Data centre glossary: power, cooling, money and AI terms in plain English

A dry cooler that sprays or wets pads to pre-cool the incoming air on hot days, so it can reject heat with less fan power and uses water only part of the year.

Why it matters: It sits between dry coolers (almost no water) and cooling towers (most water), which makes it a common choice where water is scarce.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

An engine-driven generator, usually diesel or gas, that powers the site if the grid fails for longer.

Why it matters: Generator permits and emissions are often a planning hurdle.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A large battery installation that stores electricity and releases it when needed.

Why it matters: Data centres are adding batteries to ride through faults, shave peaks and help the grid.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Money a company spends on long-lived assets such as buildings, servers and equipment.

Why it matters: Big Tech capex is the clearest public signal of how much is being built for AI.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

The unit that moves coolant to the racks and swaps heat with the building water loop through a heat exchanger, keeping the two loops separate.

Why it matters: Every liquid-cooled hall needs CDUs, so they are a key purchase in AI data centre projects.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A machine that cools water using a refrigeration cycle, so the water can then cool the data hall.

Why it matters: Chillers use a lot of power. Warmer liquid cooling loops can need less chiller time.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A data centre where many customers rent space, power and cooling for their own equipment.

Why it matters: Colocation landlords build much of the capacity that clouds and AI companies lease.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

How much text a model can take in at once, measured in tokens.

Why it matters: A bigger window lets a model read long documents in one go.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A tower that removes heat from water by letting some of it evaporate into air pulled through by a fan. The water left behind is cooler and goes back to the chillers or heat exchangers.

Why it matters: Cooling towers can cut electricity use for cooling but consume water, and their fans are among the loudest plant on a data centre site.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A form of liquid cooling where cold plates sit on the hottest chips and coolant flows through them.

Why it matters: It is the most common way to liquid-cool AI servers today.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Using cool outside air or water to cool the data centre without running the refrigeration part of a chiller.

Why it matters: It cuts energy use and improves PUE, especially in cooler climates.

Definition source: en.wikipedia.org (opens in a new tab).

A processor with thousands of small cores, first made for graphics and now the main chip for training and running AI models.

Why it matters: GPU supply and power needs shape AI data centre design.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

The waiting list of projects asking the grid operator to connect them. Many requests never get built.

Why it matters: In many markets the wait for power, not land or money, decides when a data centre can open.

Where we use it

The share of sales left after paying the direct cost of making and delivering the products, before research, sales and admin costs.

Why it matters: A high gross margin means buyers pay well above what the products cost to make, a sign of pricing power.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Stacks of memory chips placed right next to an AI chip inside the same package, so data moves to the processor much faster than from ordinary memory.

Why it matters: Memory bandwidth often limits how fast AI models answer, and HBM power is part of the chip's heat load.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A company that runs cloud infrastructure at massive scale, such as Amazon, Microsoft, Google and Meta.

Why it matters: Their spending sets the pace of the whole data centre market.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Servers are submerged in a non-conductive liquid that absorbs their heat.

Why it matters: An alternative to cold plates for very dense or harsh sites, with different maintenance and fluid questions.

Definition source: en.wikipedia.org (opens in a new tab).

How much electrical power one rack of servers draws, in kilowatts. All of that power turns into heat that the cooling system must remove.

Why it matters: AI racks draw far more than traditional racks, which is why liquid cooling and bigger power feeds are spreading. Our IT equipment page lists sourced figures for real systems.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

An AI model trained on large amounts of text that can read and write language.

Why it matters: LLMs power chat assistants and agents, and training and running them is what fills AI data centres.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Cooling servers with a liquid, usually water or a water-glycol mix, instead of only with air. Liquid carries heat away far better than air.

Why it matters: High-density AI racks are hard or impossible to cool with air alone, so liquid cooling decides which sites can host them.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

An open standard for connecting AI applications to outside tools and data, such as files, databases and apps.

Why it matters: MCP decides which tools an AI agent can use.

Definition source: modelcontextprotocol.io (opens in a new tab). Where we use it

A model design where only part of the network, the relevant experts, runs for each token.

Why it matters: It lets very large models run with less compute per answer.

Definition source: en.wikipedia.org (opens in a new tab).

Megawatt and gigawatt. 1 GW is 1,000 MW. Data centre size is usually quoted in the power it can draw.

Why it matters: Comparing projects in MW is more useful than comparing floor area.

Where we use it

A newer cloud company that mainly rents out GPUs for AI, rather than offering a full range of cloud services.

Why it matters: Many are financed with debt, so analysts watch them as an early sign of strain in the AI boom.

Where we use it

A model whose trained parameters are published so anyone can download and run it.

Why it matters: Open weights let companies run AI on their own hardware.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

What investors value a company at straight after a funding round, including the new money they put in. Pre-money valuation plus the amount raised.

Why it matters: AI labs announce their rounds this way, so it is the figure to compare between Anthropic, OpenAI and others.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A long-term contract to buy electricity from a specific generator, often a wind, solar or nuclear plant.

Why it matters: Big tech uses PPAs to back clean-energy claims and to secure supply.

Definition source: en.wikipedia.org (opens in a new tab).

Total power the whole data centre draws divided by the power that reaches the IT equipment. 1.0 would mean every watt goes to the computers. Higher means more is spent on cooling, power conversion and lighting.

Why it matters: It is the most quoted efficiency figure for a data centre, so it shows up in planning rules, sustainability reports and sales material.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A listed company that owns property and pays out most of its income to shareholders. Equinix and Digital Realty are data centre REITs.

Why it matters: Their results show demand for leased data centre space.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Recent revenue scaled up to a full year, for example one month's revenue times 12. It is the company's own measure, not audited annual revenue.

Why it matters: AI labs quote run-rate revenue when they raise money, so it is the usual yardstick for whether a valuation looks high or low.

Definition source: investopedia.com (opens in a new tab). Where we use it

Sound power is the total noise a machine gives off, whatever the distance. Sound pressure is what you hear at a given spot. Both are given in decibels, A-weighted to match the human ear.

Why it matters: Planning noise assessments start from each machine's sound power and work out the sound pressure at the nearest homes.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

The amount of heat a chip is designed to give off under heavy load, in watts.

Why it matters: Chip TDP drives rack density, and rack density drives the cooling choice.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

A small chunk of text, often part of a word, that an AI model reads or writes. Model prices are quoted per million tokens.

Why it matters: Token prices and volumes decide what an AI service costs to run.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Google's own AI chip, the Tensor Processing Unit.

Why it matters: Custom chips like TPUs compete with GPUs and change who buys what.

Definition source: en.wikipedia.org (opens in a new tab).

Batteries and power electronics that keep IT equipment running for the seconds or minutes between a grid fault and generators starting.

Why it matters: No UPS means a grid blip can take servers down.

Definition source: en.wikipedia.org (opens in a new tab). Where we use it

Litres of water a data centre uses for cooling and humidity control for each kilowatt-hour that reaches its IT equipment.

Why it matters: Water is a growing planning issue, especially where evaporative cooling is used in dry regions.

Definition source: en.wikipedia.org (opens in a new tab).

The trained AI itself. Takes text, images or audio in and produces an answer out. It does not act on its own.

Examples: Gemini 4 Argon, Claude Haiku 5.5, GPT-6.1 Terra.

More

An app or chat interface built on a model that helps a person, who stays in charge of each step.

Examples: ChatGPT, Claude, Gemini apps.

Software that replies or acts automatically, often by fixed rules. Modern bots may use a model inside.

Examples: Support chat bots, Copilot Studio standard harness agents.

Source: learn.microsoft.com (opens in a new tab)

A model plus tools and a loop: it plans, calls tools, checks results and repeats until the task is done.

Examples: Claude Code, Antigravity agent, Agents SDK agents.

Source: docs.anthropic.com (opens in a new tab)