Power and cooling (the inconvenient physics)¶

Modern AI infrastructure has discovered what the Alchemists’ Guild learned centuries ago: performing impressive feats requires consuming alarming quantities of energy and produces heat in proportions that make everyone nearby uncomfortably aware of the Second Law of Thermodynamics. A data centre training large models draws electricity at the rate of a small town while generating enough waste heat to keep that town warm through winter, if only the heat could be delivered rather than desperately removed before the hardware melts.

The physics is straightforward and unhelpful. Every computation requires energy, the energy becomes heat, and AI involves staggering amounts of computation. None of this can be optimised away in software; it can only be managed with engineering, money, and locations where power and cooling are abundant.

Power at scale¶

A single high-end accelerator draws roughly 700 watts at full load, a training cluster of ten thousand draws megawatts before counting networking, storage and cooling, and a large AI campus reaches into the tens or hundreds of megawatts, which approaches the output of small power stations and exceeds what most local grids were designed to host. At industrial rates the electricity bill for such a facility runs to tens of millions of euros a year, and over a cluster’s life the power can cost more than the hardware. Securing supply means negotiating with utilities, building substations, and occasionally explaining to local authorities why an activity producing no tangible goods needs a town’s worth of electricity. Those conversations reportedly do not always go smoothly.

The thermodynamics problem¶

Heat removal is not optional. Racks drawing 40 to 80 kilowatts have outrun air cooling, so liquid arrives at the cold plate, with immersion at the exotic end; and the heat, once collected, still has to go somewhere, which means chillers, cooling towers, and in many designs prodigious quantities of water. Evaporative cooling can consume millions of litres a day, which becomes a political question in any region that also drinks. Waste heat reuse for district heating is occasionally discussed, mostly discussed; the temperatures are awkward, the consumers are far away, and the economics rarely close.

Sustainability contradictions¶

The industry’s environmental claims and its infrastructure reality point in different directions. Carbon neutrality is announced while power draw grows; renewable certificates are purchased while the grid supplies whatever it has, which at peak is often gas; and each generation of hardware is more efficient per operation while total consumption rises anyway, because the workload grows faster than the efficiency. The honest statement of the tension is short: AI requires computation, computation requires energy, and the claim that this is sustainable rests either on the value of the computation or on renewable supply that could not have displaced fossil fuel elsewhere. Neither is obvious, and both are argued mostly by people with a financial interest in the answer.

The constraints that bind¶

Grid connections of tens to hundreds of megawatts take years to grant and build; cooling water is available where it is available; and suitable sites, flat, connected, politically stable, near power and water, are scarcer than capital. This is why data centres cluster, why announcements precede operations by years, and why location decisions are increasingly made by the electricity price rather than the customer base. The realistic outlook is not a solution but a management regime: better hardware, more renewables, cleverer cooling, and geography, while the physics remains completely indifferent to preferences, budgets, and commitments.

The clerk’s brief¶

From the clerks, for the Patrician’s eyes

Compiled July 2026. Newest first; the accumulating baseline is kept in the meter readings at the end. The clerks confess a fondness for this file: the numbers in it are the only ones in the section that obey conservation laws.

November 2025: The price signal arrives¶

Market reporting in November 2025 recorded capacity prices in the PJM electricity market at 329.17 dollars per megawatt-day for 2026 to 2027, more than ten times the year before, with study estimates of average US household bills rising 8 per cent by 2030 and far more in data centre country. The grid has started charging the boom for its externalities, or at least redistributing them. The clerks note that when the meter becomes political, siting decisions follow, and Europe’s utilities are reading the same reports.

Pew Research’s review of US data centre energy, published October 2025, put data centres at 4 per cent of US electricity in 2024, 183 terawatt-hours, projected to more than double by 2030, with direct water consumption around 17 billion gallons in 2023 and rising. The numbers are American but the pattern is exportable. The clerks file the water figures next to the electricity ones, since droughts, unlike regulators, do not grant exemptions.

August 2025: Demand outruns the grid¶

Analysis in Semiconductor Engineering, published August 2025, estimated AI data centres consuming energy at roughly four times the rate at which electricity is being added to grids, with US data centre consumption projected at anywhere from 6.7 to 12 per cent of the national total by 2028. A ratio of four to one between demand growth and supply growth is not a forecast dispute; it is a queue forming. The clerks observe that queues at fixed resources are where the section’s other files usually begin.

January 2025: The official baseline¶

The Lawrence Berkeley National Laboratory report on US data centre energy, published January 2025 under DOE commission, established the historical series from 2014 and scenario ranges to 2028, giving the argument an official baseline after years of estimates that disagreed by factors of two. The clerks approve of baselines. Arguments with baselines end sooner, or at least end citing the same document.

The meter readings¶

The readings accumulate in one direction. The International Energy Agency’s projection that global data centre consumption would pass a petawatt-hour in 2026, circulating since late 2024, roughly Japan’s entire usage, no longer looks conservative. Official baselines exist as of January 2025, the demand-versus-grid ratio was put near four to one in August 2025, the US share was measured at 4 per cent in October 2025, and the capacity price signal arrived in November 2025. The clerks’ standing assessment: the physics has not changed and will not, the costs are now visible on meters rather than in models, and the section’s usual remedy, hoping, performs especially poorly against thermodynamics.