Data centres: Science and limits

Why every watt a server draws ends up as heat

What this tab covers What physically happens in the conversion, step by step, the science behind it, and the hard limits that physics sets on how well it can ever work.

A data centre turns electricity into computation, and then all of it into low-grade heat that must be removed. Power passes through transformers, an uninterruptible power supply, server power supplies and on-board regulators before reaching the chips; cooling and these losses are measured together by PUE (power usage effectiveness).

The conversion, step by step

Grid to site substation and switchgear
Electrical (high-voltage AC) → Electrical (medium and low-voltage AC) · Small transformer losses, counted in facility overhead; not quantified in the sources usedUninterruptible power supply (UPS)
Electrical AC → Conditioned AC (via DC and battery in double conversion) · Up to 97% in double conversion and above 99% in eco mode (manufacturer claim for one Vertiv model)⁶Server power supply unit (PSU)
Electrical AC (230 V) → Electrical DC (12 V or 48 V) · 80 PLUS Titanium, 230 V redundant: at least 90% at 10% load, 94% at 20%, 96% at 50% and 91% at full load⁵Voltage regulators to chips
Electrical DC (12 or 48 V) → Electrical DC (below about 1 V at very high current) · Typically around 90% (assumption, not from the sources used)Processors and GPUs compute
Electrical → Information processing, ending entirely as heat · H100 GPU about 2.8 TFLOPS per watt at FP16 with sparsity, half that for dense maths (calculated: 1,979 TFLOPS at 700 W); a whole GB200 NVL72 rack about 2.7 TFLOPS per watt on the same basis (calculated: 360 PFLOPS at up to 132 kW)⁷¹¹¹²Cooling removes the heat
Heat in the IT room plus electrical (fans, pumps, chillers) → Heat rejected outdoors (sometimes reused) · Average PUE 1.54 in 2025, so 54% extra energy on top of IT; Google fleet 1.09²³End to end
Grid electricity bought → Electricity reaching the chips · About 56% at PUE 1.54 and about 79% at PUE 1.09 (calculated: 1/PUE x 96% PSU x an assumed 90% regulator)²³⁵

The science

Computer chips are made of billions of transistors that switch on and off. Each switch charges and discharges tiny capacitances in wires and gates, and the energy stored in them is dissipated as heat in the resistance of the circuit; leakage currents add more heat even when nothing switches. So a server is, in energy terms, an electric heater that does useful computing along the way: practically every watt drawn ends up as heat. Removing that heat is a refrigeration problem. Air cooling blows cold air over heatsinks; liquid cooling pumps water or a coolant through cold plates fixed to the chips, which carries far more heat per unit volume because water's heat capacity per litre is about 3,500 times that of air (calculated). Where the outdoor air or water is cool enough, heat can be dumped without running a compressor (free cooling); otherwise a chiller, which is a heat pump, lifts the heat to a temperature high enough to reject it outdoors. The higher the coolant temperature the chips tolerate, the less work the chiller does.

Limits set by science

Limit	Value	Why it exists	Best today
PUE floor	1.0 means every unit of electricity goes to IT equipment²³	PUE is total facility energy divided by IT energy. At 1.54, IT gets about 65% of the energy; at 1.09, about 92% (calculated).	Google fleet average 1.09 in 2025, against an industry average of 1.54 (Uptime Institute)
Carnot limit for chillers	COP max = Tc / (Th - Tc) with temperatures in kelvin; for 20 C coolant and 35 C outdoor air, 293 / 15 = about 19.5 (calculated)	A chiller's coefficient of performance (heat moved per unit of electricity) rises sharply as the gap between coolant and outdoor temperature shrinks, and a chiller is not needed at all when outdoor air is colder than the coolant.	Warm-water liquid cooling and free cooling avoid chillers for much of the year
Landauer limit for erasing a bit	kT ln 2, about 2.9 x 10^-21 J at 300 K (calculated)⁷¹⁰	Physics sets a minimum heat for irreversibly erasing one bit. An H100 uses roughly 4 to 7 x 10^-13 J per floating-point operation (calculated from its rated FP₁₆ throughput and power), about a hundred million times higher, so the limit is far from binding; wiring and moving data to and from memory dominate.	Gains come from smaller transistors, lower voltages, lower-precision arithmetic and moving data less
Heat density of air cooling	Over 80% of surveyed operators have no racks above 30 kW; a liquid-cooled GB200 NVL72 rack is designed for up to 132 kW²¹²	Air can only carry so much heat before fan power, airflow and noise become impractical, so very dense AI racks move to liquid cooling.	Direct-to-chip liquid cooling at 100 kW or more per rack

Sources

Uptime Institute Global Data Center Survey 2025 Uptime Institute, 2025
Data centers: efficiency (power usage effectiveness) Google, 2026
80 PLUS: why you need Titanium power supplies in your data center (white paper restating the 80 PLUS thresholds) Bel Fuse, 2023
Liebert EXL S1 UPS (product page) Vertiv, 2025
NVIDIA H100 Tensor Core GPU (specifications) Nvidia, 2025
Irreversibility and heat generation in the computing process (Landauer, IBM Journal of Research and Development 5(3)) IBM, 1961
NVIDIA GB200 NVL72 (specifications) Nvidia, 2025
Vertiv and Nvidia collaborate on 7MW reference architecture for GB200 NVL72 DatacenterDynamics, 2024