AI Data Center Power, Cooling and Water Calculator

The calculation uses a constant mean operating point. IT power is $P_{IT}=N P_{rack} f_{load}$. Support electricity is $P_{aux}=f_{aux}P_{IT}$, and all of that heat is conservatively assigned to the equivalent cooling loop: $Q_{loop}=P_{IT}+P_{aux}$. A chiller uses $P_{ch}=Q_{loop}/COP$; the dry-cooler option has no compressor. The grid meter supplies $(P_{IT}+P_{aux}+P_{ch})/\eta_{el}$. PUE is grid energy divided by IT energy for the same period, following The Green Grid WP49 (2012). Distribution losses are rejected separately, so $Q_{reject}+P_{loss}=P_{grid}$.

The net condenser/tower duty includes compressor input: $Q_{reject}=Q_{loop}+P_{ch}$. Water latent heat is approximated as 2430 kJ/kg near 37 °C (OpenStax, §1.5, Table 1.4 footnote), with density 1 kg/L. This gives 1.481 L/kWh of latent heat, approximately the 1.5 L/kWh rule of thumb. Multiply by the entered latent-to-net heat ratio and net tower duty to obtain evaporation. A ratio above one allows sensible heat from ambient air to supply additional evaporation. The signed sensible transfer is $Q_{sens}=Q_{reject}-Q_{lat}$; a negative value means heat gained from the air. This follows the heat-transfer interpretation in ASHRAE Handbook—HVAC Systems and Equipment (2020), chapter 40, §1. Net site heat rejection still equals grid input. The non-evaporative bleed budget is $B=E/(COC-1)$ and makeup is $M=E+B$, from water and dissolved-solids conservation. Drift is included in B. The cooling site WUE estimate is makeup divided by IT energy, consistent with the site-water boundary in The Green Grid WP35 (2011), pp.6–7. Other site water uses must be added for a complete measured facility WUE.

Equivalent freshwater-loop flow follows $Q=\rho c_p \dot V\Delta T$, with $c_p=4.186$ kJ/(kg·K), OpenStax §1.4, Table 1.3. Wet bulb uses Stull (2011), equation (1), with temperature in °C, humidity in percent inside the correlation, and inverse tangents in radians. The result is capped at dry bulb to remove a small supersaturation artifact of the empirical fit near 99% RH. Wet-tower ambient is wet bulb; dry heat rejection uses dry bulb. Add the entered approach for an indicative heat-sink temperature. Compare dry bulb plus approach against required supply temperature for dry-cooler feasibility only.

Electricity water is the entered consumptive factor times metered grid energy. Combined modeled water-use intensity is cooling makeup/IT energy plus PUE times that factor (WP35, equation 4). Evaporation plus electricity water is also computed separately as modeled consumption; it excludes bleed. The first-law balances are derived here; published reference cases test the PUE aggregation and tower salt balance.

Chiller COP is screened against OpenStax §4.5, equation (4.6): $COP_C=T_c/(T_h-T_c)$ for positive lift, with absolute temperatures in K. The screening uses return-water temperature as a generous upper bound for $T_c$, and ambient dry bulb (air cooling) or wet bulb (wet cooling) as a lower bound for $T_h$. Finite exchanger approaches are omitted, making this a deliberately loose necessary limit. The error condition requires $COP\,\max(T_h-T_c,0)/T_c<1$. With no positive lift this screen imposes no finite ceiling. Passing the screen does not predict an achievable COP; real equipment needs performance data. The sensitivity chart omits COP values at or above that ceiling.

• Rack power is whole-rack AC input, including internal conversion losses, CPUs, GPUs, memory and networking. External electrical efficiency must not count rack PSU losses twice. The editable 120 kW default is illustrative; NVIDIA and Supermicro configurations differ. Vendor information checked 25 October 2026.
• Mean electrical load fraction is not GPU utilization. Idle draw must be included in the mean. All non-compressor support loads are lumped into the support fraction; their heat enters the modeled loop. This is conservative if some reject heat directly outdoors.
• Electrical losses upstream of the IT/plant boundary reject heat separately. If those losses enter the same cooled space, include their additional cooling duty in a more detailed facility model.
• COP, electrical efficiency, support fraction and latent-to-net heat ratio stay constant for the selected run. COP 6, support fraction 0.04, latent-to-net heat ratio 0.8 and a 5 K approach are transparent scenario assumptions, not equipment guarantees or LBNL averages.
• Water-loop flow is an equivalent combined-duty flow. Real air and liquid circuits, coolant mixtures, CDUs and heat exchangers need separate selections. Water properties are constant; no glycol correction.
• Drift and blowdown share the dissolved-solids discharge budget. No leaks, overflow, recovery, reclaim credit, humidification, domestic or maintenance water are included. Zero tower makeup for air or dry cooling covers the modeled cooling process only.
• A zero electricity water factor means that upstream water is excluded. Use a locally justified consumptive-water factor; withdrawal factors and historical national averages are not interchangeable.
• Latent-to-net heat ratio is a tower-performance input, not a fraction capped at one. Hot, dry air can cool sensibly in the tower and drive evaporation whose latent heat exceeds the net facility duty. The permitted 0.1–10 input range is a screening range, not a universal physical bound.

For preliminary comparisons and education, not detailed plant selection or measured annual PUE certification. The ambient range is restricted to 5–45 °C and RH 0.10–0.99, inside the useful warm-weather part of Stull's near-sea-level fit. Its mean absolute wet-bulb error is below about 0.3 °C over the published validity range, with larger local errors possible. It is not a psychrometric sizing model at altitude.

A dry cooler can only meet the specified supply temperature when dry bulb plus approach does not exceed it. Passing this temperature screen does not establish heat-exchanger capacity, fan power, redundancy or cold-plate acceptability. A chiller can lift heat across a temperature difference; its COP must be entered for the actual supply and rejection conditions. The model does not derive a COP or latent-to-net heat ratio from ambient weather. For annual studies, run representative seasonal operating points and sum energy and water before dividing by annual IT energy.

COC must exceed one and must be supported by water chemistry and treatment. DOE FEMP's statement that going from three to six cycles reduces blowdown by 50% conflicts with its own balance: E/2 becomes E/5, a 60% reduction. The Illinois first worked table agrees with the 60% result; the 20% makeup reduction agrees in both sources. The calculator retains the conserved balance.

Local supply demand (makeup), consumptive evaporation and electricity-generation consumption have different boundaries. LBNL 2024 regional/national estimates and IEA Energy and AI (2025) provide context, not a site forecast. No 2026 capacity prices, tariffs, regulations or opinion polls enter the equations. Check utility, water and environmental requirements in force for the site. Installation and commissioning work belongs to appropriately qualified people.