Technology and media (6 of 8)

Data centres, cloud and AI compute

About 9 minutes to read in full, or 1 minute for the short version belowFacts checked 30 September 2026

In one minute

Companies build buildings full of computers, fill them with power and cooling, and rent out space or computing time to businesses, cloud users and AI labs.

The big idea: A data centre turns electricity into computing. The building is a long-lived property asset rented on long leases, while the chips inside lose value in a few years, so the two halves have very different economics. For AI, the scarce input is a large, reliable grid connection, and whoever has power, chips and signed customers at the same time wins.

One unit, in numbers: One AI GPU rented out by the hour for a year: USD 15,330 comes in, and USD 5,235 (34%) is left after its own costs.What is left is the unit's contribution, before the costs of the whole company. See the worked example
Typical margin: About 20 percent operating margin for colocation leaders (EBITDA about 50 percent); GPU clouds often lose money after interestRoughly how much of every 100 of sales (or income) is left as profit after the running costs. More on margin
Capital intensity: Very highVery large sums must be tied up before the business earns anything, so the cost of that money weighs heavily on profit. More on capital intensity
The number to watch: IT capacity (MW)The power the servers can draw; the main measure of size, capacity sold and capacity under construction.

Ask this first in a case

How much power is available, when can the grid deliver it, at what price, and how clean is it?

Words used above (5)

Data centre:: A secure building full of servers, with power supply, cooling and network connections.
Colocation:: Renting space, power and cooling in a shared data centre for your own servers.
Neocloud (GPU cloud):: A newer company that rents out AI chips by the hour, such as CoreWeave or Lambda.
GPU:: Graphics processing unit: a chip that does many small calculations at once, used for AI.
EBITDA:: Profit before interest, tax, depreciation and amortisation: a rough measure of cash profit from operations.

The industry's other words are explained in Words to know (12).

On this page (17 sections)

How money is made

Colocation rent per kW of IT capacity per month, with electricity usually billed on top at cost or with a small margin.
Cross-connect and interconnection fees for linking customers to networks and clouds inside the building.
Build-to-suit leases: a whole building or campus for one hyperscaler over 10 to 15 years or more.
Cloud usage fees: per hour of computing, per gigabyte stored, per request.
GPU hours: renting AI chips by the hour on demand, or on contracts of one to five years at a lower rate.
Take-or-pay capacity contracts with AI labs, which build large backlogs of contracted revenue.

Worked example: one unit

Unit economics means the money in and out for one unit of the business. Start from the revenue, take away the unit's own costs, and what is left is its contribution. More on unit economics

The unit: One AI GPU rented out by the hour for a year (busy 70 percent of the time at USD 2.50 an hour). Illustrative, rounded figures.
Line	Amount	ShareShare of revenue
GPU hours sold: 8,760 hours x 70 percent x USD 2.50	USD 15,330	100%
Minus Depreciation: USD 35,000 GPU and server share over five years	USD 7,000	46%
Minus Electricity: 1.25 kW including cooling x 8,760 hours x USD 0.10 per kWh	USD 1,095	7.1%
Minus Colocation rent, network and support	USD 2,000	13%
What is left (contribution)	USD 5,235	34%

Check: USD 15,330 minus USD 10,095 of costs leaves USD 5,235.

So what: The GPU breaks even at about 46 percent utilisation, and profit swings far more than revenue. Price and chip life move it most: at USD 1.80 an hour profit falls below USD 1,000, and writing the chip off over three years instead of five cuts it below USD 600, before any interest on the money borrowed to buy it.

Key measures(9)

Key measures (also called KPIs, key performance indicators) are the numbers people in this industry track. Ask for the first one or two early in a case.

IT capacity (MW)
The power the servers can draw; the main measure of size, capacity sold and capacity under construction.
Leased share or utilisation
Share of built capacity that is rented (colocation) or busy (cloud and GPUs).
Typical: primary North American markets had record-low vacancy of about 1.4 percent in the first half of 2026[3]
PUE (power usage effectiveness)
Total site power divided by the power used by the servers; lower is better, 1.0 is the floor. Glossary: PUE (power usage effectiveness)
Typical: industry average about 1.54 in 2025; new AI sites aim for about 1.1 to 1.3[2]
Rack density (kW per rack)
Power drawn by one cabinet of servers; it decides whether a building can host AI and whether liquid cooling is needed.
Typical: most racks 10 to 30 kW; the newest AI racks about 120 kW, with power supplies rated for about 132 kW[6]
Rent per kW per month
The usual colocation price for space, power delivery and cooling.
Typical: about USD 180 to 275 in Silicon Valley in late 2025; far lower per kW for large hyperscale leases[4]
Price per GPU hour
What customers pay to rent one AI chip for one hour; falls as newer chips arrive.
Typical: about USD 4 on-demand for an H100 on a specialist GPU cloud list price in September 2026; long contracts cost less[7]
Contracted backlog
Revenue customers have committed to pay in future; shows how much new capacity is already sold.
Typical: CoreWeave reported about USD 104 billion in mid 2026[11]
Capex as a share of revenue
How much of each year's revenue is spent on new buildings and equipment.
Typical: nearly half (about 47 percent) for Equinix in 2025[12]
Time to power
How long until the grid can deliver the power a new site needs; often the deciding factor.

First questions to ask

When a case lands in this industry, these questions get you to the numbers that matter.

How much power is available, when can the grid deliver it, at what price, and how clean is it?
How much capacity is already signed with customers, for how long, and with whom?
Is this colocation (rent for space and power) or compute (renting chips by the hour)?
What is the utilisation or leased share today, and what happens if prices fall by a third?
Over how many years are the chips written off, and how is the investment funded?

Value chain: where the margin sits

The value chain is the steps a product or service passes through, from the first supplier to the customer. Each step below shows how much of the value it keeps. More on value chains

Step 1: Secure power and land
Margin varies
Developers, utilities and power sellers through grid connections and power purchase agreements
The slowest step: a large grid connection can take years.
Step 2: Supply electrical and cooling equipment
Medium margin
Schneider Electric, Vertiv, Eaton, transformer and generator makers
Long order books when everyone builds at once.
Step 3: Design and build the facility
Thin margin
Specialist developers and builders, or the hyperscaler itself
Build costs rose about 21 percent per MW from late 2024 to 2026.
Step 4: Make the chips and servers
Fat margin
Nvidia, AMD and custom chips from Google, Amazon and Microsoft; servers from Foxconn, Quanta, Dell, Supermicro
Nvidia earned USD 193.7 billion of data centre revenue in fiscal 2026.
Step 5: Run the facility and rent space (colocation)
Medium margin
Equinix, Digital Realty, NTT Data, AirTrunk, Khazna, Yotta
Rent per kW per month on leases of 10 to 15 years or more for large tenants.
Step 6: Sell computing (cloud and GPU clouds)
Margin varies
Amazon Web Services, Microsoft Azure, Google Cloud, Oracle, CoreWeave, Nebius
Hyperscalers earn strong margins; GPU clouds carry heavy debt.
Step 7: Build AI models and software on top
Margin varies
AI labs, software companies and enterprises

Profit pool: who keeps the money

Where in the value chain the profit ends up, which is often not where most of the sales are. More on profit pools

Today most profit sits with the chip makers, above all Nvidia, whose data centre revenue reached USD 193.7 billion in fiscal 2026, and with the hyperscalers that sell cloud at scale. Colocation earns a steady, property-like return. GPU clouds and new builders earn the thinnest and riskiest margins, because they borrow to buy chips that lose value quickly.

Cost structure(4)

The main costs, each as a share of revenue (the money from sales).

Operating costs before depreciation for a colocation leader (staff, maintenance, power not passed through, rent, sales): about 50 percent (Equinix 2025: adjusted EBITDA margin 49 percent)[12]
Depreciation and other costs below EBITDA for a colocation leader: about 30 percent (Equinix 2025: operating income of about USD 1.8 billion on revenue of about USD 9.2 billion)[12]
Interest on borrowing for a GPU cloud: about 25 percent of revenue (CoreWeave, second quarter 2026: about USD 640 million of interest on about USD 2.6 billion of revenue)[11]
Build cost split of a new site (share of build cost, not of revenue): power infrastructure about 21 percent, core and shell and sitework about 17 percent, contingency about 16 percent, land about 7 percent[5]

Benchmarks(6)

Typical figures for the industry, to check a client's numbers against.

Adjusted EBITDA margin, colocation leader: about 49 percent (Equinix 2025), guided to about 51 percent for 2026[12]
Operating margin, colocation leader: about 20 percent (Equinix 2025)[12]Worked out from operating income of about USD 1.8 billion on about USD 9.2 billion of revenue.
Adjusted EBITDA margin, GPU cloud: about 59 percent (CoreWeave, second quarter 2026), yet a net loss after interest[11]
All-in build cost of a new site, US and Canada, excluding chips: about USD 8.9 million to 23.3 million per MW[5]
Data centres' share of world electricity use: about 1.5 percent in 2024 (about 415 TWh), heading for about 945 TWh by 2030[1]
Data centres' share of Ireland's metered electricity: 23 percent in 2025, up from 5 percent in 2015[15]

Typical cases(7)

Case prompts you might hear in this industry.

Where should we build a 300 MW AI data centre campus?
Should a telecom operator in Southeast Asia enter the colocation business?
Our GPU cloud is growing fast but losing money. What should we do?
Should a Gulf government fund a national AI compute cluster, and how?
A private equity fund wants to buy a data centre developer. Is it a good investment?
How should we price GPU capacity: on demand, or on multi-year contracts?
Should a hyperscaler build its own sites or lease from colocation developers?

Common traps(5)

Mistakes candidates make in this industry, and what to do instead.

Starting a site case with land price or tax. Start with power: amount, date, price and source.
Treating the building and the chips as one asset. The building lasts decades; the chips a few years.
Assuming electricity is the colocation operator's main cost. Customers usually pay for their own power on top of rent.
Assuming today's GPU prices last. Prices fell sharply after 2024 on many platforms, and newer chips keep arriving.
Counting announced capacity as if it were built, powered and leased.

What changed, 2024 to 2026(6)

Recent changes a case could turn on.

Cloud demand accelerated: cloud infrastructure spending reached about USD 143 billion in the second quarter of 2026, up about 43 percent on a year earlier, with Amazon at about 28 percent share, Microsoft about 20 percent and Google about 15 percent.[9]
Record capital spending: for 2026 Amazon guided to about USD 200 billion, Alphabet to about USD 195 to 205 billion and Microsoft to about USD 175 billion (after a lease accounting change; about USD 190 billion on the old basis), much of it for AI data centres, and investors are asking when it will pay back.[10]
Capacity ran short: primary North American vacancy fell to a record low of about 1.4 percent in the first half of 2026, with a record 7,481 MW under construction, and rents kept rising.[3]
GPU prices fell from their 2024 peak on many platforms; AWS cut on-demand H100 instance prices by up to about 45 percent in June 2025, which squeezed smaller GPU clouds.[8]
Sovereign AI: India's IndiaAI Mission offers more than 38,000 GPUs at about INR 65 per GPU hour to startups and researchers; the EU plans EUR 20 billion for AI gigafactories; the UAE's Stargate UAE targets a first 200 MW phase in 2026.[16]
Power became the limit: the IEA expects data centre electricity use to roughly double from 2024 to 2030, so sites now follow available power, not just customers.[1]

Players by region(7)

Well-known companies in each region. You do not need to learn them by heart; they help you picture the market.

Global: Amazon Web Services
Microsoft Azure
Google Cloud
Oracle
Equinix and Digital Realty (colocation)
Nvidia (AI chips)
United States: CoreWeave, Lambda and Crusoe (GPU clouds)
Meta (builds for its own use)
Vantage and QTS (hyperscale developers)
Europe: Nebius (Netherlands, GPU cloud)
Data4 (France)
atNorth (Nordics)
Mistral AI compute (France)
Middle East: G42 and Khazna (UAE)
Humain (Saudi Arabia, owned by the Public Investment Fund)
center3 (stc, Saudi Arabia)
India: Yotta
CtrlS
Sify
AdaniConneX
Reliance Jio
Southeast Asia: STT GDC and Keppel (Singapore)
Princeton Digital Group
AirTrunk
large campuses in Johor (Malaysia) and Batam (Indonesia)
China: Alibaba Cloud
Tencent Cloud
Huawei Cloud
GDS

Words to know(12)

Linked words have a fuller entry in the glossary.

Data centre: A secure building full of servers, with power supply, cooling and network connections.
MW of IT load: Megawatts of power the servers can draw; how data centres are sized and sold.
Colocation: Renting space, power and cooling in a shared data centre for your own servers.
Hyperscaler: A very large cloud company, such as Amazon, Microsoft or Google, that builds huge data centres.
Neocloud (GPU cloud): A newer company that rents out AI chips by the hour, such as CoreWeave or Lambda.
GPU: Graphics processing unit: a chip that does many small calculations at once, used for AI.
Training and inference: Training builds an AI model on huge data; inference uses the model to answer each question.
PUE (glossary entry): Power usage effectiveness: total site power divided by server power; lower is better.
Rack density: Power drawn by one cabinet of servers, in kW.
Sovereign AI: AI computing built and controlled inside a country for security and independence.
Power purchase agreement (PPA) (glossary entry): A long contract to buy electricity from a producer at an agreed price.
EBITDA (glossary entry): Profit before interest, tax, depreciation and amortisation: a rough measure of cash profit from operations.

Business model patterns

The ways of making money this industry follows. Spot the pattern in a new industry and you already know the first questions to ask.

Sources(18)

Facts checked on 30 September 2026. Worked examples are illustrative, shaped by these sources rather than one company's figures.

Go deeper and practise

Go deeper

The full lessons behind this brief, with sources and worked cases.

Practise it

Same pattern elsewhere

Industries that make money in a similar way. What you learned here carries over.