A cutaway of a generic AI data hall. Set how many GPU racks you run, how many kilowatts each draws, how it is cooled and what the weather and the power grid are like, then read off the megawatts, the water that evaporates, the carbon, the cost and whether the chips stay below their temperature limit. All default numbers are typical or illustrative engineering values, not measurements of any company, site or product. Drag to orbit, pinch or wheel to zoom, use View to jump to a zone.
Litres of water per kWh of IT energy (WUE). Dry designs sit on zero but spend more electricity.
Junction temperature at full load from a simple thermal-resistance model, for air and for direct-to-chip liquid, against a generic 85 C limit. Dot = your setting.
Runs a full typical year (12 months x 8 times a day) for all four designs with your other settings and compares electricity, water, carbon and chip temperature. It also prices the water saved by going dry in kWh per cubic metre.
Re-computes key results by a separate route (unit conversions, the PUE definition, the first law, a textbook reference and a published wet-bulb test value).
What is modelled. One generic AI data hall seen as a steady-state energy and water balance. Electricity flows grid, transformer, UPS (uninterruptible power supply), PDU (power distribution unit) and rack. Every watt that enters is turned into heat, which must be carried out of the building. The picture shows 48 racks at most; the numbers scale to the rack count you set. All parameters are typical round values for illustration, not data from any real facility, operator, chip or product, and nothing here describes a named company or model.
IT load. PIT = Nracks × Prack × (0.30 + 0.70 u) where u is average utilisation. The 0.30 is an idle floor (accelerators and servers draw a large share of their peak when idle). Rack power ranges here from 10 kW (older air-cooled halls) to 130 kW (dense liquid-cooled AI racks).
Power chain. Transformer 99 %, double-conversion UPS 96 %, PDU and cabling 98.5 %, so about 6.7 % of the input is lost before it reaches the racks: Pin = PIT / (0.99 × 0.96 × 0.985) Lighting, controls and offices add 1 % of IT.
PUE. Power Usage Effectiveness is facility power divided by IT power:
PUE = Pfacility / PIT = 1 + (losses + cooling + other) / PITA PUE of 1.0 would mean no overhead at all. Typical modern figures are roughly 1.1 to 1.5 depending on climate and design; the value shown for "now" uses one outdoor condition, the annual value averages a typical year of temperature swings (seasonal +-9 K, daily +-4 K around your setting).
Cooling. The heat load on the plant is the electrical input minus any heat exported for reuse. Air designs use a chilled-water loop at 18 C; direct-to-chip liquid designs capture 75 % of the heat in a warm loop at 32 C and leave 25 % to air. Heat is rejected through an evaporative tower (water temperature limited by the wet-bulb temperature, reachable without a chiller if it is low enough) or a dry cooler (limited by the dry-bulb temperature). Between -3 K and +3 K around the loop set-point a growing share of the load needs a chiller, whose coefficient of performance is taken as half the Carnot value COP = 0.5 × Tevap / (Tcond − Tevap) clamped to 2.5 to 12. Warm-water liquid cooling therefore needs the chiller far less often, which is the main reason its PUE is lower.
Water. An evaporative tower rejects heat by evaporating water. One kilogram absorbs the latent heat of vaporisation, about 2.26 MJ:
mevap = φ Qrej / Lv, Lv = 2.26 MJ/kgSo 1 MW rejected purely by evaporation is 0.4425 kg/s, or 1 593 litres per hour. φ, the evaporative share of the tower heat rejection, is an empirical estimate (the rest goes as sensible heat to the air): it rises with outdoor temperature and humidity, from about 0.5 at 10 C to 0.95 in the heat. The water you must add (make-up) is evaporation plus blowdown, the bleed that stops minerals building up: mmakeup = mevap × C / (C − 1) with C the cycles of concentration (default 4). Water Usage Effectiveness is WUE = make-up water (L) / IT energy (kWh) This is on-site water only. Water used by power stations to make the electricity ("indirect" water) is not counted, and dry cooling shifts use from water to electricity, so a fair comparison should include both.
Chip temperature. A one-line thermal-resistance model at nameplate load: Tchip = Tcoolant or air in + Rth × Pchip with Pchip = 70 % of rack power spread over 72 accelerators, Rth = 0.10 K/W for an air heat sink (plus recirculation above 30 kW per rack) and 0.025 K/W for a cold plate. The 85 C limit is a generic round number, not the specification of any device. It shows why air cooling runs out of room somewhere between roughly 40 and 70 kW per rack in this model while liquid does not.
Carbon, cost, households. CO2 = annual MWh × grid carbon intensity (g/kWh); cost = annual kWh × price. "Households" divide by round illustrative figures of 2.7 MWh of electricity, 122 m3 of water (about 335 litres a day) and 12 MWh of heating per home per year. Real homes vary a lot.
Heat reuse. Warm liquid-loop water can supply district heating. Exporting a fraction of the liquid-captured heat removes it from the plant, lowering the chiller and tower load and, for evaporative designs, the water. ERE (Energy Reuse Effectiveness) = (facility power - exported heat) / IT power. Air-cooled return air is too cool to export without a heat pump, so reuse is only offered for liquid designs, and the electricity a heat pump or heat network would use is not modelled.
Independent checks. Experiment B tests (1) evaporated water x 2.26 MJ/kg against the evaporative share of the rejected heat, (2) the PUE definition, (3) the first law (electricity in = heat out + exported heat), (4) the 1 MW = 1 593 L/h reference, (5) the wet-bulb formula against a published test value (20 C, 50 % RH gives 13.7 C), and (6) the annual routine against steady power x 8 760 h.
Simplifications. Steady state at each weather sample; no part-load chiller curves, no humidity control, no adiabatic assist, no redundancy or standby plant, no on-site generation, no water treatment details, and a single average utilisation. The tower evaporation share, power-chain efficiencies and thermal resistances are round values for teaching, not design data. A real facility differs and depends on its design, climate and operations. This is an independent educational model, not affiliated with or describing any company.