Lesson 2 of 3 Math checked Last reviewed 30 September 2026 12 min

Data centre economics: per MW, per rack and per GPU hour

Work out rent and return per megawatt for colocation, racks from a power budget, and profit per GPU hour for an AI cloud, and see why utilisation, price and chip life decide everything.

Industry brief, with a one-minute summary: Data centres, cloud and AI compute

Firm 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

Two businesses sit in one building. Colocation earns rent per kW for 10 to 15 years on a building that lasts decades, so it looks like property.
Data centres are built in waves.
Depreciation spreads the purchase price over the chip's useful life. A shorter life means a larger charge each year, and it matches the risk that newer chips make older ones less valuable.

Key idea

Two businesses sit in one building. Colocation earns rent per kW for 10 to 15 years on a building that lasts decades, so it looks like property. A GPU cloud earns per GPU hour on chips that lose value in a few years, so it lives or dies on utilisation (the share of hours its chips are rented out), price and how fast the chips must be paid back.

Reference prices and costs (checked 30 September 2026)

Reference prices and costs (checked 30 September 2026)
Item	Range	Source
Build cost, all in, greenfield, US and Canada, excluding chips	About USD 8.9 million to 23.3 million per MW, up about 21 percent since late 2024	Cushman & Wakefield 2026 cost guide, as reported
Colocation asking rent, Silicon Valley	About USD 180 to 275 per kW per month	CBRE, second half of 2025
Primary US market vacancy	About 1.4 percent, a record low	CBRE, H1 2026
H100 GPU, on-demand, specialist GPU cloud list price	About USD 4 per GPU hour	Lambda price list, September 2026
H100 GPU, AWS on-demand price cut in June 2025	Up to about 45 percent	AWS, June 2025

So-what

Prices differ a lot by market, contract length and size: a 20 MW hyperscale lease pays far less per kW than a few racks, and long GPU contracts cost less per hour than on-demand. Treat these as reference points, not a price list.

Worked case

Return per MW for a colocation campus near Mumbai

The prompt

A developer plans a 20 MW colocation campus near Mumbai and reports in USD. It costs USD 10 million per MW of IT load to build, excluding the customers' servers. It expects to lease 90 percent of the capacity at an average rent of USD 150 per kW per month; customers pay for their own electricity on top. Operating costs that are not passed through (staff, maintenance, insurance, security) are USD 400,000 per MW of built capacity a year. What are yearly revenue and operating profit, the yield on cost, and the simple payback? (Figures are illustrative.)

Open this case to practice it with a partner

The structure

Return per MW = (rent from leased MW minus operating costs) / build cost
- Revenue = MW x 1,000 kW x leased share x rent per kW x 12 months
- Operating profit = revenue minus operating costs
- Yield on cost = operating profit / build cost
- Payback = build cost / operating profit, in years

Working it through

1. Build cost
20 MW at USD 10 million each, in USD millions.
Build cost (USD millions):20 × 10 = 200
2. Yearly rent
20,000 kW, 90 percent leased, at USD 150 a month for 12 months, in USD millions.
Yearly revenue (USD millions):20 × 1,000 × 0.9 × 150 × 12 ÷ 1,000,000 = 32.4
3. Operating costs
USD 0.4 million per MW on all 20 MW built.
Yearly operating costs (USD millions):20 × 0.4 = 8
4. Operating profit
Revenue minus operating costs, before depreciation, interest and tax.
Yearly operating profit before depreciation (USD millions):32.4 - 8 = 24.4
5. Yield on cost
Operating profit divided by build cost.
Yield on cost (percent):24.4 ÷ 200 × 100 = 12.2
6. Simple payback
Build cost divided by yearly operating profit.
Simple payback (years):200 ÷ 24.4 = 8.2

The recommendation

The campus is attractive if it can be leased quickly and the power arrives on time, because it earns a yield on cost of about 12.2 percent and pays back in about 8.2 years on a building that lasts decades. First, 90 percent leasing at USD 150 per kW per month brings USD 32.4 million a year against USD 8 million of costs. Second, the return depends most on how fast the halls fill: every empty MW still costs its build money and its share of operating costs. The risk is that the grid connection or transformers arrive late, so the building sits empty. As a next step, sign an anchor tenant for the first phase before committing to build all 20 MW.

Risks: A late grid connection leaves capacity built but unleasable; A large tenant can push the rent down in exchange for signing early; Rupee and dollar exchange rates affect a cost base partly in rupees.

Timed math drill

A new AI site in the UAE has a 100 MW grid connection. The design PUE is 1.25, so part of the power goes to cooling and losses. Each AI rack draws 125 kW. How many racks can the site power?

Worked case

Profit per GPU for a GPU cloud in Singapore

The prompt

A GPU cloud (a neocloud) in Singapore rents out AI GPUs by the hour and reports in USD. Each GPU costs USD 35,000 all in, including its share of the server and network, and is written off over five years. It sells GPU time at USD 2.50 an hour and expects the GPUs to be busy 70 percent of the 8,760 hours in a year. Each GPU draws 1.25 kW including cooling, at USD 0.10 per kWh, around the clock. Colocation rent and support cost USD 2,000 per GPU a year. What is yearly profit per GPU before interest and tax, and at what utilisation does it break even? (Figures are illustrative.)

Open this case to practice it with a partner

The structure

Profit per GPU = hours sold x price minus (depreciation + power + facility)
- Hours sold = 8,760 x utilisation
- Depreciation = GPU cost / useful life
- Power = kW x 8,760 hours x price per kWh
- Break-even utilisation = total costs / (8,760 x price)

Working it through

1. Hours sold
70 percent of 8,760 hours.
GPU hours sold a year:8,760 × 0.7 = 6,132
2. Revenue
6,132 hours at USD 2.50.
Revenue per GPU a year (USD):8,760 × 0.7 × 2.5 = 15,330
3. Depreciation
USD 35,000 over five years.
Depreciation per GPU a year (USD):35,000 ÷ 5 = 7,000
4. Power
1.25 kW for 8,760 hours at USD 0.10 per kWh (power is drawn even when idle, a simplification).
Power cost per GPU a year (USD):1.25 × 8,760 × 0.1 = 1,095
5. Profit per GPU
Revenue minus depreciation, power and USD 2,000 of facility costs.
Profit per GPU a year before interest and tax (USD):15,330 - 7,000 - 1,095 - 2,000 = 5,235
6. Break-even utilisation
Total costs of USD 10,095 divided by the revenue at full use, 8,760 hours at USD 2.50.
Break-even utilisation (percent):(7,000 + 1,095 + 2,000) ÷ (8,760 × 2.5) × 100 = 46.1

The recommendation

The GPU cloud makes money at these numbers, about USD 5,235 per GPU a year, but it should lock in long contracts, because profit is very sensitive to price and chip life. First, it breaks even at about 46 percent utilisation, so a fall in demand hurts quickly. Second, if the price drops to USD 1.80 an hour at the same 70 percent utilisation, profit falls to under USD 1,000 per GPU, and if the chips must be written off over three years instead of five, depreciation rises to about USD 11,700 and profit falls to under USD 600 per GPU. The risk is that newer chips make these GPUs cheaper to rent within two or three years. As a next step, compare the length of customer contracts with the life assumed for the chips.

Risks: Newer, faster chips push down the price customers will pay for older GPUs; Interest on the debt used to buy GPUs is not in this profit figure; A single large customer leaving would cut utilisation sharply.

Timed math drill

In the Singapore GPU case, the price falls from USD 2.50 to USD 1.80 an hour, utilisation stays at 70 percent of 8,760 hours, and yearly costs stay at USD 10,095 per GPU. What is the new yearly profit per GPU, in USD? Round to the nearest dollar.

The capex cycle

Data centres are built in waves. When demand is hot, everyone orders chips, transformers and cooling at once, prices and delivery times rise, and capacity arrives two or three years later, sometimes all at the same time. If demand then slows, rents and GPU prices fall and the newest capacity sits empty. When you judge a big investment, ask what happens if prices fall by a third just as the capacity comes online.

Check your understanding

A GPU cloud writes its chips off over three years instead of five. What happens to its yearly profit per GPU, other things equal?

Terms in this lesson

Capex (capital expenditure)Depreciation and amortization Payback period Capacity utilization Breakeven Unit economics PUE (power usage effectiveness)Operating leverage Capex intensity Fixed cost

Sources for this lesson (8)

My notes on this lesson

Write anything you want to remember. Only you can see it.

0 of 5,000 characters. Saves automatically.

Try the 3 remaining checks and drills above to complete this lesson (0 of 3 done).

Spotted something wrong or out of date? Report a mistake. We check every report and correct the page.