Dominant failure modes | Staffing / tooling |
|---|---|---|---|---|---|
| Room air | Legacy 1–2; investigation often begins 10–20 kW | Air; CRAC/CRAH plant may use refrigerant/chilled water | Highest | Recirculation, bypass, fan/CRAH loss, filter loading | HVAC + airflow measurement |
| Containment | Screen roughly 10–20 kW | Air | High | Door/opening breach, pressure imbalance, sensor blind spot | Air balancing + controls |
| In-row | Product-specific; screen roughly 15–30 kW | Air plus refrigerant/water at row | High–medium | Local fan/valve/condensate failure, short-circuit airflow | Distributed HVAC service |
| RDHx | 20–80 kW cited products | Water/refrigerant at rack; air through server | Medium–high | Leak, hose/hinge damage, fouling, excess pressure drop | Hydronics + rack service |
| Single-phase DTC | OCP 120 kW rack / 85% capture anchor | Water/glycol or dielectric secondary loop | Medium | Leak, corrosion/chemistry, pump/CDU, cold-plate blockage | CDU, chemistry, QD and OEM procedures |
| Two-phase DTC | ZutaCore 20–120 kW current units | Dielectric phase-change fluid | Medium | Charge loss, non-condensables, condenser/control fault | Specialized fluid + sealed-loop service |
| Immersion | Submer 100–361 kW examples | Bulk dielectric fluid | Low–medium | Pump/CDU, contamination, material incompatibility, lift incident | Tank, fluid, lifting and adapted IT workflow |
05 / Edge & advanced
Design the loop, failure state, and exit path
High-density cooling is an interface problem: IT, controls, facility water, heat rejection, electrical redundancy, and service procedures have to agree on the same boundary conditions.
ASHRAE W classes: the name is the upper facility-supply temperature
As of the ASHRAE framework checked July 2026, all classes share a 2°C lower bound. W17, W27, W32, W40, W45 name their upper facility-supply temperature; W+ means above 45°C. Select equipment certified for the entire chosen class—not merely one nominal setpoint.
Decision: W17/W27 commonly implies colder infrastructure; W32/W40/W45 expands economizer/dry-cooler hours and heat-reuse quality, but only qualified IT can accept it. Check approach temperature from facility loop → CDU → technology loop → cold plate.
Hybrid load split: calculate residual air instead of quoting “fully liquid”
Use Qair = Prack × (1 − liquid heat-capture ratio). For the OCP MGX anchor, 120 kW × (1 − 0.85) = 18 kW remains outside the liquid-captured load. Then add any rack elements outside the stated boundary and validate the OEM thermal report. NVIDIA’s current GB200 documentation explicitly describes hybrid trays: major processors and switches are liquid-cooled; remaining components are air-cooled.
N+1: redundancy applies to flow and heat rejection, not just nameplate tons
Test loss of one CDU pump, one CDU, one facility pump, one heat-rejection cell, and utility-water interruption. Confirm check-valve behavior, automatic isolation, ride-through time, thermal mass, controls fail-position, and which racks shed load. A 1 MW CDU does not provide N+1 to a 900 kW load if it is the only CDU.
Climate and water: Phoenix and Minneapolis do not get the same “free cooling”
ASHRAE notes many locations can economize for as much as 50% of annual hours within the 18–27°C recommended envelope, but actual hours depend on dry-bulb, wet-bulb, humidity, air quality, and selected water class. Price annual energy and water from an hourly weather file; never transplant a vendor PUE from a different climate.
Heat reuse: higher-grade heat helps, but a customer and timing still matter
Track Energy Reuse Effectiveness (ERE) when exported heat displaces useful energy; PUE alone does not credit reuse. Warm W40/W45/W+ return can reduce temperature lift, yet value requires a nearby coincident sink, heat exchanger, backup, metering boundary, and contract. “100% recoverable” is not “100% economically reused.”
Two-phase vs single-phase: density alone does not decide
Two-phase can remove high chip heat flux at low mass flow; ZutaCore states 0.3 L/min per 1,000 W for its system and current in-rack units from 20 to 120 kW. Single-phase offers a broader water-cooling ecosystem and simpler fluid accounting. Compare qualified chip list, approach temperature, pump/compressor energy, fluid regulation and replacement, service recovery, control response, and supplier exit plan.
06 / Common mistakes
Failure begins in the assumptions
A useful design review attacks the input boundary before debating brands.
Using room average
Failure: 600 kW across 100 racks looks like 6 kW/rack while two GPU racks each peak at 80 kW.
Fix: size distribution from per-rack and per-row peaks plus coincidence; size plant from diversified load with explicit margin.
Adding PSU nameplates
Failure: nameplate maximum is mistaken for simultaneous heat. DOE’s Air Management Tool explicitly says not to substitute nameplate for measured power.
Fix: use branch/rack meters or vendor thermal reports; preserve a documented growth case.
Confusing IT and facility losses
Failure: UPS/PDU losses are blindly added to rack heat even when the equipment is outside the white space.
Fix: draw boundaries. Rack cooling uses delivered IT heat; facility plant must reject electrical and mechanical losses where they physically occur.
Believing “up to”
Failure: an 80 kW door rating is used without its entering-water temperature, flow, air volume, pressure drop, fouling, or redundancy conditions.
Fix: require a certified duty point and fan/pump curves at normal and failure states.
Assuming 100% heat capture
Failure: liquid-cooled CPUs/GPUs lead to deletion of room airflow.
Fix: use OEM component mapping and heat-capture ratio; calculate residual kW and verify inlet temperatures.
Ignoring maintenance geometry
Failure: doors cannot swing, hoses kink, tanks block egress, CDUs steal rack U, or lifted wet servers have no drip/service zone.
Fix: model service clearances, floor loading, pipe strain, containment, rigging, drainage, and concurrent maintenance.
Comparing PUE claims
Failure: a subsystem projection, vendor lab number, and annual whole-facility PUE are put in one ranking.
Fix: normalize boundary, annual weather, load factor, redundancy mode, water, and what parasitic energy is included.
Skipping controls commissioning
Failure: pumps fight, valves hunt, condensation alarms are untested, or facility and IT loops use incompatible setpoints.
Fix: functional-test sensor failure, low flow, high dew point, leak, power loss, restart, and degraded heat rejection.
No exit strategy
Failure: proprietary fluids, cold plates, manifolds, or firmware lock a fleet to one supplier.
Fix: document compatible alternatives, drain/flush plan, spares horizon, data/API ownership, and decommissioning cost.
07 / Primary evidence
Sources & boundaries
Vendor sources are used only for attributed product claims. Standards, government guidance, and published case studies anchor physics and operating practice.
- ASHRAE AI Data Center Energy Performance Framework — current W17/W27/W32/W40/W45/W+ naming and 2°C lower bound.
- ASHRAE TC 9.9 contamination guidelines — 18–27°C recommended inlet envelope.
- DOE FEMP Best Practices Guide (2024) — airflow efficiency, RCI, liquid cooling, metering, heat recovery.
- DOE/LBNL Air Management Tool manual — measured power, temperature rise, and warning against nameplate substitution.
- LBNL liquid-cooling case study — RDHx lessons and projected PUE ~1.4 → ~1.1.
- DOE Jefferson Lab case study — PUE >2 → 1.27 after containment and other upgrades.
- Uptime Institute Global Survey 2024 — 1.56 average annual PUE and fleet-density context.
- OCP MGX rack specification 1.1 — 120 kW rack, 85% capture, 130 L/min target flow, pressure and filtration requirements.
- NVIDIA DGX GB rack hardware guide — approximately 120 kW NVL72 rack and hybrid cooling hardware.
- Vertiv RDHx engineering article (2025) — attributed 20 kW transition and 20–80 kW range.
- Vertiv CoolLoop launch (2025) — up-to-80 kW active RDHx claim.
- Vertiv Liebert XD brochure — 20 kW passive XDR rating.
- Schneider Electric room/row/rack article — attributed 1–2 kW legacy room design claim.
- ZutaCore HyperCool product page — attributed two-phase DTC 20–120 kW current in-rack range.
- Submer SmartPod EXO product page — attributed up-to-361 kW immersion capacity claim.
- The Green Grid ERE white paper — metric for useful reuse energy.
Revalidation trigger: re-check product capacities, NVIDIA/OCP rack examples, PUE evidence, and W-class naming at least annually or whenever the selected IT bill of materials changes. Thermodynamic equations remain stable; fluid properties must match the actual mixture and operating temperature.