Data Centres & Power: The Physical Side of AI

We've priced the chips, the training runs, and the tokens. One layer remains, and it's the most physical one: the buildings. "The cloud" has always been a warehouse full of machines somewhere — but AI has taken that warehouse and turned it into some of the most power-hungry infrastructure humanity builds. This chapter is about where AI physically happens, and why electricity is quietly becoming the industry's binding constraint.

Your prompt goes to a warehouse

When you hit send, your message travels across the internet to a data centre: a warehouse of server racks, somewhere chosen very deliberately — usually where land and electricity are cheap and the climate helps with cooling. Inside, GPU clusters run your prefill and decode, and the reply streams back.

Data centres long predate AI — they run websites, streaming, and everything else the internet does. What AI changed is density. An ordinary server rack draws power like a handful of homes. A rack stuffed with AI GPUs draws around ten times that — and the newest generations push higher still. Same building shape, ten times the appetite, which breaks every assumption the old buildings were designed around: the power feeds, the cooling, the floor layout, everything.

A sense of scale

Power is measured in watts, and AI's story is watts multiplying:

Laptop

~50 W

One AI GPU

~700 W

One GPU rack

~100 kW

Training cluster

~30 MW

AI campus (planned)

GW-scale

Read that bottom row again. A gigawatt is roughly the output of a full-size nuclear reactor — and the largest announced AI campuses are planned in multiples of it. When the overview chapter said AI power deals are signed in units previously reserved for national planning, this is what it meant. A frontier training run is, energetically, a small town that exists for a few months; the inference fleet serving a popular AI product is a permanent one.

The heat problem

Physics offers no exceptions: every watt a chip consumes becomes heat, and heat must leave the building or the chips throttle and die. Cooling has therefore become AI's second engineering discipline:

• Air cooling — the classic hot-aisle/cold-aisle fan setup — tops out well below what a modern GPU rack produces. It simply can't move heat away fast enough at AI densities.
• Liquid cooling is the answer: coolant piped directly across the chips, like a car engine. Most new AI data centres are designed liquid-first, and older facilities are being retrofitted.
• Water enters the story because many designs evaporate it for cheap cooling — which is why data-centre water usage has become a genuine local politics issue in dry regions, alongside the electricity itself.

Power is the new bottleneck

Here's the twist of the mid-2020s: for years the scarce resource was the GPUs themselves. Increasingly, it's the electricity to run them. You can order chips; you cannot order a grid connection — hooking a gigawatt-scale campus to the power grid takes years of planning, permitting, and transmission build-out. The consequences are reshaping the industry's map:

• Siting follows power. New AI data centres go where electricity is abundant and cheap — near hydro dams, gas fields, wind corridors — rather than near users. Your prompt is answered wherever the power is.
• Tech companies became energy dealmakers. The hyperscalers now sign decade-long power purchase agreements, fund new generation outright, and — most strikingly — contract with nuclear plants, including deals to restart shuttered reactors and orders for small modular reactors. AI has done what decades of advocacy couldn't: made big tech desperate for nuclear power.
• The grid feels it. In data-centre-heavy regions, AI demand is now visible in national electricity statistics and local rates — which makes the buildout a political story, not just a technical one.

The wager underneath it all

Add it up — land, buildings, chips by the hundred thousand, cooling plants, power contracts — and the AI infrastructure buildout runs to hundreds of billions of dollars per year across the big players, one of the largest private infrastructure bets in history, in the same conversation as railways and telecoms booms past.

And it is a bet, because of a fact from earlier in this guide: the GPUs inside depreciate in just a few years, and the models they produce go stale even faster. A dam earns for a century; an AI data centre's contents must pay for themselves before their own obsolescence. The wager is that token demand keeps growing fast enough to fill every building being poured today. Whether that wager pays — and who collects — is exactly where this guide ends up two chapters from now.

Recap

• AI physically happens in data centres, and AI racks draw ~10× the power of ordinary server racks — density is what changed.
• The scale ladder: one GPU ≈ a microwave; a rack ≈ dozens of homes; a training cluster ≈ a town; the newest campuses are planned in gigawatts — reactor-scale units.
• Every watt becomes heat, pushing the industry from air to liquid cooling, and making water and efficiency (PUE) first-order concerns.
• Power, not chips, is becoming the bottleneck — siting follows cheap electricity, and AI companies now sign nuclear-scale, decade-long energy deals.
• The buildout costs hundreds of billions a year against hardware that depreciates in a few years — a historic infrastructure wager on token demand.

Given all this — costlier chips, hungrier buildings, reactor-scale power — the next chapter's fact should feel impossible: the price of using AI keeps collapsing. Why AI keeps getting cheaper.