Power and cooling

A GPU is an electrical device that turns power into arithmetic and heat. Two systems therefore set the limits of any AI datacentre: how much power can be delivered to the chips, and how much heat can be carried away from them. Between them they decide how many GPUs a site can host, how soon it can start, and how much of the electricity it pays for becomes useful work.

This piece follows the electricity from the grid to the chip, explains why air cooling ran out, and describes the liquid-cooled hall that replaced it.

From the grid to the chip

Electricity arrives at a campus at transmission voltage and has to fall by several orders of magnitude before a GPU's transistors can use it, without once interrupting a training run that may represent months of work. The classical chain has seven conversion stages.

• Transformers at the utility or campus substation step the transmission voltage down to a medium-voltage ring around the campus.
• A unit substation at each data hall steps medium voltage down to low voltage.
• Uninterruptible power supplies condition the supply and bridge any gap until the generators start.
• Busways and power distribution units carry the power overhead to the racks.
• Power shelves in each rack convert alternating current to direct current on a busbar.
• Converters on each server tray step that voltage down again.
• Voltage regulators packed around each GPU deliver the final, very low voltage at very high current.

Each stage is efficient on its own. Together they leak, and the largest single loss sits in the last centimetre, beside the chip, which is the silicon vendor's territory rather than the building's. That is one reason each new GPU generation brings changes to the power design as well as to the chip.

Redundancy, rethought

Datacentre resilience is usually described in three grades: no spare equipment, one spare unit for each set, or two complete and independent paths. Cloud halls were built with two full paths almost everywhere, because a web service cannot stop.

AI training changed the sums. Training is a batch workload, and a job that saves checkpoints can survive a short outage in a hall and resume. Many operators therefore build GPU halls with less redundancy and keep full duplication for the storage, the control systems and the network core. Backup power is still mostly diesel generators, started behind batteries or flywheels that carry the load for the seconds in between, and battery storage increasingly supplements both and earns money from the grid as well. The design question has moved from how many nines of availability a hall offers to the cheapest design that keeps a cluster's goodput where the customer needs it.

A load that fights back

Synchronous training makes a GPU hall electrically unlike anything utilities have served before. Thousands of GPUs compute and then communicate in lock-step, so the hall's power draw swings sharply and rhythmically, and a software crash can drop a very large load in an instant. The swings travel into the grid, and grid disturbances travel back into the run.

Grid operators have responded with rules on how fast a large load may ramp and how it must ride through faults. Operators respond in layers: energy storage in the rack power shelves, software that smooths the load at the moments the GPUs synchronise, batteries and flywheels for the whole facility, and ramp limits written into contracts. Operators who manage power quality well win their grid approvals faster, which is a quiet but real advantage.

What PUE measures, and what it hides

Power usage effectiveness, or PUE, divides the energy a whole facility uses by the energy delivered to its IT equipment. The closer it is to one, the less is spent on cooling and other overhead. Liquid-cooled AI halls run much closer to one than older enterprise sites, especially in cold climates where outside air can do much of the cooling for free.

PUE has a blind spot. Most of the losses in the electrical chain happen after the IT meter, inside the racks, and so count as IT load. A facility can report an excellent PUE while still losing a meaningful share of its power before it reaches the chips. Sophisticated buyers now ask for efficiency from the grid all the way to the token.

The next electrical design

The long-run answer to those losses is to remove conversion stages. NVIDIA has described a high-voltage direct-current design for its coming rack generations: the supply is converted to direct current once, at the edge of the hall, distributed as direct current to the racks and converted once more beside the GPU. The low-voltage transformers, the central uninterruptible supplies and the alternating-current shelves in each rack disappear, and energy storage connects straight to the direct-current bus. The large power-equipment makers have announced products for it. For investors the point is that facilities built for the next generations are a different electrical design, and that the demand to refit existing halls reaches beyond the GPU cycle itself.

Why air ran out

Air is a poor coolant. At practical speeds it can carry only a limited amount of heat from a rack before the fans, the noise and the hot spots become unmanageable, while water carries far more heat for the same volume. Once GPU packages and racks passed a certain power, direct liquid cooling stopped being an efficiency option and became the only way to deploy the product at all. NVIDIA's Blackwell rack-scale systems send most of their heat to water through cold plates, and its Rubin racks are cooled entirely by liquid, with no fans in the trays.

Inside a liquid-cooled hall

Heat leaves a GPU through a copper cold plate mounted on the package. Tubes on each tray join the rack's manifolds through drip-free quick-disconnect couplings, and the rack loops join a secondary circuit for the hall, filled with treated coolant, usually water mixed with propylene glycol and inhibitors.

The coolant distribution unit is the heart of the system: a skid of pumps and heat exchangers that separates the clean loop serving the racks from the facility's own water loop. It holds the supply temperature above the dew point, balances the flow, and filters and monitors the coolant. The facility loop then carries the heat to dry coolers where the climate allows, or to cooling towers or chillers where it does not.

The temperatures are the commercial insight. The newest systems accept warm supply water and return it warmer still, warm enough that in most climates, for most of the year, no refrigeration is needed at all. Other designs exist. Liquid-to-air units push the heat into a hall's existing air system, which helps a refit but is capped by the room. Rear-door heat exchangers capture the exhaust of air-cooled racks. Immersion, with whole servers in a bath of fluid, remains a niche for AI training.

Water and operating risk

Two operating realities deserve attention. The first is water. Evaporative cooling towers consume a great deal of it, permits increasingly depend on water efficiency, and policy now binds in places: Johor, in Malaysia, suspended the expansion of water-cooled datacentres until mid-2027. Closed-loop dry coolers use almost no water, at the cost of a little efficiency in a heatwave, and are easier to site.

The second is the liquid itself. Thousands of fluid connections now sit above some of the most expensive equipment in the building. Mature operators treat coolant chemistry as a laboratory discipline, instrument every rack with leak detection and flow and temperature sensors, and drill their drain-and-replace procedures. Insurers now price that maturity.

What it means for investors

Power and cooling decide how many GPUs a site can host and how soon, which is why the datacentre itself has become an investment question. The equipment makers for switchgear, uninterruptible power, busways, power shelves and coolant distribution, among them Vertiv, Schneider Electric, Eaton, ABB and Delta, sit on multi-year order books, and each change in the electrical or cooling design brings a round of refits as well as new builds.

Sources

• AI Infrastructure Primer 2026, Compute Terminal: the chapters "Power: from the grid to the chip", "Cooling: getting the heat back out", "The value chain" and "Site selection and energy economics", with their dated sources.
• Johor's suspension of water-cooled datacentre expansion: CNA, 29 January 2026, as cited in the Primer.