How data centres, cloud and AI compute work
What a data centre is, who builds and rents it, how cloud and AI compute sit on top, and why power is the scarce input.
Industry brief, with a one-minute summary: Data centres, cloud and AI computeFirm processes and online tests change from year to year and differ by office. Use this to prepare, and confirm the exact current steps on the firm's own careers page.
Key takeaways
- A data centre turns electricity into computing.
- A data centre is a secure building full of servers (computers without screens), storage and network equipment, with backup power, cooling and fibre connections.
- A traditional server rack (a cabinet of servers about two metres tall) drew a few kW.
Key idea
A data centre turns electricity into computing. The building, the power supply and the cooling are a property business with long leases; the chips inside are a technology business that goes out of date in a few years. For AI, the scarce input is no longer land or money but a large, reliable grid connection.
A data centre is a secure building full of servers (computers without screens), storage and network equipment, with backup power, cooling and fibre connections. Its size is measured in megawatts (MW) of IT load: the power the computers themselves can draw. A small enterprise site might be 1 MW; a large AI campus can be hundreds of MW or even a gigawatt (GW, 1,000 MW). Nearly all the electricity a server uses turns into heat, so every MW of computing needs cooling to remove about a MW of heat. Two words come up in every case: colocation means renting space, power and cooling in a shared data centre for servers the customer owns, and a hyperscaler is one of the very large cloud companies (Amazon, Microsoft, Google) that build huge data centres mainly for their own cloud services.
| Model | What the customer buys | How it is priced | Examples |
|---|---|---|---|
| Enterprise (own use) | Nothing: a bank or government runs its own site | Internal cost | Banks, telecoms, governments |
| Retail colocation | Space, power and cooling for a few racks, plus connections to many networks | Per kW per month, plus cross-connect fees | Equinix, Digital Realty, NTT, STT GDC |
| Wholesale or hyperscale colocation | A whole hall or building, often built to order for one large tenant, on a lease of 10 to 15 years or more | Per kW per month or per MW per year, lower unit price | AirTrunk, Vantage, Khazna, Yotta, CtrlS |
| Cloud and AI compute | Computing by the hour or by use: virtual servers, storage, GPU hours, AI model calls | Per hour, per gigabyte, per token | AWS, Microsoft Azure, Google Cloud, Oracle, CoreWeave |
So-what
The further up the table you go, the more the business looks like property with long, steady leases. The further down, the more it looks like a technology service with higher margins and faster-changing prices.
- From electricity to AI answers
- Key: Power and landGrid connection, power purchase agreements (PPAs), sometimes on-site gas or nuclear deals. Years to secure.
- Design and buildDevelopers and builders; electrical gear (transformers, switchgear, backup generators, batteries) and cooling from firms such as Schneider Electric, Vertiv and Eaton.
- Key: Chips and serversGPUs and AI accelerators (Nvidia, AMD, custom chips from Google, Amazon, Microsoft), memory, networking; servers assembled by firms such as Foxconn, Quanta, Dell, Supermicro.
- Operate the facilityColocation firms or the hyperscaler itself: keep power, cooling and security running.
- Sell computeCloud providers and GPU clouds (neoclouds) rent computing by the hour or by use.
- Software and AI servicesAI labs, software companies and enterprises that build products on top of the compute.
Power and chips are the two big inputs; the money is made where they are combined and sold.
Training versus inference
AI compute has two jobs. Training teaches a model from huge amounts of data: it runs for weeks on thousands of GPUs (graphics processing units, chips that do many small calculations at once) packed tightly together, and it can sit far from users because nobody waits for the answer in real time. Inference is using the trained model to answer a question: each answer is small, but there are billions of them, and they need to be fast and close enough to users. Training favours giant campuses wherever power is cheap and plentiful; inference spreads across more sites nearer cities. As AI products reach more users, inference becomes a larger share of demand.
Why power and cooling are the binding constraint
A traditional server rack (a cabinet of servers about two metres tall) drew a few kW. Uptime Institute's 2025 survey finds average densities still rising slowly, driven by racks of 10 to 30 kW; racks above 30 kW are still the exception. An AI rack is in a different league: a rack of 72 of Nvidia's newest GPUs (the GB200 NVL72) draws about 120 kW, with power supplies rated for about 132 kW, which air cannot cool, so it needs liquid cooling piped to the chips. That means older buildings often cannot host AI at all, and new ones need far more power per square metre. The IEA estimates data centres used about 415 terawatt hours (TWh) of electricity in 2024, about 1.5 percent of the world total, and projects about 945 TWh by 2030 in its base case. In some places the share is already large: Ireland's statistics office reports data centres used 23 percent of the country's metered electricity in 2025, up from 5 percent in 2015.
| Measure | Plain definition |
|---|---|
| IT load (MW) | Power the servers can draw; the main measure of capacity and of what is leased |
| PUE (power usage effectiveness) | Total site power divided by IT power; 1.0 would mean no power for cooling or losses. Uptime Institute reports an industry average of about 1.54 in 2025, flat for six years; new AI sites aim for about 1.1 to 1.3 |
| Rack density (kW per rack) | Power drawn by one cabinet of servers; a few kW for old IT, over 100 kW for the newest AI racks |
| Leased or utilised share | Share of built capacity that is rented (colocation) or busy (cloud and GPUs) |
| Price per kW per month | The usual colocation rent, for space, power delivery and cooling |
| Price per GPU hour | The usual AI compute price: one GPU rented for one hour |
| Time to power | How long until the grid can deliver the power; often the deciding factor for a site |
So-what
In a data centre case, ask for MW first: MW available, MW leased, MW under construction, and when the grid can deliver more.
Why are AI data centres more often limited by power than by land?
What is the main difference between training and inference?
Sources for this lesson (6)
- International Energy Agency: Energy and AI, executive summary (official, April 2025)
- Uptime Institute: Global Data Center Survey 2025 (July 2025)
- Business Wire: Uptime Institute 15th annual Global Data Center Survey results, average PUE 1.54 (July 2025)
- Supermicro: SuperServer SRS-GB200-NVL72 rack specification, power supplies rated at a total of 132 kW (official product page)
- Central Statistics Office Ireland: Data Centres Metered Electricity Consumption 2025 (official, July 2026)
- Recognized public explanations of case-interview concepts and frameworks
My notes on this lesson
0 of 5,000 characters. Saves automatically.
Try the 2 remaining checks and drills above to complete this lesson (0 of 2 done).
Spotted something wrong or out of date? Report a mistake. We check every report and correct the page.